~ubuntu-branches/ubuntu/utopic/fftw3/utopic

« back to all changes in this revision

Viewing changes to rdft/scalar/r2cf/r2cf_11.c

  • Committer: Package Import Robot
  • Author(s): Matthias Klose
  • Date: 2011-12-14 13:21:22 UTC
  • mfrom: (3.1.5 sid)
  • Revision ID: package-import@ubuntu.com-20111214132122-l4avyl2kkr7vq5aj
Tags: 3.3-1ubuntu1
* Merge with Debian; remaining changes:
  - Revert the ARM workaround.

Show diffs side-by-side

added added

removed removed

Lines of Context:
1
1
/*
2
 
 * Copyright (c) 2003, 2007-8 Matteo Frigo
3
 
 * Copyright (c) 2003, 2007-8 Massachusetts Institute of Technology
 
2
 * Copyright (c) 2003, 2007-11 Matteo Frigo
 
3
 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
4
4
 *
5
5
 * This program is free software; you can redistribute it and/or modify
6
6
 * it under the terms of the GNU General Public License as published by
19
19
 */
20
20
 
21
21
/* This file was automatically generated --- DO NOT EDIT */
22
 
/* Generated on Sun Jul 12 06:43:41 EDT 2009 */
 
22
/* Generated on Wed Jul 27 06:16:29 EDT 2011 */
23
23
 
24
24
#include "codelet-rdft.h"
25
25
 
26
26
#ifdef HAVE_FMA
27
27
 
28
 
/* Generated by: ../../../genfft/gen_r2cf -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 11 -name r2cf_11 -include r2cf.h */
 
28
/* Generated by: ../../../genfft/gen_r2cf.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 11 -name r2cf_11 -include r2cf.h */
29
29
 
30
30
/*
31
31
 * This function contains 60 FP additions, 50 FP multiplications,
46
46
     DK(KP634356270, +0.634356270682424498893150776899916060542806975);
47
47
     DK(KP342584725, +0.342584725681637509502641509861112333758894680);
48
48
     DK(KP521108558, +0.521108558113202722944698153526659300680427422);
49
 
     INT i;
50
 
     for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(csr), MAKE_VOLATILE_STRIDE(csi)) {
51
 
