|
1
|
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 dct64.c version for 'normal' processor may be faster
|
|
|
7 * even for Intel processors.
|
|
|
8 */
|
|
|
9
|
|
|
10 static void dct64_1(real *out0,real *out1,real *b1,real *b2,real *samples)
|
|
|
11 {
|
|
|
12
|
|
|
13 {
|
|
|
14 register real *costab = pnts[0];
|
|
|
15
|
|
|
16 b1[0x00] = samples[0x00] + samples[0x1F];
|
|
|
17 b1[0x1F] = (samples[0x00] - samples[0x1F]) * costab[0x0];
|
|
|
18
|
|
|
19 b1[0x01] = samples[0x01] + samples[0x1E];
|
|
|
20 b1[0x1E] = (samples[0x01] - samples[0x1E]) * costab[0x1];
|
|
|
21
|
|
|
22 b1[0x02] = samples[0x02] + samples[0x1D];
|
|
|
23 b1[0x1D] = (samples[0x02] - samples[0x1D]) * costab[0x2];
|
|
|
24
|
|
|
25 b1[0x03] = samples[0x03] + samples[0x1C];
|
|
|
26 b1[0x1C] = (samples[0x03] - samples[0x1C]) * costab[0x3];
|
|
|
27
|
|
|
28 b1[0x04] = samples[0x04] + samples[0x1B];
|
|
|
29 b1[0x1B] = (samples[0x04] - samples[0x1B]) * costab[0x4];
|
|
|
30
|
|
|
31 b1[0x05] = samples[0x05] + samples[0x1A];
|
|
|
32 b1[0x1A] = (samples[0x05] - samples[0x1A]) * costab[0x5];
|
|
|
33
|
|
|
34 b1[0x06] = samples[0x06] + samples[0x19];
|
|
|
35 b1[0x19] = (samples[0x06] - samples[0x19]) * costab[0x6];
|
|
|
36
|
|
|
37 b1[0x07] = samples[0x07] + samples[0x18];
|
|
|
38 b1[0x18] = (samples[0x07] - samples[0x18]) * costab[0x7];
|
|
|
39
|
|
|
40 b1[0x08] = samples[0x08] + samples[0x17];
|
|
|
41 b1[0x17] = (samples[0x08] - samples[0x17]) * costab[0x8];
|
|
|
42
|
|
|
43 b1[0x09] = samples[0x09] + samples[0x16];
|
|
|
44 b1[0x16] = (samples[0x09] - samples[0x16]) * costab[0x9];
|
|
|
45
|
|
|
46 b1[0x0A] = samples[0x0A] + samples[0x15];
|
|
|
47 b1[0x15] = (samples[0x0A] - samples[0x15]) * costab[0xA];
|
|
|
48
|
|
|
49 b1[0x0B] = samples[0x0B] + samples[0x14];
|
|
|
50 b1[0x14] = (samples[0x0B] - samples[0x14]) * costab[0xB];
|
|
|
51
|
|
|
52 b1[0x0C] = samples[0x0C] + samples[0x13];
|
|
|
53 b1[0x13] = (samples[0x0C] - samples[0x13]) * costab[0xC];
|
|
|
54
|
|
|
55 b1[0x0D] = samples[0x0D] + samples[0x12];
|
|
|
56 b1[0x12] = (samples[0x0D] - samples[0x12]) * costab[0xD];
|
|
|
57
|
|
|
58 b1[0x0E] = samples[0x0E] + samples[0x11];
|
|
|
59 b1[0x11] = (samples[0x0E] - samples[0x11]) * costab[0xE];
|
|
|
60
|
|
|
61 b1[0x0F] = samples[0x0F] + samples[0x10];
|
|
|
62 b1[0x10] = (samples[0x0F] - samples[0x10]) * costab[0xF];
|
|
|
63 }
