Mercurial > mplayer.hg

comparison mp3lib/dct64_i386.c @ 1:3b5f5d1c5041

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Initial revision
author arpi_esp
date Sat, 24 Feb 2001 20:28:24 +0000
parents
children d155623271e3
comparison
equal deleted inserted replaced
0:c1bb2c071d63 1:3b5f5d1c5041
1
2 /*
3 * Discrete Cosine Tansform (DCT) for subband synthesis
4 * optimized for machines with no auto-increment.
5 * The performance is highly compiler dependend. Maybe
6 * the mpg123_dct64.c version for 'normal' processor may be faster
7 * even for Intel processors.
8 */
9
10 #include "mpg123.h"
11
12 static void mpg123_dct64_1(real * out0, real * out1, real * b1, real * b2, real * samples)
13 {
14
15 {
16 register real *costab = pnts[0];
17
18 b1[0x00] = samples[0x00] + samples[0x1F];
19 b1[0x1F] = (samples[0x00] - samples[0x1F]) * costab[0x0];
20
21 b1[0x01] = samples[0x01] + samples[0x1E];
22 b1[0x1E] = (samples[0x01] - samples[0x1E]) * costab[0x1];
23
24 b1[0x02] = samples[0x02] + samples[0x1D];
25 b1[0x1D] = (samples[0x02] - samples[0x1D]) * costab[0x2];
26
27 b1[0x03] = samples[0x03] + samples[0x1C];
28 b1[0x1C] = (samples[0x03] - samples[0x1C]) * costab[0x3];
29
30 b1[0x04] = samples[0x04] + samples[0x1B];
31 b1[0x1B] = (samples[0x04] - samples[0x1B]) * costab[0x4];
32
33 b1[0x05] = samples[0x05] + samples[0x1A];
34 b1[0x1A] = (samples[0x05] - samples[0x1A]) * costab[0x5];
35
36 b1[0x06] = samples[0x06] + samples[0x19];
37 b1[0x19] = (samples[0x06] - samples[0x19]) * costab[0x6];
38
39 b1[0x07] = samples[0x07] + samples[0x18];
40 b1[0x18] = (samples[0x07] - samples[0x18]) * costab[0x7];
41
42 b1[0x08] = samples[0x08] + samples[0x17];
43 b1[0x17] = (samples[0x08] - samples[0x17]) * costab[0x8];
44
45 b1[0x09] = samples[0x09] + samples[0x16];
46 b1[0x16] = (samples[0x09] - samples[0x16]) * costab[0x9];
47
48 b1[0x0A] = samples[0x0A] + samples[0x15];
49 b1[0x15] = (samples[0x0A] - samples[0x15]) * costab[0xA];
50
51 b1[0x0B] = samples[0x0B] + samples[0x14];
52 b1[0x14] = (samples[0x0B] - samples[0x14]) * costab[0xB];
53
54 b1[0x0C] = samples[0x0C] + samples[0x13];
55 b1[0x13] = (samples[0x0C] - samples[0x13]) * costab[0xC];
56
57 b1[0x0D] = samples[0x0D] + samples[0x12];
58 b1[0x12] = (samples[0x0D] - samples[0x12]) * costab[0xD];
59
60 b1[0x0E] = samples[0x0E] + samples[0x11];
61 b1[0x11] = (samples[0x0E] - samples[0x11]) * costab[0xE];
62
63 b1[0x0F] = samples[0x0F] + samples[0x10];
64 b1[0x10] = (samples[0x0F] - samples[0x10]) * costab[0xF];
65 }
66
67 {
68 register real *costab = pnts[1];
69
70 b2[0x00] = b1[0x00] + b1[0x0F];
71 b2[0x0F] = (b1[0x00] - b1[0x0F]) * costab[0];
