← Back to branch summary

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

« back to all changes in this revision

Viewing changes to rdft/scalar/r2cb/r2cbIII_12.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:
rdft/scalar/r2cb/r2cbIII_12.c
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 06:46:24 EDT 2009 */
 
22
/* Generated on Wed Jul 27 06:19:06 EDT 2011 */
23
23
 
24
24
#include "codelet-rdft.h"
25
25
 
26
26
#ifdef HAVE_FMA
27
27
 
28
 
/* Generated by: ../../../genfft/gen_r2cb -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 12 -name r2cbIII_12 -dft-III -include r2cbIII.h */
 
28
/* Generated by: ../../../genfft/gen_r2cb.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 12 -name r2cbIII_12 -dft-III -include r2cbIII.h */
29
29
 
30
30
/*
31
31
 * This function contains 42 FP additions, 20 FP multiplications,
40
40
     DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
41
41
     DK(KP1_732050807, +1.732050807568877293527446341505872366942805254);
42
42
     DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
43
 
     INT i;
44
 
     for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(csr), MAKE_VOLATILE_STRIDE(csi)) {
45
 
          E TE, TD, TF, TG;
46
 
          {
47
 
               E Tx, T6, Te, Tb, T5, Tw, Ts, To, Th, Ti, T9, TA;
48
 
               {
49
 
                    E T1, Tq, Tc, Td, T4, T2, T3, T7, T8, Tr;
 
43
     {
 
44
          INT i;
 
45
          for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(csr), MAKE_VOLATILE_STRIDE(csi)) {
 
