[svn] / branches / dev-api-4 / xvidcore / src / utils / mbtransquant.c Repository:
ViewVC logotype

Annotation of /branches/dev-api-4/xvidcore/src/utils/mbtransquant.c

Parent Directory Parent Directory | Revision Log Revision Log


Revision 1014 - (view) (download)

1 : edgomez 965 /*****************************************************************************
2 :     *
3 :     * XVID MPEG-4 VIDEO CODEC
4 :     * - MB Transfert/Quantization functions -
5 :     *
6 :     * Copyright(C) 2001-2003 Peter Ross <pross@xvid.org>
7 :     * 2001-2003 Michael Militzer <isibaar@xvid.org>
8 :     * 2003 Edouard Gomez <ed.gomez@free.fr>
9 :     *
10 :     * This program is free software ; you can redistribute it and/or modify
11 :     * it under the terms of the GNU General Public License as published by
12 :     * the Free Software Foundation ; either version 2 of the License, or
13 :     * (at your option) any later version.
14 :     *
15 :     * This program is distributed in the hope that it will be useful,
16 :     * but WITHOUT ANY WARRANTY ; without even the implied warranty of
17 :     * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 :     * GNU General Public License for more details.
19 :     *
20 :     * You should have received a copy of the GNU General Public License
21 :     * along with this program ; if not, write to the Free Software
22 :     * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
23 :     *
24 : suxen_drol 1014 * $Id: mbtransquant.c,v 1.21.2.12 2003-05-12 12:33:16 suxen_drol Exp $
25 : edgomez 965 *
26 :     ****************************************************************************/
27 : Isibaar 3
28 : chl 1012 #include <stdio.h>
29 :     #include <stdlib.h>
30 : edgomez 78 #include <string.h>
31 :    
32 : Isibaar 3 #include "../portab.h"
33 :     #include "mbfunctions.h"
34 :    
35 :     #include "../global.h"
36 :     #include "mem_transfer.h"
37 :     #include "timer.h"
38 : chl 995 #include "../bitstream/mbcoding.h"
39 : chl 1011 #include "../bitstream/zigzag.h"
40 : Isibaar 3 #include "../dct/fdct.h"
41 :     #include "../dct/idct.h"
42 :     #include "../quant/quant_mpeg4.h"
43 :     #include "../quant/quant_h263.h"
44 :     #include "../encoder.h"
45 :    
46 : edgomez 851 #include "../image/reduced.h"
47 : Isibaar 3
48 : edgomez 851 MBFIELDTEST_PTR MBFieldTest;
49 : Isibaar 3
50 : edgomez 965 /*
51 :     * Skip blocks having a coefficient sum below this value. This value will be
52 :     * corrected according to the MB quantizer to avoid artifacts for quant==1
53 :     */
54 :     #define PVOP_TOOSMALL_LIMIT 1
55 :     #define BVOP_TOOSMALL_LIMIT 3
56 : Isibaar 3
57 : edgomez 965 /*****************************************************************************
58 :     * Local functions
59 :     ****************************************************************************/
60 :    
61 :     /* permute block and return field dct choice */
62 :     static __inline uint32_t
63 :     MBDecideFieldDCT(int16_t data[6 * 64])
64 : Isibaar 3 {
65 : edgomez 965 uint32_t field = MBFieldTest(data);
66 : edgomez 78
67 : edgomez 965 if (field)
68 :     MBFrameToField(data);
69 : Isibaar 3
70 : edgomez 965 return field;
71 :     }
72 : h 69
73 : edgomez 965 /* Performs Forward DCT on all blocks */
74 :     static __inline void
75 : syskin 984 MBfDCT(const MBParam * const pParam,
76 :     const FRAMEINFO * const frame,
77 :     MACROBLOCK * const pMB,
78 : edgomez 965 uint32_t x_pos,
79 :     uint32_t y_pos,
80 :     int16_t data[6 * 64])
81 : syskin 984 {
82 : edgomez 965 /* Handles interlacing */
83 : h 69 start_timer();
84 :     pMB->field_dct = 0;
85 : edgomez 949 if ((frame->vol_flags & XVID_VOL_INTERLACING) &&
86 : h 390 (x_pos>0) && (x_pos<pParam->mb_width-1) &&
87 :     (y_pos>0) && (y_pos<pParam->mb_height-1)) {
88 : h 69 pMB->field_dct = MBDecideFieldDCT(data);
89 :     }
90 :     stop_interlacing_timer();
91 :    
92 : edgomez 965 /* Perform DCT */
93 :     start_timer();
94 :     fdct(&data[0 * 64]);
95 :     fdct(&data[1 * 64]);
96 :     fdct(&data[2 * 64]);
97 :     fdct(&data[3 * 64]);
98 :     fdct(&data[4 * 64]);
99 :     fdct(&data[5 * 64]);
100 :     stop_dct_timer();
101 :     }
102 :    
103 :     /* Performs Inverse DCT on all blocks */
104 :     static __inline void
105 :     MBiDCT(int16_t data[6 * 64],
106 :     const uint8_t cbp)
107 :     {
108 :     start_timer();
109 :     if(cbp & (1 << (5 - 0))) idct(&data[0 * 64]);
110 :     if(cbp & (1 << (5 - 1))) idct(&data[1 * 64]);
111 :     if(cbp & (1 << (5 - 2))) idct(&data[2 * 64]);
112 :     if(cbp & (1 << (5 - 3))) idct(&data[3 * 64]);
113 :     if(cbp & (1 << (5 - 4))) idct(&data[4 * 64]);
114 :     if(cbp & (1 << (5 - 5))) idct(&data[5 * 64]);
115 :     stop_idct_timer();
116 :     }
117 :    
118 :     /* Quantize all blocks -- Intra mode */
119 :     static __inline void
120 :     MBQuantIntra(const MBParam * pParam,
121 : chl 995 const FRAMEINFO * const frame,
122 : edgomez 965 const MACROBLOCK * pMB,
123 : syskin 984 int16_t qcoeff[6 * 64],
124 : edgomez 965 int16_t data[6*64])
125 :     {
126 :     int i;
127 :    
128 : edgomez 195 for (i = 0; i < 6; i++) {
129 : edgomez 965 uint32_t iDcScaler = get_dc_scaler(pMB->quant, i < 4);
130 :    
131 :     /* Quantize the block */
132 :     start_timer();
133 : chl 995 if (!(pParam->vol_flags & XVID_VOL_MPEGQUANT)) {
134 : edgomez 965 quant_intra(&data[i * 64], &qcoeff[i * 64], pMB->quant, iDcScaler);
135 : chl 995 } else {
136 : edgomez 965 quant4_intra(&data[i * 64], &qcoeff[i * 64], pMB->quant, iDcScaler);
137 : chl 995 }
138 : edgomez 965 stop_quant_timer();
139 :     }
140 :     }
141 :    
142 :     /* DeQuantize all blocks -- Intra mode */
143 :     static __inline void
144 :     MBDeQuantIntra(const MBParam * pParam,
145 :     const int iQuant,
146 :     int16_t qcoeff[6 * 64],
147 :     int16_t data[6*64])
148 :     {
149 :     int i;
150 :    
151 :     for (i = 0; i < 6; i++) {
152 : Isibaar 3 uint32_t iDcScaler = get_dc_scaler(iQuant, i < 4);
153 :    
154 :     start_timer();
155 : edgomez 965 if (!(pParam->vol_flags & XVID_VOL_MPEGQUANT))
156 :     dequant_intra(&qcoeff[i * 64], &data[i * 64], iQuant, iDcScaler);
157 :     else
158 :     dequant4_intra(&qcoeff[i * 64], &data[i * 64], iQuant, iDcScaler);
159 :     stop_iquant_timer();
160 :     }
161 :     }
162 : Isibaar 3
163 : chl 1011
164 :     static int
165 :     dct_quantize_trellis_h263_c(int16_t *const Out, const int16_t *const In, int Q, const uint16_t * const Zigzag, int Non_Zero);
166 :    
167 :     static int
168 :     dct_quantize_trellis_mpeg_c(int16_t *const Out, const int16_t *const In, int Q, const uint16_t * const Zigzag, int Non_Zero);
169 :    
170 :    
171 : edgomez 965 /* Quantize all blocks -- Inter mode */
172 :     static __inline uint8_t
173 :     MBQuantInter(const MBParam * pParam,
174 : chl 995 const FRAMEINFO * const frame,
175 : edgomez 965 const MACROBLOCK * pMB,
176 :     int16_t data[6 * 64],
177 :     int16_t qcoeff[6 * 64],
178 :     int bvop,
179 :     int limit)
180 :     {
181 :    
182 :     int i;
183 :     uint8_t cbp = 0;
184 :     int sum;
185 :     int code_block;
186 :    
187 :     for (i = 0; i < 6; i++) {
188 : syskin 984
189 : edgomez 965 /* Quantize the block */
190 :     start_timer();
191 : chl 995 if (!(pParam->vol_flags & XVID_VOL_MPEGQUANT)) {
192 :     sum = quant_inter(&qcoeff[i*64], &data[i*64], pMB->quant);
193 :     if ( (sum) && (frame->vop_flags & XVID_VOP_TRELLISQUANT) ) {
194 : chl 1011 sum = dct_quantize_trellis_h263_c(&qcoeff[i*64], &data[i*64], pMB->quant, &scan_tables[0][0], 63)+1;
195 : chl 995 limit = 1;
196 :     }
197 :     } else {
198 : edgomez 965 sum = quant4_inter(&qcoeff[i * 64], &data[i * 64], pMB->quant);
199 : chl 995 // if ( (sum) && (frame->vop_flags & XVID_VOP_TRELLISQUANT) )
200 : chl 1011 // sum = dct_quantize_trellis_mpeg_c (&qcoeff[i*64], &data[i*64], pMB->quant)+1;
201 : chl 995 }
202 : edgomez 965 stop_quant_timer();
203 :    
204 :     /*
205 :     * We code the block if the sum is higher than the limit and if the first
206 :     * two AC coefficients in zig zag order are not zero.
