1 |
/****************************************************************************** |
/***************************************************************************** |
2 |
* * |
* |
3 |
* This file is part of XviD, a free MPEG-4 video encoder/decoder * |
* XVID MPEG-4 VIDEO CODEC |
4 |
* * |
* - MB Transfert/Quantization functions - |
5 |
* XviD is an implementation of a part of one or more MPEG-4 Video tools * |
* |
6 |
* as specified in ISO/IEC 14496-2 standard. Those intending to use this * |
* Copyright(C) 2001-2003 Peter Ross <pross@xvid.org> |
7 |
* software module in hardware or software products are advised that its * |
* 2001-2003 Michael Militzer <isibaar@xvid.org> |
8 |
* use may infringe existing patents or copyrights, and any such use * |
* 2003 Edouard Gomez <ed.gomez@free.fr> |
9 |
* would be at such party's own risk. The original developer of this * |
* |
10 |
* software module and his/her company, and subsequent editors and their * |
* This program is free software ; you can redistribute it and/or modify |
11 |
* companies, will have no liability for use of this software or * |
* it under the terms of the GNU General Public License as published by |
12 |
* modifications or derivatives thereof. * |
* the Free Software Foundation ; either version 2 of the License, or |
13 |
* * |
* (at your option) any later version. |
14 |
* XviD is free software; you can redistribute it and/or modify it * |
* |
15 |
* under the terms of the GNU General Public License as published by * |
* This program is distributed in the hope that it will be useful, |
16 |
* the Free Software Foundation; either version 2 of the License, or * |
* but WITHOUT ANY WARRANTY ; without even the implied warranty of |
17 |
* (at your option) any later version. * |
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
18 |
* * |
* GNU General Public License for more details. |
19 |
* XviD is distributed in the hope that it will be useful, but * |
* |
20 |
* WITHOUT ANY WARRANTY; without even the implied warranty of * |
* You should have received a copy of the GNU General Public License |
21 |
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * |
* along with this program ; if not, write to the Free Software |
22 |
* GNU General Public License for more details. * |
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
23 |
* * |
* |
24 |
* You should have received a copy of the GNU General Public License * |
* $Id$ |
25 |
* along with this program; if not, write to the Free Software * |
* |
26 |
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA * |
****************************************************************************/ |
|
* * |
|
|
******************************************************************************/ |
|
|
|
|
|
/****************************************************************************** |
|
|
* * |
|
|
* mbtransquant.c * |
|
|
* * |
|
|
* Copyright (C) 2001 - Peter Ross <pross@cs.rmit.edu.au> * |
|
|
* Copyright (C) 2001 - Michael Militzer <isibaar@xvid.org> * |
|
|
* * |
|
|
* For more information visit the XviD homepage: http://www.xvid.org * |
|
|
* * |
|
|
******************************************************************************/ |
|
|
|
|
|
/****************************************************************************** |
|
|
* * |
|
|
* Revision history: * |
|
|
* * |
|
|
* 29.03.2002 interlacing speedup - used transfer strides instead of * |
|
|
* manual field-to-frame conversion * |
|
|
* 26.03.2002 interlacing support - moved transfers outside loops * |
|
|
* 22.12.2001 get_dc_scaler() moved to common.h * |
|
|
* 19.11.2001 introduced coefficient thresholding (Isibaar) * |
|
|
* 17.11.2001 initial version * |
|
|
* * |
|
|
******************************************************************************/ |
|
27 |
|
|
28 |
|
#include <stdio.h> |
29 |
|
#include <stdlib.h> |
30 |
#include <string.h> |
#include <string.h> |
31 |
|
|
32 |
#include "../portab.h" |
#include "../portab.h" |
35 |
#include "../global.h" |
#include "../global.h" |
36 |
#include "mem_transfer.h" |
#include "mem_transfer.h" |
37 |
#include "timer.h" |
#include "timer.h" |
38 |
|