46
               E TE, TD, TF, TG;
 
47
               {
 
48
                    E Tx, T6, Te, Tb, T5, Tw, Ts, To, Th, Ti, T9, TA;
 
49
                    {
 
50
                         E T1, Tq, Tc, Td, T4, T2, T3, T7, T8, Tr;
 
51
                         T1 = Cr[WS(csr, 1)];
 
52
                         T2 = Cr[WS(csr, 5)];
 
53
                         T3 = Cr[WS(csr, 2)];
 
54
                         Tq = Ci[WS(csi, 1)];
 
55
                         Tc = Ci[WS(csi, 5)];
 
56
                         Td = Ci[WS(csi, 2)];
 
57
                         T4 = T2 + T3;
 
58
                         Tx = T2 - T3;
 
59
                         T6 = Cr[WS(csr, 4)];
 
60
                         Te = Tc + Td;
 
61
                         Tr = Td - Tc;
 
62
                         Tb = FNMS(KP2_000000000, T1, T4);
 
63
                         T5 = T1 + T4;
 
64
                         T7 = Cr[0];
 
65
                         Tw = FMA(KP2_000000000, Tq, Tr);
 
66
                         Ts = Tq - Tr;
 
67
                         T8 = Cr[WS(csr, 3)];
 
68
                         To = Ci[WS(csi, 4)];
 
69
                         Th = Ci[0];
 
70
                         Ti = Ci[WS(csi, 3)];
 
71
                         T9 = T7 + T8;
 
72
                         TA = T7 - T8;
 
73
                    }
 
74
                    {
 
75
                         E Tl, Tm, Tv, TC;
 
76
                         {
 
77
                              E Tf, Ty, Tk, TB;
 
78
                              {
 
79
                                   E Tj, Tn, Tg, Ta;
 
80
                                   Tl = FNMS(KP1_732050807, Te, Tb);
 
81
                                   Tf = FMA(KP1_732050807, Te, Tb);
 
82
                                   Tj = Th + Ti;
 
83
                                   Tn = Ti - Th;
 
84
                                   Tg = FNMS(KP2_000000000, T6, T9);
 
85
                                   Ta = T6 + T9;
 
86
                                   {
 
87
                                        E Tu, Tt, Tz, Tp;
 
88
                                        Ty = FMA(KP1_732050807, Tx, Tw);
 
89
                                        TE = FNMS(KP1_732050807, Tx, Tw);
 
90
                                        Tz = FMA(KP2_000000000, To, Tn);
 
91
                                        Tp = Tn - To;
 
92
                                        Tm = FMA(KP1_732050807, Tj, Tg);
 
93
                                        Tk = FNMS(KP1_732050807, Tj, Tg);
 
94
                                        Tu = T5 - Ta;
 
95
                                        R0[0] = KP2_000000000 * (T5 + Ta);
 
96
                                        Tt = Tp - Ts;
 
97
                                        R0[WS(rs, 3)] = KP2_000000000 * (Ts + Tp);
 
98
                                        Tv = Tk - Tf;
 
99
                                        TD = FMA(KP1_732050807, TA, Tz);
 
100
                                        TB = FNMS(KP1_732050807, TA, Tz);
 
101
                                        R1[WS(rs, 4)] = KP1_414213562 * (Tu + Tt);
 
102
                                        R1[WS(rs, 1)] = KP1_414213562 * (Tt - Tu);
 
103
                                   }
 
104
                              }
 
105
                              R0[WS(rs, 2)] = Tf + Tk;
 
106
                              TC = Ty + TB;
 
107
                              R0[WS(rs, 5)] = TB - Ty;
 
108
                         }
 
109
                         R1[WS(rs, 3)] = KP707106781 * (Tv + TC);
 
110
                         R1[0] = KP707106781 * (Tv - TC);
 
111
                         TF = Tl - Tm;
 
112
                         R0[WS(rs, 4)] = -(Tl + Tm);
 
113
                    }
 
114
               }
 
115
               R0[WS(rs, 1)] = TD - TE;
 
116
               TG = TE + TD;
 
117
               R1[WS(rs, 5)] = KP707106781 * (TF - TG);
 
118
               R1[WS(rs, 2)] = KP707106781 * (TF + TG);
 
119
          }
 
120
     }
 
121
}
 
122
 
 
123
static const kr2c_desc desc = { 12, "r2cbIII_12", {30, 8, 12, 0}, &GENUS };
 
124
 
 
125
void X(codelet_r2cbIII_12) (planner *p) {
 
126
     X(kr2c_register) (p, r2cbIII_12, &desc);
 
127
}
 
128
 
 
129
#else                           /* HAVE_FMA */
 
130
 
 
131
/* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 12 -name r2cbIII_12 -dft-III -include r2cbIII.h */
 
132
 
 
133
/*
 
134
 * This function contains 42 FP additions, 20 FP multiplications,
 
135
 * (or, 38 additions, 16 multiplications, 4 fused multiply/add),
 
136
 * 25 stack variables, 4 constants, and 24 memory accesses
 
137
 */
 
138
#include "r2cbIII.h"
 
139
 
 
140
static void r2cbIII_12(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
 
141
{
 
142
     DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
 
143
     DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
 
144
     DK(KP500000000, +0.500000000000000000000000000000000000000000000);
 
145
     DK(KP866025403, +0.866025403784438646763723170752936183471402627);
 
146
     {
 
147
          INT i;
 
148
          for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(csr), MAKE_VOLATILE_STRIDE(csi)) {
 
149
               E T5, Tw, Tb, Te, Tx, Ts, Ta, TA, Tg, Tj, Tz, Tp, Tt, Tu;
 
150
               {
 
151
                    E T1, T2, T3, T4;
50
152
                    T1 = Cr[WS(csr, 1)];
51
153
                    T2 = Cr[WS(csr, 5)];
52
154
                    T3 = Cr[WS(csr, 2)];
 
155
                    T4 = T2 + T3;
 
156
                    T5 = T1 + T4;
 
157
                    Tw = KP866025403 * (T2 - T3);
 
158
                    Tb = FNMS(KP500000000, T4, T1);
 
159
               }
 
160
               {
 
161
                    E Tq, Tc, Td, Tr;
53
162
                    Tq = Ci[WS(csi, 1)];
54
163
                    Tc = Ci[WS(csi, 5)];
55
164
                    Td = Ci[WS(csi, 2)];
56
 
