2
dct64_i386.c: DCT64, a C variant for i386
4
copyright ?-2006 by the mpg123 project - free software under the terms of the LGPL 2.1
5
see COPYING and AUTHORS files in distribution or http://mpg123.de
6
initially written by Michael Hipp
10
* Discrete Cosine Tansform (DCT) for subband synthesis
11
* optimized for machines with no auto-increment.
12
* The performance is highly compiler dependend. Maybe
13
* the dct64.c version for 'normal' processor may be faster
14
* even for Intel processors.
20
static void dct64_1(real *out0,real *out1,real *b1,real *b2,real *samples)
23
register real *costab = pnts[0];
25
b1[0x00] = samples[0x00] + samples[0x1F];
26
b1[0x01] = samples[0x01] + samples[0x1E];
27
b1[0x1F] = (samples[0x00] - samples[0x1F]) * costab[0x0];
28
b1[0x1E] = (samples[0x01] - samples[0x1E]) * costab[0x1];
30
b1[0x02] = samples[0x02] + samples[0x1D];
31
b1[0x03] = samples[0x03] + samples[0x1C];
32
b1[0x1D] = (samples[0x02] - samples[0x1D]) * costab[0x2];
33
b1[0x1C] = (samples[0x03] - samples[0x1C]) * costab[0x3];
35
b1[0x04] = samples[0x04] + samples[0x1B];
36
b1[0x05] = samples[0x05] + samples[0x1A];
37
b1[0x1B] = (samples[0x04] - samples[0x1B]) * costab[0x4];
38
b1[0x1A] = (samples[0x05] - samples[0x1A]) * costab[0x5];
40
b1[0x06] = samples[0x06] + samples[0x19];
41
b1[0x07] = samples[0x07] + samples[0x18];
42
b1[0x19] = (samples[0x06] - samples[0x19]) * costab[0x6];
43
b1[0x18] = (samples[0x07] - samples[0x18]) * costab[0x7];
45
b1[0x08] = samples[0x08] + samples[0x17];
46
b1[0x09] = samples[0x09] + samples[0x16];
47
b1[0x17] = (samples[0x08] - samples[0x17]) * costab[0x8];
48
b1[0x16] = (samples[0x09] - samples[0x16]) * costab[0x9];
50
b1[0x0A] = samples[0x0A] + samples[0x15];
51
b1[0x0B] = samples[0x0B] + samples[0x14];
52
b1[0x15] = (samples[0x0A] - samples[0x15]) * costab[0xA];
53
b1[0x14] = (samples[0x0B] - samples[0x14]) * costab[0xB];
55
b1[0x0C] = samples[0x0C] + samples[0x13];
56
b1[0x0D] = samples[0x0D] + samples[0x12];
57
b1[0x13] = (samples[0x0C] - samples[0x13]) * costab[0xC];
58
b1[0x12] = (samples[0x0D] - samples[0x12]) * costab[0xD];
60
b1[0x0E] = samples[0x0E] + samples[0x11];
61
b1[0x0F] = samples[0x0F] + samples[0x10];
62
b1[0x11] = (samples[0x0E] - samples[0x11]) * costab[0xE];
63
b1[0x10] = (samples[0x0F] - samples[0x10]) * costab[0xF];
69
register real *costab = pnts[1];
71
b2[0x00] = b1[0x00] + b1[0x0F];
72
b2[0x01] = b1[0x01] + b1[0x0E];
73
b2[0x0F] = (b1[0x00] - b1[0x0F]) * costab[0];
74
b2[0x0E] = (b1[0x01] - b1[0x0E]) * costab[1];
76
b2[0x02] = b1[0x02] + b1[0x0D];
77
b2[0x03] = b1[0x03] + b1[0x0C];
78
