[cvs] / xvidcore / src / utils / mbtransquant.c Repository:
ViewVC logotype

Annotation of /xvidcore/src/utils/mbtransquant.c

Parent Directory Parent Directory | Revision Log Revision Log


Revision 1.21.2.13 - (view) (download)

1 : edgomez 1.21.2.6 /*****************************************************************************
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 : edgomez 1.21.2.13 * $Id: mbtransquant.c,v 1.21.2.12 2003/05/12 12:33:16 suxen_drol Exp $
25 : edgomez 1.21.2.6 *
26 :     ****************************************************************************/
27 : Isibaar 1.1
28 : chl 1.21.2.11 #include <stdio.h>
29 : edgomez 1.21.2.7 #include <stdlib.h>
30 : chl 1.21.2.11 #include <string.h>
31 : edgomez 1.3
32 : Isibaar 1.1 #include "../portab.h"
33 :     #include "mbfunctions.h"
34 :    
35 :     #include "../global.h"
36 :     #include "mem_transfer.h"
37 :     #include "timer.h"
38 : chl 1.21.2.9 #include "../bitstream/mbcoding.h"
39 : chl 1.21.2.10 #include "../bitstream/zigzag.h"
40 : Isibaar 1.1 #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 1.21 #include "../image/reduced.h"
47 : Isibaar 1.1
48 : edgomez 1.21 MBFIELDTEST_PTR MBFieldTest;
49 : Isibaar 1.1
50 : edgomez 1.21.2.6 /*
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 :    
57 :     /*****************************************************************************
58 :     * Local functions
59 :     ****************************************************************************/
60 : Isibaar 1.1
61 : edgomez 1.21.2.6 /* permute block and return field dct choice */
62 :     static __inline uint32_t
63 :     MBDecideFieldDCT(int16_t data[6 * 64])
64 : Isibaar 1.1 {
65 : edgomez 1.21.2.6 uint32_t field = MBFieldTest(data);
66 : edgomez 1.3
67 : edgomez 1.21.2.6 if (field)
68 :     MBFrameToField(data);
69 : edgomez 1.21
70 : edgomez 1.21.2.6 return field;
71 :     }
72 : h 1.2
73 : edgomez 1.21.2.6 /* Performs Forward DCT on all blocks */
74 :     static __inline void
75 : syskin 1.21.2.8 MBfDCT(const MBParam * const pParam,
76 :     const FRAMEINFO * const frame,
77 :     MACROBLOCK * const pMB,
78 : edgomez 1.21.2.6 uint32_t x_pos,
79 :     uint32_t y_pos,
80 :     int16_t data[6 * 64])
81 : syskin 1.21.2.8 {
82 : edgomez 1.21.2.6 /* Handles interlacing */
83 : h 1.2 start_timer();
84 :     pMB->field_dct = 0;
85 : edgomez 1.21.2.5 if ((frame->vol_flags & XVID_VOL_INTERLACING) &&
86 : h 1.11 (x_pos>0) && (x_pos<pParam->mb_width-1) &&
87 :     (y_pos>0) && (y_pos<pParam->mb_height-1)) {
88 : h 1.2 pMB->field_dct = MBDecideFieldDCT(data);
89 :     }
90 :     stop_interlacing_timer();
91 :    
92 : edgomez 1.21.2.6 /* 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 1.21.2.9 const FRAMEINFO * const frame,
122 : edgomez 1.21.2.6 const MACROBLOCK * pMB,
123 : syskin 1.21.2.8 int16_t qcoeff[6 * 64],
124 : edgomez 1.21.2.6 int16_t data[6*64])
125 :     {
126 :     int i;
127 :    
128 :     for (i = 0; i < 6; i++) {
129 :     uint32_t iDcScaler = get_dc_scaler(pMB->quant, i < 4);
130 :    
131 :     /* Quantize the block */
132 :     start_timer();
133 : chl 1.21.2.9 if (!(pParam->vol_flags & XVID_VOL_MPEGQUANT)) {
134 : edgomez 1.21.2.6 quant_intra(&data[i * 64], &qcoeff[i * 64], pMB->quant, iDcScaler);
135 : chl 1.21.2.9 } else {
136 : edgomez 1.21.2.6 quant4_intra(&data[i * 64], &qcoeff[i * 64], pMB->quant, iDcScaler);
137 : chl 1.21.2.9 }
138 : edgomez 1.21.2.6 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 : edgomez 1.7 for (i = 0; i < 6; i++) {
152 : Isibaar 1.1 uint32_t iDcScaler = get_dc_scaler(iQuant, i < 4);
153 :    
154 :     start_timer();
155 : edgomez 1.21.2.6 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 1.1
163 : chl 1.21.2.10
164 :     static int
165 : edgomez 1.21.2.13 dct_quantize_trellis_h263_c(int16_t *const Out,
166 :     const int16_t *const In,
167 :     int Q,
168 :     const uint16_t * const Zigzag,
169 :     int Non_Zero);
170 : chl 1.21.2.10
171 : edgomez 1.21.2.13 #if 0
172 : chl 1.21.2.10 static int
173 : edgomez 1.21.2.13 dct_quantize_trellis_mpeg_c(int16_t *const Out,
174 :     const int16_t *const In,
175 :     int Q,
176 :     const uint16_t * const Zigzag,
177 :     int Non_Zero);
178 :     #endif
179 : chl 1.21.2.10
180 : edgomez 1.21.2.6 /* Quantize all blocks -- Inter mode */
181 :     static __inline uint8_t
182 :     MBQuantInter(const MBParam * pParam,
183 : chl 1.21.2.9 const FRAMEINFO * const frame,
184 : edgomez 1.21.2.6 const MACROBLOCK * pMB,
185 :     int16_t data[6 * 64],
186 :     int16_t qcoeff[6 * 64],
187 :     int bvop,
188 :     int limit)
189 :     {
190 :    
191 :     int i;
192 :     uint8_t cbp = 0;
193 :     int sum;
194 :     int code_block;
195 :    
196 :     for (i = 0; i < 6; i++) {
197 : syskin 1.21.2.8
198 : edgomez 1.21.2.6 /* Quantize the block */
199 :     start_timer();
200 : chl 1.21.2.9 if (!(pParam->vol_flags & XVID_VOL_MPEGQUANT)) {
201 :     sum = quant_inter(&qcoeff[i*64], &data[i*64], pMB->quant);
202 :     if ( (sum) && (frame->vop_flags & XVID_VOP_TRELLISQUANT) ) {
203 : chl 1.21.2.10 sum = dct_quantize_trellis_h263_c(&qcoeff[i*64], &data[i*64], pMB->quant, &scan_tables[0][0], 63)+1;
204 : chl 1.21.2.9 limit = 1;
205 :     }
206 :     } else {
207 : edgomez 1.21.2.6 sum = quant4_inter(&qcoeff[i * 64], &data[i * 64], pMB->quant);
208 : edgomez 1.21.2.13 #if 0
209 :     if ( (sum) && (frame->vop_flags & XVID_VOP_TRELLISQUANT) )
210 :     sum = dct_quantize_trellis_mpeg_c (&qcoeff[i*64], &data[i*64], pMB->quant)+1;
211 :     #endif
212 : chl 1.21.2.9 }
213 : edgomez 1.21.2.6 stop_quant_timer();
214 :    
215 :     /*
216 :     * We code the block if the sum is higher than the limit and if the first
217 :     * two AC coefficients in zig zag order are not zero.