207 :     */
208 :     code_block = 0;
209 :     if ((sum >= limit) || (qcoeff[i*64+1] != 0) || (qcoeff[i*64+8] != 0)) {
210 :     code_block = 1;
211 : edgomez 195 } else {
212 : Isibaar 3
213 : edgomez 965 if (bvop && (pMB->mode == MODE_DIRECT || pMB->mode == MODE_DIRECT_NO4V)) {
214 :     /* dark blocks prevention for direct mode */
215 :     if ((qcoeff[i*64] < -1) || (qcoeff[i*64] > 0))
216 :     code_block = 1;
217 : edgomez 851 } else {
218 : edgomez 965 /* not direct mode */
219 :     if (qcoeff[i*64] != 0)
220 :     code_block = 1;
221 : edgomez 851 }
222 : Isibaar 3 }
223 :    
224 : edgomez 965 /* Set the corresponding cbp bit */
225 :     cbp |= code_block << (5 - i);
226 :     }
227 : edgomez 851
228 : edgomez 965 return(cbp);
229 :     }
230 : Isibaar 3
231 : edgomez 965 /* DeQuantize all blocks -- Inter mode */
232 : syskin 984 static __inline void
233 : edgomez 965 MBDeQuantInter(const MBParam * pParam,
234 :     const int iQuant,
235 :     int16_t data[6 * 64],
236 :     int16_t qcoeff[6 * 64],
237 :     const uint8_t cbp)
238 :     {
239 :     int i;
240 :    
241 :     for (i = 0; i < 6; i++) {
242 : syskin 984 if (cbp & (1 << (5 - i))) {
243 : edgomez 965 start_timer();
244 :     if (!(pParam->vol_flags & XVID_VOL_MPEGQUANT))
245 :     dequant_inter(&data[i * 64], &qcoeff[i * 64], iQuant);
246 :     else
247 :     dequant4_inter(&data[i * 64], &qcoeff[i * 64], iQuant);
248 :     stop_iquant_timer();
249 : edgomez 851 }
250 : h 69 }
251 : Isibaar 3 }
252 :    
253 : edgomez 965 typedef void (transfer_operation_8to16_t) (int16_t *Dst, const uint8_t *Src, int BpS);
254 :     typedef void (transfer_operation_16to8_t) (uint8_t *Dst, const int16_t *Src, int BpS);
255 : Isibaar 3
256 : edgomez 78
257 : edgomez 965 static __inline void
258 : syskin 984 MBTrans8to16(const MBParam * const pParam,
259 :     const FRAMEINFO * const frame,
260 :     const MACROBLOCK * const pMB,
261 : edgomez 965 const uint32_t x_pos,
262 :     const uint32_t y_pos,
263 :     int16_t data[6 * 64])
264 :     {
265 : h 82 uint32_t stride = pParam->edged_width;
266 :     uint32_t stride2 = stride / 2;
267 : edgomez 965 uint32_t next_block = stride * 8;
268 : syskin 984 int32_t cst;
269 : Isibaar 3 uint8_t *pY_Cur, *pU_Cur, *pV_Cur;
270 : syskin 984 const IMAGE * const pCurrent = &frame->image;
271 : edgomez 965 transfer_operation_8to16_t *transfer_op = NULL;
272 : edgomez 195
273 : edgomez 965 if ((frame->vop_flags & XVID_VOP_REDUCED)) {
274 :    
275 :     /* Image pointers */
276 :     pY_Cur = pCurrent->y + (y_pos << 5) * stride + (x_pos << 5);
277 : edgomez 851 pU_Cur = pCurrent->u + (y_pos << 4) * stride2 + (x_pos << 4);
278 :     pV_Cur = pCurrent->v + (y_pos << 4) * stride2 + (x_pos << 4);
279 : edgomez 965
280 :     /* Block size */
281 :     cst = 16;
282 :    
283 :     /* Operation function */
284 :     transfer_op = (transfer_operation_8to16_t*)filter_18x18_to_8x8;
285 :     } else {
286 :    
287 :     /* Image pointers */
288 :     pY_Cur = pCurrent->y + (y_pos << 4) * stride + (x_pos << 4);
289 : edgomez 851 pU_Cur = pCurrent->u + (y_pos << 3) * stride2 + (x_pos << 3);
290 :     pV_Cur = pCurrent->v + (y_pos << 3) * stride2 + (x_pos << 3);
291 : edgomez 965
292 :     /* Block size */
293 :     cst = 8;
294 :    
295 :     /* Operation function */
296 :     transfer_op = (transfer_operation_8to16_t*)transfer_8to16copy;
297 : edgomez 851 }
298 : Isibaar 3
299 : edgomez 965 /* Do the transfer */
300 : h 69 start_timer();
301 : edgomez 965 transfer_op(&data[0 * 64], pY_Cur, stride);
302 :     transfer_op(&data[1 * 64], pY_Cur + cst, stride);
303 :     transfer_op(&data[2 * 64], pY_Cur + next_block, stride);
304 :     transfer_op(&data[3 * 64], pY_Cur + next_block + cst, stride);
305 :     transfer_op(&data[4 * 64], pU_Cur, stride2);
306 :     transfer_op(&data[5 * 64], pV_Cur, stride2);
307 :     stop_transfer_timer();
308 : syskin 984 }
309 : edgomez 965
310 :     static __inline void
311 : syskin 984 MBTrans16to8(const MBParam * const pParam,
312 :     const FRAMEINFO * const frame,
313 :     const MACROBLOCK * const pMB,
314 : edgomez 965 const uint32_t x_pos,
315 :     const uint32_t y_pos,
316 :     int16_t data[6 * 64],
317 :     const uint32_t add,
318 :     const uint8_t cbp)
319 :     {
320 :     uint8_t *pY_Cur, *pU_Cur, *pV_Cur;
321 :     uint32_t stride = pParam->edged_width;
322 :     uint32_t stride2 = stride / 2;
323 :     uint32_t next_block = stride * 8;
324 : syskin 984 uint32_t cst;
325 :     const IMAGE * const pCurrent = &frame->image;
326 : edgomez 965 transfer_operation_16to8_t *transfer_op = NULL;
327 :    
328 :     if (pMB->field_dct) {
329 :     next_block = stride;
330 :     stride *= 2;
331 : h 69 }
332 :    
333 : edgomez 965 if ((frame->vop_flags & XVID_VOP_REDUCED)) {
334 : edgomez 851
335 : edgomez 965 /* Image pointers */
336 :     pY_Cur = pCurrent->y + (y_pos << 5) * stride + (x_pos << 5);
337 :     pU_Cur = pCurrent->u + (y_pos << 4) * stride2 + (x_pos << 4);
338 :     pV_Cur = pCurrent->v + (y_pos << 4) * stride2 + (x_pos << 4);
339 : Isibaar 3
340 : edgomez 965 /* Block size */