b2[0x0D] = (b1[0x02] - b1[0x0D]) * costab[2];
79
b2[0x0C] = (b1[0x03] - b1[0x0C]) * costab[3];
81
b2[0x04] = b1[0x04] + b1[0x0B];
82
b2[0x05] = b1[0x05] + b1[0x0A];
83
b2[0x0B] = (b1[0x04] - b1[0x0B]) * costab[4];
84
b2[0x0A] = (b1[0x05] - b1[0x0A]) * costab[5];
86
b2[0x06] = b1[0x06] + b1[0x09];
87
b2[0x07] = b1[0x07] + b1[0x08];
88
b2[0x09] = (b1[0x06] - b1[0x09]) * costab[6];
89
b2[0x08] = (b1[0x07] - b1[0x08]) * costab[7];
93
b2[0x10] = b1[0x10] + b1[0x1F];
94
b2[0x11] = b1[0x11] + b1[0x1E];
95
b2[0x1F] = (b1[0x1F] - b1[0x10]) * costab[0];
96
b2[0x1E] = (b1[0x1E] - b1[0x11]) * costab[1];
98
b2[0x12] = b1[0x12] + b1[0x1D];
99
b2[0x13] = b1[0x13] + b1[0x1C];
100
b2[0x1D] = (b1[0x1D] - b1[0x12]) * costab[2];
101
b2[0x1C] = (b1[0x1C] - b1[0x13]) * costab[3];
103
b2[0x14] = b1[0x14] + b1[0x1B];
104
b2[0x15] = b1[0x15] + b1[0x1A];
105
b2[0x1B] = (b1[0x1B] - b1[0x14]) * costab[4];
106
b2[0x1A] = (b1[0x1A] - b1[0x15]) * costab[5];
108
b2[0x16] = b1[0x16] + b1[0x19];
109
b2[0x17] = b1[0x17] + b1[0x18];
110
b2[0x19] = (b1[0x19] - b1[0x16]) * costab[6];
111
b2[0x18] = (b1[0x18] - b1[0x17]) * costab[7];
115
register real *costab = pnts[2];
117
b1[0x00] = b2[0x00] + b2[0x07];
118
b1[0x07] = (b2[0x00] - b2[0x07]) * costab[0];
119
b1[0x01] = b2[0x01] + b2[0x06];
120
b1[0x06] = (b2[0x01] - b2[0x06]) * costab[1];
121
b1[0x02] = b2[0x02] + b2[0x05];
122
b1[0x05] = (b2[0x02] - b2[0x05]) * costab[2];
123
b1[0x03] = b2[0x03] + b2[0x04];
124
b1[0x04] = (b2[0x03] - b2[0x04]) * costab[3];
126
b1[0x08] = b2[0x08] + b2[0x0F];
127
b1[0x0F] = (b2[0x0F] - b2[0x08]) * costab[0];
128
b1[0x09] = b2[0x09] + b2[0x0E];
129
b1[0x0E] = (b2[0x0E] - b2[0x09]) * costab[1];
130
b1[0x0A] = b2[0x0A] + b2[0x0D];
131
b1[0x0D] = (b2[0x0D] - b2[0x0A]) * costab[2];
132
b1[0x0B] = b2[0x0B] + b2[0x0C];
133
b1[0x0C] = (b2[0x0C] - b2[0x0B]) * costab[3];
135
b1[0x10] = b2[0x10] + b2[0x17];
136
b1[0x17] = (b2[0x10] - b2[0x17]) * costab[0];
137
b1[0x11] = b2[0x11] + b2[0x16];
138
b1[0x16] = (b2[0x11] - b2[0x16]) * costab[1];
139
b1[0x12] = b2[0x12] + b2[0x15];
140
b1[0x15] = (b2[0x12] - b2[0x15]) * costab[2];
141
b1[0x13] = b2[0x13] + b2[0x14];
142
b1[0x14] = (b2[0x13] - b2[0x14]) * costab[3];
144
b1[0x18] = b2[0x18] + b2[0x1F];
145
b1[0x1F] = (b2[0x1F] - b2[0x18]) * costab[0];
146
b1[0x19] = b2[0x19] + b2[0x1E];
147
b1[0x1E] = (b2[0x1E] - b2[0x19]) * costab[1];
148
b1[0x1A] = b2[0x1A] + b2[0x1D];
149
b1[0x1D] = (b2[0x1D] - b2[0x1A]) * costab[2];
150
b1[0x1B] = b2[0x1B] + b2[0x1C];
151
b1[0x1C] = (b2[0x1C] - b2[0x1B]) * costab[3];
155
register real const cos0 = pnts[3][0];
156