218 :     */
219 :     code_block = 0;
220 :     if ((sum >= limit) || (qcoeff[i*64+1] != 0) || (qcoeff[i*64+8] != 0)) {
221 :     code_block = 1;
222 : edgomez 1.7 } else {
223 : edgomez 1.21
224 : edgomez 1.21.2.6 if (bvop && (pMB->mode == MODE_DIRECT || pMB->mode == MODE_DIRECT_NO4V)) {
225 :     /* dark blocks prevention for direct mode */
226 :     if ((qcoeff[i*64] < -1) || (qcoeff[i*64] > 0))
227 :     code_block = 1;
228 : edgomez 1.21 } else {
229 : edgomez 1.21.2.6 /* not direct mode */
230 :     if (qcoeff[i*64] != 0)
231 :     code_block = 1;
232 : edgomez 1.21 }
233 :     }
234 : edgomez 1.21.2.6
235 :     /* Set the corresponding cbp bit */
236 :     cbp |= code_block << (5 - i);
237 : edgomez 1.21 }
238 :    
239 : edgomez 1.21.2.6 return(cbp);
240 :     }
241 : Isibaar 1.1
242 : edgomez 1.21.2.6 /* DeQuantize all blocks -- Inter mode */
243 : syskin 1.21.2.8 static __inline void
244 : edgomez 1.21.2.6 MBDeQuantInter(const MBParam * pParam,
245 :     const int iQuant,
246 :     int16_t data[6 * 64],
247 :     int16_t qcoeff[6 * 64],
248 :     const uint8_t cbp)
249 :     {
250 :     int i;
251 :    
252 :     for (i = 0; i < 6; i++) {
253 : syskin 1.21.2.8 if (cbp & (1 << (5 - i))) {
254 : edgomez 1.21.2.6 start_timer();
255 :     if (!(pParam->vol_flags & XVID_VOL_MPEGQUANT))
256 :     dequant_inter(&data[i * 64], &qcoeff[i * 64], iQuant);
257 :     else
258 :     dequant4_inter(&data[i * 64], &qcoeff[i * 64], iQuant);
259 :     stop_iquant_timer();
260 : edgomez 1.21 }
261 : h 1.2 }
262 : Isibaar 1.1 }
263 :    
264 : edgomez 1.21.2.6 typedef void (transfer_operation_8to16_t) (int16_t *Dst, const uint8_t *Src, int BpS);
265 :     typedef void (transfer_operation_16to8_t) (uint8_t *Dst, const int16_t *Src, int BpS);
266 : Isibaar 1.1
267 : edgomez 1.3
268 : edgomez 1.21.2.6 static __inline void
269 : syskin 1.21.2.8 MBTrans8to16(const MBParam * const pParam,
270 :     const FRAMEINFO * const frame,
271 :     const MACROBLOCK * const pMB,
272 : edgomez 1.21.2.6 const uint32_t x_pos,
273 :     const uint32_t y_pos,
274 :     int16_t data[6 * 64])
275 :     {
276 : h 1.4 uint32_t stride = pParam->edged_width;
277 :     uint32_t stride2 = stride / 2;
278 : edgomez 1.21.2.6 uint32_t next_block = stride * 8;
279 : syskin 1.21.2.8 int32_t cst;
280 : Isibaar 1.1 uint8_t *pY_Cur, *pU_Cur, *pV_Cur;
281 : syskin 1.21.2.8 const IMAGE * const pCurrent = &frame->image;
282 : edgomez 1.21.2.6 transfer_operation_8to16_t *transfer_op = NULL;
283 : edgomez 1.7
284 : edgomez 1.21.2.6 if ((frame->vop_flags & XVID_VOP_REDUCED)) {
285 :    
286 :     /* Image pointers */
287 :     pY_Cur = pCurrent->y + (y_pos << 5) * stride + (x_pos << 5);
288 : edgomez 1.21 pU_Cur = pCurrent->u + (y_pos << 4) * stride2 + (x_pos << 4);
289 :     pV_Cur = pCurrent->v + (y_pos << 4) * stride2 + (x_pos << 4);
290 : edgomez 1.21.2.6
291 :     /* Block size */
292 :     cst = 16;
293 :    
294 :     /* Operation function */
295 :     transfer_op = (transfer_operation_8to16_t*)filter_18x18_to_8x8;
296 :     } else {
297 :    
298 :     /* Image pointers */
299 :     pY_Cur = pCurrent->y + (y_pos << 4) * stride + (x_pos << 4);
300 : edgomez 1.21 pU_Cur = pCurrent->u + (y_pos << 3) * stride2 + (x_pos << 3);
301 :     pV_Cur = pCurrent->v + (y_pos << 3) * stride2 + (x_pos << 3);
302 : edgomez 1.21.2.6
303 :     /* Block size */
304 :     cst = 8;
305 :    
306 :     /* Operation function */
307 :     transfer_op = (transfer_operation_8to16_t*)transfer_8to16copy;
308 : edgomez 1.21 }
309 : Isibaar 1.1
310 : edgomez 1.21.2.6 /* Do the transfer */
311 : h 1.2 start_timer();
312 : edgomez 1.21.2.6 transfer_op(&data[0 * 64], pY_Cur, stride);
313 :     transfer_op(&data[1 * 64], pY_Cur + cst, stride);
314 :     transfer_op(&data[2 * 64], pY_Cur + next_block, stride);
315 :     transfer_op(&data[3 * 64], pY_Cur + next_block + cst, stride);
316 :     transfer_op(&data[4 * 64], pU_Cur, stride2);
317 :     transfer_op(&data[5 * 64], pV_Cur, stride2);
318 :     stop_transfer_timer();
319 : syskin 1.21.2.8 }
320 : h 1.2
321 : edgomez 1.21.2.6 static __inline void
322 : syskin 1.21.2.8 MBTrans16to8(const MBParam * const pParam,
323 :     const FRAMEINFO * const frame,
324 :     const MACROBLOCK * const pMB,
325 : edgomez 1.21.2.6 const uint32_t x_pos,
326 :     const uint32_t y_pos,
327 :     int16_t data[6 * 64],
328 :     const uint32_t add,
329 :     const uint8_t cbp)
330 :     {
331 :     uint8_t *pY_Cur, *pU_Cur, *pV_Cur;
332 :     uint32_t stride = pParam->edged_width;
333 :     uint32_t stride2 = stride / 2;
334 :     uint32_t next_block = stride * 8;
335 : syskin 1.21.2.8 uint32_t cst;
336 :     const IMAGE * const pCurrent = &frame->image;
337 : edgomez 1.21.2.6 transfer_operation_16to8_t *transfer_op = NULL;
338 : edgomez 1.21
339 : edgomez 1.21.2.6 if (pMB->field_dct) {
340 :     next_block = stride;
341 :     stride *= 2;
342 :     }
343 : Isibaar 1.1
344 : edgomez 1.21.2.6 if ((frame->vop_flags & XVID_VOP_REDUCED)) {
345 :    
346 :     /* Image pointers */
347 :     pY_Cur = pCurrent->y + (y_pos << 5) * stride + (x_pos << 5);
348 :     pU_Cur = pCurrent->u + (y_pos << 4) * stride2 + (x_pos << 4);
349 :     pV_Cur = pCurrent->v + (y_pos << 4) * stride2 + (x_pos << 4);
350 : Isibaar 1.1
351 : edgomez 1.21.2.6 /* Block size */