#include "../bitstream/mbcoding.h" |
39 |
|
#include "../bitstream/zigzag.h" |
40 |
#include "../dct/fdct.h" |
#include "../dct/fdct.h" |
41 |
#include "../dct/idct.h" |
#include "../dct/idct.h" |
42 |
#include "../quant/quant_mpeg4.h" |
#include "../quant/quant_mpeg4.h" |
47 |
|
|
48 |
MBFIELDTEST_PTR MBFieldTest; |
MBFIELDTEST_PTR MBFieldTest; |
49 |
|
|
50 |
#define TOOSMALL_LIMIT 1 /* skip blocks having a coefficient sum below this value */ |
/* |
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 |
void |
/***************************************************************************** |
58 |
MBTransQuantIntra(const MBParam * pParam, |
* Local functions |
59 |
FRAMEINFO * frame, |
****************************************************************************/ |
60 |
MACROBLOCK * pMB, |
|
61 |
const uint32_t x_pos, |
/* permute block and return field dct choice */ |
62 |
const uint32_t y_pos, |
static __inline uint32_t |
63 |
int16_t data[6 * 64], |
MBDecideFieldDCT(int16_t data[6 * 64]) |
|
int16_t qcoeff[6 * 64]) |
|
64 |
{ |
{ |
65 |
|
uint32_t field = MBFieldTest(data); |
66 |
|
|
67 |
uint32_t stride = pParam->edged_width; |
if (field) |
68 |
uint32_t stride2 = stride / 2; |
MBFrameToField(data); |
|
uint32_t next_block = stride * ((frame->vop_flags & XVID_VOP_REDUCED)?16:8); |
|
|
uint32_t i; |
|
|
const uint32_t iQuant = pMB->quant; |
|
|
uint8_t *pY_Cur, *pU_Cur, *pV_Cur; |
|
|
IMAGE *pCurrent = &frame->image; |
|
69 |
|
|
70 |
start_timer(); |
return field; |
|
if ((frame->vop_flags & XVID_VOP_REDUCED)) |
|
|
{ |
|
|
pY_Cur = pCurrent->y + (y_pos << 5) * stride + (x_pos << 5); |
|
|
pU_Cur = pCurrent->u + (y_pos << 4) * stride2 + (x_pos << 4); |
|
|
pV_Cur = pCurrent->v + (y_pos << 4) * stride2 + (x_pos << 4); |
|
|
|
|
|
filter_18x18_to_8x8(&data[0 * 64], pY_Cur, stride); |
|
|
filter_18x18_to_8x8(&data[1 * 64], pY_Cur + 16, stride); |
|
|
filter_18x18_to_8x8(&data[2 * 64], pY_Cur + next_block, stride); |
|
|
filter_18x18_to_8x8(&data[3 * 64], pY_Cur + next_block + 16, stride); |
|
|
filter_18x18_to_8x8(&data[4 * 64], pU_Cur, stride2); |
|
|
filter_18x18_to_8x8(&data[5 * 64], pV_Cur, stride2); |
|
|
}else{ |
|
|
pY_Cur = pCurrent->y + (y_pos << 4) * stride + (x_pos << 4); |
|
|
pU_Cur = pCurrent->u + (y_pos << 3) * stride2 + (x_pos << 3); |
|
|
pV_Cur = pCurrent->v + (y_pos << 3) * stride2 + (x_pos << 3); |
|
|
|
|
|
transfer_8to16copy(&data[0 * 64], pY_Cur, stride); |
|
|
transfer_8to16copy(&data[1 * 64], pY_Cur + 8, stride); |
|
|
transfer_8to16copy(&data[2 * 64], pY_Cur + next_block, stride); |
|
|
transfer_8to16copy(&data[3 * 64], pY_Cur + next_block + 8, stride); |
|
|
transfer_8to16copy(&data[4 * 64], pU_Cur, stride2); |
|
|
transfer_8to16copy(&data[5 * 64], pV_Cur, stride2); |
|
71 |
} |
} |
|
stop_transfer_timer(); |
|
72 |
|
|
73 |
/* XXX: rrv+interlacing is buggy */ |
/* Performs Forward DCT on all blocks */ |
74 |
|
static __inline void |
75 |
|
MBfDCT(const MBParam * const pParam, |
76 |
|
const FRAMEINFO * const frame, |
77 |
|
MACROBLOCK * const pMB, |
78 |
|
uint32_t x_pos, |
79 |
|
uint32_t y_pos, |
80 |
|
int16_t data[6 * 64]) |
81 |
|
{ |
82 |
|
/* Handles interlacing */ |
83 |
start_timer(); |
start_timer(); |
84 |
pMB->field_dct = 0; |
pMB->field_dct = 0; |
85 |
if ((frame->vol_flags & XVID_VOL_INTERLACING) && |
if ((frame->vol_flags & XVID_VOL_INTERLACING) && |
89 |
} |
} |
90 |
stop_interlacing_timer(); |
stop_interlacing_timer(); |
91 |
|
|
92 |
for (i = 0; i < 6; i++) { |
/* Perform DCT */ |
|
uint32_t iDcScaler = get_dc_scaler(iQuant, i < 4); |
|
|
|
|
93 |
start_timer(); |
start_timer(); |
94 |
fdct(&data[i * 64]); |
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(); |
stop_dct_timer(); |
|
|
|
|
if (!(pParam->vol_flags & XVID_VOL_MPEGQUANT)) { |
|
|
start_timer(); |
|
|
quant_intra(&qcoeff[i * 64], &data[i * 64], iQuant, iDcScaler); |
|
|
stop_quant_timer(); |
|
|
} else { |
|
|
start_timer(); |
|
|
quant4_intra(&qcoeff[i * 64], &data[i * 64], iQuant, iDcScaler); |
|
|
stop_quant_timer(); |
|
101 |
} |
} |
102 |
|
|
103 |
/* speedup: dont decode when encoding only ivops */ |
/* Performs Inverse DCT on all blocks */ |
104 |