341 :     cst = 16;
342 : Isibaar 3
343 : edgomez 965 /* Operation function */
344 :     if(add)
345 :     transfer_op = (transfer_operation_16to8_t*)add_upsampled_8x8_16to8;
346 :     else
347 :     transfer_op = (transfer_operation_16to8_t*)copy_upsampled_8x8_16to8;
348 :     } else {
349 : Isibaar 3
350 : edgomez 965 /* Image pointers */
351 :     pY_Cur = pCurrent->y + (y_pos << 4) * stride + (x_pos << 4);
352 :     pU_Cur = pCurrent->u + (y_pos << 3) * stride2 + (x_pos << 3);
353 :     pV_Cur = pCurrent->v + (y_pos << 3) * stride2 + (x_pos << 3);
354 : Isibaar 3
355 : edgomez 965 /* Block size */
356 :     cst = 8;
357 : Isibaar 3
358 : edgomez 965 /* Operation function */
359 :     if(add)
360 :     transfer_op = (transfer_operation_16to8_t*)transfer_16to8add;
361 :     else
362 :     transfer_op = (transfer_operation_16to8_t*)transfer_16to8copy;
363 : Isibaar 3 }
364 : h 69
365 : edgomez 965 /* Do the operation */
366 : h 69 start_timer();
367 : edgomez 965 if (cbp&32) transfer_op(pY_Cur, &data[0 * 64], stride);
368 :     if (cbp&16) transfer_op(pY_Cur + cst, &data[1 * 64], stride);
369 :     if (cbp& 8) transfer_op(pY_Cur + next_block, &data[2 * 64], stride);
370 :     if (cbp& 4) transfer_op(pY_Cur + next_block + cst, &data[3 * 64], stride);
371 :     if (cbp& 2) transfer_op(pU_Cur, &data[4 * 64], stride2);
372 :     if (cbp& 1) transfer_op(pV_Cur, &data[5 * 64], stride2);
373 : h 69 stop_transfer_timer();
374 : Isibaar 3 }
375 : h 69
376 : edgomez 965 /*****************************************************************************
377 :     * Module functions
378 :     ****************************************************************************/
379 :    
380 : syskin 984 void
381 :     MBTransQuantIntra(const MBParam * const pParam,
382 :     const FRAMEINFO * const frame,
383 :     MACROBLOCK * const pMB,
384 : chl 368 const uint32_t x_pos,
385 :     const uint32_t y_pos,
386 :     int16_t data[6 * 64],
387 :     int16_t qcoeff[6 * 64])
388 :     {
389 : h 69
390 : edgomez 965 /* Transfer data */
391 :     MBTrans8to16(pParam, frame, pMB, x_pos, y_pos, data);
392 : chl 368
393 : edgomez 965 /* Perform DCT (and field decision) */
394 :     MBfDCT(pParam, frame, pMB, x_pos, y_pos, data);
395 : chl 368
396 : edgomez 965 /* Quantize the block */
397 : chl 995 MBQuantIntra(pParam, frame, pMB, data, qcoeff);
398 : edgomez 965
399 :     /* DeQuantize the block */
400 :     MBDeQuantIntra(pParam, pMB->quant, data, qcoeff);
401 :    
402 :     /* Perform inverse DCT*/
403 :     MBiDCT(data, 0x3F);
404 :    
405 :     /* Transfer back the data -- Don't add data */
406 :     MBTrans16to8(pParam, frame, pMB, x_pos, y_pos, data, 0, 0x3F);
407 : chl 368 }
408 :    
409 : edgomez 965
410 : chl 368 uint8_t
411 : syskin 984 MBTransQuantInter(const MBParam * const pParam,
412 :     const FRAMEINFO * const frame,
413 :     MACROBLOCK * const pMB,
414 : edgomez 914 const uint32_t x_pos,
415 :     const uint32_t y_pos,
416 : chl 368 int16_t data[6 * 64],
417 :     int16_t qcoeff[6 * 64])
418 :     {
419 :     uint8_t cbp;
420 : edgomez 965 uint32_t limit;
421 : chl 368
422 : edgomez 914 /*
423 : edgomez 965 * There is no MBTrans8to16 for Inter block, that's done in motion compensation
424 :     * already
425 : edgomez 914 */
426 : chl 368
427 : edgomez 965 /* Perform DCT (and field decision) */
428 :     MBfDCT(pParam, frame, pMB, x_pos, y_pos, data);
429 : edgomez 914
430 : edgomez 965 /* Set the limit threshold */
431 :     limit = PVOP_TOOSMALL_LIMIT + ((pMB->quant == 1)? 1 : 0);
432 : chl 368
433 : edgomez 965 /* Quantize the block */
434 : chl 995 cbp = MBQuantInter(pParam, frame, pMB, data, qcoeff, 0, limit);
435 : chl 368
436 : edgomez 965 /* DeQuantize the block */
437 :     MBDeQuantInter(pParam, pMB->quant, data, qcoeff, cbp);
438 : chl 368
439 : edgomez 965 /* Perform inverse DCT*/
440 :     MBiDCT(data, cbp);
441 : chl 368
442 : edgomez 965 /* Transfer back the data -- Add the data */
443 :     MBTrans16to8(pParam, frame, pMB, x_pos, y_pos, data, 1, cbp);
444 : syskin 984
445 : edgomez 965 return(cbp);
446 : chl 368 }
447 :    
448 : edgomez 965 uint8_t
449 :     MBTransQuantInterBVOP(const MBParam * pParam,
450 : chl 368 FRAMEINFO * frame,
451 :     MACROBLOCK * pMB,
452 :     const uint32_t x_pos,
453 :     const uint32_t y_pos,
454 :     int16_t data[6 * 64],
455 : edgomez 965 int16_t qcoeff[6 * 64])
456 : chl 368 {
457 : edgomez 965 uint8_t cbp;
458 :     uint32_t limit;
459 : syskin 984
460 : edgomez 965 /*
461 :     * There is no MBTrans8to16 for Inter block, that's done in motion compensation
462 :     * already
463 :     */
464 : chl 368
465 : edgomez 965 /* Perform DCT (and field decision) */
466 :     MBfDCT(pParam, frame, pMB, x_pos, y_pos, data);
467 : chl 368
468 : edgomez 965 /* Set the limit threshold */
469 :     limit = BVOP_TOOSMALL_LIMIT;
470 : chl 368
471 : edgomez 965 /* Quantize the block */
472 : chl 995 cbp = MBQuantInter(pParam, frame, pMB, data, qcoeff, 1, limit);
473 : chl 368
474 : edgomez 965 /*
475 :     * History comment:
476 :     * We don't have to DeQuant, iDCT and Transfer back data for B-frames.
