Mercurial > libavcodec.hg
annotate dsputil.c @ 177:53da914d6f46 libavcodec
avcodec_find_encoder_by_name() patch by Alex
author | arpi_esp |
---|---|
date | Sun, 23 Dec 2001 22:25:18 +0000 |
parents | ac5075a55488 |
children | 9e0e56869d05 |
rev | line source |
---|---|
0 | 1 /* |
2 * DSP utils | |
3 * Copyright (c) 2000, 2001 Gerard Lantau. | |
4 * | |
5 * This program is free software; you can redistribute it and/or modify | |
6 * it under the terms of the GNU General Public License as published by | |
7 * the Free Software Foundation; either version 2 of the License, or | |
8 * (at your option) any later version. | |
9 * | |
10 * This program is distributed in the hope that it will be useful, | |
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
13 * GNU General Public License for more details. | |
14 * | |
15 * You should have received a copy of the GNU General Public License | |
16 * along with this program; if not, write to the Free Software | |
17 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. | |
18 */ | |
19 #include <stdlib.h> | |
20 #include <stdio.h> | |
21 #include "avcodec.h" | |
22 #include "dsputil.h" | |
174
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
23 #include "simple_idct.h" |
0 | 24 |
19
82d4c9be9873
MMX/MMXEXT iDCT support, using external functions currently defined in libmpeg2
arpi_esp
parents:
6
diff
changeset
|
25 void (*ff_idct)(DCTELEM *block); |
0 | 26 void (*get_pixels)(DCTELEM *block, const UINT8 *pixels, int line_size); |
27 void (*put_pixels_clamped)(const DCTELEM *block, UINT8 *pixels, int line_size); | |
28 void (*add_pixels_clamped)(const DCTELEM *block, UINT8 *pixels, int line_size); | |
29 | |
30 op_pixels_abs_func pix_abs16x16; | |
31 op_pixels_abs_func pix_abs16x16_x2; | |
32 op_pixels_abs_func pix_abs16x16_y2; | |
33 op_pixels_abs_func pix_abs16x16_xy2; | |
34 | |
50 | 35 UINT8 cropTbl[256 + 2 * MAX_NEG_CROP]; |
0 | 36 UINT32 squareTbl[512]; |
37 | |
34 | 38 extern UINT16 default_intra_matrix[64]; |
39 extern UINT16 default_non_intra_matrix[64]; | |
40 | |
41 UINT8 zigzag_direct[64] = { | |
42 0, 1, 8, 16, 9, 2, 3, 10, | |
43 17, 24, 32, 25, 18, 11, 4, 5, | |
44 12, 19, 26, 33, 40, 48, 41, 34, | |
45 27, 20, 13, 6, 7, 14, 21, 28, | |
46 35, 42, 49, 56, 57, 50, 43, 36, | |
47 29, 22, 15, 23, 30, 37, 44, 51, | |
48 58, 59, 52, 45, 38, 31, 39, 46, | |
49 53, 60, 61, 54, 47, 55, 62, 63 | |
50 }; | |
51 | |
52 UINT8 ff_alternate_horizontal_scan[64] = { | |
53 0, 1, 2, 3, 8, 9, 16, 17, | |
54 10, 11, 4, 5, 6, 7, 15, 14, | |
55 13, 12, 19, 18, 24, 25, 32, 33, | |
56 26, 27, 20, 21, 22, 23, 28, 29, | |
57 30, 31, 34, 35, 40, 41, 48, 49, | |
58 42, 43, 36, 37, 38, 39, 44, 45, | |
59 46, 47, 50, 51, 56, 57, 58, 59, | |
60 52, 53, 54, 55, 60, 61, 62, 63, | |
61 }; | |
62 | |
63 UINT8 ff_alternate_vertical_scan[64] = { | |
64 0, 8, 16, 24, 1, 9, 2, 10, | |
65 17, 25, 32, 40, 48, 56, 57, 49, | |
66 41, 33, 26, 18, 3, 11, 4, 12, | |
67 19, 27, 34, 42, 50, 58, 35, 43, | |
68 51, 59, 20, 28, 5, 13, 6, 14, | |