                    T4 = T2 + T3;
57
 
                    Tx = T2 - T3;
 
165
                    Tr = Td - Tc;
 
166
                    Te = KP866025403 * (Tc + Td);
 
167
                    Tx = FMA(KP500000000, Tr, Tq);
 
168
                    Ts = Tq - Tr;
 
169
               }
 
170
               {
 
171
                    E T6, T7, T8, T9;
58
172
                    T6 = Cr[WS(csr, 4)];
59
 
                    Te = Tc + Td;
60
 
                    Tr = Td - Tc;
61
 
                    Tb = FNMS(KP2_000000000, T1, T4);
62
 
                    T5 = T1 + T4;
63
173
                    T7 = Cr[0];
64
 
                    Tw = FMA(KP2_000000000, Tq, Tr);
65
 
                    Ts = Tq - Tr;
66
174
                    T8 = Cr[WS(csr, 3)];
 
175
                    T9 = T7 + T8;
 
176
                    Ta = T6 + T9;
 
177
                    TA = KP866025403 * (T7 - T8);
 
178
                    Tg = FNMS(KP500000000, T9, T6);
 
179
               }
 
180
               {
 
181
                    E To, Th, Ti, Tn;
67
182
                    To = Ci[WS(csi, 4)];
68
183
                    Th = Ci[0];
69
184
                    Ti = Ci[WS(csi, 3)];
70
 
