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 16:49:35 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_r2cb.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 25 -name r2cb_25 -include r2cb.h */
|
cannam@127
|
29
|
cannam@127
|
30 /*
|
cannam@127
|
31 * This function contains 152 FP additions, 120 FP multiplications,
|
cannam@127
|
32 * (or, 32 additions, 0 multiplications, 120 fused multiply/add),
|
cannam@127
|
33 * 115 stack variables, 44 constants, and 50 memory accesses
|
cannam@127
|
34 */
|
cannam@127
|
35 #include "r2cb.h"
|
cannam@127
|
36
|
cannam@127
|
37 static void r2cb_25(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(KP979740652, +0.979740652857618686258237536568998933733477632);
|
cannam@127
|
40 DK(KP438153340, +0.438153340021931793654057951961031291699532119);
|
cannam@127
|
41 DK(KP1_752613360, +1.752613360087727174616231807844125166798128477);
|
cannam@127
|
42 DK(KP963507348, +0.963507348203430549974383005744259307057084020);
|
cannam@127
|
43 DK(KP1_606007150, +1.606007150877320829666881187140752009270929701);
|
cannam@127
|
44 DK(KP1_721083328, +1.721083328735889354196523361841037632825608373);
|
cannam@127
|
45 DK(KP1_011627398, +1.011627398597394192215998921771049272931807941);
|
cannam@127
|
46 DK(KP595480289, +0.595480289600000014706716770488118292997907308);
|
cannam@127
|
47 DK(KP641441904, +0.641441904830606407298806329068862424939687989);
|
cannam@127
|
48 DK(KP452413526, +0.452413526233009763856834323966348796985206956);
|
cannam@127
|
49 DK(KP1_809654104, +1.809654104932039055427337295865395187940827822);
|
cannam@127
|
50 DK(KP933137358, +0.933137358350283770603023973254446451924190884);
|
cannam@127
|
51 DK(KP1_666834356, +1.666834356657377354817925100486477686277992119);
|
cannam@127
|
52 DK(KP1_842354653, +1.842354653930286640500894870830132058718564461);
|
cannam@127
|
53 DK(KP1_082908895, +1.082908895072625554092571180165639018104066379);
|
cannam@127
|
54 DK(KP662318342, +0.662318342759882818626911127577439236802190210);
|
cannam@127
|
55 DK(KP576710603, +0.576710603632765877371579268136471017090111488);
|
cannam@127
|
56 DK(KP484291580, +0.484291580564315559745084187732367906918006201);
|
cannam@127
|
57 DK(KP1_937166322, +1.937166322257262238980336750929471627672024806);
|
cannam@127
|
58 DK(KP1_898359647, +1.898359647016882523151110931686726543423167685);
|
cannam@127
|
59 DK(KP1_386580726, +1.386580726567734802700860150804827247498955921);
|
cannam@127
|
60 DK(KP904730450, +0.904730450839922351881287709692877908104763647);
|
cannam@127
|
61 DK(KP1_115827804, +1.115827804063668528375399296931134075984874304);
|
cannam@127
|
62 DK(KP634619297, +0.634619297544148100711287640319130485732531031);
|
cannam@127
|
63 DK(KP470564281, +0.470564281212251493087595091036643380879947982);
|
cannam@127
|
64 DK(KP499013364, +0.499013364214135780976168403431725276668452610);
|
cannam@127
|
65 DK(KP1_996053456, +1.996053456856543123904673613726901106673810439);
|
cannam@127
|
66 DK(KP559154169, +0.559154169276087864842202529084232643714075927);
|
cannam@127
|
67 DK(KP683113946, +0.683113946453479238701949862233725244439656928);
|
cannam@127
|
68 DK(KP730409924, +0.730409924561256563751459444999838399157094302);
|
cannam@127
|
69 DK(KP549754652, +0.549754652192770074288023275540779861653779767);
|
cannam@127
|
70 DK(KP256756360, +0.256756360367726783319498520922669048172391148);
|
cannam@127
|
71 DK(KP451418159, +0.451418159099103183892477933432151804893354132);
|
cannam@127
|
72 DK(KP846146756, +0.846146756728608505452954290121135880883743802);
