[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 1.26 - (view) (download)

1 : edgomez 1.23 /*****************************************************************************
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 : syskin 1.26 * $Id: mbtransquant.c,v 1.25 2004/05/26 05:45:53 syskin Exp $
25 : edgomez 1.23 *
26 :     ****************************************************************************/
27 : Isibaar 1.1
28 : edgomez 1.23 #include <stdio.h>
29 :     #include <stdlib.h>
30 : edgomez 1.3 #include <string.h>
31 :    
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 : edgomez 1.23 #include "../bitstream/mbcoding.h"
39 :     #include "../bitstream/zigzag.h"
40 : Isibaar 1.1 #include "../dct/fdct.h"
41 :     #include "../dct/idct.h"
42 : edgomez 1.23 #include "../quant/quant.h"
43 : Isibaar 1.1 #include "../encoder.h"
44 :    
45 : edgomez 1.23 #include "../quant/quant_matrix.h"
46 : Isibaar 1.1
47 : edgomez 1.21 MBFIELDTEST_PTR MBFieldTest;
48 : Isibaar 1.1
49 : edgomez 1.23 /*
50 :     * Skip blocks having a coefficient sum below this value. This value will be
51 :     * corrected according to the MB quantizer to avoid artifacts for quant==1
52 :     */
53 :     #define PVOP_TOOSMALL_LIMIT 1
54 :     #define BVOP_TOOSMALL_LIMIT 3
55 :    
56 :     /*****************************************************************************
57 :     * Local functions
58 :     ****************************************************************************/
59 : Isibaar 1.1
60 : edgomez 1.23 /* permute block and return field dct choice */
61 :     static __inline uint32_t
62 :     MBDecideFieldDCT(int16_t data[6 * 64])
63 :     {
64 :     uint32_t field = MBFieldTest(data);
65 :    
66 :     if (field)
67 :     MBFrameToField(data);
68 :    
69 :     return field;
70 :     }
71 :    
72 :     /* Performs Forward DCT on all blocks */
73 : syskin 1.22 static __inline void
74 : edgomez 1.23 MBfDCT(const MBParam * const pParam,
75 :     const FRAMEINFO * const frame,
76 :     MACROBLOCK * const pMB,
77 :     uint32_t x_pos,
78 :     uint32_t y_pos,
79 :     int16_t data[6 * 64])
80 : syskin 1.22 {
81 : edgomez 1.23 /* Handles interlacing */
82 :     start_timer();
83 :     pMB->field_dct = 0;
84 :     if ((frame->vol_flags & XVID_VOL_INTERLACING) &&
85 :     (x_pos>0) && (x_pos<pParam->mb_width-1) &&
86 :     (y_pos>0) && (y_pos<pParam->mb_height-1)) {
87 :     pMB->field_dct = MBDecideFieldDCT(data);
88 :     }
89 :     stop_interlacing_timer();
90 :    
91 :     /* Perform DCT */
92 : syskin 1.22 start_timer();
93 :     fdct(&data[0 * 64]);
94 :     fdct(&data[1 * 64]);
95 :     fdct(&data[2 * 64]);
96 :     fdct(&data[3 * 64]);
97 :     fdct(&data[4 * 64]);
98 :     fdct(&data[5 * 64]);
99 :     stop_dct_timer();
100 :     }
101 :    
102 : edgomez 1.23 /* Performs Inverse DCT on all blocks */
103 :     static __inline void
104 :     MBiDCT(int16_t data[6 * 64],
105 :     const uint8_t cbp)
106 :     {
107 :     start_timer();
108 :     if(cbp & (1 << (5 - 0))) idct(&data[0 * 64]);
109 :     if(cbp & (1 << (5 - 1))) idct(&data[1 * 64]);
110 :     if(cbp & (1 << (5 - 2))) idct(&data[2 * 64]);
111 :     if(cbp & (1 << (5 - 3))) idct(&data[3 * 64]);
112 :     if(cbp & (1 << (5 - 4))) idct(&data[4 * 64]);
113 :     if(cbp & (1 << (5 - 5))) idct(&data[5 * 64]);
114 :     stop_idct_timer();
115 :     }
116 : syskin 1.22
117 : edgomez 1.23 /* Quantize all blocks -- Intra mode */
118 :     static __inline void
119 :     MBQuantIntra(const MBParam * pParam,
120 :     const FRAMEINFO * const frame,
121 :     const MACROBLOCK * pMB,
122 :     int16_t qcoeff[6 * 64],
123 :     int16_t data[6*64])
124 : syskin 1.22 {
125 : edgomez 1.23 int mpeg;
126 :     int scaler_lum, scaler_chr;
127 :    
128 :     quant_intraFuncPtr const quant[2] =
129 :     {
130 :     quant_h263_intra,
131 :     quant_mpeg_intra
132 :     };
133 :    
134 :     mpeg = !!(pParam->vol_flags & XVID_VOL_MPEGQUANT);
135 :     scaler_lum = get_dc_scaler(pMB->quant, 1);
136 :     scaler_chr = get_dc_scaler(pMB->quant, 0);
137 : syskin 1.22
138 : edgomez 1.23 /* Quantize the block */
139 : syskin 1.22 start_timer();
140 : edgomez 1.23 quant[mpeg](&data[0 * 64], &qcoeff[0 * 64], pMB->quant, scaler_lum, pParam->mpeg_quant_matrices);
141 :     quant[mpeg](&data[1 * 64], &qcoeff[1 * 64], pMB->quant, scaler_lum, pParam->mpeg_quant_matrices);
142 :     quant[mpeg](&data[2 * 64], &qcoeff[2 * 64], pMB->quant, scaler_lum, pParam->mpeg_quant_matrices);
143 :     quant[mpeg](&data[3 * 64], &qcoeff[3 * 64], pMB->quant, scaler_lum, pParam->mpeg_quant_matrices);
144 :     quant[mpeg](&data[4 * 64], &qcoeff[4 * 64], pMB->quant, scaler_chr, pParam->mpeg_quant_matrices);
145 :     quant[mpeg](&data[5 * 64], &qcoeff[5 * 64], pMB->quant, scaler_chr, pParam->mpeg_quant_matrices);
146 : syskin 1.22 stop_quant_timer();
147 :     }
148 :    
149 : edgomez 1.23 /* DeQuantize all blocks -- Intra mode */
150 :     static __inline void
151 :     MBDeQuantIntra(const MBParam * pParam,
152 :     const int iQuant,
153 :     int16_t qcoeff[6 * 64],
154 :     int16_t data[6*64])
155 : Isibaar 1.1 {
156 : edgomez 1.23 int mpeg;
157 :     int scaler_lum, scaler_chr;
158 : edgomez 1.3
159 : edgomez 1.23 quant_intraFuncPtr const dequant[2] =
160 :     {
161 :     dequant_h263_intra,
162 :     dequant_mpeg_intra
163 :     };
164 :    
165 :     mpeg = !!(pParam->vol_flags & XVID_VOL_MPEGQUANT);
166 :     scaler_lum = get_dc_scaler(iQuant, 1);
167 :     scaler_chr = get_dc_scaler(iQuant, 0);
168 : Isibaar 1.1
169 : h 1.2 start_timer();
170 : edgomez 1.23 dequant[mpeg](&qcoeff[0 * 64], &data[0 * 64], iQuant, scaler_lum, pParam->mpeg_quant_matrices);
171 :     dequant[mpeg](&qcoeff[1 * 64], &data[1 * 64], iQuant, scaler_lum, pParam->mpeg_quant_matrices);
172 :     dequant[mpeg](&qcoeff[2 * 64], &data[2 * 64], iQuant, scaler_lum, pParam->mpeg_quant_matrices);