352 :     cst = 16;
353 : Isibaar 1.1
354 : edgomez 1.21.2.6 /* Operation function */
355 :     if(add)
356 :     transfer_op = (transfer_operation_16to8_t*)add_upsampled_8x8_16to8;
357 :     else
358 :     transfer_op = (transfer_operation_16to8_t*)copy_upsampled_8x8_16to8;
359 :     } else {
360 : Isibaar 1.1
361 : edgomez 1.21.2.6 /* Image pointers */
362 :     pY_Cur = pCurrent->y + (y_pos << 4) * stride + (x_pos << 4);
363 :     pU_Cur = pCurrent->u + (y_pos << 3) * stride2 + (x_pos << 3);
364 :     pV_Cur = pCurrent->v + (y_pos << 3) * stride2 + (x_pos << 3);
365 : Isibaar 1.1
366 : edgomez 1.21.2.6 /* Block size */
367 :     cst = 8;
368 : h 1.2
369 : edgomez 1.21.2.6 /* Operation function */
370 :     if(add)
371 :     transfer_op = (transfer_operation_16to8_t*)transfer_16to8add;
372 :     else
373 :     transfer_op = (transfer_operation_16to8_t*)transfer_16to8copy;
374 : h 1.2 }
375 :    
376 : edgomez 1.21.2.6 /* Do the operation */
377 : h 1.2 start_timer();
378 : edgomez 1.21.2.6 if (cbp&32) transfer_op(pY_Cur, &data[0 * 64], stride);
379 :     if (cbp&16) transfer_op(pY_Cur + cst, &data[1 * 64], stride);
380 :     if (cbp& 8) transfer_op(pY_Cur + next_block, &data[2 * 64], stride);
381 :     if (cbp& 4) transfer_op(pY_Cur + next_block + cst, &data[3 * 64], stride);
382 :     if (cbp& 2) transfer_op(pU_Cur, &data[4 * 64], stride2);
383 :     if (cbp& 1) transfer_op(pV_Cur, &data[5 * 64], stride2);
384 : h 1.2 stop_transfer_timer();
385 :     }
386 :    
387 : edgomez 1.21.2.6 /*****************************************************************************
388 :     * Module functions
389 :     ****************************************************************************/
390 :    
391 : syskin 1.21.2.8 void
392 :     MBTransQuantIntra(const MBParam * const pParam,
393 :     const FRAMEINFO * const frame,
394 :     MACROBLOCK * const pMB,
395 : chl 1.8 const uint32_t x_pos,
396 :     const uint32_t y_pos,
397 :     int16_t data[6 * 64],
398 :     int16_t qcoeff[6 * 64])
399 :     {
400 :    
401 : edgomez 1.21.2.6 /* Transfer data */
402 :     MBTrans8to16(pParam, frame, pMB, x_pos, y_pos, data);
403 : chl 1.8
404 : edgomez 1.21.2.6 /* Perform DCT (and field decision) */
405 :     MBfDCT(pParam, frame, pMB, x_pos, y_pos, data);
406 : chl 1.8
407 : edgomez 1.21.2.6 /* Quantize the block */
408 : chl 1.21.2.9 MBQuantIntra(pParam, frame, pMB, data, qcoeff);
409 : edgomez 1.21.2.6
410 :     /* DeQuantize the block */
411 :     MBDeQuantIntra(pParam, pMB->quant, data, qcoeff);
412 :    
413 :     /* Perform inverse DCT*/
414 :     MBiDCT(data, 0x3F);
415 :    
416 :     /* Transfer back the data -- Don't add data */
417 :     MBTrans16to8(pParam, frame, pMB, x_pos, y_pos, data, 0, 0x3F);
418 : chl 1.8 }
419 :    
420 : edgomez 1.21.2.6
421 : chl 1.8 uint8_t
422 : syskin 1.21.2.8 MBTransQuantInter(const MBParam * const pParam,
423 :     const FRAMEINFO * const frame,
424 :     MACROBLOCK * const pMB,
425 : edgomez 1.21.2.2 const uint32_t x_pos,
426 :     const uint32_t y_pos,
427 : chl 1.8 int16_t data[6 * 64],
428 :     int16_t qcoeff[6 * 64])
429 :     {
430 :     uint8_t cbp;
431 : edgomez 1.21.2.6 uint32_t limit;
432 : chl 1.8
433 : edgomez 1.21.2.2 /*
434 : edgomez 1.21.2.6 * There is no MBTrans8to16 for Inter block, that's done in motion compensation
435 :     * already
436 : edgomez 1.21.2.2 */
437 :    
438 : edgomez 1.21.2.6 /* Perform DCT (and field decision) */
439 :     MBfDCT(pParam, frame, pMB, x_pos, y_pos, data);
440 : chl 1.8
441 : edgomez 1.21.2.6 /* Set the limit threshold */
442 :     limit = PVOP_TOOSMALL_LIMIT + ((pMB->quant == 1)? 1 : 0);
443 : chl 1.8
444 : edgomez 1.21.2.6 /* Quantize the block */
445 : chl 1.21.2.9 cbp = MBQuantInter(pParam, frame, pMB, data, qcoeff, 0, limit);
446 : chl 1.8
447 : edgomez 1.21.2.6 /* DeQuantize the block */
448 :     MBDeQuantInter(pParam, pMB->quant, data, qcoeff, cbp);
449 : chl 1.8
450 : edgomez 1.21.2.6 /* Perform inverse DCT*/
451 :     MBiDCT(data, cbp);
452 : chl 1.8
453 : edgomez 1.21.2.6 /* Transfer back the data -- Add the data */
454 :     MBTrans16to8(pParam, frame, pMB, x_pos, y_pos, data, 1, cbp);
455 : syskin 1.21.2.8
456 : edgomez 1.21.2.6 return(cbp);
457 : chl 1.8 }
458 :    
459 : edgomez 1.21.2.6 uint8_t
460 :     MBTransQuantInterBVOP(const MBParam * pParam,
461 : chl 1.8 FRAMEINFO * frame,
462 :     MACROBLOCK * pMB,
463 :     const uint32_t x_pos,
464 :     const uint32_t y_pos,
465 :     int16_t data[6 * 64],
466 : edgomez 1.21.2.6 int16_t qcoeff[6 * 64])
467 : chl 1.8 {
468 : edgomez 1.21.2.6 uint8_t cbp;
469 :     uint32_t limit;
470 : syskin 1.21.2.8
471 : edgomez 1.21.2.6 /*
472 :     * There is no MBTrans8to16 for Inter block, that's done in motion compensation
473 :     * already
474 :     */
475 : chl 1.8
476 : edgomez 1.21.2.6 /* Perform DCT (and field decision) */
477 :     MBfDCT(pParam, frame, pMB, x_pos, y_pos, data);
478 : chl 1.8
479 : edgomez 1.21.2.6 /* Set the limit threshold */
480 :     limit = BVOP_TOOSMALL_LIMIT;
481 : chl 1.8
482 : edgomez 1.21.2.6 /* Quantize the block */
483 : chl 1.21.2.9 cbp = MBQuantInter(pParam, frame, pMB, data, qcoeff, 1, limit);
484 : h 1.2
485 : edgomez 1.21.2.6 /*
486 :     * History comment:
487 :     * We don't have to DeQuant, iDCT and Transfer back data for B-frames.