|
cannam@127
|
73 DK(KP1_902113032, +1.902113032590307144232878666758764286811397268);
|
cannam@127
|
74 DK(KP062914667, +0.062914667253649757225485955897349402364686947);
|
cannam@127
|
75 DK(KP939062505, +0.939062505817492352556001843133229685779824606);
|
cannam@127
|
76 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
cannam@127
|
77 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
cannam@127
|
78 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
cannam@127
|
79 DK(KP1_118033988, +1.118033988749894848204586834365638117720309180);
|
cannam@127
|
80 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
|
cannam@127
|
81 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
cannam@127
|
82 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
|
cannam@127
|
83 {
|
cannam@127
|
84 INT i;
|
cannam@127
|
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)) {
|
cannam@127
|
86 E T1H, T24, T22, T1W, T1Y, T1X, T1Z, T23;
|
cannam@127
|
87 {
|
cannam@127
|
88 E T1G, Tu, T5, T1F, Tr, Te, T2o, T1N, T2a, T1t, TR, T1K, T29, T1u, TG;
|
cannam@127
|
89 E TU, TT, Tn, T1d, T1Q, T2p, T1T, T12, T1P, T1a;
|
cannam@127
|
90 {
|
cannam@127
|
91 E T1, T2, T3, Ts, Tt;
|
cannam@127
|
92 Ts = Ci[WS(csi, 5)];
|
cannam@127
|
93 Tt = Ci[WS(csi, 10)];
|
cannam@127
|
94 T1 = Cr[0];
|
cannam@127
|
95 T2 = Cr[WS(csr, 5)];
|
cannam@127
|
96 T3 = Cr[WS(csr, 10)];
|
cannam@127
|
97 T1G = FMS(KP618033988, Ts, Tt);
|
cannam@127
|
98 Tu = FMA(KP618033988, Tt, Ts);
|
cannam@127
|
99 {
|
cannam@127
|
100 E Tx, Tw, T1M, TQ, TM, T1J, TF, TL;
|
cannam@127
|
101 {
|
cannam@127
|
102 E T6, TH, TO, TP, TB, TI, Td, TJ, TE, T4, Tq, TK;
|
cannam@127
|
103 T6 = Cr[WS(csr, 1)];
|
cannam@127
|
104 T4 = T2 + T3;
|
cannam@127
|
105 Tq = T2 - T3;
|
cannam@127
|
106 TH = Ci[WS(csi, 1)];
|
cannam@127
|
107 {
|
cannam@127
|
108 E Ta, T9, Tb, T7, T8, Tp;
|
cannam@127
|
109 T7 = Cr[WS(csr, 6)];
|
cannam@127
|
110 T8 = Cr[WS(csr, 4)];
|
cannam@127
|
111 Tp = FNMS(KP500000000, T4, T1);
|
cannam@127
|
112 T5 = FMA(KP2_000000000, T4, T1);
|
cannam@127
|
113 Ta = Cr[WS(csr, 11)];
|
cannam@127
|
114 TO = T7 - T8;
|
cannam@127
|
115 T9 = T7 + T8;
|
cannam@127
|
116 T1F = FNMS(KP1_118033988, Tq, Tp);
|
cannam@127
|
117 Tr = FMA(KP1_118033988, Tq, Tp);
|
cannam@127
|
118 Tb = Cr[WS(csr, 9)];
|
cannam@127
|
119 {
|
cannam@127
|
120 E TC, TD, Tz, TA, Tc;
|
cannam@127
|
121 Tz = Ci[WS(csi, 6)];
|
cannam@127
|
122 TA = Ci[WS(csi, 4)];
|
cannam@127
|
123 TP = Tb - Ta;
|
cannam@127
|
124 Tc = Ta + Tb;
|
cannam@127
|
125 TC = Ci[WS(csi, 11)];
|
cannam@127
|
126 TB = Tz + TA;
|
cannam@127
|
127 TI = Tz - TA;
|
cannam@127
|
128 TD = Ci[WS(csi, 9)];
|
cannam@127
|
129 Td = T9 + Tc;
|
cannam@127
|
130 Tx = T9 - Tc;
|
cannam@127
|
131 TJ = TC - TD;
|
cannam@127
|
132 TE = TC + TD;
|
cannam@127
|
133 }
|
cannam@127
|
134 }
|
cannam@127
|
135 Te = T6 + Td;
|
cannam@127
|
136 Tw = FNMS(KP250000000, Td, T6);
|
cannam@127
|
137 T1M = FMA(KP618033988, TO, TP);
|
cannam@127
|
138 TQ = FNMS(KP618033988, TP, TO);
|
cannam@127
|
139 TK = TI + TJ;
|
cannam@127
|
140 TM = TI - TJ;
|
cannam@127
|
141 T1J = FNMS(KP618033988, TB, TE);
|
cannam@127
|
142 TF = FMA(KP618033988, TE, TB);
|
cannam@127
|
143 TL = FNMS(KP250000000, TK, TH);
|
cannam@127
|
144 T2o = TK + TH;
|
cannam@127
|
145 }
|
cannam@127
|
146 {
|
cannam@127
|
147 E Tf, T14, T1b, T1c, Tm, TY, T15, T16, T11, T17, T19, T18;
|
cannam@127
|
148 Tf = Cr[WS(csr, 2)];
|
cannam@127
|
149 {
|
cannam@127
|
150 E T1L, TN, T1I, Ty;
|
cannam@127
|
151 T1L = FNMS(KP559016994, TM, TL);
|
cannam@127
|