173 :     dequant[mpeg](&qcoeff[3 * 64], &data[3 * 64], iQuant, scaler_lum, pParam->mpeg_quant_matrices);
174 :     dequant[mpeg](&qcoeff[4 * 64], &data[4 * 64], iQuant, scaler_chr, pParam->mpeg_quant_matrices);
175 :     dequant[mpeg](&qcoeff[5 * 64], &data[5 * 64], iQuant, scaler_chr, pParam->mpeg_quant_matrices);
176 :     stop_iquant_timer();
177 :     }
178 :    
179 :     static int
180 :     dct_quantize_trellis_c(int16_t *const Out,
181 :     const int16_t *const In,
182 :     int Q,
183 :     const uint16_t * const Zigzag,
184 :     const uint16_t * const QuantMatrix,
185 : edgomez 1.24 int Non_Zero,
186 :     int Sum);
187 : edgomez 1.23
188 :     /* Quantize all blocks -- Inter mode */
189 :     static __inline uint8_t
190 :     MBQuantInter(const MBParam * pParam,
191 :     const FRAMEINFO * const frame,
192 :     const MACROBLOCK * pMB,
193 :     int16_t data[6 * 64],
194 :     int16_t qcoeff[6 * 64],
195 :     int bvop,
196 :     int limit)
197 :     {
198 :    
199 :     int i;
200 :     uint8_t cbp = 0;
201 :     int sum;
202 :     int code_block, mpeg;
203 : h 1.2
204 : edgomez 1.23 quant_interFuncPtr const quant[2] =
205 :     {
206 :     quant_h263_inter,
207 :     quant_mpeg_inter
208 :     };
209 : h 1.2
210 : edgomez 1.23 mpeg = !!(pParam->vol_flags & XVID_VOL_MPEGQUANT);
211 : syskin 1.22
212 : edgomez 1.7 for (i = 0; i < 6; i++) {
213 : Isibaar 1.1
214 : edgomez 1.23 /* Quantize the block */
215 : Isibaar 1.1 start_timer();
216 : edgomez 1.23
217 :     sum = quant[mpeg](&qcoeff[i*64], &data[i*64], pMB->quant, pParam->mpeg_quant_matrices);
218 :    
219 :     if(sum && (frame->vop_flags & XVID_VOP_TRELLISQUANT)) {
220 :     const uint16_t *matrix;
221 :     const static uint16_t h263matrix[] =
222 :     {
223 :     16, 16, 16, 16, 16, 16, 16, 16,
224 :     16, 16, 16, 16, 16, 16, 16, 16,
225 :     16, 16, 16, 16, 16, 16, 16, 16,
226 :     16, 16, 16, 16, 16, 16, 16, 16,
227 :     16, 16, 16, 16, 16, 16, 16, 16,
228 :     16, 16, 16, 16, 16, 16, 16, 16,
229 :     16, 16, 16, 16, 16, 16, 16, 16,
230 :     16, 16, 16, 16, 16, 16, 16, 16
231 :     };
232 :    
233 :     matrix = (mpeg)?get_inter_matrix(pParam->mpeg_quant_matrices):h263matrix;
234 :     sum = dct_quantize_trellis_c(&qcoeff[i*64], &data[i*64],
235 :     pMB->quant, &scan_tables[0][0],
236 :     matrix,
237 : edgomez 1.24 63,
238 :     sum);
239 : edgomez 1.23 }
240 : syskin 1.22 stop_quant_timer();
241 : edgomez 1.21
242 : edgomez 1.23 /*
243 :     * We code the block if the sum is higher than the limit and if the first
244 :     * two AC coefficients in zig zag order are not zero.
245 :     */
246 :     code_block = 0;
247 :     if ((sum >= limit) || (qcoeff[i*64+1] != 0) || (qcoeff[i*64+8] != 0)) {
248 :     code_block = 1;
249 :     } else {
250 : Isibaar 1.1
251 : edgomez 1.23 if (bvop && (pMB->mode == MODE_DIRECT || pMB->mode == MODE_DIRECT_NO4V)) {
252 :     /* dark blocks prevention for direct mode */
253 :     if ((qcoeff[i*64] < -1) || (qcoeff[i*64] > 0))
254 :     code_block = 1;
255 :     } else {
256 :     /* not direct mode */
257 :     if (qcoeff[i*64] != 0)
258 :     code_block = 1;
259 :     }
260 : edgomez 1.21 }
261 : edgomez 1.23
262 :     /* Set the corresponding cbp bit */
263 :     cbp |= code_block << (5 - i);
264 : edgomez 1.21 }
265 :    
266 : edgomez 1.23 return(cbp);
267 :     }
268 :    
269 :     /* DeQuantize all blocks -- Inter mode */
270 :     static __inline void
271 :     MBDeQuantInter(const MBParam * pParam,
272 :     const int iQuant,
273 :     int16_t data[6 * 64],
274 :     int16_t qcoeff[6 * 64],
275 :     const uint8_t cbp)
276 :     {
277 :     int mpeg;
278 : edgomez 1.21
279 : edgomez 1.23 quant_interFuncPtr const dequant[2] =
280 :     {
281 :     dequant_h263_inter,
282 :     dequant_mpeg_inter
283 :     };
284 : Isibaar 1.1
285 : edgomez 1.23 mpeg = !!(pParam->vol_flags & XVID_VOL_MPEGQUANT);
286 : Isibaar 1.1
287 : edgomez 1.23 start_timer();
288 :     if(cbp & (1 << (5 - 0))) dequant[mpeg](&data[0 * 64], &qcoeff[0 * 64], iQuant, pParam->mpeg_quant_matrices);
289 :     if(cbp & (1 << (5 - 1))) dequant[mpeg](&data[1 * 64], &qcoeff[1 * 64], iQuant, pParam->mpeg_quant_matrices);
290 :     if(cbp & (1 << (5 - 2))) dequant[mpeg](&data[2 * 64], &qcoeff[2 * 64], iQuant, pParam->mpeg_quant_matrices);
291 :     if(cbp & (1 << (5 - 3))) dequant[mpeg](&data[3 * 64], &qcoeff[3 * 64], iQuant, pParam->mpeg_quant_matrices);
292 :     if(cbp & (1 << (5 - 4))) dequant[mpeg](&data[4 * 64], &qcoeff[4 * 64], iQuant, pParam->mpeg_quant_matrices);
293 :     if(cbp & (1 << (5 - 5))) dequant[mpeg](&data[5 * 64], &qcoeff[5 * 64], iQuant, pParam->mpeg_quant_matrices);
294 :     stop_iquant_timer();
295 : Isibaar 1.1 }
296 :    
297 : edgomez 1.23 typedef void (transfer_operation_8to16_t) (int16_t *Dst, const uint8_t *Src, int BpS);
298 :     typedef void (transfer_operation_16to8_t) (uint8_t *Dst, const int16_t *Src, int BpS);
299 :    
300 :    
301 :     static __inline void
302 :     MBTrans8to16(const MBParam * const pParam,
303 :     const FRAMEINFO * const frame,
304 :     const MACROBLOCK * const pMB,
305 :     const uint32_t x_pos,
306 :     const uint32_t y_pos,
307 :     int16_t data[6 * 64])
308 : Isibaar 1.1 {
309 : h 1.4 uint32_t stride = pParam->edged_width;
310 : edgomez 1.23 uint32_t stride2 = stride / 2;
311 :     uint32_t next_block = stride * 8;
312 : Isibaar 1.1 uint8_t *pY_Cur, *pU_Cur, *pV_Cur;
313 : syskin 1.22 const IMAGE * const pCurrent = &frame->image;
314 :    
315 : edgomez 1.23 /* Image pointers */
316 : syskin 1.26 pY_Cur = pCurrent->y + (y_pos << 4) * stride + (x_pos << 4);
317 :     pU_Cur = pCurrent->u + (y_pos << 3) * stride2 + (x_pos << 3);
318 :     pV_Cur = pCurrent->v + (y_pos << 3) * stride2 + (x_pos << 3);
319 : Isibaar 1.1
320 : edgomez 1.23 /* Do the transfer */
321 :     start_timer();
322 : syskin 1.26 transfer_8to16copy(&data[0 * 64], pY_Cur, stride);
323 :     transfer_8to16copy(&data[1 * 64], pY_Cur + 8, stride);