                    T9 = T7 + T8;
71
 
                    TA = T7 - T8;
72
 
               }
73
 
               {
74
 
                    E Tl, Tm, Tv, TC;
75
 
                    {
76
 
                         E Tf, Ty, Tk, TB;
77
 
                         {
78
 
                              E Tj, Tn, Tg, Ta;
79
 
                              Tl = FNMS(KP1_732050807, Te, Tb);
80
 
                              Tf = FMA(KP1_732050807, Te, Tb);
81
 
                              Tj = Th + Ti;
82
 
                              Tn = Ti - Th;
83
 
                              Tg = FNMS(KP2_000000000, T6, T9);
84
 
                              Ta = T6 + T9;
85
 
                              {
86
 
                                   E Tu, Tt, Tz, Tp;
87
 
                                   Ty = FMA(KP1_732050807, Tx, Tw);
88
 
                                   TE = FNMS(KP1_732050807, Tx, Tw);
89
 
                                   Tz = FMA(KP2_000000000, To, Tn);
90
 
                                   Tp = Tn - To;
91
 
                                   Tm = FMA(KP1_732050807, Tj, Tg);
92
 
                                   Tk = FNMS(KP1_732050807, Tj, Tg);
93
 
                                   Tu = T5 - Ta;
94
 
                                   R0[0] = KP2_000000000 * (T5 + Ta);
95
 
                                   Tt = Tp - Ts;
96
 
                                   R0[WS(rs, 3)] = KP2_000000000 * (Ts + Tp);
97
 
                                   Tv = Tk - Tf;
98
 
                                   TD = FMA(KP1_732050807, TA, Tz);
99
 
                                   TB = FNMS(KP1_732050807, TA, Tz);
100
 
                                   R1[WS(rs, 4)] = KP1_414213562 * (Tu + Tt);
101
 
                                   R1[WS(rs, 1)] = KP1_414213562 * (Tt - Tu);
102
 
                              }
103
 
                         }
104
 
                         R0[WS(rs, 2)] = Tf + Tk;
105
 
                         TC = Ty + TB;
106
 
                         R0[WS(rs, 5)] = TB - Ty;
107
 
                    }
108
 
                    R1[WS(rs, 3)] = KP707106781 * (Tv + TC);
109
 
                    R1[0] = KP707106781 * (Tv - TC);
110
 
                    TF = Tl - Tm;
111
 
                    R0[WS(rs, 4)] = -(Tl + Tm);
112
 
               }
113
 
          }
114
 
          R0[WS(rs, 1)] = TD - TE;
115
 
          TG = TE + TD;
116
 
          R1[WS(rs, 5)] = KP707106781 * (TF - TG);
117
 
          R1[WS(rs, 2)] = KP707106781 * (TF + TG);
118
 
     }
119
 
}
120
 
 
121
 
static const kr2c_desc desc = { 12, "r2cbIII_12", {30, 8, 12, 0}, &GENUS };
122
 
 
123
 
void X(codelet_r2cbIII_12) (planner *p) {
124
 
     X(kr2c_register) (p, r2cbIII_12, &desc);
125
 
}
126
 
 
127
 
#else                           /* HAVE_FMA */
128
 
 
129
 
/* Generated by: ../../../genfft/gen_r2cb -compact -variables 4 -pipeline-latency 4 -sign 1 -n 12 -name r2cbIII_12 -dft-III -include r2cbIII.h */
130
 
 
131
 
/*
132
 
 * This function contains 42 FP additions, 20 FP multiplications,
133
 
 * (or, 38 additions, 16 multiplications, 4 fused multiply/add),
134
 
 * 25 stack variables, 4 constants, and 24 memory accesses
135
 
 */
136
 
#include "r2cbIII.h"
137
 
 
138
 
static void r2cbIII_12(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
139
 
{
140
 
     DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
141
 
     DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
142
 
     DK(KP500000000, +0.500000000000000000000000000000000000000000000);
143
 
     DK(KP866025403, +0.866025403784438646763723170752936183471402627);
144
 
     INT i;
145
 
     for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(csr), MAKE_VOLATILE_STRIDE(csi)) {
146
 