69 21, 29, 36, 44, 52, 60, 37, 45, | |
70 53, 61, 22, 30, 7, 15, 23, 31, | |
71 38, 46, 54, 62, 39, 47, 55, 63, | |
72 }; | |
73 | |
0 | 74 void get_pixels_c(DCTELEM *block, const UINT8 *pixels, int line_size) |
75 { | |
76 DCTELEM *p; | |
77 const UINT8 *pix; | |
78 int i; | |
79 | |
80 /* read the pixels */ | |
81 p = block; | |
82 pix = pixels; | |
83 for(i=0;i<8;i++) { | |
84 p[0] = pix[0]; | |
85 p[1] = pix[1]; | |
86 p[2] = pix[2]; | |
87 p[3] = pix[3]; | |
88 p[4] = pix[4]; | |
89 p[5] = pix[5]; | |
90 p[6] = pix[6]; | |
91 p[7] = pix[7]; | |
92 pix += line_size; | |
93 p += 8; | |
94 } | |
95 } | |
96 | |
97 void put_pixels_clamped_c(const DCTELEM *block, UINT8 *pixels, int line_size) | |
98 { | |
99 const DCTELEM *p; | |
100 UINT8 *pix; | |
101 int i; | |
102 UINT8 *cm = cropTbl + MAX_NEG_CROP; | |
103 | |
104 /* read the pixels */ | |
105 p = block; | |
106 pix = pixels; | |
107 for(i=0;i<8;i++) { | |
108 pix[0] = cm[p[0]]; | |
109 pix[1] = cm[p[1]]; | |
110 pix[2] = cm[p[2]]; | |
111 pix[3] = cm[p[3]]; | |
112 pix[4] = cm[p[4]]; | |
113 pix[5] = cm[p[5]]; | |
114 pix[6] = cm[p[6]]; | |
115 pix[7] = cm[p[7]]; | |
116 pix += line_size; | |
117 p += 8; | |
118 } | |
119 } | |
120 | |
121 void add_pixels_clamped_c(const DCTELEM *block, UINT8 *pixels, int line_size) | |
122 { | |
123 const DCTELEM *p; | |
124 UINT8 *pix; | |
125 int i; | |
126 UINT8 *cm = cropTbl + MAX_NEG_CROP; | |
127 | |
128 /* read the pixels */ | |
129 p = block; | |
130 pix = pixels; | |
131 for(i=0;i<8;i++) { | |
132 pix[0] = cm[pix[0] + p[0]]; | |
133 pix[1] = cm[pix[1] + p[1]]; | |
134 pix[2] = cm[pix[2] + p[2]]; | |
135 pix[3] = cm[pix[3] + p[3]]; | |
136 pix[4] = cm[pix[4] + p[4]]; | |
137 pix[5] = cm[pix[5] + p[5]]; | |
138 pix[6] = cm[pix[6] + p[6]]; | |
139 pix[7] = cm[pix[7] + p[7]]; | |
140 pix += line_size; | |
141 p += 8; | |
142 } | |
143 } | |
144 | |
145 #define PIXOP(BTYPE, OPNAME, OP, INCR) \ | |
146 \ | |
147 static void OPNAME ## _pixels(BTYPE *block, const UINT8 *pixels, int line_size, int h) \ | |
148 { \ | |
149 BTYPE *p; \ | |
150 const UINT8 *pix; \ | |
151 \ | |
152 p = block; \ | |
153 pix = pixels; \ | |
154 do { \ | |
155 OP(p[0], pix[0]); \ | |
156 OP(p[1], pix[1]); \ | |
157 OP(p[2], pix[2]); \ | |
158 OP(p[3], pix[3]); \ | |
159 OP(p[4], pix[4]); \ | |
160 OP(p[5], pix[5]); \ | |
161 OP(p[6], pix[6]); \ | |
162 OP(p[7], pix[7]); \ | |
163 pix += line_size; \ | |
164 p += INCR; \ | |
165 } while (--h);; \ | |
166 } \ | |
167 \ | |
168 static void OPNAME ## _pixels_x2(BTYPE *block, const UINT8 *pixels, int line_size, int h) \ | |
169 { \ | |
170 BTYPE *p; \ | |
171 const UINT8 *pix; \ | |
172 \ | |
173 p = block; \ | |
174 pix = pixels; \ | |
175 do { \ | |
176 OP(p[0], avg2(pix[0], pix[1])); \ | |
177 OP(p[1], avg2(pix[1], pix[2])); \ | |
178 OP(p[2], avg2(pix[2], pix[3])); \ | |
179 OP(p[3], avg2(pix[3], pix[4])); \ | |
180 OP(p[4], avg2(pix[4], pix[5])); \ | |
181 OP(p[5], avg2(pix[5], pix[6])); \ | |
182 OP(p[6], avg2(pix[6], pix[7])); \ | |
183 OP(p[7], avg2(pix[7], pix[8])); \ | |