324 :     transfer_8to16copy(&data[2 * 64], pY_Cur + next_block, stride);
325 :     transfer_8to16copy(&data[3 * 64], pY_Cur + next_block + 8, stride);
326 :     transfer_8to16copy(&data[4 * 64], pU_Cur, stride2);
327 :     transfer_8to16copy(&data[5 * 64], pV_Cur, stride2);
328 : edgomez 1.23 stop_transfer_timer();
329 :     }
330 : Isibaar 1.1
331 : edgomez 1.23 static __inline void
332 :     MBTrans16to8(const MBParam * const pParam,
333 :     const FRAMEINFO * const frame,
334 :     const MACROBLOCK * const pMB,
335 :     const uint32_t x_pos,
336 :     const uint32_t y_pos,
337 :     int16_t data[6 * 64],
338 :     const uint32_t add, /* Must be 1 or 0 */
339 :     const uint8_t cbp)
340 :     {
341 :     uint8_t *pY_Cur, *pU_Cur, *pV_Cur;
342 :     uint32_t stride = pParam->edged_width;
343 :     uint32_t stride2 = stride / 2;
344 :     uint32_t next_block = stride * 8;
345 :     const IMAGE * const pCurrent = &frame->image;
346 : Isibaar 1.1
347 : syskin 1.26 /* Array of function pointers, indexed by [add] */
348 :     transfer_operation_16to8_t * const functions[2] =
349 : edgomez 1.23 {
350 :     (transfer_operation_16to8_t*)transfer_16to8copy,
351 :     (transfer_operation_16to8_t*)transfer_16to8add,
352 :     };
353 :    
354 :     transfer_operation_16to8_t *transfer_op = NULL;
355 :    
356 :     /* Image pointers */
357 : syskin 1.26 pY_Cur = pCurrent->y + (y_pos << 4) * stride + (x_pos << 4);
358 :     pU_Cur = pCurrent->u + (y_pos << 3) * stride2 + (x_pos << 3);
359 :     pV_Cur = pCurrent->v + (y_pos << 3) * stride2 + (x_pos << 3);
360 : h 1.2
361 : edgomez 1.7 if (pMB->field_dct) {
362 : h 1.4 next_block = stride;
363 :     stride *= 2;
364 : h 1.2 }
365 :    
366 : edgomez 1.23 /* Operation function */
367 : syskin 1.26 transfer_op = functions[add];
368 : edgomez 1.23
369 :     /* Do the operation */
370 : h 1.2 start_timer();
371 : syskin 1.26 if (cbp&32) transfer_op(pY_Cur, &data[0 * 64], stride);
372 :     if (cbp&16) transfer_op(pY_Cur + 8, &data[1 * 64], stride);
373 :     if (cbp& 8) transfer_op(pY_Cur + next_block, &data[2 * 64], stride);
374 :     if (cbp& 4) transfer_op(pY_Cur + next_block + 8, &data[3 * 64], stride);
375 :     if (cbp& 2) transfer_op(pU_Cur, &data[4 * 64], stride2);
376 :     if (cbp& 1) transfer_op(pV_Cur, &data[5 * 64], stride2);
377 : h 1.2 stop_transfer_timer();
378 : edgomez 1.23 }
379 :    
380 :     /*****************************************************************************
381 :     * Module functions
382 :     ****************************************************************************/
383 :    
384 :     void
385 :     MBTransQuantIntra(const MBParam * const pParam,
386 :     const FRAMEINFO * const frame,
387 :     MACROBLOCK * const pMB,
388 :     const uint32_t x_pos,
389 :     const uint32_t y_pos,
390 :     int16_t data[6 * 64],
391 :     int16_t qcoeff[6 * 64])
392 :     {
393 :    
394 :     /* Transfer data */
395 :     MBTrans8to16(pParam, frame, pMB, x_pos, y_pos, data);
396 :    
397 :     /* Perform DCT (and field decision) */
398 :     MBfDCT(pParam, frame, pMB, x_pos, y_pos, data);
399 :    
400 :     /* Quantize the block */
401 :     MBQuantIntra(pParam, frame, pMB, data, qcoeff);
402 : h 1.2
403 : edgomez 1.23 /* DeQuantize the block */
404 :     MBDeQuantIntra(pParam, pMB->quant, data, qcoeff);
405 :    
406 :     /* Perform inverse DCT*/
407 :     MBiDCT(data, 0x3F);
408 :    
409 :     /* Transfer back the data -- Don't add data */
410 :     MBTrans16to8(pParam, frame, pMB, x_pos, y_pos, data, 0, 0x3F);
411 : chl 1.8 }
412 :    
413 : edgomez 1.23
414 : chl 1.8 uint8_t
415 : edgomez 1.23 MBTransQuantInter(const MBParam * const pParam,
416 :     const FRAMEINFO * const frame,
417 :     MACROBLOCK * const pMB,
418 :     const uint32_t x_pos,
419 :     const uint32_t y_pos,
420 : chl 1.8 int16_t data[6 * 64],
421 :     int16_t qcoeff[6 * 64])
422 :     {
423 : edgomez 1.23 uint8_t cbp;
424 :     uint32_t limit;
425 :    
426 :     /* There is no MBTrans8to16 for Inter block, that's done in motion compensation
427 :     * already */
428 :    
429 :     /* Perform DCT (and field decision) */
430 :     MBfDCT(pParam, frame, pMB, x_pos, y_pos, data);
431 :    
432 :     /* Set the limit threshold */
433 :     limit = PVOP_TOOSMALL_LIMIT + ((pMB->quant == 1)? 1 : 0);
434 :    
435 :     if (frame->vop_flags & XVID_VOP_CARTOON)
436 :     limit *= 3;
437 :    
438 :     /* Quantize the block */
439 :     cbp = MBQuantInter(pParam, frame, pMB, data, qcoeff, 0, limit);
440 : chl 1.8
441 : edgomez 1.23 /* DeQuantize the block */
442 :     MBDeQuantInter(pParam, pMB->quant, data, qcoeff, cbp);
443 : chl 1.8
444 : edgomez 1.23 /* Perform inverse DCT*/
445 :     MBiDCT(data, cbp);
446 : chl 1.8
447 : edgomez 1.23 /* Transfer back the data -- Add the data */
448 :     MBTrans16to8(pParam, frame, pMB, x_pos, y_pos, data, 1, cbp);
449 : chl 1.8
450 : edgomez 1.23 return(cbp);
451 : chl 1.8 }
452 :    
453 : edgomez 1.23 uint8_t
454 :     MBTransQuantInterBVOP(const MBParam * pParam,
455 :     FRAMEINFO * frame,
456 :     MACROBLOCK * pMB,
457 :     const uint32_t x_pos,
458 :     const uint32_t y_pos,
459 :     int16_t data[6 * 64],
460 :     int16_t qcoeff[6 * 64])
461 :     {
462 :     uint8_t cbp;
463 :     uint32_t limit;
464 :    
465 :     /* There is no MBTrans8to16 for Inter block, that's done in motion compensation
466 :     * already */
467 :    
468 :     /* Perform DCT (and field decision) */
469 :     MBfDCT(pParam, frame, pMB, x_pos, y_pos, data);
470 :    
471 :     /* Set the limit threshold */
472 :     limit = BVOP_TOOSMALL_LIMIT;
473 :    
474 :     if (frame->vop_flags & XVID_VOP_CARTOON)
475 :     limit *= 2;
476 :    
477 :     /* Quantize the block */
478 :     cbp = MBQuantInter(pParam, frame, pMB, data, qcoeff, 1, limit);
479 :    
480 :     /*
481 :     * History comment:
482 :     * We don't have to DeQuant, iDCT and Transfer back data for B-frames.