477 :     *
478 :     * BUT some plugins require the original frame to be passed so we have
479 :     * to take care of that here
480 :     */
481 :     if((pParam->plugin_flags & XVID_REQORIGINAL)) {
482 : chl 368
483 : edgomez 965 /* DeQuantize the block */
484 :     MBDeQuantInter(pParam, pMB->quant, data, qcoeff, cbp);
485 : chl 368
486 : edgomez 965 /* Perform inverse DCT*/
487 :     MBiDCT(data, cbp);
488 : h 69
489 : edgomez 965 /* Transfer back the data -- Add the data */
490 :     MBTrans16to8(pParam, frame, pMB, x_pos, y_pos, data, 1, cbp);
491 : edgomez 851 }
492 :    
493 : edgomez 965 return(cbp);
494 : edgomez 851 }
495 :    
496 :     /* if sum(diff between field lines) < sum(diff between frame lines), use field dct */
497 :     uint32_t
498 :     MBFieldTest_c(int16_t data[6 * 64])
499 :     {
500 : edgomez 195 const uint8_t blocks[] =
501 :     { 0 * 64, 0 * 64, 0 * 64, 0 * 64, 2 * 64, 2 * 64, 2 * 64, 2 * 64 };
502 :     const uint8_t lines[] = { 0, 16, 32, 48, 0, 16, 32, 48 };
503 : edgomez 78
504 : h 69 int frame = 0, field = 0;
505 :     int i, j;
506 :    
507 : edgomez 195 for (i = 0; i < 7; ++i) {
508 :     for (j = 0; j < 8; ++j) {
509 :     frame +=
510 : edgomez 982 abs(data[0 * 64 + (i + 1) * 8 + j] - data[0 * 64 + i * 8 + j]);
511 : edgomez 195 frame +=
512 : edgomez 982 abs(data[1 * 64 + (i + 1) * 8 + j] - data[1 * 64 + i * 8 + j]);
513 : edgomez 195 frame +=
514 : edgomez 982 abs(data[2 * 64 + (i + 1) * 8 + j] - data[2 * 64 + i * 8 + j]);
515 : edgomez 195 frame +=
516 : edgomez 982 abs(data[3 * 64 + (i + 1) * 8 + j] - data[3 * 64 + i * 8 + j]);
517 : h 69
518 : edgomez 195 field +=
519 : edgomez 982 abs(data[blocks[i + 1] + lines[i + 1] + j] -
520 : edgomez 195 data[blocks[i] + lines[i] + j]);
521 :     field +=
522 : edgomez 982 abs(data[blocks[i + 1] + lines[i + 1] + 8 + j] -
523 : edgomez 195 data[blocks[i] + lines[i] + 8 + j]);
524 :     field +=
525 : edgomez 982 abs(data[blocks[i + 1] + 64 + lines[i + 1] + j] -
526 : edgomez 195 data[blocks[i] + 64 + lines[i] + j]);
527 :     field +=
528 : edgomez 982 abs(data[blocks[i + 1] + 64 + lines[i + 1] + 8 + j] -
529 : edgomez 195 data[blocks[i] + 64 + lines[i] + 8 + j]);
530 : h 69 }
531 :     }
532 :    
533 : edgomez 851 return (frame >= (field + 350));
534 : h 69 }
535 :    
536 :    
537 :     /* deinterlace Y blocks vertically */
538 :    
539 :     #define MOVLINE(X,Y) memcpy(X, Y, sizeof(tmp))
540 : syskin 984 #define LINE(X,Y) &data[X*64 + Y*8]
541 : h 69
542 : edgomez 195 void
543 :     MBFrameToField(int16_t data[6 * 64])
544 : h 69 {
545 :     int16_t tmp[8];
546 :    
547 :     /* left blocks */
548 :    
549 : edgomez 851 // 1=2, 2=4, 4=8, 8=1
550 : edgomez 195 MOVLINE(tmp, LINE(0, 1));
551 :     MOVLINE(LINE(0, 1), LINE(0, 2));
552 :     MOVLINE(LINE(0, 2), LINE(0, 4));
553 :     MOVLINE(LINE(0, 4), LINE(2, 0));
554 :     MOVLINE(LINE(2, 0), tmp);
555 : h 69
556 : edgomez 851 // 3=6, 6=12, 12=9, 9=3
557 : edgomez 195 MOVLINE(tmp, LINE(0, 3));
558 :     MOVLINE(LINE(0, 3), LINE(0, 6));
559 :     MOVLINE(LINE(0, 6), LINE(2, 4));
560 :     MOVLINE(LINE(2, 4), LINE(2, 1));
561 :     MOVLINE(LINE(2, 1), tmp);
562 : h 69
563 : edgomez 851 // 5=10, 10=5
564 : edgomez 195 MOVLINE(tmp, LINE(0, 5));
565 :     MOVLINE(LINE(0, 5), LINE(2, 2));
566 :     MOVLINE(LINE(2, 2), tmp);
567 : h 69
568 : edgomez 851 // 7=14, 14=13, 13=11, 11=7
569 : edgomez 195 MOVLINE(tmp, LINE(0, 7));
570 :     MOVLINE(LINE(0, 7), LINE(2, 6));
571 :     MOVLINE(LINE(2, 6), LINE(2, 5));
572 :     MOVLINE(LINE(2, 5), LINE(2, 3));
573 :     MOVLINE(LINE(2, 3), tmp);
574 : h 69
575 :     /* right blocks */
576 :    
577 : edgomez 851 // 1=2, 2=4, 4=8, 8=1
578 : edgomez 195 MOVLINE(tmp, LINE(1, 1));
579 :     MOVLINE(LINE(1, 1), LINE(1, 2));
580 :     MOVLINE(LINE(1, 2), LINE(1, 4));
581 :     MOVLINE(LINE(1, 4), LINE(3, 0));
582 :     MOVLINE(LINE(3, 0), tmp);
583 : h 69
584 : edgomez 851 // 3=6, 6=12, 12=9, 9=3
585 : edgomez 195 MOVLINE(tmp, LINE(1, 3));
586 :     MOVLINE(LINE(1, 3), LINE(1, 6));
587 :     MOVLINE(LINE(1, 6), LINE(3, 4));
588 :     MOVLINE(LINE(3, 4), LINE(3, 1));
589 :     MOVLINE(LINE(3, 1), tmp);
590 : h 69
591 : edgomez 851 // 5=10, 10=5
592 : edgomez 195 MOVLINE(tmp, LINE(1, 5));
593 :     MOVLINE(LINE(1, 5), LINE(3, 2));
594 :     MOVLINE(LINE(3, 2), tmp);
595 : h 69
596 : edgomez 851 // 7=14, 14=13, 13=11, 11=7
597 : edgomez 195 MOVLINE(tmp, LINE(1, 7));
598 :     MOVLINE(LINE(1, 7), LINE(3, 6));
599 :     MOVLINE(LINE(3, 6), LINE(3, 5));
600 :     MOVLINE(LINE(3, 5), LINE(3, 3));
601 :     MOVLINE(LINE(3, 3), tmp);
602 : h 69 }
603 : chl 1011
604 :    
605 :    
606 :    
607 :    
608 :     /************************************************************************
609 :     * Trellis based R-D optimal quantization *
610 :     * *
611 :     * Trellis Quant code (C) 2003 Pascal Massimino skal(at)planet-d.net *
612 :     * *
613 :     ************************************************************************/
614 :    
615 :    
616 : chl 1012 static int
617 :     dct_quantize_trellis_mpeg_c(int16_t *const Out, const int16_t *const In, int Q,
618 :     const uint16_t * const Zigzag, int Non_Zero)
619 : chl 1011 { return 63; }
620 :    
621 :    
622 :     //////////////////////////////////////////////////////////
623 :     //
624 :     // Trellis-Based quantization
625 :     //
626 :     // So far I understand this paper:
627 :     //
628 :     // "Trellis-Based R-D Optimal Quantization in H.263+"
629 :     // J.Wen, M.Luttrell, J.Villasenor
630 :     // IEEE Transactions on Image Processing, Vol.9, No.8, Aug. 2000.