          E T5, Tw, Tb, Te, Tx, Ts, Ta, TA, Tg, Tj, Tz, Tp, Tt, Tu;
147
 
          {
148
 
               E T1, T2, T3, T4;
149
 
               T1 = Cr[WS(csr, 1)];
150
 
               T2 = Cr[WS(csr, 5)];
151
 
               T3 = Cr[WS(csr, 2)];
152
 
               T4 = T2 + T3;
153
 
               T5 = T1 + T4;
154
 
               Tw = KP866025403 * (T2 - T3);
155
 
               Tb = FNMS(KP500000000, T4, T1);
156
 
          }
157
 
          {
158
 
               E Tq, Tc, Td, Tr;
159
 
               Tq = Ci[WS(csi, 1)];
160
 
               Tc = Ci[WS(csi, 5)];
161
 
               Td = Ci[WS(csi, 2)];
162
 
               Tr = Td - Tc;
163
 
               Te = KP866025403 * (Tc + Td);
164
 
               Tx = FMA(KP500000000, Tr, Tq);
165
 
               Ts = Tq - Tr;
166
 
          }
167
 
          {
168
 
               E T6, T7, T8, T9;
169
 
               T6 = Cr[WS(csr, 4)];
170
 
               T7 = Cr[0];
171
 
               T8 = Cr[WS(csr, 3)];
172
 
               T9 = T7 + T8;
173
 
               Ta = T6 + T9;
174
 
               TA = KP866025403 * (T7 - T8);
175
 
               Tg = FNMS(KP500000000, T9, T6);
176
 
          }
177
 
          {
178
 
               E To, Th, Ti, Tn;
179
 
               To = Ci[WS(csi, 4)];
180
 
               Th = Ci[0];
181
 
               Ti = Ci[WS(csi, 3)];
182
 
               Tn = Ti - Th;
183
 
               Tj = KP866025403 * (Th + Ti);
184
 
               Tz = FMA(KP500000000, Tn, To);
185
 
               Tp = Tn - To;
186
 
          }
187
 
          R0[0] = KP2_000000000 * (T5 + Ta);
188
 
          R0[WS(rs, 3)] = KP2_000000000 * (Ts + Tp);
189
 
          Tt = Tp - Ts;
190
 
          Tu = T5 - Ta;
191
 
          R1[WS(rs, 1)] = KP1_414213562 * (Tt - Tu);
192
 
          R1[WS(rs, 4)] = KP1_414213562 * (Tu + Tt);
193
 
          {
194
 
               E Tf, Tk, Tv, Ty, TB, TC;
195
 
               Tf = Tb - Te;
196
 
               Tk = Tg + Tj;
197
 
               Tv = Tf - Tk;
198
 
               Ty = Tw + Tx;
199
 
               TB = Tz - TA;
200
 
               TC = Ty + TB;
201
 
               R0[WS(rs, 2)] = -(KP2_000000000 * (Tf + Tk));
202
 
               R0[WS(rs, 5)] = KP2_000000000 * (TB - Ty);
203
 
               R1[0] = KP1_414213562 * (Tv - TC);
204
 
               R1[WS(rs, 3)] = KP1_414213562 * (Tv + TC);
205
 
          }
206
 
          {
207
 
               E Tl, Tm, TF, TD, TE, TG;
208
 
               Tl = Tb + Te;
209
 
               Tm = Tg - Tj;
210
 
               TF = Tm - Tl;
211
 
               TD = TA + Tz;
212
 
               TE = Tx - Tw;
213
 
               TG = TE + TD;
214
 
               R0[WS(rs, 4)] = KP2_000000000 * (Tl + Tm);
215
 
               R1[WS(rs, 2)] = KP1_414213562 * (TF + TG);
216
 
               R0[WS(rs, 1)] = KP2_000000000 * (TD - TE);
217
 
               R1[WS(rs, 5)] = KP1_414213562 * (TF - TG);
 
185
                    Tn = Ti - Th;
 
186
                    Tj = KP866025403 * (Th + Ti);
 
187
                    Tz = FMA(KP500000000, Tn, To);
 
188
                    Tp = Tn - To;
 
189
               }
 
190
               R0[0] = KP2_000000000 * (T5 + Ta);
 
191
               R0[WS(rs, 3)] = KP2_000000000 * (Ts + Tp);
 
192
               Tt = Tp - Ts;
 
193
               Tu = T5 - Ta;
 
194
               R1[WS(rs, 1)] = KP1_414213562 * (Tt - Tu);
 
195
               R1[WS(rs, 4)] = KP1_414213562 * (Tu + Tt);
 
196
               {
 
197
                    E Tf, Tk, Tv, Ty, TB, TC;
 
198
                    Tf = Tb - Te;
 
199
                    Tk = Tg + Tj;
 
200
                    Tv = Tf - Tk;
 
201
                    Ty = Tw + Tx;
 
202
                    TB = Tz - TA;
 
203
                    TC = Ty + TB;
 
204
                    R0[WS(rs, 2)] = -(KP2_000000000 * (Tf + Tk));
 
205
                    R0[WS(rs, 5)] = KP2_000000000 * (TB - Ty);
 
206
                    R1[0] = KP1_414213562 * (Tv - TC);
 
207
                    R1[WS(rs, 3)] = KP1_414213562 * (Tv + TC);
 
208
               }
 
209
               {
 
210
                    E Tl, Tm, TF, TD, TE, TG;
 
211
                    Tl = Tb + Te;
 
212
                    Tm = Tg - Tj;
 
213
                    TF = Tm - Tl;
 
214
                    TD = TA + Tz;
 
215
                    TE = Tx - Tw;
 
216
                    TG = TE + TD;
 
217
                    R0[WS(rs, 4)] = KP2_000000000 * (Tl + Tm);
 
218
                    R1[WS(rs, 2)] = KP1_414213562 * (TF + TG);
 
219
                    R0[WS(rs, 1)] = KP2_000000000 * (TD - TE);
 
220
                    R1[WS(rs, 5)] = KP1_414213562 * (TF - TG);
 
221
               }
218
222
          }
219
223
     }
220
224
}