          E T1, Tg, TF, TB, TI, TL, Tz, TA;
52
 
          {
53
 
               E T4, TC, TE, T7, TD, Ta, TS, TG, TJ, Td, TP, TM, Ty, Tq, Th;
54
 
               E Tt, Tl;
55
 
               T1 = R0[0];
56
 
               {
57
 
                    E Tb, Tc, Tx, Tp;
58
 
                    {
59
 
                         E T2, T3, Te, Tf;
60
 
                         T2 = R1[0];
61
 
                         T3 = R0[WS(rs, 5)];
62
 
                         Te = R1[WS(rs, 2)];
63
 
                         Tf = R0[WS(rs, 3)];
64
 
                         {
65
 
                              E T5, T6, T8, T9;
66
 
                              T5 = R0[WS(rs, 1)];
67
 
                              T4 = T2 + T3;
68
 
                              TC = T3 - T2;
69
 
                              Tg = Te + Tf;
70
 
                              TE = Tf - Te;
71
 
                              T6 = R1[WS(rs, 4)];
72
 
                              T8 = R1[WS(rs, 1)];
73
 
                              T9 = R0[WS(rs, 4)];
74
 
                              Tb = R0[WS(rs, 2)];
75
 
                              T7 = T5 + T6;
76
 
                              TD = T5 - T6;
77
 
                              Ta = T8 + T9;
78
 
                              TF = T9 - T8;
79
 
                              Tc = R1[WS(rs, 3)];
80
 
                         }
81
 
                    }
82
 
                    TS = FMA(KP521108558, TC, TD);
83
 
                    TG = FMA(KP521108558, TF, TE);
84
 
                    TJ = FMA(KP521108558, TE, TC);
85
 
                    Td = Tb + Tc;
86
 
                    TB = Tb - Tc;
87
 
                    Tx = FNMS(KP342584725, Ta, T7);
88
 
                    Tp = FNMS(KP342584725, T4, Ta);
89
 
                    TP = FNMS(KP521108558, TB, TF);
90
 
                    TM = FNMS(KP521108558, TD, TB);
91
 
                    Ty = FNMS(KP634356270, Tx, Td);
92
 
                    Tq = FNMS(KP634356270, Tp, Tg);
93
 
                    Th = FNMS(KP342584725, Tg, Td);
94
 
                    Tt = FNMS(KP342584725, Td, T4);
95
 
                    Tl = FNMS(KP342584725, T7, Tg);
96
 
               }
97
 
               {
98
 
                    E Tu, Ts, TN, Tv;
99
 
                    {
100
 
                         E Tm, TU, Tj, Ti, TT;
101
 
                         TT = FMA(KP715370323, TS, TF);
102
 
                         Ti = FNMS(KP634356270, Th, Ta);
103
 
                         Tu = FNMS(KP634356270, Tt, T7);
104
 
                         Tm = FNMS(KP634356270, Tl, T4);
105
 
                         TU = FMA(KP830830026, TT, TB);
106
 
                         Tj = FNMS(KP778434453, Ti, T7);
107
 
                         {
108
 
                              E Tk, TR, To, Tn, TQ, Tr;
109
 
                              TQ = FMA(KP715370323, TP, TC);
110
 
                              Tn = FNMS(KP778434453, Tm, Ta);
111
 
                              Ci[WS(csi, 5)] = KP989821441 * (FMA(KP918985947, TU, TE));
112
 
                              Tk = FNMS(KP876768831, Tj, T4);
113
 
                              TR = FNMS(KP830830026, TQ, TE);
114
 
                              To = FNMS(KP876768831, Tn, Td);
115
 
                              Tr = FNMS(KP778434453, Tq, Td);
116
 
                              Cr[WS(csr, 5)] = FNMS(KP959492973, Tk, T1);
117
 
                              Ci[WS(csi, 4)] = KP989821441 * (FNMS(KP918985947, TR, TD));
118
 
                              Cr[WS(csr, 4)] = FNMS(KP959492973, To, T1);
119
 
                              Ts = FNMS(KP876768831, Tr, T7);
120
 
                         }
121
 
                    }
122
 
                    TN = FNMS(KP715370323, TM, TE);
123
 
                    Tv = FNMS(KP778434453, Tu, Tg);
124
 
                    Cr[0] = T1 + T4 + T7 + Ta + Td + Tg;
125
 
                    Cr[WS(csr, 3)] = FNMS(KP959492973, Ts, T1);
126
 
                    {
127
 
                         E TO, Tw, TH, TK;
128
 
                         TO = FNMS(KP830830026, TN, TF);
129
 
                         Tw = FNMS(KP876768831, Tv, Ta);
130
 
                         TH = FMA(KP715370323, TG, TD);
131
 
                         TK = FNMS(KP715370323, TJ, TB);
132
 
                         Ci[WS(csi, 3)] = KP989821441 * (FNMS(KP918985947, TO, TC));
133
 
                         Cr[WS(csr, 2)] = FNMS(KP959492973, Tw, T1);
134
 
                         TI = FNMS(KP830830026, TH, TC);
135
 
                         TL = FMA(KP830830026, TK, TD);
136
 
                         Tz = FNMS(KP778434453, Ty, T4);
137
 
                    }
138
 
               }
 
49
     {
 
50
          INT i;
 
51
          for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(csr), MAKE_VOLATILE_STRIDE(csi)) {
 