631 :     //
632 :     // we are at stake with a simplified Bellmand-Ford / Dijkstra Single
633 :     // Source Shorted Path algo. But due to the underlying graph structure
634 :     // ("Trellis"), it can be turned into a dynamic programming algo,
635 :     // partially saving the explicit graph's nodes representation. And
636 :     // without using a heap, since the open frontier of the DAG is always
637 :     // known, and of fixed sized.
638 :     //
639 :     //////////////////////////////////////////////////////////
640 :    
641 :    
642 :     //////////////////////////////////////////////////////////
643 :     // Codes lengths for relevant levels.
644 :    
645 :     // let's factorize:
646 :     static const uint8_t Code_Len0[64] = {
647 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
648 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
649 :     static const uint8_t Code_Len1[64] = {
650 :     20,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
651 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
652 :     static const uint8_t Code_Len2[64] = {
653 :     19,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
654 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
655 :     static const uint8_t Code_Len3[64] = {
656 :     18,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
657 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
658 :     static const uint8_t Code_Len4[64] = {
659 :     17,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
660 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
661 :     static const uint8_t Code_Len5[64] = {
662 :     16,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
663 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
664 :     static const uint8_t Code_Len6[64] = {
665 :     15,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
666 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
667 :     static const uint8_t Code_Len7[64] = {
668 :     13,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
669 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
670 :     static const uint8_t Code_Len8[64] = {
671 :     11,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
672 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
673 :     static const uint8_t Code_Len9[64] = {
674 :     12,21,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
675 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
676 :     static const uint8_t Code_Len10[64] = {
677 :     12,20,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
678 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
679 :     static const uint8_t Code_Len11[64] = {
680 :     12,19,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
681 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
682 :     static const uint8_t Code_Len12[64] = {
683 :     11,17,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
684 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
685 :     static const uint8_t Code_Len13[64] = {
686 :     11,15,21,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
687 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
688 :     static const uint8_t Code_Len14[64] = {
689 :     10,12,19,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
690 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
691 :     static const uint8_t Code_Len15[64] = {
692 :     10,13,17,19,21,21,21,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
693 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
694 :     static const uint8_t Code_Len16[64] = {
695 :     9,12,13,18,18,19,19,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
696 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30};
697 :     static const uint8_t Code_Len17[64] = {
698 :     8,11,13,14,14,14,15,19,19,19,21,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
699 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
700 :     static const uint8_t Code_Len18[64] = {
701 :     7, 9,11,11,13,13,13,15,15,15,16,22,22,22,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
702 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
703 :     static const uint8_t Code_Len19[64] = {
704 :     5, 7, 9,10,10,11,11,11,11,11,13,14,16,17,17,18,18,18,18,18,18,18,18,20,20,21,21,30,30,30,30,30,
705 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
706 :     static const uint8_t Code_Len20[64] = {
707 :     3, 4, 5, 6, 6, 6, 7, 7, 7, 7, 8, 8, 8, 9, 9,10,10,10,10,10,10,10,10,12,12,13,13,12,13,14,15,15,
708 :     15,16,16,16,16,17,17,17,18,18,19,19,19,19,19,19,19,19,21,21,22,22,30,30,30,30,30,30,30,30,30,30 };
709 :    
710 :     // a few more table for LAST table:
711 :     static const uint8_t Code_Len21[64] = {
712 :     13,20,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
713 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30};
714 :     static const uint8_t Code_Len22[64] = {
715 :     12,15,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
716 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30};
717 :     static const uint8_t Code_Len23[64] = {
718 :     10,12,15,15,15,16,16,16,16,17,17,17,17,17,17,17,17,18,18,18,18,18,18,18,18,19,19,19,19,20,20,20,
719 :     20,21,21,21,21,21,21,21,21,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30};
720 :     static const uint8_t Code_Len24[64] = {
721 :     5, 7, 7, 7, 7, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9,10,10,10,10,10,10,10,10,11,11,11,11,12,12,12,
722 :     12,13,13,13,13,13,13,13,13,14,16,16,16,16,17,17,17,17,18,18,18,18,18,18,18,18,19,19,19,19,19,19};
723 :    
724 :    
725 :     static const uint8_t * const B16_17_Code_Len[24] = { // levels [1..24]
726 :     Code_Len20,Code_Len19,Code_Len18,Code_Len17,
727 :     Code_Len16,Code_Len15,Code_Len14,Code_Len13,
728 :     Code_Len12,Code_Len11,Code_Len10,Code_Len9,
729 :     Code_Len8, Code_Len7 ,Code_Len6 ,Code_Len5,
730 :     Code_Len4, Code_Len3, Code_Len3 ,Code_Len2,
731 :     Code_Len2, Code_Len1, Code_Len1, Code_Len1,
732 :     };
733 :    
734 :     static const uint8_t * const B16_17_Code_Len_Last[6] = { // levels [1..6]
735 :     Code_Len24,Code_Len23,Code_Len22,Code_Len21, Code_Len3, Code_Len1,
736 :     };
737 :    
738 :     #define TL(q) 0xfe00/(q*q)
739 :    
740 :     static const int Trellis_Lambda_Tabs[31] = {
741 :     TL( 1),TL( 2),TL( 3),TL( 4),TL( 5),TL( 6), TL( 7),
742 :     TL( 8),TL( 9),TL(10),TL(11),TL(12),TL(13),TL(14), TL(15),
743 :     TL(16),TL(17),TL(18),TL(19),TL(20),TL(21),TL(22), TL(23),
744 :     TL(24),TL(25),TL(26),TL(27),TL(28),TL(29),TL(30), TL(31)
745 :     };
746 :     #undef TL
747 :    
748 : suxen_drol 1014 static __inline int Find_Last(const int16_t *C, const uint16_t *Zigzag, int i)
749 : chl 1011 {
750 :     while(i>=0)
751 :     if (C[Zigzag[i]])
752 :     return i;
753 :     else i--;
754 :     return -1;
755 :     }
756 :    
757 :     //////////////////////////////////////////////////////////
758 : chl 1012 // this routine has been strippen of all debug code
759 :     //////////////////////////////////////////////////////////
760 : chl 1011
761 : chl 1012 static int
762 :     dct_quantize_trellis_h263_c(int16_t *const Out, const int16_t *const In, int Q, const uint16_t * const Zigzag, int Non_Zero)
763 :     {
764 :    
765 :     // Note: We should search last non-zero coeffs on *real* DCT input coeffs (In[]),
766 :     // not quantized one (Out[]). However, it only improves the result *very*
767 :     // slightly (~0.01dB), whereas speed drops to crawling level :)
768 :     // Well, actually, taking 1 more coeff past Non_Zero into account sometimes helps,
769 :    
770 :     typedef struct { int16_t Run, Level; } NODE;
771 :    
772 :     NODE Nodes[65], Last;
773 :     uint32_t Run_Costs0[64+1];
774 :     uint32_t * const Run_Costs = Run_Costs0 + 1;
775 :     const int Mult = 2*Q;
776 :     const int Bias = (Q-1) | 1;
777 :     const int Lev0 = Mult + Bias;
778 :     const int Lambda = Trellis_Lambda_Tabs[Q-1]; // it's 1/lambda, actually
779 :    
780 :     int Run_Start = -1;
781 :     uint32_t Min_Cost = 2<<16;
782 :    
783 :     int Last_Node = -1;
784 :     uint32_t Last_Cost = 0;
785 :    
786 :     int i, j;
787 : suxen_drol 1014 Run_Costs[-1] = 2<<16; // source (w/ CBP penalty)
788 : chl 1012
789 :     Non_Zero = Find_Last(Out, Zigzag, Non_Zero);
790 :     if (Non_Zero<0)
791 :     return -1;
792 :    
793 :     for(i=0; i<=Non_Zero; i++)
794 :     {
795 :     const int AC = In[Zigzag[i]];
796 :     const int Level1 = Out[Zigzag[i]];
797 :     const int Dist0 = Lambda* AC*AC;
798 :     uint32_t Best_Cost = 0xf0000000;
799 :     Last_Cost += Dist0;
800 :    
801 :     if ((uint32_t)(Level1+1)<3) // very specialized loop for -1,0,+1
802 :     {
803 :     int dQ;
804 :     int Run;
805 :     uint32_t Cost0;
806 :    
807 :     if (AC<0) {
808 :     Nodes[i].Level = -1;
809 :     dQ = Lev0 + AC;
810 :     } else {
811 :     Nodes[i].Level = 1;
812 :     dQ = Lev0 - AC;
813 :     }
814 :     Cost0 = Lambda*dQ*dQ;
815 :    
816 :     Nodes[i].Run = 1;
817 :     Best_Cost = (Code_Len20[0]<<16) + Run_Costs[i-1]+Cost0;
818 :     for(Run=i-Run_Start; Run>0; --Run)
819 :     {
820 :     const uint32_t Cost_Base = Cost0 + Run_Costs[i-Run];
821 :     const uint32_t Cost = Cost_Base + (Code_Len20[Run-1]<<16);
822 :     const uint32_t lCost = Cost_Base + (Code_Len24[Run-1]<<16);
823 :    
824 :     // TODO: what about tie-breaks? Should we favor short runs or
825 :     // long runs? Although the error is the same, it would not be
826 :     // spread the same way along high and low frequencies...