if (pParam->iMaxKeyInterval != 1 || pParam->max_bframes > 0) |
static __inline void |
105 |
|
MBiDCT(int16_t data[6 * 64], |
106 |
|
const uint8_t cbp) |
107 |
{ |
{ |
|
if (!(pParam->vol_flags & XVID_VOL_MPEGQUANT)) { |
|
|
start_timer(); |
|
|
dequant_intra(&data[i * 64], &qcoeff[i * 64], iQuant, iDcScaler); |
|
|
stop_iquant_timer(); |
|
|
} else { |
|
|
start_timer(); |
|
|
dequant4_intra(&data[i * 64], &qcoeff[i * 64], iQuant, iDcScaler); |
|
|
stop_iquant_timer(); |
|
|
} |
|
|
|
|
108 |
start_timer(); |
start_timer(); |
109 |
idct(&data[i * 64]); |
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(); |
stop_idct_timer(); |
116 |
} |
} |
|
} |
|
117 |
|
|
118 |
/* speedup: dont decode when encoding only ivops */ |
/* Quantize all blocks -- Intra mode */ |
119 |
if (pParam->iMaxKeyInterval != 1 || pParam->max_bframes > 0) |
static __inline void |
120 |
|
MBQuantIntra(const MBParam * pParam, |
121 |
|
const FRAMEINFO * const frame, |
122 |
|
const MACROBLOCK * pMB, |
123 |
|
int16_t qcoeff[6 * 64], |
124 |
|
int16_t data[6*64]) |
125 |
{ |
{ |
126 |
|
int mpeg; |
127 |
|
int scaler_lum, scaler_chr; |
128 |
|
|
129 |
if (pMB->field_dct) { |
quanth263_intraFuncPtr const quant[2] = |
130 |
next_block = stride; |
{ |
131 |
stride *= 2; |
(quanth263_intraFuncPtr)quant_intra, |
132 |
|
(quanth263_intraFuncPtr)quant4_intra |
133 |
|
}; |
134 |
|
|
135 |
|
mpeg = !!(pParam->vol_flags & XVID_VOL_MPEGQUANT); |
136 |
|
scaler_lum = get_dc_scaler(pMB->quant, 1); |
137 |
|
scaler_chr = get_dc_scaler(pMB->quant, 0); |
138 |
|
|
139 |
|
/* Quantize the block */ |
140 |
|
start_timer(); |
141 |
|
quant[mpeg](&data[0 * 64], &qcoeff[0 * 64], pMB->quant, scaler_lum); |
142 |
|
quant[mpeg](&data[1 * 64], &qcoeff[1 * 64], pMB->quant, scaler_lum); |
143 |
|
quant[mpeg](&data[2 * 64], &qcoeff[2 * 64], pMB->quant, scaler_lum); |
144 |
|
quant[mpeg](&data[3 * 64], &qcoeff[3 * 64], pMB->quant, scaler_lum); |
145 |
|
quant[mpeg](&data[4 * 64], &qcoeff[4 * 64], pMB->quant, scaler_chr); |
146 |
|
quant[mpeg](&data[5 * 64], &qcoeff[5 * 64], pMB->quant, scaler_chr); |
147 |
|
stop_quant_timer(); |
148 |
} |
} |
149 |
|
|
150 |
start_timer(); |
/* DeQuantize all blocks -- Intra mode */ |
151 |
if ((frame->vop_flags & XVID_VOP_REDUCED)) |
static __inline void |
152 |
|
MBDeQuantIntra(const MBParam * pParam, |
153 |
|
const int iQuant, |
154 |
|
int16_t qcoeff[6 * 64], |
155 |
|
int16_t data[6*64]) |
156 |
{ |
{ |
157 |
copy_upsampled_8x8_16to8(pY_Cur, &data[0 * 64], stride); |
int mpeg; |
158 |
copy_upsampled_8x8_16to8(pY_Cur + 16, &data[1 * 64], stride); |
int scaler_lum, scaler_chr; |
|
copy_upsampled_8x8_16to8(pY_Cur + next_block, &data[2 * 64], stride); |
|
|
copy_upsampled_8x8_16to8(pY_Cur + next_block + 16, &data[3 * 64], stride); |
|
|
copy_upsampled_8x8_16to8(pU_Cur, &data[4 * 64], stride2); |
|
|
copy_upsampled_8x8_16to8(pV_Cur, &data[5 * 64], stride2); |
|
159 |
|
|
160 |
}else{ |
quanth263_intraFuncPtr const dequant[2] = |
161 |
transfer_16to8copy(pY_Cur, &data[0 * 64], stride); |
{ |
162 |
transfer_16to8copy(pY_Cur + 8, &data[1 * 64], stride); |
(quanth263_intraFuncPtr)dequant_intra, |
163 |
transfer_16to8copy(pY_Cur + next_block, &data[2 * 64], stride); |
(quanth263_intraFuncPtr)dequant4_intra |
164 |
transfer_16to8copy(pY_Cur + next_block + 8, &data[3 * 64], stride); |
}; |
165 |
transfer_16to8copy(pU_Cur, &data[4 * 64], stride2); |
|
166 |
transfer_16to8copy(pV_Cur, &data[5 * 64], stride2); |
mpeg = !!(pParam->vol_flags & XVID_VOL_MPEGQUANT); |
167 |
} |
scaler_lum = get_dc_scaler(iQuant, 1); |
168 |
stop_transfer_timer(); |
scaler_chr = get_dc_scaler(iQuant, 0); |
169 |
|
|
170 |
|
start_timer(); |
171 |
|
dequant[mpeg](&qcoeff[0 * 64], &data[0 * 64], iQuant, scaler_lum); |
172 |
|
dequant[mpeg](&qcoeff[1 * 64], &data[1 * 64], iQuant, scaler_lum); |
173 |
|
dequant[mpeg](&qcoeff[2 * 64], &data[2 * 64], iQuant, scaler_lum); |
174 |
|
dequant[mpeg](&qcoeff[3 * 64], &data[3 * 64], iQuant, scaler_lum); |
175 |
|
dequant[mpeg](&qcoeff[4 * 64], &data[4 * 64], iQuant, scaler_chr); |
176 |
|
dequant[mpeg](&qcoeff[5 * 64], &data[5 * 64], iQuant, scaler_chr); |
177 |
|
stop_iquant_timer(); |
178 |
} |
} |
179 |
|
|