483 :     *
484 :     * BUT some plugins require the rebuilt original frame to be passed so we
485 :     * have to take care of that here
486 :     */
487 :     if((pParam->plugin_flags & XVID_REQORIGINAL)) {
488 :    
489 :     /* DeQuantize the block */
490 :     MBDeQuantInter(pParam, pMB->quant, data, qcoeff, cbp);
491 :    
492 :     /* Perform inverse DCT*/
493 :     MBiDCT(data, cbp);
494 : h 1.2
495 : edgomez 1.23 /* Transfer back the data -- Add the data */
496 :     MBTrans16to8(pParam, frame, pMB, x_pos, y_pos, data, 1, cbp);
497 :     }
498 : edgomez 1.21
499 : edgomez 1.23 return(cbp);
500 : edgomez 1.21 }
501 : edgomez 1.3
502 : edgomez 1.21 /* if sum(diff between field lines) < sum(diff between frame lines), use field dct */
503 :     uint32_t
504 :     MBFieldTest_c(int16_t data[6 * 64])
505 :     {
506 : edgomez 1.7 const uint8_t blocks[] =
507 :     { 0 * 64, 0 * 64, 0 * 64, 0 * 64, 2 * 64, 2 * 64, 2 * 64, 2 * 64 };
508 :     const uint8_t lines[] = { 0, 16, 32, 48, 0, 16, 32, 48 };
509 : h 1.2
510 :     int frame = 0, field = 0;
511 :     int i, j;
512 :    
513 : edgomez 1.7 for (i = 0; i < 7; ++i) {
514 :     for (j = 0; j < 8; ++j) {
515 :     frame +=
516 : edgomez 1.23 abs(data[0 * 64 + (i + 1) * 8 + j] - data[0 * 64 + i * 8 + j]);
517 : edgomez 1.7 frame +=
518 : edgomez 1.23 abs(data[1 * 64 + (i + 1) * 8 + j] - data[1 * 64 + i * 8 + j]);
519 : edgomez 1.7 frame +=
520 : edgomez 1.23 abs(data[2 * 64 + (i + 1) * 8 + j] - data[2 * 64 + i * 8 + j]);
521 : edgomez 1.7 frame +=
522 : edgomez 1.23 abs(data[3 * 64 + (i + 1) * 8 + j] - data[3 * 64 + i * 8 + j]);
523 : edgomez 1.7
524 :     field +=
525 : edgomez 1.23 abs(data[blocks[i + 1] + lines[i + 1] + j] -
526 : edgomez 1.7 data[blocks[i] + lines[i] + j]);
527 :     field +=
528 : edgomez 1.23 abs(data[blocks[i + 1] + lines[i + 1] + 8 + j] -
529 : edgomez 1.7 data[blocks[i] + lines[i] + 8 + j]);
530 :     field +=
531 : edgomez 1.23 abs(data[blocks[i + 1] + 64 + lines[i + 1] + j] -
532 : edgomez 1.7 data[blocks[i] + 64 + lines[i] + j]);
533 :     field +=
534 : edgomez 1.23 abs(data[blocks[i + 1] + 64 + lines[i + 1] + 8 + j] -
535 : edgomez 1.7 data[blocks[i] + 64 + lines[i] + 8 + j]);
536 : Isibaar 1.1 }
537 :     }
538 : h 1.2
539 : edgomez 1.21 return (frame >= (field + 350));
540 : h 1.2 }
541 :    
542 :    
543 :     /* deinterlace Y blocks vertically */
544 :    
545 :     #define MOVLINE(X,Y) memcpy(X, Y, sizeof(tmp))
546 : edgomez 1.23 #define LINE(X,Y) &data[X*64 + Y*8]
547 : h 1.2
548 : edgomez 1.7 void
549 :     MBFrameToField(int16_t data[6 * 64])
550 : h 1.2 {
551 :     int16_t tmp[8];
552 :    
553 :     /* left blocks */
554 :    
555 : edgomez 1.23 /* 1=2, 2=4, 4=8, 8=1 */
556 : edgomez 1.7 MOVLINE(tmp, LINE(0, 1));
557 :     MOVLINE(LINE(0, 1), LINE(0, 2));
558 :     MOVLINE(LINE(0, 2), LINE(0, 4));
559 :     MOVLINE(LINE(0, 4), LINE(2, 0));
560 :     MOVLINE(LINE(2, 0), tmp);
561 : h 1.2
562 : edgomez 1.23 /* 3=6, 6=12, 12=9, 9=3 */
563 : edgomez 1.7 MOVLINE(tmp, LINE(0, 3));
564 :     MOVLINE(LINE(0, 3), LINE(0, 6));
565 :     MOVLINE(LINE(0, 6), LINE(2, 4));
566 :     MOVLINE(LINE(2, 4), LINE(2, 1));
567 :     MOVLINE(LINE(2, 1), tmp);
568 : h 1.2
569 : edgomez 1.23 /* 5=10, 10=5 */
570 : edgomez 1.7 MOVLINE(tmp, LINE(0, 5));
571 :     MOVLINE(LINE(0, 5), LINE(2, 2));
572 :     MOVLINE(LINE(2, 2), tmp);
573 : h 1.2
574 : edgomez 1.23 /* 7=14, 14=13, 13=11, 11=7 */
575 : edgomez 1.7 MOVLINE(tmp, LINE(0, 7));
576 :     MOVLINE(LINE(0, 7), LINE(2, 6));
577 :     MOVLINE(LINE(2, 6), LINE(2, 5));
578 :     MOVLINE(LINE(2, 5), LINE(2, 3));
579 :     MOVLINE(LINE(2, 3), tmp);
580 : h 1.2
581 :     /* right blocks */
582 :    
583 : edgomez 1.23 /* 1=2, 2=4, 4=8, 8=1 */
584 : edgomez 1.7 MOVLINE(tmp, LINE(1, 1));
585 :     MOVLINE(LINE(1, 1), LINE(1, 2));
586 :     MOVLINE(LINE(1, 2), LINE(1, 4));
587 :     MOVLINE(LINE(1, 4), LINE(3, 0));
588 :     MOVLINE(LINE(3, 0), tmp);
589 : h 1.2
590 : edgomez 1.23 /* 3=6, 6=12, 12=9, 9=3 */
591 : edgomez 1.7 MOVLINE(tmp, LINE(1, 3));
592 :     MOVLINE(LINE(1, 3), LINE(1, 6));
593 :     MOVLINE(LINE(1, 6), LINE(3, 4));
594 :     MOVLINE(LINE(3, 4), LINE(3, 1));
595 :     MOVLINE(LINE(3, 1), tmp);
596 : h 1.2
597 : edgomez 1.23 /* 5=10, 10=5 */
598 : edgomez 1.7 MOVLINE(tmp, LINE(1, 5));
599 :     MOVLINE(LINE(1, 5), LINE(3, 2));
600 :     MOVLINE(LINE(3, 2), tmp);
601 : h 1.2
602 : edgomez 1.23 /* 7=14, 14=13, 13=11, 11=7 */
603 : edgomez 1.7 MOVLINE(tmp, LINE(1, 7));
604 :     MOVLINE(LINE(1, 7), LINE(3, 6));
605 :     MOVLINE(LINE(3, 6), LINE(3, 5));
606 :     MOVLINE(LINE(3, 5), LINE(3, 3));
607 :     MOVLINE(LINE(3, 3), tmp);
608 : Isibaar 1.1 }
609 : edgomez 1.23
610 :     /*****************************************************************************
611 :     * Trellis based R-D optimal quantization
612 :     *
613 :     * Trellis Quant code (C) 2003 Pascal Massimino skal(at)planet-d.net
614 :     *
615 :     ****************************************************************************/
616 :    
617 :     /*----------------------------------------------------------------------------
618 :     *
619 :     * Trellis-Based quantization
620 :     *
621 :     * So far I understand this paper:
622 :     *
623 :     * "Trellis-Based R-D Optimal Quantization in H.263+"
624 :     * J.Wen, M.Luttrell, J.Villasenor
625 :     * IEEE Transactions on Image Processing, Vol.9, No.8, Aug. 2000.