72 b2[0x01] = b1[0x01] + b1[0x0E];
73 b2[0x0E] = (b1[0x01] - b1[0x0E]) * costab[1];
74 b2[0x02] = b1[0x02] + b1[0x0D];
75 b2[0x0D] = (b1[0x02] - b1[0x0D]) * costab[2];
76 b2[0x03] = b1[0x03] + b1[0x0C];
77 b2[0x0C] = (b1[0x03] - b1[0x0C]) * costab[3];
78 b2[0x04] = b1[0x04] + b1[0x0B];
79 b2[0x0B] = (b1[0x04] - b1[0x0B]) * costab[4];
80 b2[0x05] = b1[0x05] + b1[0x0A];
81 b2[0x0A] = (b1[0x05] - b1[0x0A]) * costab[5];
82 b2[0x06] = b1[0x06] + b1[0x09];
83 b2[0x09] = (b1[0x06] - b1[0x09]) * costab[6];
84 b2[0x07] = b1[0x07] + b1[0x08];
85 b2[0x08] = (b1[0x07] - b1[0x08]) * costab[7];
86
87 b2[0x10] = b1[0x10] + b1[0x1F];
88 b2[0x1F] = (b1[0x1F] - b1[0x10]) * costab[0];
89 b2[0x11] = b1[0x11] + b1[0x1E];
90 b2[0x1E] = (b1[0x1E] - b1[0x11]) * costab[1];
91 b2[0x12] = b1[0x12] + b1[0x1D];
92 b2[0x1D] = (b1[0x1D] - b1[0x12]) * costab[2];
93 b2[0x13] = b1[0x13] + b1[0x1C];
94 b2[0x1C] = (b1[0x1C] - b1[0x13]) * costab[3];
95 b2[0x14] = b1[0x14] + b1[0x1B];
96 b2[0x1B] = (b1[0x1B] - b1[0x14]) * costab[4];
97 b2[0x15] = b1[0x15] + b1[0x1A];
98 b2[0x1A] = (b1[0x1A] - b1[0x15]) * costab[5];
99 b2[0x16] = b1[0x16] + b1[0x19];
100 b2[0x19] = (b1[0x19] - b1[0x16]) * costab[6];
101 b2[0x17] = b1[0x17] + b1[0x18];
102 b2[0x18] = (b1[0x18] - b1[0x17]) * costab[7];
103 }
104
105 {
106 register real *costab = pnts[2];
107
108 b1[0x00] = b2[0x00] + b2[0x07];
109 b1[0x07] = (b2[0x00] - b2[0x07]) * costab[0];
110 b1[0x01] = b2[0x01] + b2[0x06];
111 b1[0x06] = (b2[0x01] - b2[0x06]) * costab[1];
112 b1[0x02] = b2[0x02] + b2[0x05];
113 b1[0x05] = (b2[0x02] - b2[0x05]) * costab[2];
114 b1[0x03] = b2[0x03] + b2[0x04];
115 b1[0x04] = (b2[0x03] - b2[0x04]) * costab[3];
116
117 b1[0x08] = b2[0x08] + b2[0x0F];
118 b1[0x0F] = (b2[0x0F] - b2[0x08]) * costab[0];
119 b1[0x09] = b2[0x09] + b2[0x0E];
120 b1[0x0E] = (b2[0x0E] - b2[0x09]) * costab[1];
121 b1[0x0A] = b2[0x0A] + b2[0x0D];
122 b1[0x0D] = (b2[0x0D] - b2[0x0A]) * costab[2];
123 b1[0x0B] = b2[0x0B] + b2[0x0C];
124 b1[0x0C] = (b2[0x0C] - b2[0x0B]) * costab[3];
125
126 b1[0x10] = b2[0x10] + b2[0x17];
127 b1[0x17] = (b2[0x10] - b2[0x17]) * costab[0];
128 b1[0x11] = b2[0x11] + b2[0x16];
129 b1[0x16] = (b2[0x11] - b2[0x16]) * costab[1];
130 b1[0x12] = b2[0x12] + b2[0x15];
131 b1[0x15] = (b2[0x12] - b2[0x15]) * costab[2];
132 b1[0x13] = b2[0x13] + b2[0x14];
133 b1[0x14] = (b2[0x13] - b2[0x14]) * costab[3];
134
135 b1[0x18] = b2[0x18] + b2[0x1F];
136 b1[0x1F] = (b2[0x1F] - b2[0x18]) * costab[0];
137 b1[0x19] = b2[0x19] + b2[0x1E];
138 b1[0x1E] = (b2[0x1E] - b2[0x19]) * costab[1];
139 b1[0x1A] = b2[0x1A] + b2[0x1D];
140 b1[0x1D] = (b2[0x1D] - b2[0x1A]) * costab[2];
141 b1[0x1B] = b2[0x1B] + b2[0x1C];