152 TN = FMA(KP559016994, TM, TL);
|
cannam@127
|
153 T1I = FNMS(KP559016994, Tx, Tw);
|
cannam@127
|
154 Ty = FMA(KP559016994, Tx, Tw);
|
cannam@127
|
155 T1N = FMA(KP951056516, T1M, T1L);
|
cannam@127
|
156 T2a = FNMS(KP951056516, T1M, T1L);
|
cannam@127
|
157 T1t = FNMS(KP951056516, TQ, TN);
|
cannam@127
|
158 TR = FMA(KP951056516, TQ, TN);
|
cannam@127
|
159 T1K = FMA(KP951056516, T1J, T1I);
|
cannam@127
|
160 T29 = FNMS(KP951056516, T1J, T1I);
|
cannam@127
|
161 T1u = FMA(KP951056516, TF, Ty);
|
cannam@127
|
162 TG = FNMS(KP951056516, TF, Ty);
|
cannam@127
|
163 T14 = Ci[WS(csi, 2)];
|
cannam@127
|
164 }
|
cannam@127
|
165 {
|
cannam@127
|
166 E Tg, Th, Tj, Tk;
|
cannam@127
|
167 Tg = Cr[WS(csr, 7)];
|
cannam@127
|
168 Th = Cr[WS(csr, 3)];
|
cannam@127
|
169 Tj = Cr[WS(csr, 12)];
|
cannam@127
|
170 Tk = Cr[WS(csr, 8)];
|
cannam@127
|
171 {
|
cannam@127
|
172 E TW, Ti, Tl, TX, TZ, T10;
|
cannam@127
|
173 TW = Ci[WS(csi, 7)];
|
cannam@127
|
174 T1b = Th - Tg;
|
cannam@127
|
175 Ti = Tg + Th;
|
cannam@127
|
176 T1c = Tj - Tk;
|
cannam@127
|
177 Tl = Tj + Tk;
|
cannam@127
|
178 TX = Ci[WS(csi, 3)];
|
cannam@127
|
179 TZ = Ci[WS(csi, 12)];
|
cannam@127
|
180 T10 = Ci[WS(csi, 8)];
|
cannam@127
|
181 Tm = Ti + Tl;
|
cannam@127
|
182 TU = Tl - Ti;
|
cannam@127
|
183 TY = TW + TX;
|
cannam@127
|
184 T15 = TW - TX;
|
cannam@127
|
185 T16 = TZ - T10;
|
cannam@127
|
186 T11 = TZ + T10;
|
cannam@127
|
187 }
|
cannam@127
|
188 }
|
cannam@127
|
189 TT = FNMS(KP250000000, Tm, Tf);
|
cannam@127
|
190 Tn = Tf + Tm;
|
cannam@127
|
191 T17 = T15 + T16;
|
cannam@127
|
192 T19 = T16 - T15;
|
cannam@127
|
193 T1d = FNMS(KP618033988, T1c, T1b);
|
cannam@127
|
194 T1Q = FMA(KP618033988, T1b, T1c);
|
cannam@127
|
195 T18 = FNMS(KP250000000, T17, T14);
|
cannam@127
|
196 T2p = T17 + T14;
|
cannam@127
|
197 T1T = FNMS(KP618033988, TY, T11);
|
cannam@127
|
198 T12 = FMA(KP618033988, T11, TY);
|
cannam@127
|
199 T1P = FMA(KP559016994, T19, T18);
|
cannam@127
|
200 T1a = FNMS(KP559016994, T19, T18);
|
cannam@127
|
201 }
|
cannam@127
|
202 }
|
cannam@127
|
203 }
|
cannam@127
|
204 {
|
cannam@127
|
205 E T1R, T1e, T1q, T1U, T13, T1r, T2b, T28, T25, T2i, T2k;
|
cannam@127
|
206 {
|
cannam@127
|
207 E T2m, To, T26, T27, TV, T1S;
|
cannam@127
|
208 T2m = Te - Tn;
|
cannam@127
|
209 To = Te + Tn;
|
cannam@127
|
210 TV = FNMS(KP559016994, TU, TT);
|
cannam@127
|
211 T1S = FMA(KP559016994, TU, TT);
|
cannam@127
|
212 T26 = FMA(KP951056516, T1Q, T1P);
|
cannam@127
|
213 T1R = FNMS(KP951056516, T1Q, T1P);
|
cannam@127
|
214 T1e = FNMS(KP951056516, T1d, T1a);
|
cannam@127
|
215 T1q = FMA(KP951056516, T1d, T1a);
|
cannam@127
|
216 T27 = FNMS(KP951056516, T1T, T1S);
|
cannam@127
|
217 T1U = FMA(KP951056516, T1T, T1S);
|
cannam@127
|
218 T13 = FNMS(KP951056516, T12, TV);
|
cannam@127
|
219 T1r = FMA(KP951056516, T12, TV);
|
cannam@127
|
220 {
|
cannam@127
|
221 E T2g, T2q, T2s, T2h, T2n, T2r, T2l;
|
cannam@127
|
222 T2g = FMA(KP939062505, T29, T2a);
|
cannam@127
|
223 T2b = FNMS(KP939062505, T2a, T29);
|
cannam@127
|
224 R0[0] = FMA(KP2_000000000, To, T5);
|
cannam@127
|
225 T2l = FNMS(KP500000000, To, T5);
|
cannam@127
|
226 T2q = FMA(KP618033988, T2p, T2o);
|
cannam@127
|
227 T2s = FNMS(KP618033988, T2o, T2p);
|
cannam@127
|
228 T28 = FNMS(KP062914667, T27, T26);
|
cannam@127
|
229 T2h = FMA(KP062914667, T26, T27);
|
cannam@127
|
230 T2n = FMA(KP1_118033988, T2m, T2l);
|
cannam@127
|
231 T2r = FNMS(KP1_118033988, T2m, T2l);
|
cannam@127
|
232 T25 = FMA(KP1_902113032, T1G, T1F);
|
cannam@127
|
233 T1H = FNMS(KP1_902113032, T1G, T1F);
|
cannam@127