626 :     *
627 :     * we are at stake with a simplified Bellmand-Ford / Dijkstra Single
628 :     * Source Shortest Path algo. But due to the underlying graph structure
629 :     * ("Trellis"), it can be turned into a dynamic programming algo,
630 :     * partially saving the explicit graph's nodes representation. And
631 :     * without using a heap, since the open frontier of the DAG is always
632 :     * known, and of fixed size.
633 :     *--------------------------------------------------------------------------*/
634 :    
635 :    
636 :    
637 :     /* Codes lengths for relevant levels. */
638 :    
639 :     /* let's factorize: */
640 :     static const uint8_t Code_Len0[64] = {
641 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,
642 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
643 :     static const uint8_t Code_Len1[64] = {
644 :     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,
645 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 };
646 :     static const uint8_t Code_Len2[64] = {
647 :     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,
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_Len3[64] = {
650 :     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,
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_Len4[64] = {
653 :     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,
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_Len5[64] = {
656 :     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,
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_Len6[64] = {
659 :     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,
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_Len7[64] = {
662 :     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,
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_Len8[64] = {
665 :     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,
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_Len9[64] = {
668 :     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,
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_Len10[64] = {
671 :     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,
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_Len11[64] = {
674 :     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,
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_Len12[64] = {
677 :     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,
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_Len13[64] = {
680 :     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,
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_Len14[64] = {
683 :     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,
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_Len15[64] = {
686 :     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,
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_Len16[64] = {
689 :     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,
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_Len17[64] = {
692 :     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,
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_Len18[64] = {
695 :     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,
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_Len19[64] = {
698 :     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,
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_Len20[64] = {
701 :     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,
702 :     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 };
703 :    
704 :     /* a few more table for LAST table: */
705 :     static const uint8_t Code_Len21[64] = {
706 :     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,
707 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30};
708 :     static const uint8_t Code_Len22[64] = {
709 :     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,
710 :     30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30};
711 :     static const uint8_t Code_Len23[64] = {
712 :     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,
713 :     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};
714 :     static const uint8_t Code_Len24[64] = {
715 :     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,
716 :     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};
717 :    
718 :    
719 :     static const uint8_t * const B16_17_Code_Len[24] = { /* levels [1..24] */
720 :     Code_Len20,Code_Len19,Code_Len18,Code_Len17,
721 :     Code_Len16,Code_Len15,Code_Len14,Code_Len13,
722 :     Code_Len12,Code_Len11,Code_Len10,Code_Len9,
723 :     Code_Len8, Code_Len7 ,Code_Len6 ,Code_Len5,
724 :     Code_Len4, Code_Len3, Code_Len3 ,Code_Len2,
725 :     Code_Len2, Code_Len1, Code_Len1, Code_Len1,
726 :     };
727 :    
728 :     static const uint8_t * const B16_17_Code_Len_Last[6] = { /* levels [1..6] */
729 :     Code_Len24,Code_Len23,Code_Len22,Code_Len21, Code_Len3, Code_Len1,
730 :     };
731 :    
732 :     /* TL_SHIFT controls the precision of the RD optimizations in trellis
733 :     * valid range is [10..16]. The bigger, the more trellis is vulnerable
734 :     * to overflows in cost formulas.
735 :     * - 10 allows ac values up to 2^11 == 2048
736 :     * - 16 allows ac values up to 2^8 == 256
737 :     */
738 :     #define TL_SHIFT 11
739 :     #define TL(q) ((0xfe00>>(16-TL_SHIFT))/(q*q))
740 :    
741 :     static const int Trellis_Lambda_Tabs[31] = {
742 :     TL( 1),TL( 2),TL( 3),TL( 4),TL( 5),TL( 6), TL( 7),
743 :     TL( 8),TL( 9),TL(10),TL(11),TL(12),TL(13),TL(14), TL(15),
744 :     TL(16),TL(17),TL(18),TL(19),TL(20),TL(21),TL(22), TL(23),
745 :     TL(24),TL(25),TL(26),TL(27),TL(28),TL(29),TL(30), TL(31)
746 :     };
747 :     #undef TL
748 :    
749 :     static int __inline
750 :     Find_Last(const int16_t *C, const uint16_t *Zigzag, int i)
751 :     {
752 :     while(i>=0)
753 :     if (C[Zigzag[i]])
754 :     return i;
755 :     else i--;
756 :     return -1;
757 :     }
758 :    
759 :     /* this routine has been strippen of all debug code */
760 :     static int