184 pix += line_size; \ | |
185 p += INCR; \ | |
186 } while (--h); \ | |
187 } \ | |
188 \ | |
189 static void OPNAME ## _pixels_y2(BTYPE *block, const UINT8 *pixels, int line_size, int h) \ | |
190 { \ | |
191 BTYPE *p; \ | |
192 const UINT8 *pix; \ | |
193 const UINT8 *pix1; \ | |
194 \ | |
195 p = block; \ | |
196 pix = pixels; \ | |
197 pix1 = pixels + line_size; \ | |
198 do { \ | |
199 OP(p[0], avg2(pix[0], pix1[0])); \ | |
200 OP(p[1], avg2(pix[1], pix1[1])); \ | |
201 OP(p[2], avg2(pix[2], pix1[2])); \ | |
202 OP(p[3], avg2(pix[3], pix1[3])); \ | |
203 OP(p[4], avg2(pix[4], pix1[4])); \ | |
204 OP(p[5], avg2(pix[5], pix1[5])); \ | |
205 OP(p[6], avg2(pix[6], pix1[6])); \ | |
206 OP(p[7], avg2(pix[7], pix1[7])); \ | |
207 pix += line_size; \ | |
208 pix1 += line_size; \ | |
209 p += INCR; \ | |
210 } while(--h); \ | |
211 } \ | |
212 \ | |
213 static void OPNAME ## _pixels_xy2(BTYPE *block, const UINT8 *pixels, int line_size, int h) \ | |
214 { \ | |
215 BTYPE *p; \ | |
216 const UINT8 *pix; \ | |
217 const UINT8 *pix1; \ | |
218 \ | |
219 p = block; \ | |
220 pix = pixels; \ | |
221 pix1 = pixels + line_size; \ | |
222 do { \ | |
223 OP(p[0], avg4(pix[0], pix[1], pix1[0], pix1[1])); \ | |
224 OP(p[1], avg4(pix[1], pix[2], pix1[1], pix1[2])); \ | |
225 OP(p[2], avg4(pix[2], pix[3], pix1[2], pix1[3])); \ | |
226 OP(p[3], avg4(pix[3], pix[4], pix1[3], pix1[4])); \ | |
227 OP(p[4], avg4(pix[4], pix[5], pix1[4], pix1[5])); \ | |
228 OP(p[5], avg4(pix[5], pix[6], pix1[5], pix1[6])); \ | |
229 OP(p[6], avg4(pix[6], pix[7], pix1[6], pix1[7])); \ | |
230 OP(p[7], avg4(pix[7], pix[8], pix1[7], pix1[8])); \ | |
231 pix += line_size; \ | |
232 pix1 += line_size; \ | |
233 p += INCR; \ | |
234 } while(--h); \ | |
235 } \ | |
236 \ | |
237 void (*OPNAME ## _pixels_tab[4])(BTYPE *block, const UINT8 *pixels, int line_size, int h) = { \ | |
238 OPNAME ## _pixels, \ | |
239 OPNAME ## _pixels_x2, \ | |
240 OPNAME ## _pixels_y2, \ | |
241 OPNAME ## _pixels_xy2, \ | |
242 }; | |
243 | |
244 | |
245 /* rounding primitives */ | |
246 #define avg2(a,b) ((a+b+1)>>1) | |
247 #define avg4(a,b,c,d) ((a+b+c+d+2)>>2) | |
248 | |
249 #define op_put(a, b) a = b | |
250 #define op_avg(a, b) a = avg2(a, b) | |
251 #define op_sub(a, b) a -= b | |
252 | |
253 PIXOP(UINT8, put, op_put, line_size) | |
254 PIXOP(UINT8, avg, op_avg, line_size) | |
255 | |
256 PIXOP(DCTELEM, sub, op_sub, 8) | |
257 | |
258 /* not rounding primitives */ | |
259 #undef avg2 | |
260 #undef avg4 | |
261 #define avg2(a,b) ((a+b)>>1) | |
262 #define avg4(a,b,c,d) ((a+b+c+d+1)>>2) | |
263 | |
264 PIXOP(UINT8, put_no_rnd, op_put, line_size) | |
265 PIXOP(UINT8, avg_no_rnd, op_avg, line_size) | |
266 | |
267 /* motion estimation */ | |
268 | |
269 #undef avg2 | |
270 #undef avg4 | |
271 #define avg2(a,b) ((a+b+1)>>1) | |
272 #define avg4(a,b,c,d) ((a+b+c+d+2)>>2) | |
273 | |
274 int pix_abs16x16_c(UINT8 *pix1, UINT8 *pix2, int line_size, int h) | |
275 { | |
276 int s, i; | |
277 | |
278 s = 0; | |
279 for(i=0;i<h;i++) { | |