827 :    
828 :     // (I'd say: favour short runs => hifreq errors (HVS) -- gruel )
829 :    
830 :     if (Cost<Best_Cost) {
831 :     Best_Cost = Cost;
832 :     Nodes[i].Run = Run;
833 :     }
834 :    
835 :     if (lCost<Last_Cost) {
836 :     Last_Cost = lCost;
837 :     Last.Run = Run;
838 :     Last_Node = i;
839 :     }
840 :     }
841 :     if (Last_Node==i)
842 :     Last.Level = Nodes[i].Level;
843 :     }
844 :     else // "big" levels
845 :     {
846 :     const uint8_t *Tbl_L1, *Tbl_L2, *Tbl_L1_Last, *Tbl_L2_Last;
847 :     int Level2;
848 :     int dQ1, dQ2;
849 :     int Run;
850 :     uint32_t Dist1,Dist2;
851 :     int dDist21;
852 :    
853 :     if (Level1>1) {
854 :     dQ1 = Level1*Mult-AC + Bias;
855 :     dQ2 = dQ1 - Mult;
856 :     Level2 = Level1-1;
857 :     Tbl_L1 = (Level1<=24) ? B16_17_Code_Len[Level1-1] : Code_Len0;
858 :     Tbl_L2 = (Level2<=24) ? B16_17_Code_Len[Level2-1] : Code_Len0;
859 :     Tbl_L1_Last = (Level1<=6) ? B16_17_Code_Len_Last[Level1-1] : Code_Len0;
860 :     Tbl_L2_Last = (Level2<=6) ? B16_17_Code_Len_Last[Level2-1] : Code_Len0;
861 :     } else { // Level1<-1
862 :     dQ1 = Level1*Mult-AC - Bias;
863 :     dQ2 = dQ1 + Mult;
864 :     Level2 = Level1 + 1;
865 :     Tbl_L1 = (Level1>=-24) ? B16_17_Code_Len[Level1^-1] : Code_Len0;
866 :     Tbl_L2 = (Level2>=-24) ? B16_17_Code_Len[Level2^-1] : Code_Len0;
867 :     Tbl_L1_Last = (Level1>=- 6) ? B16_17_Code_Len_Last[Level1^-1] : Code_Len0;
868 :     Tbl_L2_Last = (Level2>=- 6) ? B16_17_Code_Len_Last[Level2^-1] : Code_Len0;
869 :     }
870 :     Dist1 = Lambda*dQ1*dQ1;
871 :     Dist2 = Lambda*dQ2*dQ2;
872 :     dDist21 = Dist2-Dist1;
873 :    
874 :     for(Run=i-Run_Start; Run>0; --Run)
875 :     {
876 :     const uint32_t Cost_Base = Dist1 + Run_Costs[i-Run];
877 :     uint32_t Cost1, Cost2;
878 :     int bLevel;
879 :    
880 :     // for sub-optimal (but slightly worth it, speed-wise) search, uncomment the following:
881 :     // if (Cost_Base>=Best_Cost) continue;
882 :     // (? doesn't seem to have any effect -- gruel )
883 :    
884 :     Cost1 = Cost_Base + (Tbl_L1[Run-1]<<16);
885 :     Cost2 = Cost_Base + (Tbl_L2[Run-1]<<16) + dDist21;
886 :    
887 :     if (Cost2<Cost1) {
888 :     Cost1 = Cost2;
889 :     bLevel = Level2;
890 :     } else
891 :     bLevel = Level1;
892 :    
893 :     if (Cost1<Best_Cost) {
894 :     Best_Cost = Cost1;
895 :     Nodes[i].Run = Run;
896 :     Nodes[i].Level = bLevel;
897 :     }
898 :    
899 :     Cost1 = Cost_Base + (Tbl_L1_Last[Run-1]<<16);
900 :     Cost2 = Cost_Base + (Tbl_L2_Last[Run-1]<<16) + dDist21;
901 :    
902 :     if (Cost2<Cost1) {
903 :     Cost1 = Cost2;
904 :     bLevel = Level2;
905 :     } else
906 :     bLevel = Level1;
907 :    
908 :     if (Cost1<Last_Cost) {
909 :     Last_Cost = Cost1;
910 :     Last.Run = Run;
911 :     Last.Level = bLevel;
912 :     Last_Node = i;
913 :     }
914 :     } //end of "for Run"
915 :    
916 :     }
917 :    
918 :     Run_Costs[i] = Best_Cost;
919 :    
920 :     if (Best_Cost < Min_Cost + Dist0) {
921 :     Min_Cost = Best_Cost;
922 :     Run_Start = i;
923 :     }
924 :     else
925 :     {
926 :     // as noticed by Michael Niedermayer (michaelni at gmx.at), there's
927 :     // a code shorter by 1 bit for a larger run (!), same level. We give
928 :     // it a chance by not moving the left barrier too much.