|
|
|
64
|
|
|
65
|
|
|
66 {
|
|
|
67 register real *costab = pnts[1];
|
|
|
68
|
|
|
69 b2[0x00] = b1[0x00] + b1[0x0F];
|
|
|
70 b2[0x0F] = (b1[0x00] - b1[0x0F]) * costab[0];
|
|
|
71 b2[0x01] = b1[0x01] + b1[0x0E];
|
|
|
72 b2[0x0E] = (b1[0x01] - b1[0x0E]) * costab[1];
|
|
|
73 b2[0x02] = b1[0x02] + b1[0x0D];
|
|
|
74 b2[0x0D] = (b1[0x02] - b1[0x0D]) * costab[2];
|
|
|
75 b2[0x03] = b1[0x03] + b1[0x0C];
|
|
|
76 b2[0x0C] = (b1[0x03] - b1[0x0C]) * costab[3];
|
|
|
77 b2[0x04] = b1[0x04] + b1[0x0B];
|
|
|
78 b2[0x0B] = (b1[0x04] - b1[0x0B]) * costab[4];
|
|
|
79 b2[0x05] = b1[0x05] + b1[0x0A];
|
|
|
80 b2[0x0A] = (b1[0x05] - b1[0x0A]) * costab[5];
|
|
|
81 b2[0x06] = b1[0x06] + b1[0x09];
|
|
|
82 b2[0x09] = (b1[0x06] - b1[0x09]) * costab[6];
|
|
|
83 b2[0x07] = b1[0x07] + b1[0x08];
|
|
|
84 b2[0x08] = (b1[0x07] - b1[0x08]) * costab[7];
|
|
|
85
|
|
|
86 b2[0x10] = b1[0x10] + b1[0x1F];
|
|
|
87 b2[0x1F] = (b1[0x1F] - b1[0x10]) * costab[0];
|
|
|
88 b2[0x11] = b1[0x11] + b1[0x1E];
|
|
|
89 b2[0x1E] = (b1[0x1E] - b1[0x11]) * costab[1];
|
|
|
90 b2[0x12] = b1[0x12] + b1[0x1D];
|
|
|
91 b2[0x1D] = (b1[0x1D] - b1[0x12]) * costab[2];
|
|
|
92 b2[0x13] = b1[0x13] + b1[0x1C];
|
|
|
93 b2[0x1C] = (b1[0x1C] - b1[0x13]) * costab[3];
|
|
|
94 b2[0x14] = b1[0x14] + b1[0x1B];
|
|
|
95 b2[0x1B] = (b1[0x1B] - b1[0x14]) * costab[4];
|
|
|
96 b2[0x15] = b1[0x15] + b1[0x1A];
|
|
|
97 b2[0x1A] = (b1[0x1A] - b1[0x15]) * costab[5];
|
|
|
98 b2[0x16] = b1[0x16] + b1[0x19];
|
|
|
99 b2[0x19] = (b1[0x19] - b1[0x16]) * costab[6];
|
|
|
100 b2[0x17] = b1[0x17] + b1[0x18];
|
|
|
101 b2[0x18] = (b1[0x18] - b1[0x17]) * costab[7];
|
|
|
102 }
|
|
|
103
|
|
|
104 {
|
|
|
105 register real *costab = pnts[2];
|
|
|
106
|
|
|
107 b1[0x00] = b2[0x00] + b2[0x07];
|
|
|
108 b1[0x07] = (b2[0x00] - b2[0x07]) * costab[0];
|
|
|
109 b1[0x01] = b2[0x01] + b2[0x06];
|
|
|
110 b1[0x06] = (b2[0x01] - b2[0x06]) * costab[1];
|
|
|
111 b1[0x02] = b2[0x02] + b2[0x05];
|
|
|
112 b1[0x05] = (b2[0x02] - b2[0x05]) * costab[2];
|
|
|
113 b1[0x03] = b2[0x03] + b2[0x04];
|
|
|
114 b1[0x04] = (b2[0x03] - b2[0x04]) * costab[3];
|
|
|
115
|
|
|
116 b1[0x08] = b2[0x08] + b2[0x0F];
|
|
|
117 b1[0x0F] = (b2[0x0F] - b2[0x08]) * costab[0];
|
|
|
118 b1[0x09] = b2[0x09] + b2[0x0E];
|
|
|
119 b1[0x0E] = (b2[0x0E] - b2[0x09]) * costab[1];
|
|
|
120 b1[0x0A] = b2[0x0A] + b2[0x0D];
|
|
|