|
234 T2i = FMA(KP846146756, T2h, T2g);
|
cannam@127
|
235 T2k = FNMS(KP451418159, T2g, T2h);
|
cannam@127
|
236 R0[WS(rs, 10)] = FMA(KP1_902113032, T2q, T2n);
|
cannam@127
|
237 R1[WS(rs, 2)] = FNMS(KP1_902113032, T2q, T2n);
|
cannam@127
|
238 R0[WS(rs, 5)] = FMA(KP1_902113032, T2s, T2r);
|
cannam@127
|
239 R1[WS(rs, 7)] = FNMS(KP1_902113032, T2s, T2r);
|
cannam@127
|
240 }
|
cannam@127
|
241 }
|
cannam@127
|
242 {
|
cannam@127
|
243 E TS, T1f, T1p, Tv, T2e, T1o, T1m, T2d, T1k, T1l, T2c;
|
cannam@127
|
244 TS = FNMS(KP256756360, TR, TG);
|
cannam@127
|
245 T1k = FMA(KP256756360, TG, TR);
|
cannam@127
|
246 T1l = FMA(KP549754652, T13, T1e);
|
cannam@127
|
247 T1f = FNMS(KP549754652, T1e, T13);
|
cannam@127
|
248 T1p = FMA(KP1_902113032, Tu, Tr);
|
cannam@127
|
249 Tv = FNMS(KP1_902113032, Tu, Tr);
|
cannam@127
|
250 T2e = FMA(KP730409924, T2b, T28);
|
cannam@127
|
251 T2c = FNMS(KP730409924, T2b, T28);
|
cannam@127
|
252 T1o = FNMS(KP683113946, T1k, T1l);
|
cannam@127
|
253 T1m = FMA(KP559154169, T1l, T1k);
|
cannam@127
|
254 R1[WS(rs, 1)] = FNMS(KP1_996053456, T2c, T25);
|
cannam@127
|
255 T2d = FMA(KP499013364, T2c, T25);
|
cannam@127
|
256 {
|
cannam@127
|
257 E T1C, T1E, T1y, T1w;
|
cannam@127
|
258 {
|
cannam@127
|
259 E T1s, T1v, T1i, T1h, T1n, T1j;
|
cannam@127
|
260 {
|
cannam@127
|
261 E T1A, T1B, T2f, T2j, T1g;
|
cannam@127
|
262 T1A = FNMS(KP470564281, T1q, T1r);
|
cannam@127
|
263 T1s = FMA(KP470564281, T1r, T1q);
|
cannam@127
|
264 T1v = FNMS(KP634619297, T1u, T1t);
|
cannam@127
|
265 T1B = FMA(KP634619297, T1t, T1u);
|
cannam@127
|
266 T2f = FMA(KP1_115827804, T2e, T2d);
|
cannam@127
|
267 T2j = FNMS(KP1_115827804, T2e, T2d);
|
cannam@127
|
268 T1i = FNMS(KP904730450, T1f, TS);
|
cannam@127
|
269 T1g = FMA(KP904730450, T1f, TS);
|
cannam@127
|
270 R1[WS(rs, 11)] = FMA(KP1_386580726, T2i, T2f);
|
cannam@127
|
271 R0[WS(rs, 4)] = FNMS(KP1_386580726, T2i, T2f);
|
cannam@127
|
272 R1[WS(rs, 6)] = FMA(KP1_898359647, T2k, T2j);
|
cannam@127
|
273 R0[WS(rs, 9)] = FNMS(KP1_898359647, T2k, T2j);
|
cannam@127
|
274 R1[0] = FMA(KP1_937166322, T1g, Tv);
|
cannam@127
|
275 T1h = FNMS(KP484291580, T1g, Tv);
|
cannam@127
|
276 T1C = FNMS(KP576710603, T1B, T1A);
|
cannam@127
|
277 T1E = FMA(KP662318342, T1A, T1B);
|
cannam@127
|
278 }
|
cannam@127
|
279 T1n = FNMS(KP1_082908895, T1i, T1h);
|
cannam@127
|
280 T1j = FMA(KP1_082908895, T1i, T1h);
|
cannam@127
|
281 R1[WS(rs, 10)] = FMA(KP1_842354653, T1m, T1j);
|
cannam@127
|
282 R0[WS(rs, 3)] = FNMS(KP1_842354653, T1m, T1j);
|
cannam@127
|
283 R1[WS(rs, 5)] = FMA(KP1_666834356, T1o, T1n);
|
cannam@127
|
284 R0[WS(rs, 8)] = FNMS(KP1_666834356, T1o, T1n);
|
cannam@127
|
285 T1y = FNMS(KP933137358, T1v, T1s);
|
cannam@127
|
286 T1w = FMA(KP933137358, T1v, T1s);
|
cannam@127
|
287 }
|
cannam@127
|
288 {
|
cannam@127
|
289 E T1O, T20, T21, T1V, T1x, T1z, T1D;
|
cannam@127
|
290 T1O = FNMS(KP549754652, T1N, T1K);
|
cannam@127
|
291 T20 = FMA(KP549754652, T1K, T1N);
|
cannam@127
|
292 T21 = FMA(KP634619297, T1R, T1U);
|
cannam@127
|
293 T1V = FNMS(KP634619297, T1U, T1R);
|
cannam@127
|
294 R0[WS(rs, 2)] = FNMS(KP1_809654104, T1w, T1p);
|
cannam@127
|
295 T1x = FMA(KP452413526, T1w, T1p);
|
cannam@127
|
296 T24 = FNMS(KP641441904, T20, T21);
|
cannam@127
|
297 T22 = FMA(KP595480289, T21, T20);
|
cannam@127
|
298 T1z = FNMS(KP1_011627398, T1y, T1x);
|
cannam@127
|
299 T1D = FMA(KP1_011627398, T1y, T1x);
|
cannam@127
|
300 R1[WS(rs, 9)] = FNMS(KP1_721083328, T1C, T1z);
|
cannam@127
|
301 R0[WS(rs, 7)] = FMA(KP1_721083328, T1C, T1z);
|
cannam@127
|