488 :     *
489 :     * BUT some plugins require the original frame to be passed so we have
490 :     * to take care of that here
491 :     */
492 :     if((pParam->plugin_flags & XVID_REQORIGINAL)) {
493 : h 1.2
494 : edgomez 1.21.2.6 /* DeQuantize the block */
495 :     MBDeQuantInter(pParam, pMB->quant, data, qcoeff, cbp);
496 : Isibaar 1.1
497 : edgomez 1.21.2.6 /* Perform inverse DCT*/
498 :     MBiDCT(data, cbp);
499 : edgomez 1.21
500 : edgomez 1.21.2.6 /* Transfer back the data -- Add the data */
501 :     MBTrans16to8(pParam, frame, pMB, x_pos, y_pos, data, 1, cbp);
502 : edgomez 1.21 }
503 :    
504 : edgomez 1.21.2.6 return(cbp);
505 : edgomez 1.21 }
506 :    
507 :     /* if sum(diff between field lines) < sum(diff between frame lines), use field dct */
508 :     uint32_t
509 :     MBFieldTest_c(int16_t data[6 * 64])
510 :     {
511 : edgomez 1.7 const uint8_t blocks[] =
512 :     { 0 * 64, 0 * 64, 0 * 64, 0 * 64, 2 * 64, 2 * 64, 2 * 64, 2 * 64 };
513 :     const uint8_t lines[] = { 0, 16, 32, 48, 0, 16, 32, 48 };
514 : h 1.2
515 :     int frame = 0, field = 0;
516 :     int i, j;
517 :    
518 : edgomez 1.7 for (i = 0; i < 7; ++i) {
519 :     for (j = 0; j < 8; ++j) {
520 :     frame +=
521 : edgomez 1.21.2.7 abs(data[0 * 64 + (i + 1) * 8 + j] - data[0 * 64 + i * 8 + j]);
522 : edgomez 1.7 frame +=
523 : edgomez 1.21.2.7 abs(data[1 * 64 + (i + 1) * 8 + j] - data[1 * 64 + i * 8 + j]);
524 : edgomez 1.7 frame +=
525 : edgomez 1.21.2.7 abs(data[2 * 64 + (i + 1) * 8 + j] - data[2 * 64 + i * 8 + j]);
526 : edgomez 1.7 frame +=
527 : edgomez 1.21.2.7 abs(data[3 * 64 + (i + 1) * 8 + j] - data[3 * 64 + i * 8 + j]);
528 : edgomez 1.7
529 :     field +=
530 : edgomez 1.21.2.7 abs(data[blocks[i + 1] + lines[i + 1] + j] -
531 : edgomez 1.7 data[blocks[i] + lines[i] + j]);
532 :     field +=
533 : edgomez 1.21.2.7 abs(data[blocks[i + 1] + lines[i + 1] + 8 + j] -
534 : edgomez 1.7 data[blocks[i] + lines[i] + 8 + j]);
535 :     field +=
536 : edgomez 1.21.2.7 abs(data[blocks[i + 1] + 64 + lines[i + 1] + j] -
537 : edgomez 1.7 data[blocks[i] + 64 + lines[i] + j]);
538 :     field +=
539 : edgomez 1.21.2.7 abs(data[blocks[i + 1] + 64 + lines[i + 1] + 8 + j] -
540 : edgomez 1.7 data[blocks[i] + 64 + lines[i] + 8 + j]);
541 : Isibaar 1.1 }
542 :     }
543 : h 1.2
544 : edgomez 1.21 return (frame >= (field + 350));
545 : h 1.2 }
546 :    
547 :    
548 :     /* deinterlace Y blocks vertically */
549 :    
550 :     #define MOVLINE(X,Y) memcpy(X, Y, sizeof(tmp))
551 : syskin 1.21.2.8 #define LINE(X,Y) &data[X*64 + Y*8]
552 : h 1.2
553 : edgomez 1.7 void
554 :     MBFrameToField(int16_t data[6 * 64])
555 : h 1.2 {
556 :     int16_t tmp[8];
557 :    
558 :     /* left blocks */
559 :    
560 : edgomez 1.21.2.13 /* 1=2, 2=4, 4=8, 8=1 */
561 : edgomez 1.7 MOVLINE(tmp, LINE(0, 1));
562 :     MOVLINE(LINE(0, 1), LINE(0, 2));
563 :     MOVLINE(LINE(0, 2), LINE(0, 4));
564 :     MOVLINE(LINE(0, 4), LINE(2, 0));
565 :     MOVLINE(LINE(2, 0), tmp);
566 : h 1.2
567 : edgomez 1.21.2.13 /* 3=6, 6=12, 12=9, 9=3 */
568 : edgomez 1.7 MOVLINE(tmp, LINE(0, 3));
569 :     MOVLINE(LINE(0, 3), LINE(0, 6));
570 :     MOVLINE(LINE(0, 6), LINE(2, 4));
571 :     MOVLINE(LINE(2, 4), LINE(2, 1));
572 :     MOVLINE(LINE(2, 1), tmp);
573 : h 1.2
574 : edgomez 1.21.2.13 /* 5=10, 10=5 */
575 : edgomez 1.7 MOVLINE(tmp, LINE(0, 5));
576 :     MOVLINE(LINE(0, 5), LINE(2, 2));
577 :     MOVLINE(LINE(2, 2), tmp);
578 : h 1.2
579 : edgomez 1.21.2.13 /* 7=14, 14=13, 13=11, 11=7 */
580 : edgomez 1.7 MOVLINE(tmp, LINE(0, 7));
581 :     MOVLINE(LINE(0, 7), LINE(2, 6));
582 :     MOVLINE(LINE(2, 6), LINE(2, 5));
583 :     MOVLINE(LINE(2, 5), LINE(2, 3));
584 :     MOVLINE(LINE(2, 3), tmp);
585 : h 1.2
586 :     /* right blocks */
587 :    
588 : edgomez 1.21.2.13 /* 1=2, 2=4, 4=8, 8=1 */
589 : edgomez 1.7 MOVLINE(tmp, LINE(1, 1));
590 :     MOVLINE(LINE(1, 1), LINE(1, 2));
591 :     MOVLINE(LINE(1, 2), LINE(1, 4));
592 :     MOVLINE(LINE(1, 4), LINE(3, 0));
593 :     MOVLINE(LINE(3, 0), tmp);
594 : h 1.2
595 : edgomez 1.21.2.13 /* 3=6, 6=12, 12=9, 9=3 */
596 : edgomez 1.7 MOVLINE(tmp, LINE(1, 3));
597 :     MOVLINE(LINE(1, 3), LINE(1, 6));
598 :     MOVLINE(LINE(1, 6), LINE(3, 4));
599 :     MOVLINE(LINE(3, 4), LINE(3, 1));
600 :     MOVLINE(LINE(3, 1), tmp);
601 : h 1.2
602 : edgomez 1.21.2.13 /* 5=10, 10=5 */
603 : edgomez 1.7 MOVLINE(tmp, LINE(1, 5));
604 :     MOVLINE(LINE(1, 5), LINE(3, 2));
605 :     MOVLINE(LINE(3, 2), tmp);
606 : h 1.2
607 : edgomez 1.21.2.13 /* 7=14, 14=13, 13=11, 11=7 */
608 : edgomez 1.7 MOVLINE(tmp, LINE(1, 7));
609 :     MOVLINE(LINE(1, 7), LINE(3, 6));
610 :     MOVLINE(LINE(3, 6), LINE(3, 5));
611 :     MOVLINE(LINE(3, 5), LINE(3, 3));
612 :     MOVLINE(LINE(3, 3), tmp);
613 : Isibaar 1.1 }
614 : chl 1.21.2.10
615 :    
616 :    
617 :    
618 :    
619 : edgomez 1.21.2.13 /*****************************************************************************
620 :     * Trellis based R-D optimal quantization
621 :     *
622 :     * Trellis Quant code (C) 2003 Pascal Massimino skal(at)planet-d.net
623 :     *
624 :     ****************************************************************************/
625 : chl 1.21.2.10
626 :    
627 : edgomez 1.21.2.13 #if 0
628 : chl 1.21.2.11 static int
629 : edgomez 1.21.2.13 dct_quantize_trellis_mpeg_c(int16_t *const Out,
630 :     const int16_t *const In,
631 :     int Q,
632 :     const uint16_t * const Zigzag,
633 :     int Non_Zero)
634 :     {
635 :     return 63;
636 :     }
637 :     #endif
638 :    
639 :     /*----------------------------------------------------------------------------
640 :     *
641 :     * Trellis-Based quantization
642 :     *
643 :     * So far I understand this paper:
644 :     *
645 :     * "Trellis-Based R-D Optimal Quantization in H.263+"
646 :     * J.Wen, M.Luttrell, J.Villasenor
647 :     * IEEE Transactions on Image Processing, Vol.9, No.8, Aug. 2000.
648 :     *
649 :     * we are at stake with a simplified Bellmand-Ford / Dijkstra Single
650 :     * Source Shorted Path algo. But due to the underlying graph structure
651 :     * ("Trellis"), it can be turned into a dynamic programming algo,
652 :     * partially saving the explicit graph's nodes representation. And
653 :     * without using a heap, since the open frontier of the DAG is always
654 :     * known, and of fixed sized.