52
               E T1, Tg, TF, TB, TI, TL, Tz, TA;
 
53
               {
 
54
                    E T4, TC, TE, T7, TD, Ta, TS, TG, TJ, Td, TP, TM, Ty, Tq, Th;
 
55
                    E Tt, Tl;
 
56
                    T1 = R0[0];
 
57
                    {
 
58
                         E Tb, Tc, Tx, Tp;
 
59
                         {
 
60
                              E T2, T3, Te, Tf;
 
61
                              T2 = R1[0];
 
62
                              T3 = R0[WS(rs, 5)];
 
63
                              Te = R1[WS(rs, 2)];
 
64
                              Tf = R0[WS(rs, 3)];
 
65
                              {
 
66
                                   E T5, T6, T8, T9;
 
67
                                   T5 = R0[WS(rs, 1)];
 
68
                                   T4 = T2 + T3;
 
69
                                   TC = T3 - T2;
 
70
                                   Tg = Te + Tf;
 
71
                                   TE = Tf - Te;
 
72
                                   T6 = R1[WS(rs, 4)];
 
73
                                   T8 = R1[WS(rs, 1)];
 
74
                                   T9 = R0[WS(rs, 4)];
 
75
                                   Tb = R0[WS(rs, 2)];
 
76
                                   T7 = T5 + T6;
 
77
                                   TD = T5 - T6;
 
78
                                   Ta = T8 + T9;
 
79
                                   TF = T9 - T8;
 
80
                                   Tc = R1[WS(rs, 3)];
 
81
                              }
 
82
                         }
 
83
                         TS = FMA(KP521108558, TC, TD);
 
84
                         TG = FMA(KP521108558, TF, TE);
 
85
                         TJ = FMA(KP521108558, TE, TC);
 
86
                         Td = Tb + Tc;
 
87
                         TB = Tb - Tc;
 
88
                         Tx = FNMS(KP342584725, Ta, T7);
 
89
                         Tp = FNMS(KP342584725, T4, Ta);
 
90
                         TP = FNMS(KP521108558, TB, TF);
 
91
                         TM = FNMS(KP521108558, TD, TB);
 
92
                         Ty = FNMS(KP634356270, Tx, Td);
 
93
                         Tq = FNMS(KP634356270, Tp, Tg);
 
94
                         Th = FNMS(KP342584725, Tg, Td);
 
95
                         Tt = FNMS(KP342584725, Td, T4);
 
96
                         Tl = FNMS(KP342584725, T7, Tg);
 
97
                    }
 
98
                    {
 
99
                         E Tu, Ts, TN, Tv;
 
100
                         {
 
101
                              E Tm, TU, Tj, Ti, TT;
 
102
                              TT = FMA(KP715370323, TS, TF);
 
103
                              Ti = FNMS(KP634356270, Th, Ta);
 
104
                              Tu = FNMS(KP634356270, Tt, T7);
 
105
                              Tm = FNMS(KP634356270, Tl, T4);
 
106
                              TU = FMA(KP830830026, TT, TB);
 
107
                              Tj = FNMS(KP778434453, Ti, T7);
 
108
                              {
 
109
                                   E Tk, TR, To, Tn, TQ, Tr;
 
110
                                   TQ = FMA(KP715370323, TP, TC);
 
111
                                   Tn = FNMS(KP778434453, Tm, Ta);
 
112
                                   Ci[WS(csi, 5)] = KP989821441 * (FMA(KP918985947, TU, TE));
 
113
                                   Tk = FNMS(KP876768831, Tj, T4);
 
114
                                   TR = FNMS(KP830830026, TQ, TE);
 
115
                                   To = FNMS(KP876768831, Tn, Td);
 
116
                                   Tr = FNMS(KP778434453, Tq, Td);
 
117
                                   Cr[WS(csr, 5)] = FNMS(KP959492973, Tk, T1);
 
118
                                   Ci[WS(csi, 4)] = KP989821441 * (FNMS(KP918985947, TR, TD));
 
119
                                   Cr[WS(csr, 4)] = FNMS(KP959492973, To, T1);
 
120
                                   Ts = FNMS(KP876768831, Tr, T7);
 
121
                              }
 
122
                         }
 
123
                         TN = FNMS(KP715370323, TM, TE);
 
124
                         Tv = FNMS(KP778434453, Tu, Tg);
 
125
                         Cr[0] = T1 + T4 + T7 + Ta + Td + Tg;
 
126
                         Cr[WS(csr, 3)] = FNMS(KP959492973, Ts, T1);
 