121 b1[0x0D] = (b2[0x0D] - b2[0x0A]) * costab[2];
|
|
|
122 b1[0x0B] = b2[0x0B] + b2[0x0C];
|
|
|
123 b1[0x0C] = (b2[0x0C] - b2[0x0B]) * costab[3];
|
|
|
124
|
|
|
125 b1[0x10] = b2[0x10] + b2[0x17];
|
|
|
126 b1[0x17] = (b2[0x10] - b2[0x17]) * costab[0];
|
|
|
127 b1[0x11] = b2[0x11] + b2[0x16];
|
|
|
128 b1[0x16] = (b2[0x11] - b2[0x16]) * costab[1];
|
|
|
129 b1[0x12] = b2[0x12] + b2[0x15];
|
|
|
130 b1[0x15] = (b2[0x12] - b2[0x15]) * costab[2];
|
|
|
131 b1[0x13] = b2[0x13] + b2[0x14];
|
|
|
132 b1[0x14] = (b2[0x13] - b2[0x14]) * costab[3];
|
|
|
133
|
|
|
134 b1[0x18] = b2[0x18] + b2[0x1F];
|
|
|
135 b1[0x1F] = (b2[0x1F] - b2[0x18]) * costab[0];
|
|
|
136 b1[0x19] = b2[0x19] + b2[0x1E];
|
|
|
137 b1[0x1E] = (b2[0x1E] - b2[0x19]) * costab[1];
|
|
|
138 b1[0x1A] = b2[0x1A] + b2[0x1D];
|
|
|
139 b1[0x1D] = (b2[0x1D] - b2[0x1A]) * costab[2];
|
|
|
140 b1[0x1B] = b2[0x1B] + b2[0x1C];
|
|
|
141 b1[0x1C] = (b2[0x1C] - b2[0x1B]) * costab[3];
|
|
|
142 }
|
|
|
143
|
|
|
144 {
|
|
|
145 register real const cos0 = pnts[3][0];
|
|
|
146 register real const cos1 = pnts[3][1];
|
|
|
147
|
|
|
148 b2[0x00] = b1[0x00] + b1[0x03];
|
|
|
149 b2[0x03] = (b1[0x00] - b1[0x03]) * cos0;
|
|
|
150 b2[0x01] = b1[0x01] + b1[0x02];
|
|
|
151 b2[0x02] = (b1[0x01] - b1[0x02]) * cos1;
|
|
|
152
|
|
|
153 b2[0x04] = b1[0x04] + b1[0x07];
|
|
|
154 b2[0x07] = (b1[0x07] - b1[0x04]) * cos0;
|
|
|
155 b2[0x05] = b1[0x05] + b1[0x06];
|
|
|
156 b2[0x06] = (b1[0x06] - b1[0x05]) * cos1;
|
|
|
157
|
|
|
158 b2[0x08] = b1[0x08] + b1[0x0B];
|
|
|
159 b2[0x0B] = (b1[0x08] - b1[0x0B]) * cos0;
|
|
|
160 b2[0x09] = b1[0x09] + b1[0x0A];
|
|
|
161 b2[0x0A] = (b1[0x09] - b1[0x0A]) * cos1;
|
|
|
162
|
|
|
163 b2[0x0C] = b1[0x0C] + b1[0x0F];
|
|
|
164 b2[0x0F] = (b1[0x0F] - b1[0x0C]) * cos0;
|
|
|
165 b2[0x0D] = b1[0x0D] + b1[0x0E];
|
|
|
166 b2[0x0E] = (b1[0x0E] - b1[0x0D]) * cos1;
|
|
|
167
|
|
|
168 b2[0x10] = b1[0x10] + b1[0x13];
|
|
|
169 b2[0x13] = (b1[0x10] - b1[0x13]) * cos0;
|
|
|
170 b2[0x11] = b1[0x11] + b1[0x12];
|
|
|
171 b2[0x12] = (b1[0x11] - b1[0x12]) * cos1;
|
|
|
172
|
|
|
173 b2[0x14] = b1[0x14] + b1[0x17];
|
|
|
174 b2[0x17] = (b1[0x17] - b1[0x14]) * cos0;
|
|
|
175 b2[0x15] = b1[0x15] + b1[0x16];
|
|
|
176 b2[0x16] = (b1[0x16] - b1[0x15]) * cos1;
|
|
|
177
|
|
|
178 b2[0x18] = b1[0x18] + b1[0x1B];
|
|
|
179 b2[0x1B] = (b1[0x18] - b1[0x1B]) * cos0;
|
|
|