|
} |
|
180 |
|
|
181 |
|
typedef int (*trellis_func_ptr_t)(int16_t *const Out, |
182 |
|
const int16_t *const In, |
183 |
|
int Q, |
184 |
|
const uint16_t * const Zigzag, |
185 |
|
int Non_Zero); |
186 |
|
|
187 |
|
static int |
188 |
|
dct_quantize_trellis_h263_c(int16_t *const Out, |
189 |
|
const int16_t *const In, |
190 |
|
int Q, |
191 |
|
const uint16_t * const Zigzag, |
192 |
|
int Non_Zero); |
193 |
|
|
194 |
|
static int |
195 |
|
dct_quantize_trellis_mpeg_c(int16_t *const Out, |
196 |
|
const int16_t *const In, |
197 |
|
int Q, |
198 |
|
const uint16_t * const Zigzag, |
199 |
|
int Non_Zero); |
200 |
|
|
201 |
uint8_t |
/* Quantize all blocks -- Inter mode */ |
202 |
MBTransQuantInter(const MBParam * pParam, |
static __inline uint8_t |
203 |
FRAMEINFO * frame, |
MBQuantInter(const MBParam * pParam, |
204 |
MACROBLOCK * pMB, |
const FRAMEINFO * const frame, |
205 |
const uint32_t x_pos, |
const MACROBLOCK * pMB, |
|
const uint32_t y_pos, |
|
206 |
int16_t data[6 * 64], |
int16_t data[6 * 64], |
207 |
int16_t qcoeff[6 * 64]) |
int16_t qcoeff[6 * 64], |
208 |
|
int bvop, |
209 |
|
int limit) |
210 |
{ |
{ |
211 |
|
|
212 |
uint32_t stride = pParam->edged_width; |
int i; |
|
uint32_t stride2 = stride / 2; |
|
|
uint32_t next_block = stride * ((frame->vop_flags & XVID_VOP_REDUCED)?16:8); |
|
|
uint32_t i; |
|
|
const uint32_t iQuant = pMB->quant; |
|
|
uint8_t *pY_Cur, *pU_Cur, *pV_Cur; |
|
213 |
uint8_t cbp = 0; |
uint8_t cbp = 0; |
214 |
uint32_t sum; |
int sum; |
215 |
IMAGE *pCurrent = &frame->image; |
int code_block, mpeg; |
216 |
|
|
217 |
if ((frame->vop_flags & XVID_VOP_REDUCED)) |
quanth263_interFuncPtr const quant[2] = |
218 |
{ |
{ |
219 |
pY_Cur = pCurrent->y + (y_pos << 5) * stride + (x_pos << 5); |
(quanth263_interFuncPtr)quant_inter, |
220 |
pU_Cur = pCurrent->u + (y_pos << 4) * stride2 + (x_pos << 4); |
(quanth263_interFuncPtr)quant4_inter |
221 |
pV_Cur = pCurrent->v + (y_pos << 4) * stride2 + (x_pos << 4); |
}; |
|
}else{ |
|
|
pY_Cur = pCurrent->y + (y_pos << 4) * stride + (x_pos << 4); |
|
|
pU_Cur = pCurrent->u + (y_pos << 3) * stride2 + (x_pos << 3); |
|
|
pV_Cur = pCurrent->v + (y_pos << 3) * stride2 + (x_pos << 3); |
|
|
} |
|
222 |
|
|
223 |
start_timer(); |
trellis_func_ptr_t const trellis[2] = |
224 |
pMB->field_dct = 0; |
{ |
225 |
if ((frame->vol_flags & XVID_VOL_INTERLACING) && |
(trellis_func_ptr_t)dct_quantize_trellis_h263_c, |
226 |
(x_pos>0) && (x_pos<pParam->mb_width-1) && |
(trellis_func_ptr_t)dct_quantize_trellis_mpeg_c |
227 |
(y_pos>0) && (y_pos<pParam->mb_height-1)) { |
}; |
228 |
pMB->field_dct = MBDecideFieldDCT(data); |
|
229 |
} |
mpeg = !!(pParam->vol_flags & XVID_VOL_MPEGQUANT); |
|
stop_interlacing_timer(); |
|
230 |
|
|
231 |
for (i = 0; i < 6; i++) { |
for (i = 0; i < 6; i++) { |
|
uint32_t increase_limit = (iQuant == 1) ? 1 : 0; |
|
232 |
|
|
233 |
/* |
/* Quantize the block */ |
|
* no need to transfer 8->16-bit |
|
|
* (this is performed already in motion compensation) |
|
|
*/ |
|
234 |
start_timer(); |
start_timer(); |
|
fdct(&data[i * 64]); |
|
|
stop_dct_timer(); |
|
235 |
|
|
236 |
if (!(pParam->vol_flags & XVID_VOL_MPEGQUANT)) { |
sum = quant[mpeg](&qcoeff[i*64], &data[i*64], pMB->quant); |
237 |
start_timer(); |
|
238 |
sum = quant_inter(&qcoeff[i * 64], &data[i * 64], iQuant); |
if(sum && (frame->vop_flags & XVID_VOP_TRELLISQUANT)) { |
239 |
stop_quant_timer(); |
sum = trellis[mpeg](&qcoeff[i*64], &data[i*64], pMB->quant, &scan_tables[0][0], 63); |
|
} else { |
|
|
start_timer(); |
|
|
sum = quant4_inter(&qcoeff[i * 64], &data[i * 64], iQuant); |
|
|
stop_quant_timer(); |
|
240 |
} |
} |
241 |
|
stop_quant_timer(); |
242 |
|
|
243 |
if ((sum >= TOOSMALL_LIMIT + increase_limit) || (qcoeff[i*64] != 0) || |
/* |
244 |
(qcoeff[i*64+1] != 0) || (qcoeff[i*64+8] != 0)) { |
* We code the block if the sum is higher than the limit and if the first |
245 |
|
* two AC coefficients in zig zag order are not zero. |
246 |
|
*/ |
247 |
|
code_block = 0; |
248 |
|
if ((sum >= limit) || (qcoeff[i*64+1] != 0) || (qcoeff[i*64+8] != 0)) { |
249 |
|
code_block = 1; |
250 |
|
} else { |
251 |
|
|
252 |
if (!(pParam->vol_flags & XVID_VOL_MPEGQUANT)) { |