655 :     *--------------------------------------------------------------------------*/
656 :    
657 : chl 1.21.2.10
658 :    
659 : edgomez 1.21.2.13 /* Codes lengths for relevant levels. */
660 : chl 1.21.2.10
661 : edgomez 1.21.2.13 /* let's factorize: */
662 : chl 1.21.2.10 static const uint8_t Code_Len0[64] = {
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 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
665 :     static const uint8_t Code_Len1[64] = {
666 :     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,
667 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
668 :     static const uint8_t Code_Len2[64] = {
669 :     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,
670 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
671 :     static const uint8_t Code_Len3[64] = {
672 :     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,
673 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
674 :     static const uint8_t Code_Len4[64] = {
675 :     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,
676 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
677 :     static const uint8_t Code_Len5[64] = {
678 :     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,
679 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
680 :     static const uint8_t Code_Len6[64] = {
681 :     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,
682 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
683 :     static const uint8_t Code_Len7[64] = {
684 :     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,
685 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
686 :     static const uint8_t Code_Len8[64] = {
687 :     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,
688 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
689 :     static const uint8_t Code_Len9[64] = {
690 :     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,
691 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
692 :     static const uint8_t Code_Len10[64] = {
693 :     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,
694 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
695 :     static const uint8_t Code_Len11[64] = {
696 :     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,
697 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
698 :     static const uint8_t Code_Len12[64] = {
699 :     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,
700 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
701 :     static const uint8_t Code_Len13[64] = {
702 :     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,
703 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
704 :     static const uint8_t Code_Len14[64] = {
705 :     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,
706 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
707 :     static const uint8_t Code_Len15[64] = {
708 :     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,
709 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
710 :     static const uint8_t Code_Len16[64] = {
711 :     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,
712 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30};
713 :     static const uint8_t Code_Len17[64] = {
714 :     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,
715 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
716 :     static const uint8_t Code_Len18[64] = {
717 :     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,
718 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
719 :     static const uint8_t Code_Len19[64] = {
720 :     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,
721 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
722 :     static const uint8_t Code_Len20[64] = {
723 :     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,
724 :     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 };
725 :    
726 : edgomez 1.21.2.13 /* a few more table for LAST table: */
727 : chl 1.21.2.10 static const uint8_t Code_Len21[64] = {
728 :     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,
729 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30};
730 :     static const uint8_t Code_Len22[64] = {
731 :     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,
732 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30};
733 :     static const uint8_t Code_Len23[64] = {
734 :     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,
735 :     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};
736 :     static const uint8_t Code_Len24[64] = {
737 :     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,
738 :     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};
739 :    
740 :    
741 : edgomez 1.21.2.13 static const uint8_t * const B16_17_Code_Len[24] = { /* levels [1..24] */
742 : chl 1.21.2.10 Code_Len20,Code_Len19,Code_Len18,Code_Len17,
743 :     Code_Len16,Code_Len15,Code_Len14,Code_Len13,
744 :     Code_Len12,Code_Len11,Code_Len10,Code_Len9,
745 :     Code_Len8, Code_Len7 ,Code_Len6 ,Code_Len5,
746 :     Code_Len4, Code_Len3, Code_Len3 ,Code_Len2,
747 :     Code_Len2, Code_Len1, Code_Len1, Code_Len1,
748 :     };
749 :    
750 : edgomez 1.21.2.13 static const uint8_t * const B16_17_Code_Len_Last[6] = { /* levels [1..6] */
751 : chl 1.21.2.10 Code_Len24,Code_Len23,Code_Len22,Code_Len21, Code_Len3, Code_Len1,
752 :     };
753 :    
754 :     #define TL(q) 0xfe00/(q*q)
755 :    
756 :     static const int Trellis_Lambda_Tabs[31] = {
757 :     TL( 1),TL( 2),TL( 3),TL( 4),TL( 5),TL( 6), TL( 7),
758 :     TL( 8),TL( 9),TL(10),TL(11),TL(12),TL(13),TL(14), TL(15),
759 :     TL(16),TL(17),TL(18),TL(19),TL(20),TL(21),TL(22), TL(23),
760 :     TL(24),TL(25),TL(26),TL(27),TL(28),TL(29),TL(30), TL(31)
761 :     };
762 :     #undef TL
763 :    
764 : suxen_drol 1.21.2.12 static __inline int Find_Last(const int16_t *C, const uint16_t *Zigzag, int i)
765 : chl 1.21.2.10 {
766 :     while(i>=0)
767 :     if (C[Zigzag[i]])
768 :     return i;
769 :     else i--;
770 :     return -1;
771 :     }
772 :    
773 : edgomez 1.21.2.13 /* this routine has been strippen of all debug code */
774 : chl 1.21.2.11
775 :     static int
776 :     dct_quantize_trellis_h263_c(int16_t *const Out, const int16_t *const In, int Q, const uint16_t * const Zigzag, int Non_Zero)
777 :     {
778 :    
779 : edgomez 1.21.2.13 /*
780 :     * Note: We should search last non-zero coeffs on *real* DCT input coeffs (In[]),
781 :     * not quantized one (Out[]). However, it only improves the result *very*
782 :     * slightly (~0.01dB), whereas speed drops to crawling level :)
783 :     * Well, actually, taking 1 more coeff past Non_Zero into account sometimes helps.
784 :     */
785 : chl 1.21.2.11 typedef struct { int16_t Run, Level; } NODE;
786 :    
787 :     NODE Nodes[65], Last;
788 :     uint32_t Run_Costs0[64+1];
789 :     uint32_t * const Run_Costs = Run_Costs0 + 1;
790 :     const int Mult = 2*Q;
791 :     const int Bias = (Q-1) | 1;
792 :     const int Lev0 = Mult + Bias;
793 : edgomez 1.21.2.13 const int Lambda = Trellis_Lambda_Tabs[Q-1]; /* it's 1/lambda, actually */
794 : chl 1.21.2.11
795 :     int Run_Start = -1;
796 :     uint32_t Min_Cost = 2<<16;
797 :    
798 :     int Last_Node = -1;
799 :     uint32_t Last_Cost = 0;
800 :    
801 :     int i, j;
802 : edgomez 1.21.2.13 Run_Costs[-1] = 2<<16; /* source (w/ CBP penalty) */
803 : chl 1.21.2.11
804 :     Non_Zero = Find_Last(Out, Zigzag, Non_Zero);
805 :     if (Non_Zero<0)
806 :     return -1;
807 :    
808 :     for(i=0; i<=Non_Zero; i++)
809 :     {
810 :     const int AC = In[Zigzag[i]];
811 :     const int Level1 = Out[Zigzag[i]];
812 :     const int Dist0 = Lambda* AC*AC;
813 :     uint32_t Best_Cost = 0xf0000000;
814 :     Last_Cost += Dist0;
815 :    
816 : edgomez 1.21.2.13 if ((uint32_t)(Level1+1)<3) /* very specialized loop for -1,0,+1 */
817 : chl 1.21.2.11 {
818 :     int dQ;
819 :     int Run;
820 :     uint32_t Cost0;
821 :    
822 :     if (AC<0) {
823 :     Nodes[i].Level = -1;
824 :     dQ = Lev0 + AC;
825 :     } else {
826 :     Nodes[i].Level = 1;
827 :     dQ = Lev0 - AC;
828 :     }
829 :     Cost0 = Lambda*dQ*dQ;
830 :    
831 :     Nodes[i].Run = 1;
832 :     Best_Cost = (Code_Len20[0]<<16) + Run_Costs[i-1]+Cost0;
833 :     for(Run=i-Run_Start; Run>0; --Run)
834 :     {
835 :     const uint32_t Cost_Base = Cost0 + Run_Costs[i-Run];
836 :     const uint32_t Cost = Cost_Base + (Code_Len20[Run-1]<<16);
837 :     const uint32_t lCost = Cost_Base + (Code_Len24[Run-1]<<16);
838 :    
839 : edgomez 1.21.2.13 /*
840 :     * TODO: what about tie-breaks? Should we favor short runs or
841 :     * long runs? Although the error is the same, it would not be
842 :     * spread the same way along high and low frequencies...