127
                         {
 
128
                              E TO, Tw, TH, TK;
 
129
                              TO = FNMS(KP830830026, TN, TF);
 
130
                              Tw = FNMS(KP876768831, Tv, Ta);
 
131
                              TH = FMA(KP715370323, TG, TD);
 
132
                              TK = FNMS(KP715370323, TJ, TB);
 
133
                              Ci[WS(csi, 3)] = KP989821441 * (FNMS(KP918985947, TO, TC));
 
134
                              Cr[WS(csr, 2)] = FNMS(KP959492973, Tw, T1);
 
135
                              TI = FNMS(KP830830026, TH, TC);
 
136
                              TL = FMA(KP830830026, TK, TD);
 
137
                              Tz = FNMS(KP778434453, Ty, T4);
 
138
                         }
 
139
                    }
 
140
               }
 
141
               Ci[WS(csi, 2)] = KP989821441 * (FMA(KP918985947, TI, TB));
 
142
               Ci[WS(csi, 1)] = KP989821441 * (FNMS(KP918985947, TL, TF));
 
143
               TA = FNMS(KP876768831, Tz, Tg);
 
144
               Cr[WS(csr, 1)] = FNMS(KP959492973, TA, T1);
139
145
          }
140
 
          Ci[WS(csi, 2)] = KP989821441 * (FMA(KP918985947, TI, TB));
141
 
          Ci[WS(csi, 1)] = KP989821441 * (FNMS(KP918985947, TL, TF));
142
 
          TA = FNMS(KP876768831, Tz, Tg);
143
 
          Cr[WS(csr, 1)] = FNMS(KP959492973, TA, T1);
144
146
     }
145
147
}
146
148
 
152
154
 
153
155
#else                           /* HAVE_FMA */
154
156
 
155
 
/* Generated by: ../../../genfft/gen_r2cf -compact -variables 4 -pipeline-latency 4 -n 11 -name r2cf_11 -include r2cf.h */
 
157
/* Generated by: ../../../genfft/gen_r2cf.native -compact -variables 4 -pipeline-latency 4 -n 11 -name r2cf_11 -include r2cf.h */
156
158
 
157
159
/*
158
160
 * This function contains 60 FP additions, 50 FP multiplications,
173
175
     DK(KP281732556, +0.281732556841429697711417915346616899035777899);
174
176
     DK(KP540640817, +0.540640817455597582107635954318691695431770608);
175
177
     DK(KP755749574, +0.755749574354258283774035843972344420179717445);
176
 
     INT i;
177
 
     for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(csr), MAKE_VOLATILE_STRIDE(csi)) {
178
 
          E T1, T4, Tl, Tg, Th, Td, Ti, Ta, Tk, T7, Tj, Tb, Tc;
179
 
          T1 = R0[0];
180
 
          {
181
 
               E T2, T3, Te, Tf;
182
 
               T2 = R0[WS(rs, 1)];
183
 
               T3 = R1[WS(rs, 4)];
184
 
               T4 = T2 + T3;
185
 
               Tl = T3 - T2;
186
 
               Te = R1[0];
187
 
               Tf = R0[WS(rs, 5)];
188
 
               Tg = Te + Tf;
189
 
               Th = Tf - Te;
190
 
          }
191
 
          Tb = R1[WS(rs, 1)];
192
 
          Tc = R0[WS(rs, 4)];
193
 
          Td = Tb + Tc;
194
 
          Ti = Tc - Tb;
195
 
          {
196
 
               E T8, T9, T5, T6;
197
 
               T8 = R1[WS(rs, 2)];
198
 
               T9 = R0[WS(rs, 3)];
199
 
               Ta = T8 + T9;
200
 
               Tk = T9 - T8;
201
 
               T5 = R0[WS(rs, 2)];
202
 
               T6 = R1[WS(rs, 3)];
203
 
               T7 = T5 + T6;
204
 
               Tj = T6 - T5;
205
 
          }
206
 
          Ci[WS(csi, 4)] = FMA(KP755749574, Th, KP540640817 * Ti) + FNMS(KP909631995, Tk, KP281732556 * Tj) - (KP989821441 * Tl);
207
 
          Cr[WS(csr, 4)] = FMA(KP841253532, Td, T1) + FNMS(KP959492973, T7, KP415415013 * Ta) + FNMA(KP142314838, T4, KP654860733 * Tg);
208
 
          Ci[WS(csi, 2)] = FMA(KP909631995, Th, KP755749574 * Tl) + FNMA(KP540640817, Tk, KP989821441 * Tj) - (KP281732556 * Ti);
209
 
          Ci[WS(csi, 5)] = FMA(KP281732556, Th, KP755749574 * Ti) + FNMS(KP909631995, Tj, KP989821441 * Tk) - (KP540640817 * Tl);
210
 
          Ci[WS(csi, 1)] = FMA(KP540640817, Th, KP909631995 * Tl) + FMA(KP989821441, Ti, KP755749574 * Tj) + (KP281732556 * Tk);
211
 
          Ci[WS(csi, 3)] = FMA(KP989821441, Th, KP540640817 * Tj) + FNMS(KP909631995, Ti, KP755749574 * Tk) - (KP281732556 * Tl);
212
 
          Cr[WS(csr, 3)] = FMA(KP415415013, Td, T1) + FNMS(KP654860733, Ta, KP841253532 * T7) + FNMA(KP959492973, T4, KP142314838 * Tg);
213
 
          Cr[WS(csr, 1)] = FMA(KP841253532, Tg, T1) + FNMS(KP959492973, Ta, KP415415013 * T4) + FNMA(KP654860733, T7, KP142314838 * Td);
214
 