142 b1[0x1C] = (b2[0x1C] - b2[0x1B]) * costab[3];
143 }
144
145 {
146 register real const cos0 = pnts[3][0];
147 register real const cos1 = pnts[3][1];
148
149 b2[0x00] = b1[0x00] + b1[0x03];
150 b2[0x03] = (b1[0x00] - b1[0x03]) * cos0;
151 b2[0x01] = b1[0x01] + b1[0x02];
152 b2[0x02] = (b1[0x01] - b1[0x02]) * cos1;
153
154 b2[0x04] = b1[0x04] + b1[0x07];
155 b2[0x07] = (b1[0x07] - b1[0x04]) * cos0;
156 b2[0x05] = b1[0x05] + b1[0x06];
157 b2[0x06] = (b1[0x06] - b1[0x05]) * cos1;
158
159 b2[0x08] = b1[0x08] + b1[0x0B];
160 b2[0x0B] = (b1[0x08] - b1[0x0B]) * cos0;
161 b2[0x09] = b1[0x09] + b1[0x0A];
162 b2[0x0A] = (b1[0x09] - b1[0x0A]) * cos1;
163
164 b2[0x0C] = b1[0x0C] + b1[0x0F];
165 b2[0x0F] = (b1[0x0F] - b1[0x0C]) * cos0;
166 b2[0x0D] = b1[0x0D] + b1[0x0E];
167 b2[0x0E] = (b1[0x0E] - b1[0x0D]) * cos1;
168
169 b2[0x10] = b1[0x10] + b1[0x13];
170 b2[0x13] = (b1[0x10] - b1[0x13]) * cos0;
171 b2[0x11] = b1[0x11] + b1[0x12];
172 b2[0x12] = (b1[0x11] - b1[0x12]) * cos1;
173
174 b2[0x14] = b1[0x14] + b1[0x17];
175 b2[0x17] = (b1[0x17] - b1[0x14]) * cos0;
176 b2[0x15] = b1[0x15] + b1[0x16];
177 b2[0x16] = (b1[0x16] - b1[0x15]) * cos1;
178
179 b2[0x18] = b1[0x18] + b1[0x1B];
180 b2[0x1B] = (b1[0x18] - b1[0x1B]) * cos0;
181 b2[0x19] = b1[0x19] + b1[0x1A];
182 b2[0x1A] = (b1[0x19] - b1[0x1A]) * cos1;
183
184 b2[0x1C] = b1[0x1C] + b1[0x1F];
185 b2[0x1F] = (b1[0x1F] - b1[0x1C]) * cos0;
186 b2[0x1D] = b1[0x1D] + b1[0x1E];
187 b2[0x1E] = (b1[0x1E] - b1[0x1D]) * cos1;
188 }
189
190 {
191 register real const cos0 = pnts[4][0];
192
193 b1[0x00] = b2[0x00] + b2[0x01];
194 b1[0x01] = (b2[0x00] - b2[0x01]) * cos0;
195 b1[0x02] = b2[0x02] + b2[0x03];
196 b1[0x03] = (b2[0x03] - b2[0x02]) * cos0;
197 b1[0x02] += b1[0x03];
198
199 b1[0x04] = b2[0x04] + b2[0x05];
200 b1[0x05] = (b2[0x04] - b2[0x05]) * cos0;
201 b1[0x06] = b2[0x06] + b2[0x07];
202 b1[0x07] = (b2[0x07] - b2[0x06]) * cos0;
203 b1[0x06] += b1[0x07];
204 b1[0x04] += b1[0x06];
205 b1[0x06] += b1[0x05];
206 b1[0x05] += b1[0x07];
207
208 b1[0x08] = b2[0x08] + b2[0x09];
209 b1[0x09] = (b2[0x08] - b2[0x09]) * cos0;
210 b1[0x0A] = b2[0x0A] + b2[0x0B];
211 b1[0x0B] = (b2[0x0B] - b2[0x0A]) * cos0;
212 b1[0x0A] += b1[0x0B];
213
214 b1[0x0C] = b2[0x0C] + b2[0x0D];
215 b1[0x0D] = (b2[0x0C] - b2[0x0D]) * cos0;
216 b1[0x0E] = b2[0x0E] + b2[0x0F];
217 b1[0x0F] = (b2[0x0F] - b2[0x0E]) * cos0;
218 b1[0x0E] += b1[0x0F];
219 b1[0x0C] += b1[0x0E];
220 b1[0x0E] += b1[0x0D];
221 b1[0x0D] += b1[0x0F];
222
223 b1[0x10] = b2[0x10] + b2[0x11];
224 b1[0x11] = (b2[0x10] - b2[0x11]) * cos0;
225 b1[0x12] = b2[0x12] + b2[0x13];
226 b1[0x13] = (b2[0x13] - b2[0x12]) * cos0;