843 :     */
844 : chl 1.21.2.11
845 : edgomez 1.21.2.13 /* (I'd say: favour short runs => hifreq errors (HVS) -- gruel ) */
846 : chl 1.21.2.11
847 :     if (Cost<Best_Cost) {
848 :     Best_Cost = Cost;
849 :     Nodes[i].Run = Run;
850 :     }
851 :    
852 :     if (lCost<Last_Cost) {
853 :     Last_Cost = lCost;
854 :     Last.Run = Run;
855 :     Last_Node = i;
856 :     }
857 :     }
858 :     if (Last_Node==i)
859 :     Last.Level = Nodes[i].Level;
860 :     }
861 : edgomez 1.21.2.13 else /* "big" levels */
862 : chl 1.21.2.11 {
863 :     const uint8_t *Tbl_L1, *Tbl_L2, *Tbl_L1_Last, *Tbl_L2_Last;
864 :     int Level2;
865 :     int dQ1, dQ2;
866 :     int Run;
867 :     uint32_t Dist1,Dist2;
868 :     int dDist21;
869 :    
870 :     if (Level1>1) {
871 :     dQ1 = Level1*Mult-AC + Bias;
872 :     dQ2 = dQ1 - Mult;
873 :     Level2 = Level1-1;
874 :     Tbl_L1 = (Level1<=24) ? B16_17_Code_Len[Level1-1] : Code_Len0;
875 :     Tbl_L2 = (Level2<=24) ? B16_17_Code_Len[Level2-1] : Code_Len0;
876 :     Tbl_L1_Last = (Level1<=6) ? B16_17_Code_Len_Last[Level1-1] : Code_Len0;
877 :     Tbl_L2_Last = (Level2<=6) ? B16_17_Code_Len_Last[Level2-1] : Code_Len0;
878 : edgomez 1.21.2.13 } else { /* Level1<-1 */
879 : chl 1.21.2.11 dQ1 = Level1*Mult-AC - Bias;
880 :     dQ2 = dQ1 + Mult;
881 :     Level2 = Level1 + 1;
882 :     Tbl_L1 = (Level1>=-24) ? B16_17_Code_Len[Level1^-1] : Code_Len0;
883 :     Tbl_L2 = (Level2>=-24) ? B16_17_Code_Len[Level2^-1] : Code_Len0;
884 :     Tbl_L1_Last = (Level1>=- 6) ? B16_17_Code_Len_Last[Level1^-1] : Code_Len0;
885 :     Tbl_L2_Last = (Level2>=- 6) ? B16_17_Code_Len_Last[Level2^-1] : Code_Len0;
886 :     }
887 :     Dist1 = Lambda*dQ1*dQ1;
888 :     Dist2 = Lambda*dQ2*dQ2;
889 :     dDist21 = Dist2-Dist1;
890 :    
891 :     for(Run=i-Run_Start; Run>0; --Run)
892 :     {
893 :     const uint32_t Cost_Base = Dist1 + Run_Costs[i-Run];
894 :     uint32_t Cost1, Cost2;
895 :     int bLevel;
896 :    
897 : edgomez 1.21.2.13 /*
898 :     * for sub-optimal (but slightly worth it, speed-wise) search, uncomment the following:
899 :     * if (Cost_Base>=Best_Cost) continue;
900 :     * (? doesn't seem to have any effect -- gruel )
901 :     */
902 : chl 1.21.2.11
903 :     Cost1 = Cost_Base + (Tbl_L1[Run-1]<<16);
904 :     Cost2 = Cost_Base + (Tbl_L2[Run-1]<<16) + dDist21;
905 :    
906 :     if (Cost2<Cost1) {
907 :     Cost1 = Cost2;
908 :     bLevel = Level2;
909 :     } else
910 :     bLevel = Level1;
911 :    
912 :     if (Cost1<Best_Cost) {
913 :     Best_Cost = Cost1;
914 :     Nodes[i].Run = Run;
915 :     Nodes[i].Level = bLevel;
916 :     }
917 :    
918 :     Cost1 = Cost_Base + (Tbl_L1_Last[Run-1]<<16);
919 :     Cost2 = Cost_Base + (Tbl_L2_Last[Run-1]<<16) + dDist21;
920 :    
921 :     if (Cost2<Cost1) {
922 :     Cost1 = Cost2;
923 :     bLevel = Level2;
924 :     } else
925 :     bLevel = Level1;
926 :    
927 :     if (Cost1<Last_Cost) {
928 :     Last_Cost = Cost1;
929 :     Last.Run = Run;
930 :     Last.Level = bLevel;
931 :     Last_Node = i;
932 :     }
933 : edgomez 1.21.2.13 } /* end of "for Run" */
934 : chl 1.21.2.11
935 :     }
936 :    
937 :     Run_Costs[i] = Best_Cost;
938 :    
939 :     if (Best_Cost < Min_Cost + Dist0) {
940 :     Min_Cost = Best_Cost;
941 :     Run_Start = i;
942 :     }
943 :     else
944 :     {
945 : edgomez 1.21.2.13 /*
946 :     * as noticed by Michael Niedermayer (michaelni at gmx.at), there's
947 :     * a code shorter by 1 bit for a larger run (!), same level. We give
948 :     * it a chance by not moving the left barrier too much.
949 :     */
950 : chl 1.21.2.11
951 :     while( Run_Costs[Run_Start]>Min_Cost+(1<<16) )
952 :     Run_Start++;
953 :    
954 : edgomez 1.21.2.13 /* spread on preceding coeffs the cost incurred by skipping this one */
955 : chl 1.21.2.11 for(j=Run_Start; j<i; ++j) Run_Costs[j] += Dist0;
956 :     Min_Cost += Dist0;
957 :     }
958 :     }
959 :    
960 :     if (Last_Node<0)
961 :     return -1;
962 :    
963 : edgomez 1.21.2.13 /* reconstruct optimal sequence backward with surviving paths */
964 : chl 1.21.2.11 memset(Out, 0x00, 64*sizeof(*Out));
965 :     Out[Zigzag[Last_Node]] = Last.Level;
966 :     i = Last_Node - Last.Run;
967 :     while(i>=0) {
968 :     Out[Zigzag[i]] = Nodes[i].Level;
969 :     i -= Nodes[i].Run;
970 :     }
971 :     return Last_Node;
972 :     }
973 :    
974 :    
975 :    
976 :    
977 :    
978 :    
979 :    
980 :    
981 :    
982 :    
983 :    
984 : edgomez 1.21.2.13 /* original version including heavy debugging info */
985 : chl 1.21.2.11
986 :     #ifdef DBGTRELL
987 : chl 1.21.2.10
988 :     #define DBG 0
989 :    
990 : suxen_drol 1.21.2.12 static __inline uint32_t Evaluate_Cost(const int16_t *C, int Mult, int Bias,
991 : chl 1.21.2.10 const uint16_t * Zigzag, int Max, int Lambda)
992 :     {
993 :     #if (DBG>0)
994 :     const int16_t * const Ref = C + 6*64;
995 :     int Last = Max;
996 :     int Bits = 0;
997 : chl 1.21.2.11 int Dist = 0;
998 :     int i;
999 :     uint32_t Cost;
1000 :    
1001 :     while(Last>=0 && C[Zigzag[Last]]==0)
1002 :     Last--;
1003 :    
1004 : chl 1.21.2.10 if (Last>=0) {
1005 :     int j=0, j0=0;
1006 :     int Run, Level;
1007 : chl 1.21.2.11
1008 : edgomez 1.21.2.13 Bits = 2; /* CBP */
1009 : chl 1.21.2.10 while(j<Last) {
1010 : chl 1.21.2.11 while(!C[Zigzag[j]])
1011 :     j++;
1012 :     if (j==Last)
1013 :     break;
1014 : chl 1.21.2.10 Level=C[Zigzag[j]];
1015 :     Run = j - j0;
1016 :     j0 = ++j;
1017 : chl 1.21.2.11 if (Level>=-24 && Level<=24)
1018 :     Bits += B16_17_Code_Len[(Level<0) ? -Level-1 : Level-1][Run];
1019 :     else
1020 :     Bits += 30;
1021 : chl 1.21.2.10 }
1022 :     Level = C[Zigzag[Last]];
1023 :     Run = j - j0;
1024 : chl 1.21.2.11 if (Level>=-6 && Level<=6)
1025 :     Bits += B16_17_Code_Len_Last[(Level<0) ? -Level-1 : Level-1][Run];
1026 :     else
1027 :     Bits += 30;
1028 : chl 1.21.2.10 }
1029 :    
1030 :     for(i=0; i<=Last; ++i) {
1031 :     int V = C[Zigzag[i]]*Mult;
1032 : chl 1.21.2.11 if (V>0)
1033 :     V += Bias;
1034 :     else
1035 :     if (V<0)
1036 :     V -= Bias;
1037 : chl 1.21.2.10 V -= Ref[Zigzag[i]];
1038 :     Dist += V*V;
1039 :     }
1040 : chl 1.21.2.11 Cost = Lambda*Dist + (Bits<<16);
1041 : chl 1.21.2.10 if (DBG==1)
1042 :     printf( " Last:%2d/%2d Cost = [(Bits=%5.0d) + Lambda*(Dist=%6.0d) = %d ] >>12= %d ", Last,Max, Bits, Dist, Cost, Cost>>12 );
1043 :     return Cost;
1044 :    
1045 :     #else
1046 :     return 0;
1047 :     #endif
1048 :     }
1049 :    
1050 :    
1051 :     static int
1052 :     dct_quantize_trellis_h263_c(int16_t *const Out, const int16_t *const In, int Q, const uint16_t * const Zigzag, int Non_Zero)
1053 :     {
1054 :    
1055 : edgomez 1.21.2.13 /*
1056 :     * Note: We should search last non-zero coeffs on *real* DCT input coeffs (In[]),
1057 :     * not quantized one (Out[]). However, it only improves the result *very*
1058 :     * slightly (~0.01dB), whereas speed drops to crawling level :)
1059 :     * Well, actually, taking 1 more coeff past Non_Zero into account sometimes helps.