          Cr[0] = T1 + Tg + T4 + Td + T7 + Ta;
215
 
          Cr[WS(csr, 2)] = FMA(KP415415013, Tg, T1) + FNMS(KP142314838, T7, KP841253532 * Ta) + FNMA(KP959492973, Td, KP654860733 * T4);
216
 
          Cr[WS(csr, 5)] = FMA(KP841253532, T4, T1) + FNMS(KP142314838, Ta, KP415415013 * T7) + FNMA(KP654860733, Td, KP959492973 * Tg);
 
178
     {
 
179
          INT i;
 
180
          for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(csr), MAKE_VOLATILE_STRIDE(csi)) {
 
181
               E T1, T4, Tl, Tg, Th, Td, Ti, Ta, Tk, T7, Tj, Tb, Tc;
 
182
               T1 = R0[0];
 
183
               {
 
184
                    E T2, T3, Te, Tf;
 
185
                    T2 = R0[WS(rs, 1)];
 
186
                    T3 = R1[WS(rs, 4)];
 
187
                    T4 = T2 + T3;
 
188
                    Tl = T3 - T2;
 
189
                    Te = R1[0];
 
190
                    Tf = R0[WS(rs, 5)];
 
191
                    Tg = Te + Tf;
 
192
                    Th = Tf - Te;
 
193
               }
 
194
               Tb = R1[WS(rs, 1)];
 
195
               Tc = R0[WS(rs, 4)];
 
196
               Td = Tb + Tc;
 
197
               Ti = Tc - Tb;
 
198
               {
 
199
                    E T8, T9, T5, T6;
 
200
                    T8 = R1[WS(rs, 2)];
 
201
                    T9 = R0[WS(rs, 3)];
 
202
                    Ta = T8 + T9;
 
203
                    Tk = T9 - T8;
 
204
                    T5 = R0[WS(rs, 2)];
 
205
                    T6 = R1[WS(rs, 3)];
 
206
                    T7 = T5 + T6;
 
207
                    Tj = T6 - T5;
 
208
               }
 
209
               Ci[WS(csi, 4)] = FMA(KP755749574, Th, KP540640817 * Ti) + FNMS(KP909631995, Tk, KP281732556 * Tj) - (KP989821441 * Tl);
 
210
               Cr[WS(csr, 4)] = FMA(KP841253532, Td, T1) + FNMS(KP959492973, T7, KP415415013 * Ta) + FNMA(KP142314838, T4, KP654860733 * Tg);
 
211
               Ci[WS(csi, 2)] = FMA(KP909631995, Th, KP755749574 * Tl) + FNMA(KP540640817, Tk, KP989821441 * Tj) - (KP281732556 * Ti);
 
212
               Ci[WS(csi, 5)] = FMA(KP281732556, Th, KP755749574 * Ti) + FNMS(KP909631995, Tj, KP989821441 * Tk) - (KP540640817 * Tl);
 
213
               Ci[WS(csi, 1)] = FMA(KP540640817, Th, KP909631995 * Tl) + FMA(KP989821441, Ti, KP755749574 * Tj) + (KP281732556 * Tk);
 
214
               Ci[WS(csi, 3)] = FMA(KP989821441, Th, KP540640817 * Tj) + FNMS(KP909631995, Ti, KP755749574 * Tk) - (KP281732556 * Tl);
 
215
               Cr[WS(csr, 3)] = FMA(KP415415013, Td, T1) + FNMS(KP654860733, Ta, KP841253532 * T7) + FNMA(KP959492973, T4, KP142314838 * Tg);
 
216
               Cr[WS(csr, 1)] = FMA(KP841253532, Tg, T1) + FNMS(KP959492973, Ta, KP415415013 * T4) + FNMA(KP654860733, T7, KP142314838 * Td);
 
217
               Cr[0] = T1 + Tg + T4 + Td + T7 + Ta;
 
218
               Cr[WS(csr, 2)] = FMA(KP415415013, Tg, T1) + FNMS(KP142314838, T7, KP841253532 * Ta) + FNMA(KP959492973, Td, KP654860733 * T4);
 
219
               Cr[WS(csr, 5)] = FMA(KP841253532, T4, T1) + FNMS(KP142314838, Ta, KP415415013 * T7) + FNMA(KP654860733, Td, KP959492973 * Tg);
 
220
          }
217
221
     }
218
222
}
219
223