929 :    
930 :     while( Run_Costs[Run_Start]>Min_Cost+(1<<16) )
931 :     Run_Start++;
932 :    
933 :     // spread on preceding coeffs the cost incurred by skipping this one
934 :     for(j=Run_Start; j<i; ++j) Run_Costs[j] += Dist0;
935 :     Min_Cost += Dist0;
936 :     }
937 :     }
938 :    
939 :     if (Last_Node<0)
940 :     return -1;
941 :    
942 :     // reconstruct optimal sequence backward with surviving paths
943 :     memset(Out, 0x00, 64*sizeof(*Out));
944 :     Out[Zigzag[Last_Node]] = Last.Level;
945 :     i = Last_Node - Last.Run;
946 :     while(i>=0) {
947 :     Out[Zigzag[i]] = Nodes[i].Level;
948 :     i -= Nodes[i].Run;
949 :     }
950 :     return Last_Node;
951 :     }
952 :    
953 :    
954 :    
955 :    
956 :    
957 :    
958 :    
959 :    
960 :    
961 :    
962 :    
963 :     //////////////////////////////////////////////////////////
964 :     // original version including heavy debugging info
965 :     //////////////////////////////////////////////////////////
966 :    
967 :    
968 :     #ifdef DBGTRELL
969 :    
970 : chl 1011 #define DBG 0
971 :    
972 : suxen_drol 1014 static __inline uint32_t Evaluate_Cost(const int16_t *C, int Mult, int Bias,
973 : chl 1011 const uint16_t * Zigzag, int Max, int Lambda)
974 :     {
975 :     #if (DBG>0)
976 :     const int16_t * const Ref = C + 6*64;
977 :     int Last = Max;
978 :     int Bits = 0;
979 : chl 1012 int Dist = 0;
980 :     int i;
981 :     uint32_t Cost;
982 :    
983 :     while(Last>=0 && C[Zigzag[Last]]==0)
984 :     Last--;
985 :    
986 : chl 1011 if (Last>=0) {
987 :     int j=0, j0=0;
988 :     int Run, Level;
989 : chl 1012
990 :     Bits = 2; // CBP
991 : chl 1011 while(j<Last) {
992 : chl 1012 while(!C[Zigzag[j]])
993 :     j++;
994 :     if (j==Last)
995 :     break;
996 : chl 1011 Level=C[Zigzag[j]];
997 :     Run = j - j0;
998 :     j0 = ++j;
999 : chl 1012 if (Level>=-24 && Level<=24)
1000 :     Bits += B16_17_Code_Len[(Level<0) ? -Level-1 : Level-1][Run];
1001 :     else
1002 :     Bits += 30;
1003 : chl 1011 }
1004 :     Level = C[Zigzag[Last]];
1005 :     Run = j - j0;
1006 : chl 1012 if (Level>=-6 && Level<=6)
1007 :     Bits += B16_17_Code_Len_Last[(Level<0) ? -Level-1 : Level-1][Run];
1008 :     else
1009 :     Bits += 30;
1010 : chl 1011 }
1011 :    
1012 :     for(i=0; i<=Last; ++i) {
1013 :     int V = C[Zigzag[i]]*Mult;
1014 : chl 1012 if (V>0)
1015 :     V += Bias;
1016 :     else
1017 :     if (V<0)
1018 :     V -= Bias;
1019 : chl 1011 V -= Ref[Zigzag[i]];
1020 :     Dist += V*V;
1021 :     }
1022 : chl 1012 Cost = Lambda*Dist + (Bits<<16);
1023 : chl 1011 if (DBG==1)
1024 :     printf( " Last:%2d/%2d Cost = [(Bits=%5.0d) + Lambda*(Dist=%6.0d) = %d ] >>12= %d ", Last,Max, Bits, Dist, Cost, Cost>>12 );
1025 :     return Cost;
1026 :    
1027 :     #else
1028 :     return 0;
1029 :     #endif
1030 :     }
1031 :    
1032 :    
1033 :     static int
1034 :     dct_quantize_trellis_h263_c(int16_t *const Out, const int16_t *const In, int Q, const uint16_t * const Zigzag, int Non_Zero)
1035 :     {
1036 :    
1037 :     // Note: We should search last non-zero coeffs on *real* DCT input coeffs (In[]),
1038 :     // not quantized one (Out[]). However, it only improves the result *very*
1039 :     // slightly (~0.01dB), whereas speed drops to crawling level :)
1040 :     // Well, actually, taking 1 more coeff past Non_Zero into account sometimes helps,
1041 :    
1042 :     typedef struct { int16_t Run, Level; } NODE;
1043 :    
1044 :     NODE Nodes[65], Last;
1045 : chl 1012 uint32_t Run_Costs0[64+1];
1046 :     uint32_t * const Run_Costs = Run_Costs0 + 1;
1047 : chl 1011 const int Mult = 2*Q;
1048 :     const int Bias = (Q-1) | 1;
1049 :     const int Lev0 = Mult + Bias;
1050 :     const int Lambda = Trellis_Lambda_Tabs[Q-1]; // it's 1/lambda, actually
1051 :    
1052 :     int Run_Start = -1;
1053 :     Run_Costs[-1] = 2<<16; // source (w/ CBP penalty)
1054 :     uint32_t Min_Cost = 2<<16;
1055 :    
1056 :     int Last_Node = -1;
1057 :     uint32_t Last_Cost = 0;
1058 :    
1059 : chl 1012 int i, j;
1060 :    
1061 : chl 1011 #if (DBG>0)
1062 :     Last.Level = 0; Last.Run = -1; // just initialize to smthg
1063 :     #endif
1064 :    
1065 :     Non_Zero = Find_Last(Out, Zigzag, Non_Zero);
1066 :     if (Non_Zero<0)
1067 :     return -1;
1068 :    
1069 :     for(i=0; i<=Non_Zero; i++)
1070 :     {
1071 :     const int AC = In[Zigzag[i]];
1072 :     const int Level1 = Out[Zigzag[i]];
1073 :     const int Dist0 = Lambda* AC*AC;
1074 :     uint32_t Best_Cost = 0xf0000000;
1075 :     Last_Cost += Dist0;
1076 :    
1077 :     if ((uint32_t)(Level1+1)<3) // very specialized loop for -1,0,+1
1078 :     {
1079 : chl 1012 int dQ;
1080 :     int Run;
1081 :     uint32_t Cost0;
1082 : chl 1011
1083 :     if (AC<0) {
1084 :     Nodes[i].Level = -1;
1085 :     dQ = Lev0 + AC;
1086 :     } else {
1087 :     Nodes[i].Level = 1;
1088 :     dQ = Lev0 - AC;
1089 :     }
1090 : chl 1012 Cost0 = Lambda*dQ*dQ;
1091 :    
1092 : chl 1011 Nodes[i].Run = 1;
1093 :     Best_Cost = (Code_Len20[0]<<16) + Run_Costs[i-1]+Cost0;
1094 :     for(Run=i-Run_Start; Run>0; --Run)
1095 :     {
1096 :     const uint32_t Cost_Base = Cost0 + Run_Costs[i-Run];
1097 :     const uint32_t Cost = Cost_Base + (Code_Len20[Run-1]<<16);
1098 : chl 1012 const uint32_t lCost = Cost_Base + (Code_Len24[Run-1]<<16);
1099 :    
1100 : chl 1011 // TODO: what about tie-breaks? Should we favor short runs or
1101 :     // long runs? Although the error is the same, it would not be
1102 :     // spread the same way along high and low frequencies...