1060 :     */
1061 : chl 1.21.2.10 typedef struct { int16_t Run, Level; } NODE;
1062 :    
1063 :     NODE Nodes[65], Last;
1064 : chl 1.21.2.11 uint32_t Run_Costs0[64+1];
1065 :     uint32_t * const Run_Costs = Run_Costs0 + 1;
1066 : chl 1.21.2.10 const int Mult = 2*Q;
1067 :     const int Bias = (Q-1) | 1;
1068 :     const int Lev0 = Mult + Bias;
1069 : edgomez 1.21.2.13 const int Lambda = Trellis_Lambda_Tabs[Q-1]; /* it's 1/lambda, actually */
1070 : chl 1.21.2.10
1071 :     int Run_Start = -1;
1072 : edgomez 1.21.2.13 Run_Costs[-1] = 2<<16; /* source (w/ CBP penalty) */
1073 : chl 1.21.2.10 uint32_t Min_Cost = 2<<16;
1074 :    
1075 :     int Last_Node = -1;
1076 :     uint32_t Last_Cost = 0;
1077 :    
1078 : chl 1.21.2.11 int i, j;
1079 :    
1080 : chl 1.21.2.10 #if (DBG>0)
1081 : edgomez 1.21.2.13 Last.Level = 0; Last.Run = -1; /* just initialize to smthg */
1082 : chl 1.21.2.10 #endif
1083 :    
1084 :     Non_Zero = Find_Last(Out, Zigzag, Non_Zero);
1085 :     if (Non_Zero<0)
1086 :     return -1;
1087 :    
1088 :     for(i=0; i<=Non_Zero; i++)
1089 :     {
1090 :     const int AC = In[Zigzag[i]];
1091 :     const int Level1 = Out[Zigzag[i]];
1092 :     const int Dist0 = Lambda* AC*AC;
1093 :     uint32_t Best_Cost = 0xf0000000;
1094 :     Last_Cost += Dist0;
1095 :    
1096 : edgomez 1.21.2.13 if ((uint32_t)(Level1+1)<3) /* very specialized loop for -1,0,+1 */
1097 : chl 1.21.2.10 {
1098 : chl 1.21.2.11 int dQ;
1099 :     int Run;
1100 :     uint32_t Cost0;
1101 : chl 1.21.2.10
1102 :     if (AC<0) {
1103 :     Nodes[i].Level = -1;
1104 :     dQ = Lev0 + AC;
1105 :     } else {
1106 :     Nodes[i].Level = 1;
1107 :     dQ = Lev0 - AC;
1108 :     }
1109 : chl 1.21.2.11 Cost0 = Lambda*dQ*dQ;
1110 :    
1111 : chl 1.21.2.10 Nodes[i].Run = 1;
1112 :     Best_Cost = (Code_Len20[0]<<16) + Run_Costs[i-1]+Cost0;
1113 :     for(Run=i-Run_Start; Run>0; --Run)
1114 :     {
1115 :     const uint32_t Cost_Base = Cost0 + Run_Costs[i-Run];
1116 :     const uint32_t Cost = Cost_Base + (Code_Len20[Run-1]<<16);
1117 : chl 1.21.2.11 const uint32_t lCost = Cost_Base + (Code_Len24[Run-1]<<16);
1118 :    
1119 : edgomez 1.21.2.13 /*
1120 :     * TODO: what about tie-breaks? Should we favor short runs or
1121 :     * long runs? Although the error is the same, it would not be
1122 :     * spread the same way along high and low frequencies...