280 s += abs(pix1[0] - pix2[0]); | |
281 s += abs(pix1[1] - pix2[1]); | |
282 s += abs(pix1[2] - pix2[2]); | |
283 s += abs(pix1[3] - pix2[3]); | |
284 s += abs(pix1[4] - pix2[4]); | |
285 s += abs(pix1[5] - pix2[5]); | |
286 s += abs(pix1[6] - pix2[6]); | |
287 s += abs(pix1[7] - pix2[7]); | |
288 s += abs(pix1[8] - pix2[8]); | |
289 s += abs(pix1[9] - pix2[9]); | |
290 s += abs(pix1[10] - pix2[10]); | |
291 s += abs(pix1[11] - pix2[11]); | |
292 s += abs(pix1[12] - pix2[12]); | |
293 s += abs(pix1[13] - pix2[13]); | |
294 s += abs(pix1[14] - pix2[14]); | |
295 s += abs(pix1[15] - pix2[15]); | |
296 pix1 += line_size; | |
297 pix2 += line_size; | |
298 } | |
299 return s; | |
300 } | |
301 | |
302 int pix_abs16x16_x2_c(UINT8 *pix1, UINT8 *pix2, int line_size, int h) | |
303 { | |
304 int s, i; | |
305 | |
306 s = 0; | |
307 for(i=0;i<h;i++) { | |
308 s += abs(pix1[0] - avg2(pix2[0], pix2[1])); | |
309 s += abs(pix1[1] - avg2(pix2[1], pix2[2])); | |
310 s += abs(pix1[2] - avg2(pix2[2], pix2[3])); | |
311 s += abs(pix1[3] - avg2(pix2[3], pix2[4])); | |
312 s += abs(pix1[4] - avg2(pix2[4], pix2[5])); | |
313 s += abs(pix1[5] - avg2(pix2[5], pix2[6])); | |
314 s += abs(pix1[6] - avg2(pix2[6], pix2[7])); | |
315 s += abs(pix1[7] - avg2(pix2[7], pix2[8])); | |
316 s += abs(pix1[8] - avg2(pix2[8], pix2[9])); | |
317 s += abs(pix1[9] - avg2(pix2[9], pix2[10])); | |
318 s += abs(pix1[10] - avg2(pix2[10], pix2[11])); | |
319 s += abs(pix1[11] - avg2(pix2[11], pix2[12])); | |
320 s += abs(pix1[12] - avg2(pix2[12], pix2[13])); | |
321 s += abs(pix1[13] - avg2(pix2[13], pix2[14])); | |
322 s += abs(pix1[14] - avg2(pix2[14], pix2[15])); | |
323 s += abs(pix1[15] - avg2(pix2[15], pix2[16])); | |
324 pix1 += line_size; | |
325 pix2 += line_size; | |
326 } | |
327 return s; | |
328 } | |
329 | |
330 int pix_abs16x16_y2_c(UINT8 *pix1, UINT8 *pix2, int line_size, int h) | |
331 { | |
332 int s, i; | |
333 UINT8 *pix3 = pix2 + line_size; | |
334 | |
335 s = 0; | |
336 for(i=0;i<h;i++) { | |
337 s += abs(pix1[0] - avg2(pix2[0], pix3[0])); | |
338 s += abs(pix1[1] - avg2(pix2[1], pix3[1])); | |
339 s += abs(pix1[2] - avg2(pix2[2], pix3[2])); | |
340 s += abs(pix1[3] - avg2(pix2[3], pix3[3])); | |
341 s += abs(pix1[4] - avg2(pix2[4], pix3[4])); | |
342 s += abs(pix1[5] - avg2(pix2[5], pix3[5])); | |
343 s += abs(pix1[6] - avg2(pix2[6], pix3[6])); | |
344 s += abs(pix1[7] - avg2(pix2[7], pix3[7])); | |
345 s += abs(pix1[8] - avg2(pix2[8], pix3[8])); | |
346 s += abs(pix1[9] - avg2(pix2[9], pix3[9])); | |
347 s += abs(pix1[10] - avg2(pix2[10], pix3[10])); | |
348 s += abs(pix1[11] - avg2(pix2[11], pix3[11])); | |
349 s += abs(pix1[12] - avg2(pix2[12], pix3[12])); | |
350 s += abs(pix1[13] - avg2(pix2[13], pix3[13])); | |
351 s += abs(pix1[14] - avg2(pix2[14], pix3[14])); | |
352 s += abs(pix1[15] - avg2(pix2[15], pix3[15])); | |
353 pix1 += line_size; | |
354 pix2 += line_size; | |
355 pix3 += line_size; | |
356 } | |
357 return s; | |
358 } | |
359 | |
360 int pix_abs16x16_xy2_c(UINT8 *pix1, UINT8 *pix2, int line_size, int h) | |