register real const cos1 = pnts[3][1];
158
b2[0x00] = b1[0x00] + b1[0x03];
159
b2[0x03] = (b1[0x00] - b1[0x03]) * cos0;
160
b2[0x01] = b1[0x01] + b1[0x02];
161
b2[0x02] = (b1[0x01] - b1[0x02]) * cos1;
163
b2[0x04] = b1[0x04] + b1[0x07];
164
b2[0x07] = (b1[0x07] - b1[0x04]) * cos0;
165
b2[0x05] = b1[0x05] + b1[0x06];
166
b2[0x06] = (b1[0x06] - b1[0x05]) * cos1;
168
b2[0x08] = b1[0x08] + b1[0x0B];
169
b2[0x0B] = (b1[0x08] - b1[0x0B]) * cos0;
170
b2[0x09] = b1[0x09] + b1[0x0A];
171
b2[0x0A] = (b1[0x09] - b1[0x0A]) * cos1;
173
b2[0x0C] = b1[0x0C] + b1[0x0F];
174
b2[0x0F] = (b1[0x0F] - b1[0x0C]) * cos0;
175
b2[0x0D] = b1[0x0D] + b1[0x0E];
176
b2[0x0E] = (b1[0x0E] - b1[0x0D]) * cos1;
178
b2[0x10] = b1[0x10] + b1[0x13];
179
b2[0x13] = (b1[0x10] - b1[0x13]) * cos0;
180
b2[0x11] = b1[0x11] + b1[0x12];
181
b2[0x12] = (b1[0x11] - b1[0x12]) * cos1;
183
b2[0x14] = b1[0x14] + b1[0x17];
184
b2[0x17] = (b1[0x17] - b1[0x14]) * cos0;
185
b2[0x15] = b1[0x15] + b1[0x16];
186
b2[0x16] = (b1[0x16] - b1[0x15]) * cos1;
188
b2[0x18] = b1[0x18] + b1[0x1B];
189
b2[0x1B] = (b1[0x18] - b1[0x1B]) * cos0;
190
b2[0x19] = b1[0x19] + b1[0x1A];
191
b2[0x1A] = (b1[0x19] - b1[0x1A]) * cos1;
193
b2[0x1C] = b1[0x1C] + b1[0x1F];
194
b2[0x1F] = (b1[0x1F] - b1[0x1C]) * cos0;
195
b2[0x1D] = b1[0x1D] + b1[0x1E];
196
b2[0x1E] = (b1[0x1E] - b1[0x1D]) * cos1;
200
register real const cos0 = pnts[4][0];
202
b1[0x00] = b2[0x00] + b2[0x01];
203
b1[0x01] = (b2[0x00] - b2[0x01]) * cos0;
204
b1[0x02] = b2[0x02] + b2[0x03];
205
b1[0x03] = (b2[0x03] - b2[0x02]) * cos0;
206
b1[0x02] += b1[0x03];
208
b1[0x04] = b2[0x04] + b2[0x05];
209
b1[0x05] = (b2[0x04] - b2[0x05]) * cos0;
210
b1[0x06] = b2[0x06] + b2[0x07];
211
b1[0x07] = (b2[0x07] - b2[0x06]) * cos0;
212
b1[0x06] += b1[0x07];
213
b1[0x04] += b1[0x06];
214
b1[0x06] += b1[0x05];
215
b1[0x05] += b1[0x07];
217
b1[0x08] = b2[0x08] + b2[0x09];
218
b1[0x09] = (b2[0x08] - b2[0x09]) * cos0;
219
b1[0x0A] = b2[0x0A] + b2[0x0B];
220
b1[0x0B] = (b2[0x0B] - b2[0x0A]) * cos0;
221
b1[0x0A] += b1[0x0B];
223
b1[0x0C] = b2[0x0C] + b2[0x0D];
224
b1[0x0D] = (b2[0x0C] - b2[0x0D]) * cos0;
225
b1[0x0E] = b2[0x0E] + b2[0x0F];
226
b1[0x0F] = (b2[0x0F] - b2[0x0E]) * cos0;
227
b1[0x0E] += b1[0x0F];
228
b1[0x0C] += b1[0x0E];
229
b1[0x0E] += b1[0x0D];
230
b1[0x0D] += b1[0x0F];
232
b1[0x10] = b2[0x10] + b2[0x11];
233
b1[0x11] = (b2[0x10] - b2[0x11]) * cos0;
234
b1[0x12] = b2[0x12] + b2[0x13];
235
b1[0x13] = (b2[0x13] - b2[0x12]) * cos0;
236
b1[0x12] += b1[0x13];
238
b1[0x14] = b2[0x14] + b2[0x15];
239
b1[0x15] = (b2[0x14] - b2[0x15]) * cos0;
240
b1[0x16] = b2[0x16] + b2[0x17];
241