180 b2[0x19] = b1[0x19] + b1[0x1A];
|
|
|
181 b2[0x1A] = (b1[0x19] - b1[0x1A]) * cos1;
|
|
|
182
|
|
|
183 b2[0x1C] = b1[0x1C] + b1[0x1F];
|
|
|
184 b2[0x1F] = (b1[0x1F] - b1[0x1C]) * cos0;
|
|
|
185 b2[0x1D] = b1[0x1D] + b1[0x1E];
|
|
|
186 b2[0x1E] = (b1[0x1E] - b1[0x1D]) * cos1;
|
|
|
187 }
|
|
|
188
|
|
|
189 {
|
|
|
190 register real const cos0 = pnts[4][0];
|
|
|
191
|
|
|
192 b1[0x00] = b2[0x00] + b2[0x01];
|
|
|
193 b1[0x01] = (b2[0x00] - b2[0x01]) * cos0;
|
|
|
194 b1[0x02] = b2[0x02] + b2[0x03];
|
|
|
195 b1[0x03] = (b2[0x03] - b2[0x02]) * cos0;
|
|
|
196 b1[0x02] += b1[0x03];
|
|
|
197
|
|
|
198 b1[0x04] = b2[0x04] + b2[0x05];
|
|
|
199 b1[0x05] = (b2[0x04] - b2[0x05]) * cos0;
|
|
|
200 b1[0x06] = b2[0x06] + b2[0x07];
|
|
|
201 b1[0x07] = (b2[0x07] - b2[0x06]) * cos0;
|
|
|
202 b1[0x06] += b1[0x07];
|
|
|
203 b1[0x04] += b1[0x06];
|
|
|
204 b1[0x06] += b1[0x05];
|
|
|
205 b1[0x05] += b1[0x07];
|
|
|
206
|
|
|
207 b1[0x08] = b2[0x08] + b2[0x09];
|
|
|
208 b1[0x09] = (b2[0x08] - b2[0x09]) * cos0;
|
|
|
209 b1[0x0A] = b2[0x0A] + b2[0x0B];
|
|
|
210 b1[0x0B] = (b2[0x0B] - b2[0x0A]) * cos0;
|
|
|
211 b1[0x0A] += b1[0x0B];
|
|
|
212
|
|
|
213 b1[0x0C] = b2[0x0C] + b2[0x0D];
|
|
|
214 b1[0x0D] = (b2[0x0C] - b2[0x0D]) * cos0;
|
|
|
215 b1[0x0E] = b2[0x0E] + b2[0x0F];
|
|
|
216 b1[0x0F] = (b2[0x0F] - b2[0x0E]) * cos0;
|
|
|
217 b1[0x0E] += b1[0x0F];
|
|
|
218 b1[0x0C] += b1[0x0E];
|
|
|
219 b1[0x0E] += b1[0x0D];
|
|
|
220 b1[0x0D] += b1[0x0F];
|
|
|
221
|
|
|
222 b1[0x10] = b2[0x10] + b2[0x11];
|
|
|
223 b1[0x11] = (b2[0x10] - b2[0x11]) * cos0;
|
|
|
224 b1[0x12] = b2[0x12] + b2[0x13];
|
|
|
225 b1[0x13] = (b2[0x13] - b2[0x12]) * cos0;
|
|
|
226 b1[0x12] += b1[0x13];
|
|
|
227
|
|
|
228 b1[0x14] = b2[0x14] + b2[0x15];
|
|
|
229 b1[0x15] = (b2[0x14] - b2[0x15]) * cos0;
|
|
|
230 b1[0x16] = b2[0x16] + b2[0x17];
|
|
|
231 b1[0x17] = (b2[0x17] - b2[0x16]) * cos0;
|
|
|
232 b1[0x16] += b1[0x17];
|
|
|
233 b1[0x14] += b1[0x16];
|
|
|
234 b1[0x16] += b1[0x15];
|
|
|
235 b1[0x15] += b1[0x17];
|
|
|
236
|
|
|
237 b1[0x18] = b2[0x18] + b2[0x19];
|
|
|
238 b1[0x19] = (b2[0x18] - b2[0x19]) * cos0;
|
|
|
239 b1[0x1A] = b2[0x1A] + b2[0x1B];
|
|
|
240 b1[0x1B] = (b2[0x1B] - b2[0x1A]) * cos0;
|
|
|
241 b1[0x1A] += b1[0x1B];
|
|
|
242
|
|
|
243 b1[0x1C] = b2[0x1C] + b2[0x1D];
|
|
|
244 b1[0x1D] = (b2[0x1C] - b2[0x1D]) * cos0;
|
|
|
245 b1[0x1E] = b2[0x1E] + b2[0x1F];