if (bvop && (pMB->mode == MODE_DIRECT || pMB->mode == MODE_DIRECT_NO4V)) { |
253 |
start_timer(); |
/* dark blocks prevention for direct mode */ |
254 |
dequant_inter(&data[i * 64], &qcoeff[i * 64], iQuant); |
if ((qcoeff[i*64] < -1) || (qcoeff[i*64] > 0)) |
255 |
stop_iquant_timer(); |
code_block = 1; |
256 |
} else { |
} else { |
257 |
start_timer(); |
/* not direct mode */ |
258 |
dequant4_inter(&data[i * 64], &qcoeff[i * 64], iQuant); |
if (qcoeff[i*64] != 0) |
259 |
stop_iquant_timer(); |
code_block = 1; |
260 |
} |
} |
|
|
|
|
cbp |= 1 << (5 - i); |
|
|
|
|
|
start_timer(); |
|
|
idct(&data[i * 64]); |
|
|
stop_idct_timer(); |
|
261 |
} |
} |
262 |
|
|
263 |
|
/* Set the corresponding cbp bit */ |
264 |
|
cbp |= code_block << (5 - i); |
265 |
} |
} |
266 |
|
|
267 |
if (pMB->field_dct) { |
return(cbp); |
|
next_block = stride; |
|
|
stride *= 2; |
|
268 |
} |
} |
269 |
|
|
270 |
start_timer(); |
/* DeQuantize all blocks -- Inter mode */ |
271 |
if ((frame->vop_flags & XVID_VOP_REDUCED)) |
static __inline void |
272 |
|
MBDeQuantInter(const MBParam * pParam, |
273 |
|
const int iQuant, |
274 |
|
int16_t data[6 * 64], |
275 |
|
int16_t qcoeff[6 * 64], |
276 |
|
const uint8_t cbp) |
277 |
{ |
{ |
278 |
if (cbp & 32) |
int mpeg; |
|
add_upsampled_8x8_16to8(pY_Cur, &data[0 * 64], stride); |
|
|
if (cbp & 16) |
|
|
add_upsampled_8x8_16to8(pY_Cur + 16, &data[1 * 64], stride); |
|
|
if (cbp & 8) |
|
|
add_upsampled_8x8_16to8(pY_Cur + next_block, &data[2 * 64], stride); |
|
|
if (cbp & 4) |
|
|
add_upsampled_8x8_16to8(pY_Cur + 16 + next_block, &data[3 * 64], stride); |
|
|
if (cbp & 2) |
|
|
add_upsampled_8x8_16to8(pU_Cur, &data[4 * 64], stride2); |
|
|
if (cbp & 1) |
|
|
add_upsampled_8x8_16to8(pV_Cur, &data[5 * 64], stride2); |
|
|
}else{ |
|
|
if (cbp & 32) |
|
|
transfer_16to8add(pY_Cur, &data[0 * 64], stride); |
|
|
if (cbp & 16) |
|
|
transfer_16to8add(pY_Cur + 8, &data[1 * 64], stride); |
|
|
if (cbp & 8) |
|
|
transfer_16to8add(pY_Cur + next_block, &data[2 * 64], stride); |
|
|
if (cbp & 4) |
|
|
transfer_16to8add(pY_Cur + next_block + 8, &data[3 * 64], stride); |
|
|
if (cbp & 2) |
|
|
transfer_16to8add(pU_Cur, &data[4 * 64], stride2); |
|
|
if (cbp & 1) |
|
|
transfer_16to8add(pV_Cur, &data[5 * 64], stride2); |
|
|
} |
|
|
stop_transfer_timer(); |
|
279 |
|
|
280 |
return cbp; |
quanth263_interFuncPtr const dequant[2] = |
281 |
|
{ |
282 |
|
(quanth263_interFuncPtr)dequant_inter, |
283 |
|
(quanth263_interFuncPtr)dequant4_inter |
284 |
|
}; |
285 |
|
|
286 |
} |
mpeg = !!(pParam->vol_flags & XVID_VOL_MPEGQUANT); |
287 |
|
|
288 |
void |
start_timer(); |
289 |
MBTransQuantIntra2(const MBParam * pParam, |
if(cbp & (1 << (5 - 0))) dequant[mpeg](&data[0 * 64], &qcoeff[0 * 64], iQuant); |
290 |
FRAMEINFO * frame, |
if(cbp & (1 << (5 - 1))) dequant[mpeg](&data[1 * 64], &qcoeff[1 * 64], iQuant); |
291 |
MACROBLOCK * pMB, |
if(cbp & (1 << (5 - 2))) dequant[mpeg](&data[2 * 64], &qcoeff[2 * 64], iQuant); |
292 |
const uint32_t x_pos, |
if(cbp & (1 << (5 - 3))) dequant[mpeg](&data[3 * 64], &qcoeff[3 * 64], iQuant); |
293 |
const uint32_t y_pos, |
if(cbp & (1 << (5 - 4))) dequant[mpeg](&data[4 * 64], &qcoeff[4 * 64], iQuant); |
294 |
int16_t data[6 * 64], |
if(cbp & (1 << (5 - 5))) dequant[mpeg](&data[5 * 64], &qcoeff[5 * 64], iQuant); |
295 |
int16_t qcoeff[6 * 64]) |
stop_iquant_timer(); |
|
{ |
|
|
MBTrans(pParam,frame,pMB,x_pos,y_pos,data); |
|
|
MBfDCT(pParam,frame,pMB,data); |
|
|
MBQuantIntra(pParam,frame,pMB,data,qcoeff); |
|
|
MBDeQuantIntra(pParam,pMB->quant,data,qcoeff); |
|
|
MBiDCT(data,0x3F); |
|
|
MBTransAdd(pParam,frame,pMB,x_pos,y_pos,data,0x3F); |
|
296 |
} |
} |
297 |
|
|
298 |
|
typedef void (transfer_operation_8to16_t) (int16_t *Dst, const uint8_t *Src, int BpS); |
299 |
|
typedef void (transfer_operation_16to8_t) (uint8_t *Dst, const int16_t *Src, int BpS); |
300 |
|
|
301 |
uint8_t |
|
302 |
MBTransQuantInter2(const MBParam * pParam, |
static __inline void |
303 |
FRAMEINFO * frame, |
MBTrans8to16(const MBParam * const pParam, |
304 |
MACROBLOCK * pMB, |
const FRAMEINFO * const frame, |
305 |
|
const MACROBLOCK * const pMB, |
306 |