761 :     dct_quantize_trellis_c(int16_t *const Out,
762 :     const int16_t *const In,
763 :     int Q,
764 :     const uint16_t * const Zigzag,
765 :     const uint16_t * const QuantMatrix,
766 : edgomez 1.24 int Non_Zero,
767 :     int Sum)
768 : edgomez 1.23 {
769 :    
770 :     /* Note: We should search last non-zero coeffs on *real* DCT input coeffs
771 :     * (In[]), not quantized one (Out[]). However, it only improves the result
772 :     * *very* slightly (~0.01dB), whereas speed drops to crawling level :)
773 :     * Well, actually, taking 1 more coeff past Non_Zero into account sometimes
774 :     * helps. */
775 :     typedef struct { int16_t Run, Level; } NODE;
776 :    
777 :     NODE Nodes[65], Last;
778 :     uint32_t Run_Costs0[64+1];
779 :     uint32_t * const Run_Costs = Run_Costs0 + 1;
780 :    
781 :     /* it's 1/lambda, actually */
782 :     const int Lambda = Trellis_Lambda_Tabs[Q-1];
783 :    
784 :     int Run_Start = -1;
785 :     uint32_t Min_Cost = 2<<TL_SHIFT;
786 :    
787 :     int Last_Node = -1;
788 :     uint32_t Last_Cost = 0;
789 :    
790 : edgomez 1.24 int i, j;
791 : edgomez 1.23
792 :     /* source (w/ CBP penalty) */
793 :     Run_Costs[-1] = 2<<TL_SHIFT;
794 :    
795 :     Non_Zero = Find_Last(Out, Zigzag, Non_Zero);
796 :     if (Non_Zero<0)
797 :     return 0; /* Sum is zero if there are only zero coeffs */
798 :    
799 :     for(i=0; i<=Non_Zero; i++) {
800 :     const int q = ((Q*QuantMatrix[Zigzag[i]])>>4);
801 :     const int Mult = 2*q;
802 :     const int Bias = (q-1) | 1;
803 :     const int Lev0 = Mult + Bias;
804 :    
805 :     const int AC = In[Zigzag[i]];
806 :     const int Level1 = Out[Zigzag[i]];
807 :     const unsigned int Dist0 = Lambda* AC*AC;
808 :     uint32_t Best_Cost = 0xf0000000;
809 :     Last_Cost += Dist0;
810 :    
811 :     /* very specialized loop for -1,0,+1 */
812 :     if ((uint32_t)(Level1+1)<3) {
813 :     int dQ;
814 :     int Run;
815 :     uint32_t Cost0;
816 :    
817 :     if (AC<0) {
818 :     Nodes[i].Level = -1;
819 :     dQ = Lev0 + AC;
820 :     } else {
821 :     Nodes[i].Level = 1;
822 :     dQ = Lev0 - AC;
823 :     }
824 :     Cost0 = Lambda*dQ*dQ;
825 :    
826 :     Nodes[i].Run = 1;
827 :     Best_Cost = (Code_Len20[0]<<TL_SHIFT) + Run_Costs[i-1]+Cost0;
828 :     for(Run=i-Run_Start; Run>0; --Run) {
829 :     const uint32_t Cost_Base = Cost0 + Run_Costs[i-Run];
830 :     const uint32_t Cost = Cost_Base + (Code_Len20[Run-1]<<TL_SHIFT);
831 :     const uint32_t lCost = Cost_Base + (Code_Len24[Run-1]<<TL_SHIFT);
832 :    
833 :     /* TODO: what about tie-breaks? Should we favor short runs or
834 :     * long runs? Although the error is the same, it would not be
835 :     * spread the same way along high and low frequencies... */
836 :    
837 :     /* Gruel: I'd say, favour short runs => hifreq errors (HVS) */
838 :    
839 :     if (Cost<Best_Cost) {
840 :     Best_Cost = Cost;
841 :     Nodes[i].Run = Run;
842 :     }
843 :    
844 :     if (lCost<Last_Cost) {
845 :     Last_Cost = lCost;
846 :     Last.Run = Run;
847 :     Last_Node = i;
848 :     }
849 :     }
850 :     if (Last_Node==i)
851 :     Last.Level = Nodes[i].Level;
852 :     } else if (51U>(uint32_t)(Level1+25)) {
853 :     /* "big" levels (not less than ESC3, though) */
854 :     const uint8_t *Tbl_L1, *Tbl_L2, *Tbl_L1_Last, *Tbl_L2_Last;
855 :     int Level2;
856 :     int dQ1, dQ2;
857 :     int Run;
858 :     uint32_t Dist1,Dist2;
859 :     int dDist21;
860 :    
861 :     if (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 :     } else { /* Level1<-1 */
870 :     dQ1 = Level1*Mult-AC - Bias;
871 :     dQ2 = dQ1 + Mult;
872 :     Level2 = Level1 + 1;
873 :     Tbl_L1 = (Level1>=-24) ? B16_17_Code_Len[Level1^-1] : Code_Len0;
874 :     Tbl_L2 = (Level2>=-24) ? B16_17_Code_Len[Level2^-1] : Code_Len0;
875 :     Tbl_L1_Last = (Level1>=- 6) ? B16_17_Code_Len_Last[Level1^-1] : Code_Len0;
876 :     Tbl_L2_Last = (Level2>=- 6) ? B16_17_Code_Len_Last[Level2^-1] : Code_Len0;
877 :     }
878 :    
879 :     Dist1 = Lambda*dQ1*dQ1;
880 :     Dist2 = Lambda*dQ2*dQ2;
881 :     dDist21 = Dist2-Dist1;
882 :    
883 :     for(Run=i-Run_Start; Run>0; --Run)
884 :     {
885 :     const uint32_t Cost_Base = Dist1 + Run_Costs[i-Run];
886 :     uint32_t Cost1, Cost2;
887 :     int bLevel;
888 :    
889 :     /* for sub-optimal (but slightly worth it, speed-wise) search,
890 :     * uncomment the following:
891 :     * if (Cost_Base>=Best_Cost) continue;
892 :     * (? doesn't seem to have any effect -- gruel ) */
893 :    
894 :     Cost1 = Cost_Base + (Tbl_L1[Run-1]<<TL_SHIFT);
895 :     Cost2 = Cost_Base + (Tbl_L2[Run-1]<<TL_SHIFT) + dDist21;
896 :    
897 :     if (Cost2<Cost1) {
898 :     Cost1 = Cost2;
899 :     bLevel = Level2;
900 :     } else {
901 :     bLevel = Level1;
902 :     }
903 :    
904 :     if (Cost1<Best_Cost) {
905 :     Best_Cost = Cost1;
906 :     Nodes[i].Run = Run;
907 :     Nodes[i].Level = bLevel;
908 :     }
909 :    
910 :     Cost1 = Cost_Base + (Tbl_L1_Last[Run-1]<<TL_SHIFT);
911 :     Cost2 = Cost_Base + (Tbl_L2_Last[Run-1]<<TL_SHIFT) + dDist21;
912 :    
913 :     if (Cost2<Cost1) {
914 :     Cost1 = Cost2;
915 :     bLevel = Level2;
916 :     } else {
917 :     bLevel = Level1;
918 :     }
919 :    
920 :     if (Cost1<Last_Cost) {
921 :     Last_Cost = Cost1;
922 :     Last.Run = Run;
923 :     Last.Level = bLevel;
924 :     Last_Node = i;
925 :     }
926 :     } /* end of "for Run" */
927 :     } else {
928 :     /* Very very high levels, with no chance of being optimizable
929 :     * => Simply pick best Run. */
930 :     int Run;
931 :     for(Run=i-Run_Start; Run>0; --Run) {
932 :     /* 30 bits + no distortion */
933 :     const uint32_t Cost = (30<<TL_SHIFT) + Run_Costs[i-Run];
934 :     if (Cost<Best_Cost) {
935 :     Best_Cost = Cost;
936 :     Nodes[i].Run = Run;
937 :     Nodes[i].Level = Level1;
938 :     }
939 :    
940 :     if (Cost<Last_Cost) {
941 :     Last_Cost = Cost;
942 :     Last.Run = Run;
943 :     Last.Level = Level1;
944 :     Last_Node = i;
945 :     }