302 R0[WS(rs, 12)] = FMA(KP1_606007150, T1E, T1D);
|
cannam@127
|
303 R1[WS(rs, 4)] = FNMS(KP1_606007150, T1E, T1D);
|
cannam@127
|
304 T1W = FNMS(KP963507348, T1V, T1O);
|
cannam@127
|
305 T1Y = FMA(KP963507348, T1V, T1O);
|
cannam@127
|
306 }
|
cannam@127
|
307 }
|
cannam@127
|
308 }
|
cannam@127
|
309 }
|
cannam@127
|
310 }
|
cannam@127
|
311 R0[WS(rs, 1)] = FMA(KP1_752613360, T1W, T1H);
|
cannam@127
|
312 T1X = FNMS(KP438153340, T1W, T1H);
|
cannam@127
|
313 T1Z = FMA(KP979740652, T1Y, T1X);
|
cannam@127
|
314 T23 = FNMS(KP979740652, T1Y, T1X);
|
cannam@127
|
315 R0[WS(rs, 11)] = FMA(KP1_666834356, T22, T1Z);
|
cannam@127
|
316 R1[WS(rs, 3)] = FNMS(KP1_666834356, T22, T1Z);
|
cannam@127
|
317 R1[WS(rs, 8)] = FNMS(KP1_606007150, T24, T23);
|
cannam@127
|
318 R0[WS(rs, 6)] = FMA(KP1_606007150, T24, T23);
|
cannam@127
|
319 }
|
cannam@127
|
320 }
|
cannam@127
|
321 }
|
cannam@127
|
322
|
cannam@127
|
323 static const kr2c_desc desc = { 25, "r2cb_25", {32, 0, 120, 0}, &GENUS };
|
cannam@127
|
324
|
cannam@127
|
325 void X(codelet_r2cb_25) (planner *p) {
|
cannam@127
|
326 X(kr2c_register) (p, r2cb_25, &desc);
|
cannam@127
|
327 }
|
cannam@127
|
328
|
cannam@127
|
329 #else /* HAVE_FMA */
|
cannam@127
|
330
|
cannam@127
|
331 /* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 25 -name r2cb_25 -include r2cb.h */
|
cannam@127
|
332
|
cannam@127
|
333 /*
|
cannam@127
|
334 * This function contains 152 FP additions, 98 FP multiplications,
|
cannam@127
|
335 * (or, 100 additions, 46 multiplications, 52 fused multiply/add),
|
cannam@127
|
336 * 65 stack variables, 21 constants, and 50 memory accesses
|
cannam@127
|
337 */
|
cannam@127
|
338 #include "r2cb.h"
|
cannam@127
|
339
|
cannam@127
|
340 static void r2cb_25(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
cannam@127
|
341 {
|
cannam@127
|
342 DK(KP425779291, +0.425779291565072648862502445744251703979973042);
|
cannam@127
|
343 DK(KP904827052, +0.904827052466019527713668647932697593970413911);
|
cannam@127
|
344 DK(KP535826794, +0.535826794978996618271308767867639978063575346);
|
cannam@127
|
345 DK(KP844327925, +0.844327925502015078548558063966681505381659241);
|
cannam@127
|
346 DK(KP876306680, +0.876306680043863587308115903922062583399064238);
|
cannam@127
|
347 DK(KP481753674, +0.481753674101715274987191502872129653528542010);
|
cannam@127
|
348 DK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
cannam@127
|
349 DK(KP248689887, +0.248689887164854788242283746006447968417567406);
|
cannam@127
|
350 DK(KP062790519, +0.062790519529313376076178224565631133122484832);
|
cannam@127
|
351 DK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
cannam@127
|
352 DK(KP728968627, +0.728968627421411523146730319055259111372571664);
|
cannam@127
|
353 DK(KP684547105, +0.684547105928688673732283357621209269889519233);
|
cannam@127
|
354 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
cannam@127
|
355 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
cannam@127
|
356 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
cannam@127
|
357 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
cannam@127
|
358 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
cannam@127
|
359 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
|
cannam@127
|
360 DK(KP1_118033988, +1.118033988749894848204586834365638117720309180);
|
cannam@127
|
361 DK(KP1_175570504, +1.175570504584946258337411909278145537195304875);
|
cannam@127
|
362 DK(KP1_902113032, +1.902113032590307144232878666758764286811397268);
|
cannam@127
|
363 {
|
cannam@127
|
364 INT i;