const uint32_t x_pos, |
const uint32_t x_pos, |
307 |
const uint32_t y_pos, |
const uint32_t y_pos, |
308 |
int16_t data[6 * 64], |
int16_t data[6 * 64]) |
|
int16_t qcoeff[6 * 64]) |
|
309 |
{ |
{ |
310 |
uint8_t cbp; |
uint32_t stride = pParam->edged_width; |
311 |
|
uint32_t stride2 = stride / 2; |
312 |
/* there is no MBTrans for Inter block, that's done in motion compensation already */ |
uint32_t next_block = stride * 8; |
313 |
|
int32_t cst; |
314 |
MBfDCT(pParam,frame,pMB,data); |
int vop_reduced; |
315 |
cbp = MBQuantInter(pParam,pMB->quant,data,qcoeff); |
uint8_t *pY_Cur, *pU_Cur, *pV_Cur; |
316 |
MBDeQuantInter(pParam,pMB->quant,data,qcoeff,cbp); |
const IMAGE * const pCurrent = &frame->image; |
317 |
MBiDCT(data,cbp); |
transfer_operation_8to16_t * const functions[2] = |
318 |
MBTransAdd(pParam,frame,pMB,x_pos,y_pos,data,cbp); |
{ |
319 |
|
(transfer_operation_8to16_t *)transfer_8to16copy, |
320 |
return cbp; |
(transfer_operation_8to16_t *)filter_18x18_to_8x8 |
321 |
|
}; |
322 |
|
transfer_operation_8to16_t *transfer_op = NULL; |
323 |
|
|
324 |
|
vop_reduced = !!(frame->vop_flags & XVID_VOP_REDUCED); |
325 |
|
|
326 |
|
/* Image pointers */ |
327 |
|
pY_Cur = pCurrent->y + (y_pos << (4+vop_reduced)) * stride + (x_pos << (4+vop_reduced)); |
328 |
|
pU_Cur = pCurrent->u + (y_pos << (3+vop_reduced)) * stride2 + (x_pos << (3+vop_reduced)); |
329 |
|
pV_Cur = pCurrent->v + (y_pos << (3+vop_reduced)) * stride2 + (x_pos << (3+vop_reduced)); |
330 |
|
|
331 |
|
/* Block size */ |
332 |
|
cst = 8<<vop_reduced; |
333 |
|
|
334 |
|
/* Operation function */ |
335 |
|
transfer_op = functions[vop_reduced]; |
336 |
|
|
337 |
|
/* Do the transfer */ |
338 |
|
start_timer(); |
339 |
|
transfer_op(&data[0 * 64], pY_Cur, stride); |
340 |
|
transfer_op(&data[1 * 64], pY_Cur + cst, stride); |
341 |
|
transfer_op(&data[2 * 64], pY_Cur + next_block, stride); |
342 |
|
transfer_op(&data[3 * 64], pY_Cur + next_block + cst, stride); |
343 |
|
transfer_op(&data[4 * 64], pU_Cur, stride2); |
344 |
|
transfer_op(&data[5 * 64], pV_Cur, stride2); |
345 |
|
stop_transfer_timer(); |
346 |
} |
} |
347 |
|
|
348 |
uint8_t |
static __inline void |
349 |
MBTransQuantInterBVOP(const MBParam * pParam, |
MBTrans16to8(const MBParam * const pParam, |
350 |
FRAMEINFO * frame, |
const FRAMEINFO * const frame, |
351 |
MACROBLOCK * pMB, |
const MACROBLOCK * const pMB, |
352 |
const uint32_t x_pos, |
const uint32_t x_pos, |
353 |
const uint32_t y_pos, |
const uint32_t y_pos, |
354 |
int16_t data[6 * 64], |
int16_t data[6 * 64], |
355 |
int16_t qcoeff[6 * 64]) |
const uint32_t add, /* Must be 1 or 0 */ |
356 |
|
const uint8_t cbp) |
357 |
{ |
{ |
358 |
uint8_t cbp; |
uint8_t *pY_Cur, *pU_Cur, *pV_Cur; |
359 |
|
uint32_t stride = pParam->edged_width; |
360 |
/* there is no MBTrans for Inter block, that's done in motion compensation already */ |
uint32_t stride2 = stride / 2; |
361 |
|
uint32_t next_block = stride * 8; |
362 |
MBfDCT(pParam,frame,pMB,data); |
uint32_t cst; |
363 |
cbp = MBQuantInter(pParam,pMB->quant,data,qcoeff); |
int vop_reduced; |
364 |
|
const IMAGE * const pCurrent = &frame->image; |
365 |
/* |
/* Array of function pointers, indexed by [vop_reduced<<1+add] */ |
366 |
* History comment: |
transfer_operation_16to8_t * const functions[4] = |
367 |
* we don't have to DeQuant, iDCT and Transfer back data for B-frames |
{ |
368 |
*/ |
(transfer_operation_16to8_t*)transfer_16to8copy, |
369 |
|
(transfer_operation_16to8_t*)transfer_16to8add, |
370 |
|
(transfer_operation_16to8_t*)copy_upsampled_8x8_16to8, |
371 |
|
(transfer_operation_16to8_t*)add_upsampled_8x8_16to8 |
372 |
|
}; |
373 |
|
|
374 |
/* |
transfer_operation_16to8_t *transfer_op = NULL; |
|
* As an exception to the previous rule, if we are willing to have extra |
|
|
* stats then we have to DeQuant, iDCT and Transfer back the data :-) |
|
|
*/ |
|
|
if((pParam->plugin_flags & XVID_REQORIGINAL)) { |
|
|
MBDeQuantInter(pParam,pMB->quant,data,qcoeff,cbp); |
|
|
MBiDCT(data,cbp); |
|
|
MBTransAdd(pParam,frame,pMB,x_pos,y_pos,data,cbp); |
|
|
} |
|
375 |
|
|
376 |
return cbp; |
if (pMB->field_dct) { |
377 |
|
next_block = stride; |
378 |
|