227 b1[0x12] += b1[0x13];
228
229 b1[0x14] = b2[0x14] + b2[0x15];
230 b1[0x15] = (b2[0x14] - b2[0x15]) * cos0;
231 b1[0x16] = b2[0x16] + b2[0x17];
232 b1[0x17] = (b2[0x17] - b2[0x16]) * cos0;
233 b1[0x16] += b1[0x17];
234 b1[0x14] += b1[0x16];
235 b1[0x16] += b1[0x15];
236 b1[0x15] += b1[0x17];
237
238 b1[0x18] = b2[0x18] + b2[0x19];
239 b1[0x19] = (b2[0x18] - b2[0x19]) * cos0;
240 b1[0x1A] = b2[0x1A] + b2[0x1B];
241 b1[0x1B] = (b2[0x1B] - b2[0x1A]) * cos0;
242 b1[0x1A] += b1[0x1B];
243
244 b1[0x1C] = b2[0x1C] + b2[0x1D];
245 b1[0x1D] = (b2[0x1C] - b2[0x1D]) * cos0;
246 b1[0x1E] = b2[0x1E] + b2[0x1F];
247 b1[0x1F] = (b2[0x1F] - b2[0x1E]) * cos0;
248 b1[0x1E] += b1[0x1F];
249 b1[0x1C] += b1[0x1E];
250 b1[0x1E] += b1[0x1D];
251 b1[0x1D] += b1[0x1F];
252 }
253
254 out0[0x10 * 16] = b1[0x00];
255 out0[0x10 * 12] = b1[0x04];
256 out0[0x10 * 8] = b1[0x02];
257 out0[0x10 * 4] = b1[0x06];
258 out0[0x10 * 0] = b1[0x01];
259 out1[0x10 * 0] = b1[0x01];
260 out1[0x10 * 4] = b1[0x05];
261 out1[0x10 * 8] = b1[0x03];
262 out1[0x10 * 12] = b1[0x07];
263
264 b1[0x08] += b1[0x0C];
265 out0[0x10 * 14] = b1[0x08];
266 b1[0x0C] += b1[0x0a];
267 out0[0x10 * 10] = b1[0x0C];
268 b1[0x0A] += b1[0x0E];
269 out0[0x10 * 6] = b1[0x0A];
270 b1[0x0E] += b1[0x09];
271 out0[0x10 * 2] = b1[0x0E];
272 b1[0x09] += b1[0x0D];
273 out1[0x10 * 2] = b1[0x09];
274 b1[0x0D] += b1[0x0B];
275 out1[0x10 * 6] = b1[0x0D];
276 b1[0x0B] += b1[0x0F];
277 out1[0x10 * 10] = b1[0x0B];
278 out1[0x10 * 14] = b1[0x0F];
279
280 b1[0x18] += b1[0x1C];
281 out0[0x10 * 15] = b1[0x10] + b1[0x18];
282 out0[0x10 * 13] = b1[0x18] + b1[0x14];
283 b1[0x1C] += b1[0x1a];
284 out0[0x10 * 11] = b1[0x14] + b1[0x1C];
285 out0[0x10 * 9] = b1[0x1C] + b1[0x12];
286 b1[0x1A] += b1[0x1E];
287 out0[0x10 * 7] = b1[0x12] + b1[0x1A];
288 out0[0x10 * 5] = b1[0x1A] + b1[0x16];
289 b1[0x1E] += b1[0x19];
290 out0[0x10 * 3] = b1[0x16] + b1[0x1E];
291 out0[0x10 * 1] = b1[0x1E] + b1[0x11];
292 b1[0x19] += b1[0x1D];
293 out1[0x10 * 1] = b1[0x11] + b1[0x19];
294 out1[0x10 * 3] = b1[0x19] + b1[0x15];
295 b1[0x1D] += b1[0x1B];
296 out1[0x10 * 5] = b1[0x15] + b1[0x1D];
297 out1[0x10 * 7] = b1[0x1D] + b1[0x13];
298 b1[0x1B] += b1[0x1F];
299 out1[0x10 * 9] = b1[0x13] + b1[0x1B];
300 out1[0x10 * 11] = b1[0x1B] + b1[0x17];
301 out1[0x10 * 13] = b1[0x17] + b1[0x1F];
302 out1[0x10 * 15] = b1[0x1F];
303 }
304
305 /*
306 * the call via mpg123_dct64 is a trick to force GCC to use
307 * (new) registers for the b1,b2 pointer to the bufs[xx] field
308 */
309 void mpg123_dct64(real * a, real * b, real * c)
310 {
311 real bufs[0x40];
312
313 mpg123_dct64_1(a, b, bufs, bufs + 0x20, c);
314 }