|
cannam@127
|
365 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)) {
|
cannam@127
|
366 E Tu, T1G, T5, Tr, T1F, TN, TO, Te, TR, T27, T1r, T1N, TG, T26, T1q;
|
cannam@127
|
367 E T1K, T1a, T1b, Tn, T1e, T2a, T1u, T1U, T13, T29, T1t, T1R, Ts, Tt;
|
cannam@127
|
368 Ts = Ci[WS(csi, 5)];
|
cannam@127
|
369 Tt = Ci[WS(csi, 10)];
|
cannam@127
|
370 Tu = FMA(KP1_902113032, Ts, KP1_175570504 * Tt);
|
cannam@127
|
371 T1G = FNMS(KP1_902113032, Tt, KP1_175570504 * Ts);
|
cannam@127
|
372 {
|
cannam@127
|
373 E T1, T4, Tp, T2, T3, Tq;
|
cannam@127
|
374 T1 = Cr[0];
|
cannam@127
|
375 T2 = Cr[WS(csr, 5)];
|
cannam@127
|
376 T3 = Cr[WS(csr, 10)];
|
cannam@127
|
377 T4 = T2 + T3;
|
cannam@127
|
378 Tp = KP1_118033988 * (T2 - T3);
|
cannam@127
|
379 T5 = FMA(KP2_000000000, T4, T1);
|
cannam@127
|
380 Tq = FNMS(KP500000000, T4, T1);
|
cannam@127
|
381 Tr = Tp + Tq;
|
cannam@127
|
382 T1F = Tq - Tp;
|
cannam@127
|
383 }
|
cannam@127
|
384 {
|
cannam@127
|
385 E T6, Td, TI, Tw, TH, TB, TE, TM;
|
cannam@127
|
386 T6 = Cr[WS(csr, 1)];
|
cannam@127
|
387 TN = Ci[WS(csi, 1)];
|
cannam@127
|
388 {
|
cannam@127
|
389 E T7, T8, T9, Ta, Tb, Tc;
|
cannam@127
|
390 T7 = Cr[WS(csr, 6)];
|
cannam@127
|
391 T8 = Cr[WS(csr, 4)];
|
cannam@127
|
392 T9 = T7 + T8;
|
cannam@127
|
393 Ta = Cr[WS(csr, 11)];
|
cannam@127
|
394 Tb = Cr[WS(csr, 9)];
|
cannam@127
|
395 Tc = Ta + Tb;
|
cannam@127
|
396 Td = T9 + Tc;
|
cannam@127
|
397 TI = Ta - Tb;
|
cannam@127
|
398 Tw = KP559016994 * (T9 - Tc);
|
cannam@127
|
399 TH = T7 - T8;
|
cannam@127
|
400 }
|
cannam@127
|
401 {
|
cannam@127
|
402 E Tz, TA, TK, TC, TD, TL;
|
cannam@127
|
403 Tz = Ci[WS(csi, 6)];
|
cannam@127
|
404 TA = Ci[WS(csi, 4)];
|
cannam@127
|
405 TK = Tz - TA;
|
cannam@127
|
406 TC = Ci[WS(csi, 11)];
|
cannam@127
|
407 TD = Ci[WS(csi, 9)];
|
cannam@127
|
408 TL = TC - TD;
|
cannam@127
|
409 TB = Tz + TA;
|
cannam@127
|
410 TO = TK + TL;
|
cannam@127
|
411 TE = TC + TD;
|
cannam@127
|
412 TM = KP559016994 * (TK - TL);
|
cannam@127
|
413 }
|
cannam@127
|
414 Te = T6 + Td;
|
cannam@127
|
415 {
|
cannam@127
|
416 E TJ, T1L, TQ, T1M, TP;
|
cannam@127
|
417 TJ = FMA(KP951056516, TH, KP587785252 * TI);
|
cannam@127
|
418 T1L = FNMS(KP951056516, TI, KP587785252 * TH);
|
cannam@127
|
419 TP = FNMS(KP250000000, TO, TN);
|
cannam@127
|
420 TQ = TM + TP;
|
cannam@127
|
421 T1M = TP - TM;
|
cannam@127
|
422 TR = TJ + TQ;
|
cannam@127
|
423 T27 = T1M - T1L;
|
cannam@127
|
424 T1r = TQ - TJ;
|
cannam@127
|
425 T1N = T1L + T1M;
|
cannam@127
|
426 }
|
cannam@127
|
427 {
|
cannam@127
|
428 E TF, T1J, Ty, T1I, Tx;
|
cannam@127
|
429 TF = FMA(KP951056516, TB, KP587785252 * TE);
|
cannam@127
|
430 T1J = FNMS(KP951056516, TE, KP587785252 * TB);
|
cannam@127
|
431 Tx = FNMS(KP250000000, Td, T6);
|
cannam@127
|
432 Ty = Tw + Tx;
|
cannam@127
|
433 T1I = Tx - Tw;
|
cannam@127
|
434 TG = Ty - TF;
|
cannam@127
|
435 T26 = T1I + T1J;
|
cannam@127
|
436 T1q = Ty + TF;
|
cannam@127
|
437 T1K = T1I - T1J;
|
cannam@127
|
438 }
|
cannam@127
|
439 }
|
cannam@127
|
440 {
|
cannam@127
|
441 E Tf, Tm, T15, TT, T14, TY, T11, T19;
|
cannam@127
|
442 Tf = Cr[WS(csr, 2)];
|
cannam@127
|
443 T1a = Ci[WS(csi, 2)];
|
cannam@127
|
444 {
|
cannam@127
|
445 E Tg, Th, Ti, Tj, Tk, Tl;
|
cannam@127
|
446 Tg = Cr[WS(csr, 7)];
|
cannam@127
|
447 Th = Cr[WS(csr, 3)];
|
cannam@127
|
448 Ti = Tg + Th;
|
cannam@127
|
449 Tj = Cr[WS(csr, 12)];
|
cannam@127
|
450 Tk = Cr[WS(csr, 8)];
|
cannam@127
|
451 Tl = Tj + Tk;
|
cannam@127
|
452 Tm = Ti + Tl;
|
cannam@127
|
453 T15 = Tj - Tk;
|
cannam@127
|
454 TT = KP559016994 * (Ti - Tl);