361 { | |
362 int s, i; | |
363 UINT8 *pix3 = pix2 + line_size; | |
364 | |
365 s = 0; | |
366 for(i=0;i<h;i++) { | |
367 s += abs(pix1[0] - avg4(pix2[0], pix2[1], pix3[0], pix3[1])); | |
368 s += abs(pix1[1] - avg4(pix2[1], pix2[2], pix3[1], pix3[2])); | |
369 s += abs(pix1[2] - avg4(pix2[2], pix2[3], pix3[2], pix3[3])); | |
370 s += abs(pix1[3] - avg4(pix2[3], pix2[4], pix3[3], pix3[4])); | |
371 s += abs(pix1[4] - avg4(pix2[4], pix2[5], pix3[4], pix3[5])); | |
372 s += abs(pix1[5] - avg4(pix2[5], pix2[6], pix3[5], pix3[6])); | |
373 s += abs(pix1[6] - avg4(pix2[6], pix2[7], pix3[6], pix3[7])); | |
374 s += abs(pix1[7] - avg4(pix2[7], pix2[8], pix3[7], pix3[8])); | |
375 s += abs(pix1[8] - avg4(pix2[8], pix2[9], pix3[8], pix3[9])); | |
376 s += abs(pix1[9] - avg4(pix2[9], pix2[10], pix3[9], pix3[10])); | |
377 s += abs(pix1[10] - avg4(pix2[10], pix2[11], pix3[10], pix3[11])); | |
378 s += abs(pix1[11] - avg4(pix2[11], pix2[12], pix3[11], pix3[12])); | |
379 s += abs(pix1[12] - avg4(pix2[12], pix2[13], pix3[12], pix3[13])); | |
380 s += abs(pix1[13] - avg4(pix2[13], pix2[14], pix3[13], pix3[14])); | |
381 s += abs(pix1[14] - avg4(pix2[14], pix2[15], pix3[14], pix3[15])); | |
382 s += abs(pix1[15] - avg4(pix2[15], pix2[16], pix3[15], pix3[16])); | |
383 pix1 += line_size; | |
384 pix2 += line_size; | |
385 pix3 += line_size; | |
386 } | |
387 return s; | |
388 } | |
389 | |
34 | 390 /* permute block according so that it corresponds to the MMX idct |
391 order */ | |
174
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
392 #ifdef SIMPLE_IDCT |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
393 void block_permute(INT16 *block) |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
394 { |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
395 int i; |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
396 INT16 temp[64]; |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
397 |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
398 // for(i=0; i<64; i++) temp[i] = block[ block_permute_op(i) ]; |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
399 for(i=0; i<64; i++) temp[ block_permute_op(i) ] = block[i]; |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
400 |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
401 for(i=0; i<64; i++) block[i] = temp[i]; |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
402 /* |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
403 for(i=0; i<64; i++) |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
404 { |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
405 if((i&7)==0) printf("\n"); |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
406 printf("%2d ", block[i]); |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
407 } |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
408 */ |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
409 } |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
410 |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