b1[0x17] = (b2[0x17] - b2[0x16]) * cos0;
242
b1[0x16] += b1[0x17];
243
b1[0x14] += b1[0x16];
244
b1[0x16] += b1[0x15];
245
b1[0x15] += b1[0x17];
247
b1[0x18] = b2[0x18] + b2[0x19];
248
b1[0x19] = (b2[0x18] - b2[0x19]) * cos0;
249
b1[0x1A] = b2[0x1A] + b2[0x1B];
250
b1[0x1B] = (b2[0x1B] - b2[0x1A]) * cos0;
251
b1[0x1A] += b1[0x1B];
253
b1[0x1C] = b2[0x1C] + b2[0x1D];
254
b1[0x1D] = (b2[0x1C] - b2[0x1D]) * cos0;
255
b1[0x1E] = b2[0x1E] + b2[0x1F];
256
b1[0x1F] = (b2[0x1F] - b2[0x1E]) * cos0;
257
b1[0x1E] += b1[0x1F];
258
b1[0x1C] += b1[0x1E];
259
b1[0x1E] += b1[0x1D];
260
b1[0x1D] += b1[0x1F];
263
out0[0x10*16] = b1[0x00];
264
out0[0x10*12] = b1[0x04];
265
out0[0x10* 8] = b1[0x02];
266
out0[0x10* 4] = b1[0x06];
267
out0[0x10* 0] = b1[0x01];
268
out1[0x10* 0] = b1[0x01];
269
out1[0x10* 4] = b1[0x05];
270
out1[0x10* 8] = b1[0x03];
271
out1[0x10*12] = b1[0x07];
274
out0[0x10*14] = b1[0x08] + b1[0x0C];
275
out0[0x10*10] = b1[0x0C] + b1[0x0a];
276
out0[0x10* 6] = b1[0x0A] + b1[0x0E];
277
out0[0x10* 2] = b1[0x0E] + b1[0x09];
278
out1[0x10* 2] = b1[0x09] + b1[0x0D];
279
out1[0x10* 6] = b1[0x0D] + b1[0x0B];
280
out1[0x10*10] = b1[0x0B] + b1[0x0F];
281
out1[0x10*14] = b1[0x0F];
283
b1[0x08] += b1[0x0C];
284
out0[0x10*14] = b1[0x08];
285
b1[0x0C] += b1[0x0a];
286
out0[0x10*10] = b1[0x0C];
287
b1[0x0A] += b1[0x0E];
288
out0[0x10* 6] = b1[0x0A];
289
b1[0x0E] += b1[0x09];
290
out0[0x10* 2] = b1[0x0E];
291
b1[0x09] += b1[0x0D];
292
out1[0x10* 2] = b1[0x09];
293
b1[0x0D] += b1[0x0B];
294
out1[0x10* 6] = b1[0x0D];
295
b1[0x0B] += b1[0x0F];
296
out1[0x10*10] = b1[0x0B];
297
out1[0x10*14] = b1[0x0F];
302
tmp = b1[0x18] + b1[0x1C];
303
out0[0x10*15] = tmp + b1[0x10];
304
out0[0x10*13] = tmp + b1[0x14];
305
tmp = b1[0x1C] + b1[0x1A];
306
out0[0x10*11] = tmp + b1[0x14];
307
out0[0x10* 9] = tmp + b1[0x12];
308
tmp = b1[0x1A] + b1[0x1E];
309
out0[0x10* 7] = tmp + b1[0x12];
310
out0[0x10* 5] = tmp + b1[0x16];
311
tmp = b1[0x1E] + b1[0x19];
312
out0[0x10* 3] = tmp + b1[0x16];
313
out0[0x10* 1] = tmp + b1[0x11];
314
tmp = b1[0x19] + b1[0x1D];
315
out1[0x10* 1] = tmp + b1[0x11];
316
out1[0x10* 3] = tmp + b1[0x15];
317
tmp = b1[0x1D] + b1[0x1B];
318
out1[0x10* 5] = tmp + b1[0x15];
319
out1[0x10* 7] = tmp + b1[0x13];
320
tmp = b1[0x1B] + b1[0x1F];
321
out1[0x10* 9] = tmp + b1[0x13];
322
out1[0x10*11] = tmp + b1[0x17];
323
out1[0x10*13] = b1[0x17] + b1[0x1F];
324
out1[0x10*15] = b1[0x1F];
329
* the call via dct64 is a trick to force GCC to use
330
* (new) registers for the b1,b2 pointer to the bufs[xx] field
332
void dct64(real *a,real *b,real *c)
335
dct64_1(a,b,bufs,bufs+0x20,c);