946 :     }
947 :     }
948 :    
949 :    
950 :     Run_Costs[i] = Best_Cost;
951 :    
952 :     if (Best_Cost < Min_Cost + Dist0) {
953 :     Min_Cost = Best_Cost;
954 :     Run_Start = i;
955 :     } else {
956 :     /* as noticed by Michael Niedermayer (michaelni at gmx.at),
957 :     * there's a code shorter by 1 bit for a larger run (!), same
958 :     * level. We give it a chance by not moving the left barrier too
959 :     * much. */
960 :     while( Run_Costs[Run_Start]>Min_Cost+(1<<TL_SHIFT) )
961 :     Run_Start++;
962 :    
963 :     /* spread on preceding coeffs the cost incurred by skipping this
964 :     * one */
965 :     for(j=Run_Start; j<i; ++j) Run_Costs[j] += Dist0;
966 :     Min_Cost += Dist0;
967 :     }
968 :     }
969 :    
970 : edgomez 1.24 /* It seems trellis doesn't give good results... just leave the block untouched
971 :     * and return the original sum value */
972 : edgomez 1.23 if (Last_Node<0)
973 : edgomez 1.24 return Sum;
974 : edgomez 1.23
975 :     /* reconstruct optimal sequence backward with surviving paths */
976 :     memset(Out, 0x00, 64*sizeof(*Out));
977 :     Out[Zigzag[Last_Node]] = Last.Level;
978 :     i = Last_Node - Last.Run;
979 : syskin 1.25 Sum = abs(Last.Level);
980 : edgomez 1.23 while(i>=0) {
981 :     Out[Zigzag[i]] = Nodes[i].Level;
982 : edgomez 1.24 Sum += abs(Nodes[i].Level);
983 : edgomez 1.23 i -= Nodes[i].Run;
984 :     }
985 :    
986 : edgomez 1.24 return Sum;
987 : edgomez 1.23 }
988 :    
989 :     /* original version including heavy debugging info */
990 :    
991 :     #ifdef DBGTRELL
992 :    
993 :     #define DBG 0
994 :    
995 :     static __inline uint32_t Evaluate_Cost(const int16_t *C, int Mult, int Bias,
996 :     const uint16_t * Zigzag, int Max, int Lambda)
997 :     {
998 :     #if (DBG>0)
999 :     const int16_t * const Ref = C + 6*64;
1000 :     int Last = Max;
1001 :     int Bits = 0;
1002 :     int Dist = 0;
1003 :     int i;
1004 :     uint32_t Cost;
1005 :    
1006 :     while(Last>=0 && C[Zigzag[Last]]==0)
1007 :     Last--;
1008 :    
1009 :     if (Last>=0) {
1010 :     int j=0, j0=0;
1011 :     int Run, Level;
1012 :    
1013 :     Bits = 2; /* CBP */
1014 :     while(j<Last) {
1015 :     while(!C[Zigzag[j]])
1016 :     j++;
1017 :     if (j==Last)
1018 :     break;
1019 :     Level=C[Zigzag[j]];
1020 :     Run = j - j0;
1021 :     j0 = ++j;
1022 :     if (Level>=-24 && Level<=24)
1023 :     Bits += B16_17_Code_Len[(Level<0) ? -Level-1 : Level-1][Run];
1024 :     else
1025 :     Bits += 30;
1026 :     }
1027 :     Level = C[Zigzag[Last]];
1028 :     Run = j - j0;
1029 :     if (Level>=-6 && Level<=6)
1030 :     Bits += B16_17_Code_Len_Last[(Level<0) ? -Level-1 : Level-1][Run];
1031 :     else
1032 :     Bits += 30;
1033 :     }
1034 :    
1035 :     for(i=0; i<=Last; ++i) {
1036 :     int V = C[Zigzag[i]]*Mult;
1037 :     if (V>0)
1038 :     V += Bias;
1039 :     else
1040 :     if (V<0)
1041 :     V -= Bias;
1042 :     V -= Ref[Zigzag[i]];
1043 :     Dist += V*V;
1044 :     }
1045 :     Cost = Lambda*Dist + (Bits<<TL_SHIFT);
1046 :     if (DBG==1)
1047 :     printf( " Last:%2d/%2d Cost = [(Bits=%5.0d) + Lambda*(Dist=%6.0d) = %d ] >>12= %d ", Last,Max, Bits, Dist, Cost, Cost>>12 );
1048 :     return Cost;
1049 :    
1050 :     #else
1051 :     return 0;
1052 :     #endif
1053 :     }
1054 :    
1055 :    
1056 :     static int
1057 :     dct_quantize_trellis_h263_c(int16_t *const Out, const int16_t *const In, int Q, const uint16_t * const Zigzag, int Non_Zero)
1058 :     {
1059 :    
1060 :     /*
1061 :     * Note: We should search last non-zero coeffs on *real* DCT input coeffs (In[]),
1062 :     * not quantized one (Out[]). However, it only improves the result *very*
1063 :     * slightly (~0.01dB), whereas speed drops to crawling level :)
1064 :     * Well, actually, taking 1 more coeff past Non_Zero into account sometimes helps.
1065 :     */
1066 :     typedef struct { int16_t Run, Level; } NODE;
1067 :    
1068 :     NODE Nodes[65], Last;
1069 :     uint32_t Run_Costs0[64+1];
1070 :     uint32_t * const Run_Costs = Run_Costs0 + 1;
1071 :     const int Mult = 2*Q;
1072 :     const int Bias = (Q-1) | 1;
1073 :     const int Lev0 = Mult + Bias;
1074 :     const int Lambda = Trellis_Lambda_Tabs[Q-1]; /* it's 1/lambda, actually */
1075 :    
1076 :     int Run_Start = -1;
1077 :     Run_Costs[-1] = 2<<TL_SHIFT; /* source (w/ CBP penalty) */
1078 :     uint32_t Min_Cost = 2<<TL_SHIFT;
1079 :    
1080 :     int Last_Node = -1;
1081 :     uint32_t Last_Cost = 0;
1082 :    
1083 :     int i, j;
1084 :    
1085 :     #if (DBG>0)
1086 :     Last.Level = 0; Last.Run = -1; /* just initialize to smthg */
1087 :     #endif
1088 :    
1089 :     Non_Zero = Find_Last(Out, Zigzag, Non_Zero);
1090 :     if (Non_Zero<0)
1091 :     return -1;
1092 :    
1093 :     for(i=0; i<=Non_Zero; i++)
1094 :     {
1095 :     const int AC = In[Zigzag[i]];
1096 :     const int Level1 = Out[Zigzag[i]];
1097 :     const int Dist0 = Lambda* AC*AC;
1098 :     uint32_t Best_Cost = 0xf0000000;
1099 :     Last_Cost += Dist0;
1100 :    
1101 :     if ((uint32_t)(Level1+1)<3) /* very specialized loop for -1,0,+1 */
1102 :     {
1103 :     int dQ;
1104 :     int Run;
1105 :     uint32_t Cost0;
1106 :    
1107 :     if (AC<0) {
1108 :     Nodes[i].Level = -1;
1109 :     dQ = Lev0 + AC;
1110 :     } else {
1111 :     Nodes[i].Level = 1;
1112 :     dQ = Lev0 - AC;
1113 :     }
1114 :     Cost0 = Lambda*dQ*dQ;
1115 :    
1116 :     Nodes[i].Run = 1;
1117 :     Best_Cost = (Code_Len20[0]<<TL_SHIFT) + Run_Costs[i-1]+Cost0;
1118 :     for(Run=i-Run_Start; Run>0; --Run)
1119 :     {
1120 :     const uint32_t Cost_Base = Cost0 + Run_Costs[i-Run];
1121 :     const uint32_t Cost = Cost_Base + (Code_Len20[Run-1]<<TL_SHIFT);
1122 :     const uint32_t lCost = Cost_Base + (Code_Len24[Run-1]<<TL_SHIFT);
1123 :    
1124 :     /*
1125 :     * TODO: what about tie-breaks? Should we favor short runs or
1126 :     * long runs? Although the error is the same, it would not be
1127 :     * spread the same way along high and low frequencies...