1103 : chl 1012 if (Cost<Best_Cost) {
1104 : chl 1011 Best_Cost = Cost;
1105 :     Nodes[i].Run = Run;
1106 :     }
1107 : chl 1012
1108 :     if (lCost<Last_Cost) {
1109 : chl 1011 Last_Cost = lCost;
1110 :     Last.Run = Run;
1111 :     Last_Node = i;
1112 :     }
1113 :     }
1114 : chl 1012 if (Last_Node==i)
1115 :     Last.Level = Nodes[i].Level;
1116 : chl 1011
1117 :     if (DBG==1) {
1118 :     Run_Costs[i] = Best_Cost;
1119 :     printf( "Costs #%2d: ", i);
1120 :     for(j=-1;j<=Non_Zero;++j) {
1121 :     if (j==Run_Start) printf( " %3.0d|", Run_Costs[j]>>12 );
1122 :     else if (j>Run_Start && j<i) printf( " %3.0d|", Run_Costs[j]>>12 );
1123 :     else if (j==i) printf( "(%3.0d)", Run_Costs[j]>>12 );
1124 :     else printf( " - |" );
1125 :     }
1126 :     printf( "<%3.0d %2d %d>", Min_Cost>>12, Nodes[i].Level, Nodes[i].Run );
1127 :     printf( " Last:#%2d {%3.0d %2d %d}", Last_Node, Last_Cost>>12, Last.Level, Last.Run );
1128 :     printf( " AC:%3.0d Dist0:%3d Dist(%d)=%d", AC, Dist0>>12, Nodes[i].Level, Cost0>>12 );
1129 :     printf( "\n" );
1130 :     }
1131 :     }
1132 :     else // "big" levels
1133 :     {
1134 :     const uint8_t *Tbl_L1, *Tbl_L2, *Tbl_L1_Last, *Tbl_L2_Last;
1135 :     int Level2;
1136 :     int dQ1, dQ2;
1137 :     int Run;
1138 : chl 1012 uint32_t Dist1,Dist2;
1139 :     int dDist21;
1140 :    
1141 : chl 1011 if (Level1>1) {
1142 :     dQ1 = Level1*Mult-AC + Bias;
1143 :     dQ2 = dQ1 - Mult;
1144 :     Level2 = Level1-1;
1145 :     Tbl_L1 = (Level1<=24) ? B16_17_Code_Len[Level1-1] : Code_Len0;
1146 :     Tbl_L2 = (Level2<=24) ? B16_17_Code_Len[Level2-1] : Code_Len0;
1147 :     Tbl_L1_Last = (Level1<=6) ? B16_17_Code_Len_Last[Level1-1] : Code_Len0;
1148 :     Tbl_L2_Last = (Level2<=6) ? B16_17_Code_Len_Last[Level2-1] : Code_Len0;
1149 : chl 1012 } else { // Level1<-1
1150 : chl 1011 dQ1 = Level1*Mult-AC - Bias;
1151 :     dQ2 = dQ1 + Mult;
1152 :     Level2 = Level1 + 1;
1153 :     Tbl_L1 = (Level1>=-24) ? B16_17_Code_Len[Level1^-1] : Code_Len0;
1154 :     Tbl_L2 = (Level2>=-24) ? B16_17_Code_Len[Level2^-1] : Code_Len0;
1155 :     Tbl_L1_Last = (Level1>=- 6) ? B16_17_Code_Len_Last[Level1^-1] : Code_Len0;
1156 :     Tbl_L2_Last = (Level2>=- 6) ? B16_17_Code_Len_Last[Level2^-1] : Code_Len0;
1157 :     }
1158 : chl 1012 Dist1 = Lambda*dQ1*dQ1;
1159 :     Dist2 = Lambda*dQ2*dQ2;
1160 :     dDist21 = Dist2-Dist1;
1161 : chl 1011
1162 :     for(Run=i-Run_Start; Run>0; --Run)
1163 :     {
1164 :     const uint32_t Cost_Base = Dist1 + Run_Costs[i-Run];
1165 : chl 1012 uint32_t Cost1, Cost2;
1166 :     int bLevel;
1167 : chl 1011
1168 :     // for sub-optimal (but slightly worth it, speed-wise) search, uncomment the following:
1169 :     // if (Cost_Base>=Best_Cost) continue;
1170 :    
1171 :     Cost1 = Cost_Base + (Tbl_L1[Run-1]<<16);
1172 :     Cost2 = Cost_Base + (Tbl_L2[Run-1]<<16) + dDist21;
1173 :    
1174 : chl 1012 if (Cost2<Cost1) {
1175 :     Cost1 = Cost2;
1176 :     bLevel = Level2;
1177 :     } else
1178 :     bLevel = Level1;
1179 : chl 1011
1180 : chl 1012 if (Cost1<Best_Cost) {
1181 : chl 1011 Best_Cost = Cost1;
1182 :     Nodes[i].Run = Run;
1183 :     Nodes[i].Level = bLevel;
1184 :     }
1185 :    
1186 :     Cost1 = Cost_Base + (Tbl_L1_Last[Run-1]<<16);
1187 :     Cost2 = Cost_Base + (Tbl_L2_Last[Run-1]<<16) + dDist21;
1188 :    
1189 : chl 1012 if (Cost2<Cost1) {
1190 :     Cost1 = Cost2;
1191 :     bLevel = Level2;
1192 :     } else
1193 :     bLevel = Level1;
1194 :    
1195 :     if (Cost1<Last_Cost) {
1196 : chl 1011 Last_Cost = Cost1;
1197 :     Last.Run = Run;
1198 :     Last.Level = bLevel;
1199 :     Last_Node = i;
1200 :     }
1201 : chl 1012 } //end of "for Run"
1202 : chl 1011
1203 :     if (DBG==1) {
1204 :     Run_Costs[i] = Best_Cost;
1205 :     printf( "Costs #%2d: ", i);
1206 :     for(j=-1;j<=Non_Zero;++j) {
1207 :     if (j==Run_Start) printf( " %3.0d|", Run_Costs[j]>>12 );
1208 :     else if (j>Run_Start && j<i) printf( " %3.0d|", Run_Costs[j]>>12 );
1209 :     else if (j==i) printf( "(%3.0d)", Run_Costs[j]>>12 );
1210 :     else printf( " - |" );
1211 :     }
1212 :     printf( "<%3.0d %2d %d>", Min_Cost>>12, Nodes[i].Level, Nodes[i].Run );
1213 :     printf( " Last:#%2d {%3.0d %2d %d}", Last_Node, Last_Cost>>12, Last.Level, Last.Run );
1214 :     printf( " AC:%3.0d Dist0:%3d Dist(%2d):%3d Dist(%2d):%3d", AC, Dist0>>12, Level1, Dist1>>12, Level2, Dist2>>12 );
1215 :     printf( "\n" );
1216 :     }
1217 :     }
1218 :    
1219 :     Run_Costs[i] = Best_Cost;
1220 :    
1221 :     if (Best_Cost < Min_Cost + Dist0) {
1222 :     Min_Cost = Best_Cost;
1223 :     Run_Start = i;
1224 :     }
1225 :     else
1226 :     {
1227 :     // as noticed by Michael Niedermayer (michaelni at gmx.at), there's
1228 :     // a code shorter by 1 bit for a larger run (!), same level. We give
1229 :     // it a chance by not moving the left barrier too much.
1230 : chl 1012
1231 : chl 1011 while( Run_Costs[Run_Start]>Min_Cost+(1<<16) )
1232 :     Run_Start++;
1233 :    
1234 :     // spread on preceding coeffs the cost incurred by skipping this one
1235 :     for(j=Run_Start; j<i; ++j) Run_Costs[j] += Dist0;
1236 :     Min_Cost += Dist0;
1237 :     }
1238 :     }
1239 :    
1240 :     if (DBG) {
1241 :     Last_Cost = Evaluate_Cost(Out,Mult,Bias, Zigzag,Non_Zero, Lambda);
1242 :     if (DBG==1) {
1243 :     printf( "=> " );
1244 :     for(i=0; i<=Non_Zero; ++i) printf( "[%3.0d] ", Out[Zigzag[i]] );
1245 :     printf( "\n" );
1246 :     }
1247 :     }
1248 :    
1249 :     if (Last_Node<0)
1250 :     return -1;
1251 :    
1252 :     // reconstruct optimal sequence backward with surviving paths
1253 : chl 1012 memset(Out, 0x00, 64*sizeof(*Out));
1254 : chl 1011 Out[Zigzag[Last_Node]] = Last.Level;
1255 :     i = Last_Node - Last.Run;
1256 :     while(i>=0) {
1257 :     Out[Zigzag[i]] = Nodes[i].Level;
1258 :     i -= Nodes[i].Run;
1259 :     }
1260 :    
1261 :     if (DBG) {
1262 :     uint32_t Cost = Evaluate_Cost(Out,Mult,Bias, Zigzag,Non_Zero, Lambda);
1263 :     if (DBG==1) {
1264 :     printf( "<= " );
1265 :     for(i=0; i<=Last_Node; ++i) printf( "[%3.0d] ", Out[Zigzag[i]] );
1266 :     printf( "\n--------------------------------\n" );
1267 :     }
1268 :     if (Cost>Last_Cost) printf( "!!! %u > %u\n", Cost, Last_Cost );
1269 :     }
1270 :     return Last_Node;
1271 :     }
1272 :    
1273 :     #undef DBG
1274 : chl 1012
1275 :     #endif

No admin address has been configured
ViewVC Help
Powered by ViewVC 1.0.4