1123 :     */
1124 : chl 1.21.2.11 if (Cost<Best_Cost) {
1125 : chl 1.21.2.10 Best_Cost = Cost;
1126 :     Nodes[i].Run = Run;
1127 :     }
1128 : chl 1.21.2.11
1129 :     if (lCost<Last_Cost) {
1130 : chl 1.21.2.10 Last_Cost = lCost;
1131 :     Last.Run = Run;
1132 :     Last_Node = i;
1133 :     }
1134 :     }
1135 : chl 1.21.2.11 if (Last_Node==i)
1136 :     Last.Level = Nodes[i].Level;
1137 : chl 1.21.2.10
1138 :     if (DBG==1) {
1139 :     Run_Costs[i] = Best_Cost;
1140 :     printf( "Costs #%2d: ", i);
1141 :     for(j=-1;j<=Non_Zero;++j) {
1142 :     if (j==Run_Start) printf( " %3.0d|", Run_Costs[j]>>12 );
1143 :     else if (j>Run_Start && j<i) printf( " %3.0d|", Run_Costs[j]>>12 );
1144 :     else if (j==i) printf( "(%3.0d)", Run_Costs[j]>>12 );
1145 :     else printf( " - |" );
1146 :     }
1147 :     printf( "<%3.0d %2d %d>", Min_Cost>>12, Nodes[i].Level, Nodes[i].Run );
1148 :     printf( " Last:#%2d {%3.0d %2d %d}", Last_Node, Last_Cost>>12, Last.Level, Last.Run );
1149 :     printf( " AC:%3.0d Dist0:%3d Dist(%d)=%d", AC, Dist0>>12, Nodes[i].Level, Cost0>>12 );
1150 :     printf( "\n" );
1151 :     }
1152 :     }
1153 : edgomez 1.21.2.13 else /* "big" levels */
1154 : chl 1.21.2.10 {
1155 :     const uint8_t *Tbl_L1, *Tbl_L2, *Tbl_L1_Last, *Tbl_L2_Last;
1156 :     int Level2;
1157 :     int dQ1, dQ2;
1158 :     int Run;
1159 : chl 1.21.2.11 uint32_t Dist1,Dist2;
1160 :     int dDist21;
1161 :    
1162 : chl 1.21.2.10 if (Level1>1) {
1163 :     dQ1 = Level1*Mult-AC + Bias;
1164 :     dQ2 = dQ1 - Mult;
1165 :     Level2 = Level1-1;
1166 :     Tbl_L1 = (Level1<=24) ? B16_17_Code_Len[Level1-1] : Code_Len0;
1167 :     Tbl_L2 = (Level2<=24) ? B16_17_Code_Len[Level2-1] : Code_Len0;
1168 :     Tbl_L1_Last = (Level1<=6) ? B16_17_Code_Len_Last[Level1-1] : Code_Len0;
1169 :     Tbl_L2_Last = (Level2<=6) ? B16_17_Code_Len_Last[Level2-1] : Code_Len0;
1170 : edgomez 1.21.2.13 } else { /* Level1<-1 */
1171 : chl 1.21.2.10 dQ1 = Level1*Mult-AC - Bias;
1172 :     dQ2 = dQ1 + Mult;
1173 :     Level2 = Level1 + 1;
1174 :     Tbl_L1 = (Level1>=-24) ? B16_17_Code_Len[Level1^-1] : Code_Len0;
1175 :     Tbl_L2 = (Level2>=-24) ? B16_17_Code_Len[Level2^-1] : Code_Len0;
1176 :     Tbl_L1_Last = (Level1>=- 6) ? B16_17_Code_Len_Last[Level1^-1] : Code_Len0;
1177 :     Tbl_L2_Last = (Level2>=- 6) ? B16_17_Code_Len_Last[Level2^-1] : Code_Len0;
1178 :     }
1179 : chl 1.21.2.11 Dist1 = Lambda*dQ1*dQ1;
1180 :     Dist2 = Lambda*dQ2*dQ2;
1181 :     dDist21 = Dist2-Dist1;
1182 : chl 1.21.2.10
1183 :     for(Run=i-Run_Start; Run>0; --Run)
1184 :     {
1185 :     const uint32_t Cost_Base = Dist1 + Run_Costs[i-Run];
1186 : chl 1.21.2.11 uint32_t Cost1, Cost2;
1187 :     int bLevel;
1188 : chl 1.21.2.10
1189 : edgomez 1.21.2.13 /*
1190 :     * for sub-optimal (but slightly worth it, speed-wise) search, uncomment the following:
1191 :     * if (Cost_Base>=Best_Cost) continue;
1192 :     */
1193 : chl 1.21.2.10 Cost1 = Cost_Base + (Tbl_L1[Run-1]<<16);
1194 :     Cost2 = Cost_Base + (Tbl_L2[Run-1]<<16) + dDist21;
1195 :    
1196 : chl 1.21.2.11 if (Cost2<Cost1) {
1197 :     Cost1 = Cost2;
1198 :     bLevel = Level2;
1199 :     } else
1200 :     bLevel = Level1;
1201 : chl 1.21.2.10
1202 : chl 1.21.2.11 if (Cost1<Best_Cost) {
1203 : chl 1.21.2.10 Best_Cost = Cost1;
1204 :     Nodes[i].Run = Run;
1205 :     Nodes[i].Level = bLevel;
1206 :     }
1207 :    
1208 :     Cost1 = Cost_Base + (Tbl_L1_Last[Run-1]<<16);
1209 :     Cost2 = Cost_Base + (Tbl_L2_Last[Run-1]<<16) + dDist21;
1210 :    
1211 : chl 1.21.2.11 if (Cost2<Cost1) {
1212 :     Cost1 = Cost2;
1213 :     bLevel = Level2;
1214 :     } else
1215 :     bLevel = Level1;
1216 :    
1217 :     if (Cost1<Last_Cost) {
1218 : chl 1.21.2.10 Last_Cost = Cost1;
1219 :     Last.Run = Run;
1220 :     Last.Level = bLevel;
1221 :     Last_Node = i;
1222 :     }
1223 : edgomez 1.21.2.13 } /* end of "for Run" */
1224 : chl 1.21.2.10
1225 :     if (DBG==1) {
1226 :     Run_Costs[i] = Best_Cost;
1227 :     printf( "Costs #%2d: ", i);
1228 :     for(j=-1;j<=Non_Zero;++j) {
1229 :     if (j==Run_Start) printf( " %3.0d|", Run_Costs[j]>>12 );
1230 :     else if (j>Run_Start && j<i) printf( " %3.0d|", Run_Costs[j]>>12 );
1231 :     else if (j==i) printf( "(%3.0d)", Run_Costs[j]>>12 );
1232 :     else printf( " - |" );
1233 :     }
1234 :     printf( "<%3.0d %2d %d>", Min_Cost>>12, Nodes[i].Level, Nodes[i].Run );
1235 :     printf( " Last:#%2d {%3.0d %2d %d}", Last_Node, Last_Cost>>12, Last.Level, Last.Run );
1236 :     printf( " AC:%3.0d Dist0:%3d Dist(%2d):%3d Dist(%2d):%3d", AC, Dist0>>12, Level1, Dist1>>12, Level2, Dist2>>12 );
1237 :     printf( "\n" );
1238 :     }
1239 :     }
1240 :    
1241 :     Run_Costs[i] = Best_Cost;
1242 :    
1243 :     if (Best_Cost < Min_Cost + Dist0) {
1244 :     Min_Cost = Best_Cost;
1245 :     Run_Start = i;
1246 :     }
1247 :     else
1248 :     {
1249 : edgomez 1.21.2.13 /*
1250 :     * as noticed by Michael Niedermayer (michaelni at gmx.at), there's
1251 :     * a code shorter by 1 bit for a larger run (!), same level. We give
1252 :     * it a chance by not moving the left barrier too much.
1253 :     */
1254 : chl 1.21.2.11
1255 : chl 1.21.2.10 while( Run_Costs[Run_Start]>Min_Cost+(1<<16) )
1256 :     Run_Start++;
1257 :    
1258 : edgomez 1.21.2.13 /* spread on preceding coeffs the cost incurred by skipping this one */
1259 : chl 1.21.2.10 for(j=Run_Start; j<i; ++j) Run_Costs[j] += Dist0;
1260 :     Min_Cost += Dist0;
1261 :     }
1262 :     }
1263 :    
1264 :     if (DBG) {
1265 :     Last_Cost = Evaluate_Cost(Out,Mult,Bias, Zigzag,Non_Zero, Lambda);
1266 :     if (DBG==1) {
1267 :     printf( "=> " );
1268 :     for(i=0; i<=Non_Zero; ++i) printf( "[%3.0d] ", Out[Zigzag[i]] );
1269 :     printf( "\n" );
1270 :     }
1271 :     }
1272 :    
1273 :     if (Last_Node<0)
1274 :     return -1;
1275 :    
1276 : edgomez 1.21.2.13 /* reconstruct optimal sequence backward with surviving paths */
1277 : chl 1.21.2.11 memset(Out, 0x00, 64*sizeof(*Out));
1278 : chl 1.21.2.10 Out[Zigzag[Last_Node]] = Last.Level;
1279 :     i = Last_Node - Last.Run;
1280 :     while(i>=0) {
1281 :     Out[Zigzag[i]] = Nodes[i].Level;
1282 :     i -= Nodes[i].Run;
1283 :     }
1284 :    
1285 :     if (DBG) {
1286 :     uint32_t Cost = Evaluate_Cost(Out,Mult,Bias, Zigzag,Non_Zero, Lambda);
1287 :     if (DBG==1) {
1288 :     printf( "<= " );
1289 :     for(i=0; i<=Last_Node; ++i) printf( "[%3.0d] ", Out[Zigzag[i]] );
1290 :     printf( "\n--------------------------------\n" );
1291 :     }
1292 :     if (Cost>Last_Cost) printf( "!!! %u > %u\n", Cost, Last_Cost );
1293 :     }
1294 :     return Last_Node;
1295 :     }
1296 :    
1297 :     #undef DBG
1298 : chl 1.21.2.11
1299 :     #endif

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