411 #else |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
412 |
34 | 413 void block_permute(INT16 *block) |
414 { | |
415 int tmp1, tmp2, tmp3, tmp4, tmp5, tmp6; | |
416 int i; | |
417 | |
418 for(i=0;i<8;i++) { | |
419 tmp1 = block[1]; | |
420 tmp2 = block[2]; | |
421 tmp3 = block[3]; | |
422 tmp4 = block[4]; | |
423 tmp5 = block[5]; | |
424 tmp6 = block[6]; | |
425 block[1] = tmp2; | |
426 block[2] = tmp4; | |
427 block[3] = tmp6; | |
428 block[4] = tmp1; | |
429 block[5] = tmp3; | |
430 block[6] = tmp5; | |
431 block += 8; | |
432 } | |
433 } | |
174
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
434 #endif |
34 | 435 |
0 | 436 void dsputil_init(void) |
437 { | |
34 | 438 int i, j; |
88 | 439 int use_permuted_idct; |
0 | 440 |
441 for(i=0;i<256;i++) cropTbl[i + MAX_NEG_CROP] = i; | |
442 for(i=0;i<MAX_NEG_CROP;i++) { | |
443 cropTbl[i] = 0; | |
444 cropTbl[i + MAX_NEG_CROP + 256] = 255; | |
445 } | |
446 | |
447 for(i=0;i<512;i++) { | |
448 squareTbl[i] = (i - 256) * (i - 256); | |
449 } | |
450 | |
174
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
451 #ifdef SIMPLE_IDCT |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
452 ff_idct = simple_idct; |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
453 #else |
19
82d4c9be9873
MMX/MMXEXT iDCT support, using external functions currently defined in libmpeg2
arpi_esp
parents:
6
diff
changeset
|
454 ff_idct = j_rev_dct; |
174
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
455 #endif |
0 | 456 get_pixels = get_pixels_c; |
457 put_pixels_clamped = put_pixels_clamped_c; | |
458 add_pixels_clamped = add_pixels_clamped_c; | |
459 | |
460 pix_abs16x16 = pix_abs16x16_c; | |
461 pix_abs16x16_x2 = pix_abs16x16_x2_c; | |
462 pix_abs16x16_y2 = pix_abs16x16_y2_c; | |
463 pix_abs16x16_xy2 = pix_abs16x16_xy2_c; | |
464 av_fdct = jpeg_fdct_ifast; | |
465 | |
88 | 466 use_permuted_idct = 1; |
34 | 467 |
2 | 468 #ifdef HAVE_MMX |
0 | 469 dsputil_init_mmx(); |
470 #endif | |
62 | 471 #ifdef ARCH_ARMV4L |
472 dsputil_init_armv4l(); | |
473 #endif | |
88 | 474 #ifdef HAVE_MLIB |
475 dsputil_init_mlib(); | |
476 use_permuted_idct = 0; | |
477 #endif | |
478 | |
174
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
479 #ifdef SIMPLE_IDCT |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
480 if(ff_idct == simple_idct) use_permuted_idct=0; |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
481 #endif |
ac5075a55488
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
88
diff
changeset
|
482 |
88 | 483 if (use_permuted_idct) { |
484 /* permute for IDCT */ | |
485 for(i=0;i<64;i++) { | |
486 j = zigzag_direct[i]; | |
487 zigzag_direct[i] = block_permute_op(j); | |
488 j = ff_alternate_horizontal_scan[i]; | |
489 ff_alternate_horizontal_scan[i] = block_permute_op(j); | |
490 j = ff_alternate_vertical_scan[i]; | |
491 ff_alternate_vertical_scan[i] = block_permute_op(j); | |
492 } | |
493 block_permute(default_intra_matrix); | |
494 block_permute(default_non_intra_matrix); | |
495 } | |
0 | 496 } |