stride *= 2; |
379 |
} |
} |
380 |
|
|
381 |
|
/* Makes this vars booleans */ |
382 |
|
vop_reduced = !!(frame->vop_flags & XVID_VOP_REDUCED); |
383 |
|
|
384 |
void |
/* Image pointers */ |
385 |
MBfDCT(const MBParam * pParam, |
pY_Cur = pCurrent->y + (y_pos << (4+vop_reduced)) * stride + (x_pos << (4+vop_reduced)); |
386 |
FRAMEINFO * frame, |
pU_Cur = pCurrent->u + (y_pos << (3+vop_reduced)) * stride2 + (x_pos << (3+vop_reduced)); |
387 |
MACROBLOCK * pMB, |
pV_Cur = pCurrent->v + (y_pos << (3+vop_reduced)) * stride2 + (x_pos << (3+vop_reduced)); |
388 |
int16_t data[6 * 64]) |
|
389 |
{ |
/* Block size */ |
390 |
int i; |
cst = 8<<vop_reduced; |
391 |
|
|
392 |
start_timer(); |
/* Operation function */ |
393 |
pMB->field_dct = 0; |
transfer_op = functions[(vop_reduced<<1) + add]; |
394 |
if ((frame->vol_flags & XVID_VOL_INTERLACING)) { |
|
395 |
pMB->field_dct = MBDecideFieldDCT(data); |
/* Do the operation */ |
396 |
|
start_timer(); |
397 |
|
if (cbp&32) transfer_op(pY_Cur, &data[0 * 64], stride); |
398 |
|
if (cbp&16) transfer_op(pY_Cur + cst, &data[1 * 64], stride); |
399 |
|
if (cbp& 8) transfer_op(pY_Cur + next_block, &data[2 * 64], stride); |
400 |
|
if (cbp& 4) transfer_op(pY_Cur + next_block + cst, &data[3 * 64], stride); |
401 |
|
if (cbp& 2) transfer_op(pU_Cur, &data[4 * 64], stride2); |
402 |
|
if (cbp& 1) transfer_op(pV_Cur, &data[5 * 64], stride2); |
403 |
|
stop_transfer_timer(); |
404 |
} |
} |
|
stop_interlacing_timer(); |
|
405 |
|
|
406 |
for (i = 0; i < 6; i++) { |
/***************************************************************************** |
407 |
start_timer(); |
* Module functions |
408 |
fdct(&data[i * 64]); |
****************************************************************************/ |
|
stop_dct_timer(); |
|
|
} |
|
|
} |
|
409 |
|
|
410 |
void |
void |
411 |
MBQuantDeQuantIntra(const MBParam * pParam, |
MBTransQuantIntra(const MBParam * const pParam, |
412 |
FRAMEINFO * frame, |
const FRAMEINFO * const frame, |
413 |
MACROBLOCK * pMB, |
MACROBLOCK * const pMB, |
414 |
int16_t qcoeff[6 * 64], |
const uint32_t x_pos, |
415 |
int16_t data[6*64]) |
const uint32_t y_pos, |
416 |
|
int16_t data[6 * 64], |
417 |
|
int16_t qcoeff[6 * 64]) |
418 |
{ |
{ |
|
int i; |
|
|
int iQuant = pMB->quant; |
|
419 |
|
|
420 |
start_timer(); |
/* Transfer data */ |
421 |
pMB->field_dct = 0; |
MBTrans8to16(pParam, frame, pMB, x_pos, y_pos, data); |
|
if ((frame->vol_flags & XVID_VOL_INTERLACING)) { |
|
|
pMB->field_dct = MBDecideFieldDCT(data); |
|
|
} |
|
|
stop_interlacing_timer(); |
|
422 |
|
|
423 |
for (i = 0; i < 6; i++) { |
/* Perform DCT (and field decision) */ |
424 |
uint32_t iDcScaler = get_dc_scaler(iQuant, i < 4); |
MBfDCT(pParam, frame, pMB, x_pos, y_pos, data); |
|
|
|
|
if (!(pParam->vol_flags & XVID_VOL_MPEGQUANT)) { |
|
|
start_timer(); |
|
|
quant_intra(&qcoeff[i * 64], &data[i * 64], iQuant, iDcScaler); |
|
|
stop_quant_timer(); |
|
425 |
|
|
426 |
start_timer(); |
/* Quantize the block */ |
427 |
dequant_intra(&data[i * 64], &qcoeff[i * 64], iQuant, iDcScaler); |
MBQuantIntra(pParam, frame, pMB, data, qcoeff); |
|
stop_iquant_timer(); |
|
|
} else { |
|
|
start_timer(); |
|
|
quant4_intra(&qcoeff[i * 64], &data[i * 64], iQuant, iDcScaler); |
|
|
stop_quant_timer(); |
|
|
|
|
|
start_timer(); |
|
|
dequant4_intra(&data[i * 64], &qcoeff[i * 64], iQuant, iDcScaler); |
|
|
stop_iquant_timer(); |
|
|
} |
|
|
} |
|
|
} |
|
|
|
|
|
void |
|
|
MBQuantIntra(const MBParam * pParam, |
|
|
FRAMEINFO * frame, |
|
|
MACROBLOCK *pMB, |
|
|
int16_t qcoeff[6 * 64], |
|
|
int16_t data[6*64]) |
|
|
{ |
|
|
int i; |
|
|
int iQuant = pMB->quant; |
|
428 |
|
|
429 |
start_timer(); |
/* DeQuantize the block */ |
430 |
pMB->field_dct = 0; |
MBDeQuantIntra(pParam, pMB->quant, data, qcoeff); |
|
if ((frame->vol_flags & XVID_VOL_INTERLACING)) { |
|
|
pMB->field_dct = MBDecideFieldDCT(data); |
|
|
} |
|
|
stop_interlacing_timer(); |
|
431 |
|
|
432 |
for (i = 0; i < 6; i++) { |
/* Perform inverse DCT*/ |
433 |
uint32_t iDcScaler = get_dc_scaler(iQuant, i < 4); |
MBiDCT(data, 0x3F); |
434 |
|
|
435 |
if (!(pParam->vol_flags & XVID_VOL_MPEGQUANT)) { |