|
cannam@127
|
455 T14 = Tg - Th;
|
cannam@127
|
456 }
|
cannam@127
|
457 {
|
cannam@127
|
458 E TW, TX, T17, TZ, T10, T18;
|
cannam@127
|
459 TW = Ci[WS(csi, 7)];
|
cannam@127
|
460 TX = Ci[WS(csi, 3)];
|
cannam@127
|
461 T17 = TW - TX;
|
cannam@127
|
462 TZ = Ci[WS(csi, 12)];
|
cannam@127
|
463 T10 = Ci[WS(csi, 8)];
|
cannam@127
|
464 T18 = TZ - T10;
|
cannam@127
|
465 TY = TW + TX;
|
cannam@127
|
466 T1b = T17 + T18;
|
cannam@127
|
467 T11 = TZ + T10;
|
cannam@127
|
468 T19 = KP559016994 * (T17 - T18);
|
cannam@127
|
469 }
|
cannam@127
|
470 Tn = Tf + Tm;
|
cannam@127
|
471 {
|
cannam@127
|
472 E T16, T1S, T1d, T1T, T1c;
|
cannam@127
|
473 T16 = FMA(KP951056516, T14, KP587785252 * T15);
|
cannam@127
|
474 T1S = FNMS(KP951056516, T15, KP587785252 * T14);
|
cannam@127
|
475 T1c = FNMS(KP250000000, T1b, T1a);
|
cannam@127
|
476 T1d = T19 + T1c;
|
cannam@127
|
477 T1T = T1c - T19;
|
cannam@127
|
478 T1e = T16 + T1d;
|
cannam@127
|
479 T2a = T1T - T1S;
|
cannam@127
|
480 T1u = T1d - T16;
|
cannam@127
|
481 T1U = T1S + T1T;
|
cannam@127
|
482 }
|
cannam@127
|
483 {
|
cannam@127
|
484 E T12, T1Q, TV, T1P, TU;
|
cannam@127
|
485 T12 = FMA(KP951056516, TY, KP587785252 * T11);
|
cannam@127
|
486 T1Q = FNMS(KP951056516, T11, KP587785252 * TY);
|
cannam@127
|
487 TU = FNMS(KP250000000, Tm, Tf);
|
cannam@127
|
488 TV = TT + TU;
|
cannam@127
|
489 T1P = TU - TT;
|
cannam@127
|
490 T13 = TV - T12;
|
cannam@127
|
491 T29 = T1P + T1Q;
|
cannam@127
|
492 T1t = TV + T12;
|
cannam@127
|
493 T1R = T1P - T1Q;
|
cannam@127
|
494 }
|
cannam@127
|
495 }
|
cannam@127
|
496 {
|
cannam@127
|
497 E T2m, To, T2l, T2q, T2s, T2o, T2p, T2r, T2n;
|
cannam@127
|
498 T2m = KP1_118033988 * (Te - Tn);
|
cannam@127
|
499 To = Te + Tn;
|
cannam@127
|
500 T2l = FNMS(KP500000000, To, T5);
|
cannam@127
|
501 T2o = TO + TN;
|
cannam@127
|
502 T2p = T1b + T1a;
|
cannam@127
|
503 T2q = FNMS(KP1_902113032, T2p, KP1_175570504 * T2o);
|
cannam@127
|
504 T2s = FMA(KP1_902113032, T2o, KP1_175570504 * T2p);
|
cannam@127
|
505 R0[0] = FMA(KP2_000000000, To, T5);
|
cannam@127
|
506 T2r = T2m + T2l;
|
cannam@127
|
507 R1[WS(rs, 2)] = T2r - T2s;
|
cannam@127
|
508 R0[WS(rs, 10)] = T2r + T2s;
|
cannam@127
|
509 T2n = T2l - T2m;
|
cannam@127
|
510 R0[WS(rs, 5)] = T2n - T2q;
|
cannam@127
|
511 R1[WS(rs, 7)] = T2n + T2q;
|
cannam@127
|
512 }
|
cannam@127
|
513 {
|
cannam@127
|
514 E T2i, T2k, T25, T2c, T2d, T2e, T2j, T2f;
|
cannam@127
|
515 {
|
cannam@127
|
516 E T2g, T2h, T28, T2b;
|
cannam@127
|
517 T2g = FMA(KP684547105, T26, KP728968627 * T27);
|
cannam@127
|
518 T2h = FMA(KP998026728, T29, KP062790519 * T2a);
|
cannam@127
|
519 T2i = FNMS(KP1_902113032, T2h, KP1_175570504 * T2g);
|
cannam@127
|
520 T2k = FMA(KP1_902113032, T2g, KP1_175570504 * T2h);
|
cannam@127
|
521 T25 = T1F + T1G;
|
cannam@127
|
522 T28 = FNMS(KP684547105, T27, KP728968627 * T26);
|
cannam@127
|
523 T2b = FNMS(KP998026728, T2a, KP062790519 * T29);
|
cannam@127
|
524 T2c = T28 + T2b;
|
cannam@127
|
525 T2d = FNMS(KP500000000, T2c, T25);
|
cannam@127
|
526 T2e = KP1_118033988 * (T28 - T2b);
|
cannam@127
|
527 }
|
cannam@127
|
528 R1[WS(rs, 1)] = FMA(KP2_000000000, T2c, T25);
|
cannam@127
|
529 T2j = T2e + T2d;
|
cannam@127
|
530 R0[WS(rs, 4)] = T2j - T2k;
|
cannam@127
|
531 R1[WS(rs, 11)] = T2j + T2k;
|
cannam@127
|
532 T2f = T2d - T2e;
|
cannam@127
|
533 R1[WS(rs, 6)] = T2f - T2i;
|
cannam@127
|
534 R0[WS(rs, 9)] = T2f + T2i;
|
cannam@127
|
535 }
|
cannam@127
|
536 {
|
cannam@127
|
537 E T1m, T1o, Tv, T1g, T1h, T1i, T1n, T1j;
|
cannam@127
|
538 {
|
cannam@127