1128 :     */
1129 :     if (Cost<Best_Cost) {
1130 :     Best_Cost = Cost;
1131 :     Nodes[i].Run = Run;
1132 :     }
1133 :    
1134 :     if (lCost<Last_Cost) {
1135 :     Last_Cost = lCost;
1136 :     Last.Run = Run;
1137 :     Last_Node = i;
1138 :     }
1139 :     }
1140 :     if (Last_Node==i)
1141 :     Last.Level = Nodes[i].Level;
1142 :    
1143 :     if (DBG==1) {
1144 :     Run_Costs[i] = Best_Cost;
1145 :     printf( "Costs #%2d: ", i);
1146 :     for(j=-1;j<=Non_Zero;++j) {
1147 :     if (j==Run_Start) printf( " %3.0d|", Run_Costs[j]>>12 );
1148 :     else if (j>Run_Start && j<i) printf( " %3.0d|", Run_Costs[j]>>12 );
1149 :     else if (j==i) printf( "(%3.0d)", Run_Costs[j]>>12 );
1150 :     else printf( " - |" );
1151 :     }
1152 :     printf( "<%3.0d %2d %d>", Min_Cost>>12, Nodes[i].Level, Nodes[i].Run );
1153 :     printf( " Last:#%2d {%3.0d %2d %d}", Last_Node, Last_Cost>>12, Last.Level, Last.Run );
1154 :     printf( " AC:%3.0d Dist0:%3d Dist(%d)=%d", AC, Dist0>>12, Nodes[i].Level, Cost0>>12 );
1155 :     printf( "\n" );
1156 :     }
1157 :     }
1158 :     else /* "big" levels */
1159 :     {
1160 :     const uint8_t *Tbl_L1, *Tbl_L2, *Tbl_L1_Last, *Tbl_L2_Last;
1161 :     int Level2;
1162 :     int dQ1, dQ2;
1163 :     int Run;
1164 :     uint32_t Dist1,Dist2;
1165 :     int dDist21;
1166 :    
1167 :     if (Level1>1) {
1168 :     dQ1 = Level1*Mult-AC + Bias;
1169 :     dQ2 = dQ1 - Mult;
1170 :     Level2 = Level1-1;
1171 :     Tbl_L1 = (Level1<=24) ? B16_17_Code_Len[Level1-1] : Code_Len0;
1172 :     Tbl_L2 = (Level2<=24) ? B16_17_Code_Len[Level2-1] : Code_Len0;
1173 :     Tbl_L1_Last = (Level1<=6) ? B16_17_Code_Len_Last[Level1-1] : Code_Len0;
1174 :     Tbl_L2_Last = (Level2<=6) ? B16_17_Code_Len_Last[Level2-1] : Code_Len0;
1175 :     } else { /* Level1<-1 */
1176 :     dQ1 = Level1*Mult-AC - Bias;
1177 :     dQ2 = dQ1 + Mult;
1178 :     Level2 = Level1 + 1;
1179 :     Tbl_L1 = (Level1>=-24) ? B16_17_Code_Len[Level1^-1] : Code_Len0;
1180 :     Tbl_L2 = (Level2>=-24) ? B16_17_Code_Len[Level2^-1] : Code_Len0;
1181 :     Tbl_L1_Last = (Level1>=- 6) ? B16_17_Code_Len_Last[Level1^-1] : Code_Len0;
1182 :     Tbl_L2_Last = (Level2>=- 6) ? B16_17_Code_Len_Last[Level2^-1] : Code_Len0;
1183 :     }
1184 :     Dist1 = Lambda*dQ1*dQ1;
1185 :     Dist2 = Lambda*dQ2*dQ2;
1186 :     dDist21 = Dist2-Dist1;
1187 :    
1188 :     for(Run=i-Run_Start; Run>0; --Run)
1189 :     {
1190 :     const uint32_t Cost_Base = Dist1 + Run_Costs[i-Run];
1191 :     uint32_t Cost1, Cost2;
1192 :     int bLevel;
1193 :    
1194 :     /*
1195 :     * for sub-optimal (but slightly worth it, speed-wise) search, uncomment the following:
1196 :     * if (Cost_Base>=Best_Cost) continue;
1197 :     */
1198 :     Cost1 = Cost_Base + (Tbl_L1[Run-1]<<TL_SHIFT);
1199 :     Cost2 = Cost_Base + (Tbl_L2[Run-1]<<TL_SHIFT) + dDist21;
1200 :    
1201 :     if (Cost2<Cost1) {
1202 :     Cost1 = Cost2;
1203 :     bLevel = Level2;
1204 :     } else
1205 :     bLevel = Level1;
1206 :    
1207 :     if (Cost1<Best_Cost) {
1208 :     Best_Cost = Cost1;
1209 :     Nodes[i].Run = Run;
1210 :     Nodes[i].Level = bLevel;
1211 :     }
1212 :    
1213 :     Cost1 = Cost_Base + (Tbl_L1_Last[Run-1]<<TL_SHIFT);
1214 :     Cost2 = Cost_Base + (Tbl_L2_Last[Run-1]<<TL_SHIFT) + dDist21;
1215 :    
1216 :     if (Cost2<Cost1) {
1217 :     Cost1 = Cost2;
1218 :     bLevel = Level2;
1219 :     } else
1220 :     bLevel = Level1;
1221 :    
1222 :     if (Cost1<Last_Cost) {
1223 :     Last_Cost = Cost1;
1224 :     Last.Run = Run;
1225 :     Last.Level = bLevel;
1226 :     Last_Node = i;
1227 :     }
1228 :     } /* end of "for Run" */
1229 :    
1230 :     if (DBG==1) {
1231 :     Run_Costs[i] = Best_Cost;
1232 :     printf( "Costs #%2d: ", i);
1233 :     for(j=-1;j<=Non_Zero;++j) {
1234 :     if (j==Run_Start) printf( " %3.0d|", Run_Costs[j]>>12 );
1235 :     else if (j>Run_Start && j<i) printf( " %3.0d|", Run_Costs[j]>>12 );
1236 :     else if (j==i) printf( "(%3.0d)", Run_Costs[j]>>12 );
1237 :     else printf( " - |" );
1238 :     }
1239 :     printf( "<%3.0d %2d %d>", Min_Cost>>12, Nodes[i].Level, Nodes[i].Run );
1240 :     printf( " Last:#%2d {%3.0d %2d %d}", Last_Node, Last_Cost>>12, Last.Level, Last.Run );
1241 :     printf( " AC:%3.0d Dist0:%3d Dist(%2d):%3d Dist(%2d):%3d", AC, Dist0>>12, Level1, Dist1>>12, Level2, Dist2>>12 );
1242 :     printf( "\n" );
1243 :     }
1244 :     }
1245 :    
1246 :     Run_Costs[i] = Best_Cost;
1247 :    
1248 :     if (Best_Cost < Min_Cost + Dist0) {
1249 :     Min_Cost = Best_Cost;
1250 :     Run_Start = i;
1251 :     }
1252 :     else
1253 :     {
1254 :     /*
1255 :     * as noticed by Michael Niedermayer (michaelni at gmx.at), there's
1256 :     * a code shorter by 1 bit for a larger run (!), same level. We give
1257 :     * it a chance by not moving the left barrier too much.
1258 :     */
1259 :    
1260 :     while( Run_Costs[Run_Start]>Min_Cost+(1<<TL_SHIFT) )
1261 :     Run_Start++;
1262 :    
1263 :     /* spread on preceding coeffs the cost incurred by skipping this one */
1264 :     for(j=Run_Start; j<i; ++j) Run_Costs[j] += Dist0;
1265 :     Min_Cost += Dist0;
1266 :     }
1267 :     }
1268 :    
1269 :     if (DBG) {
1270 :     Last_Cost = Evaluate_Cost(Out,Mult,Bias, Zigzag,Non_Zero, Lambda);
1271 :     if (DBG==1) {
1272 :     printf( "=> " );
1273 :     for(i=0; i<=Non_Zero; ++i) printf( "[%3.0d] ", Out[Zigzag[i]] );
1274 :     printf( "\n" );
1275 :     }
1276 :     }
1277 :    
1278 :     if (Last_Node<0)
1279 :     return -1;
1280 :    
1281 :     /* reconstruct optimal sequence backward with surviving paths */
1282 :     memset(Out, 0x00, 64*sizeof(*Out));
1283 :     Out[Zigzag[Last_Node]] = Last.Level;
1284 :     i = Last_Node - Last.Run;
1285 :     while(i>=0) {
1286 :     Out[Zigzag[i]] = Nodes[i].Level;
1287 :     i -= Nodes[i].Run;
1288 :     }
1289 :    
1290 :     if (DBG) {
1291 :     uint32_t Cost = Evaluate_Cost(Out,Mult,Bias, Zigzag,Non_Zero, Lambda);
1292 :     if (DBG==1) {
1293 :     printf( "<= " );
1294 :     for(i=0; i<=Last_Node; ++i) printf( "[%3.0d] ", Out[Zigzag[i]] );
1295 :     printf( "\n--------------------------------\n" );
1296 :     }
1297 :     if (Cost>Last_Cost) printf( "!!! %u > %u\n", Cost, Last_Cost );
1298 :     }
1299 :     return Last_Node;
1300 :     }
1301 :    
1302 :     #undef DBG
1303 :    
1304 :     #endif
No admin address has been configured
 ViewVC Help
Powered by ViewVC 1.0.4