|
|
|
246 b1[0x1F] = (b2[0x1F] - b2[0x1E]) * cos0;
|
|
|
247 b1[0x1E] += b1[0x1F];
|
|
|
248 b1[0x1C] += b1[0x1E];
|
|
|
249 b1[0x1E] += b1[0x1D];
|
|
|
250 b1[0x1D] += b1[0x1F];
|
|
|
251 }
|
|
|
252
|
|
|
253 out0[0x10*16] = b1[0x00];
|
|
|
254 out0[0x10*12] = b1[0x04];
|
|
|
255 out0[0x10* 8] = b1[0x02];
|
|
|
256 out0[0x10* 4] = b1[0x06];
|
|
|
257 out0[0x10* 0] = b1[0x01];
|
|
|
258 out1[0x10* 0] = b1[0x01];
|
|
|
259 out1[0x10* 4] = b1[0x05];
|
|
|
260 out1[0x10* 8] = b1[0x03];
|
|
|
261 out1[0x10*12] = b1[0x07];
|
|
|
262
|
|
|
263 b1[0x08] += b1[0x0C];
|
|
|
264 out0[0x10*14] = b1[0x08];
|
|
|
265 b1[0x0C] += b1[0x0a];
|
|
|
266 out0[0x10*10] = b1[0x0C];
|
|
|
267 b1[0x0A] += b1[0x0E];
|
|
|
268 out0[0x10* 6] = b1[0x0A];
|
|
|
269 b1[0x0E] += b1[0x09];
|
|
|
270 out0[0x10* 2] = b1[0x0E];
|
|
|
271 b1[0x09] += b1[0x0D];
|
|
|
272 out1[0x10* 2] = b1[0x09];
|
|
|
273 b1[0x0D] += b1[0x0B];
|
|
|
274 out1[0x10* 6] = b1[0x0D];
|
|
|
275 b1[0x0B] += b1[0x0F];
|
|
|
276 out1[0x10*10] = b1[0x0B];
|
|
|
277 out1[0x10*14] = b1[0x0F];
|
|
|
278
|
|
|
279 b1[0x18] += b1[0x1C];
|
|
|
280 out0[0x10*15] = b1[0x10] + b1[0x18];
|
|
|
281 out0[0x10*13] = b1[0x18] + b1[0x14];
|
|
|
282 b1[0x1C] += b1[0x1a];
|
|
|
283 out0[0x10*11] = b1[0x14] + b1[0x1C];
|
|
|
284 out0[0x10* 9] = b1[0x1C] + b1[0x12];
|
|
|
285 b1[0x1A] += b1[0x1E];
|
|
|
286 out0[0x10* 7] = b1[0x12] + b1[0x1A];
|
|
|
287 out0[0x10* 5] = b1[0x1A] + b1[0x16];
|
|
|
288 b1[0x1E] += b1[0x19];
|
|
|
289 out0[0x10* 3] = b1[0x16] + b1[0x1E];
|
|
|
290 out0[0x10* 1] = b1[0x1E] + b1[0x11];
|
|
|
291 b1[0x19] += b1[0x1D];
|
|
|
292 out1[0x10* 1] = b1[0x11] + b1[0x19];
|
|
|
293 out1[0x10* 3] = b1[0x19] + b1[0x15];
|
|
|
294 b1[0x1D] += b1[0x1B];
|
|
|
295 out1[0x10* 5] = b1[0x15] + b1[0x1D];
|
|
|
296 out1[0x10* 7] = b1[0x1D] + b1[0x13];
|
|
|
297 b1[0x1B] += b1[0x1F];
|
|
|
298 out1[0x10* 9] = b1[0x13] + b1[0x1B];
|
|
|
299 out1[0x10*11] = b1[0x1B] + b1[0x17];
|
|
|
300 out1[0x10*13] = b1[0x17] + b1[0x1F];
|
|
|
301 out1[0x10*15] = b1[0x1F];
|
|
|
302 }
|
|
|
303
|
|
|
304 /*
|
|
|
305 * the call via dct64 is a trick to force GCC to use
|
|
|
306 * (new) registers for the b1,b2 pointer to the bufs[xx] field
|
|
|
307 */
|
|
|
308 void dct64(real *a,real *b,real *c)
|
|
|
309 {
|
|
|
310 real bufs[0x40];
|
|
|
311 dct64_1(a,b,bufs,bufs+0x20,c);
|
|
|
312 }
|
|
|
313
|