|
539 E T1k, T1l, TS, T1f;
|
cannam@127
|
540 T1k = FMA(KP248689887, TG, KP968583161 * TR);
|
cannam@127
|
541 T1l = FMA(KP481753674, T13, KP876306680 * T1e);
|
cannam@127
|
542 T1m = FNMS(KP1_902113032, T1l, KP1_175570504 * T1k);
|
cannam@127
|
543 T1o = FMA(KP1_902113032, T1k, KP1_175570504 * T1l);
|
cannam@127
|
544 Tv = Tr - Tu;
|
cannam@127
|
545 TS = FNMS(KP248689887, TR, KP968583161 * TG);
|
cannam@127
|
546 T1f = FNMS(KP481753674, T1e, KP876306680 * T13);
|
cannam@127
|
547 T1g = TS + T1f;
|
cannam@127
|
548 T1h = FNMS(KP500000000, T1g, Tv);
|
cannam@127
|
549 T1i = KP1_118033988 * (TS - T1f);
|
cannam@127
|
550 }
|
cannam@127
|
551 R1[0] = FMA(KP2_000000000, T1g, Tv);
|
cannam@127
|
552 T1n = T1i + T1h;
|
cannam@127
|
553 R0[WS(rs, 3)] = T1n - T1o;
|
cannam@127
|
554 R1[WS(rs, 10)] = T1n + T1o;
|
cannam@127
|
555 T1j = T1h - T1i;
|
cannam@127
|
556 R1[WS(rs, 5)] = T1j - T1m;
|
cannam@127
|
557 R0[WS(rs, 8)] = T1j + T1m;
|
cannam@127
|
558 }
|
cannam@127
|
559 {
|
cannam@127
|
560 E T1C, T1E, T1p, T1w, T1x, T1y, T1D, T1z;
|
cannam@127
|
561 {
|
cannam@127
|
562 E T1A, T1B, T1s, T1v;
|
cannam@127
|
563 T1A = FMA(KP844327925, T1q, KP535826794 * T1r);
|
cannam@127
|
564 T1B = FNMS(KP425779291, T1u, KP904827052 * T1t);
|
cannam@127
|
565 T1C = FNMS(KP1_902113032, T1B, KP1_175570504 * T1A);
|
cannam@127
|
566 T1E = FMA(KP1_902113032, T1A, KP1_175570504 * T1B);
|
cannam@127
|
567 T1p = Tr + Tu;
|
cannam@127
|
568 T1s = FNMS(KP844327925, T1r, KP535826794 * T1q);
|
cannam@127
|
569 T1v = FMA(KP425779291, T1t, KP904827052 * T1u);
|
cannam@127
|
570 T1w = T1s - T1v;
|
cannam@127
|
571 T1x = FNMS(KP500000000, T1w, T1p);
|
cannam@127
|
572 T1y = KP1_118033988 * (T1s + T1v);
|
cannam@127
|
573 }
|
cannam@127
|
574 R0[WS(rs, 2)] = FMA(KP2_000000000, T1w, T1p);
|
cannam@127
|
575 T1D = T1x + T1y;
|
cannam@127
|
576 R1[WS(rs, 4)] = T1D - T1E;
|
cannam@127
|
577 R0[WS(rs, 12)] = T1E + T1D;
|
cannam@127
|
578 T1z = T1x - T1y;
|
cannam@127
|
579 R0[WS(rs, 7)] = T1z - T1C;
|
cannam@127
|
580 R1[WS(rs, 9)] = T1C + T1z;
|
cannam@127
|
581 }
|
cannam@127
|
582 {
|
cannam@127
|
583 E T22, T24, T1H, T1W, T1X, T1Y, T23, T1Z;
|
cannam@127
|
584 {
|
cannam@127
|
585 E T20, T21, T1O, T1V;
|
cannam@127
|
586 T20 = FMA(KP481753674, T1K, KP876306680 * T1N);
|
cannam@127
|
587 T21 = FMA(KP844327925, T1R, KP535826794 * T1U);
|
cannam@127
|
588 T22 = FNMS(KP1_902113032, T21, KP1_175570504 * T20);
|
cannam@127
|
589 T24 = FMA(KP1_902113032, T20, KP1_175570504 * T21);
|
cannam@127
|
590 T1H = T1F - T1G;
|
cannam@127
|
591 T1O = FNMS(KP481753674, T1N, KP876306680 * T1K);
|
cannam@127
|
592 T1V = FNMS(KP844327925, T1U, KP535826794 * T1R);
|
cannam@127
|
593 T1W = T1O + T1V;
|
cannam@127
|
594 T1X = FNMS(KP500000000, T1W, T1H);
|
cannam@127
|
595 T1Y = KP1_118033988 * (T1O - T1V);
|
cannam@127
|
596 }
|
cannam@127
|
597 R0[WS(rs, 1)] = FMA(KP2_000000000, T1W, T1H);
|
cannam@127
|
598 T23 = T1Y + T1X;
|
cannam@127
|
599 R1[WS(rs, 3)] = T23 - T24;
|
cannam@127
|
600 R0[WS(rs, 11)] = T23 + T24;
|
cannam@127
|
601 T1Z = T1X - T1Y;
|
cannam@127
|
602 R0[WS(rs, 6)] = T1Z - T22;
|
cannam@127
|
603 R1[WS(rs, 8)] = T1Z + T22;
|
cannam@127
|
604 }
|
cannam@127
|
605 }
|
cannam@127
|
606 }
|
cannam@127
|
607 }
|
cannam@127
|
608
|
cannam@127
|
609 static const kr2c_desc desc = { 25, "r2cb_25", {100, 46, 52, 0}, &GENUS };
|
cannam@127
|
610
|
cannam@127
|
611 void X(codelet_r2cb_25) (planner *p) {
|
cannam@127
|
612 X(kr2c_register) (p, r2cb_25, &desc);
|
cannam@127
|
613 }
|
cannam@127
|
614
|
cannam@127
|
615 #endif /* HAVE_FMA */
|