/* Transfer back the data -- Don't add data */ |
436 |
start_timer(); |
MBTrans16to8(pParam, frame, pMB, x_pos, y_pos, data, 0, 0x3F); |
|
quant_intra(&qcoeff[i * 64], &data[i * 64], iQuant, iDcScaler); |
|
|
stop_quant_timer(); |
|
|
} else { |
|
|
start_timer(); |
|
|
quant4_intra(&qcoeff[i * 64], &data[i * 64], iQuant, iDcScaler); |
|
|
stop_quant_timer(); |
|
437 |
} |
} |
|
} |
|
|
} |
|
|
|
|
|
void |
|
|
MBDeQuantIntra(const MBParam * pParam, |
|
|
const int iQuant, |
|
|
int16_t qcoeff[6 * 64], |
|
|
int16_t data[6*64]) |
|
|
{ |
|
|
int i; |
|
438 |
|
|
|
for (i = 0; i < 6; i++) { |
|
|
uint32_t iDcScaler = get_dc_scaler(iQuant, i < 4); |
|
|
|
|
|
if (!(pParam->vol_flags & XVID_VOL_MPEGQUANT)) { |
|
|
start_timer(); |
|
|
dequant_intra(&data[i * 64], &qcoeff[i * 64], iQuant, iDcScaler); |
|
|
stop_iquant_timer(); |
|
|
} else { |
|
|
start_timer(); |
|
|
dequant4_intra(&data[i * 64], &qcoeff[i * 64], iQuant, iDcScaler); |
|
|
stop_iquant_timer(); |
|
|
} |
|
|
} |
|
|
} |
|
439 |
|
|
440 |
uint8_t |
uint8_t |
441 |
MBQuantInter(const MBParam * pParam, |
MBTransQuantInter(const MBParam * const pParam, |
442 |
const int iQuant, |
const FRAMEINFO * const frame, |
443 |
|
MACROBLOCK * const pMB, |
444 |
|
const uint32_t x_pos, |
445 |
|
const uint32_t y_pos, |
446 |
int16_t data[6 * 64], |
int16_t data[6 * 64], |
447 |
int16_t qcoeff[6 * 64]) |
int16_t qcoeff[6 * 64]) |
448 |
{ |
{ |
449 |
|
uint8_t cbp; |
450 |
|
uint32_t limit; |
451 |
|
|
452 |
int i; |
/* |
453 |
uint8_t cbp = 0; |
* There is no MBTrans8to16 for Inter block, that's done in motion compensation |
454 |
int sum; |
* already |
455 |
|
*/ |
456 |
|
|
457 |
for (i = 0; i < 6; i++) { |
/* Perform DCT (and field decision) */ |
458 |
|
MBfDCT(pParam, frame, pMB, x_pos, y_pos, data); |
459 |
|
|
460 |
if (!(pParam->vol_flags & XVID_VOL_MPEGQUANT)) { |
/* Set the limit threshold */ |
461 |
start_timer(); |
limit = PVOP_TOOSMALL_LIMIT + ((pMB->quant == 1)? 1 : 0); |
|
sum = quant_inter(&qcoeff[i * 64], &data[i * 64], iQuant); |
|
|
stop_quant_timer(); |
|
|
} else { |
|
|
start_timer(); |
|
|
sum = quant4_inter(&qcoeff[i * 64], &data[i * 64], iQuant); |
|
|
stop_quant_timer(); |
|
|
} |
|
462 |
|
|
463 |
if (sum >= TOOSMALL_LIMIT) { // skip block ? |
if (frame->vop_flags & XVID_VOP_CARTOON) |
464 |
cbp |= 1 << (5 - i); |
limit *= 3; |
|
} |
|
|
} |
|
|
return cbp; |
|
|
} |
|
465 |
|
|
466 |
void |
/* Quantize the block */ |
467 |
MBDeQuantInter( const MBParam * pParam, |
cbp = MBQuantInter(pParam, frame, pMB, data, qcoeff, 0, limit); |
|
const int iQuant, |
|
|
int16_t data[6 * 64], |
|
|
int16_t qcoeff[6 * 64], |
|
|
const uint8_t cbp) |
|
|
{ |
|
|
int i; |
|
468 |
|
|
469 |
for (i = 0; i < 6; i++) { |
/* DeQuantize the block */ |
470 |
if (cbp & (1 << (5 - i))) |
MBDeQuantInter(pParam, pMB->quant, data, qcoeff, cbp); |
|
{ |
|
|
if (!(pParam->vol_flags & XVID_VOL_MPEGQUANT)) { |
|
|
start_timer(); |
|
|
dequant_inter(&data[i * 64], &qcoeff[i * 64], iQuant); |
|
|
stop_iquant_timer(); |
|
|
} else { |
|
|
start_timer(); |
|
|
dequant4_inter(&data[i * 64], &qcoeff[i * 64], iQuant); |
|
|
stop_iquant_timer(); |
|
|
} |
|
|
} |
|
|
} |
|
|
} |
|
471 |
|
|
472 |
void |
/* Perform inverse DCT*/ |
473 |
MBiDCT( int16_t data[6 * 64], |
MBiDCT(data, cbp); |
|
const uint8_t cbp) |
|
|
{ |
|
|
int i; |
|
474 |
|
|
475 |
for (i = 0; i < 6; i++) { |
/* Transfer back the data -- Add the data */ |
476 |
if (cbp & (1 << (5 - i))) |
MBTrans16to8(pParam, frame, pMB, x_pos, y_pos, data, 1, cbp); |
|
{ |
|
|
start_timer(); |
|
|
idct(&data[i * 64]); |
|
|
stop_idct_timer(); |
|
477 |
|
|
478 |
} |
return(cbp); |
|
} |
|
479 |
} |
} |
480 |
|
|
481 |
|
uint8_t |
482 |
void |
MBTransQuantInterBVOP(const MBParam * pParam, |
|
MBTrans(const MBParam * pParam, |
|
|
FRAMEINFO * frame, |
|
|
MACROBLOCK * pMB, |
|
|
const uint32_t x_pos, |
|
|
const uint32_t y_pos, |
|
|
int16_t data[6 * 64]) |
|
|
{ |
|
|
uint32_t stride = pParam->edged_width; |
|
|
uint32_t stride2 = stride / 2; |
|
|
uint32_t next_block = stride * 8; |
|
|
uint8_t *pY_Cur, *pU_Cur, *pV_Cur; |
|
|
IMAGE *pCurrent = &frame->image; |
|
|
|
|
|
pY_Cur = pCurrent->y + (y_pos << 4) * stride + (x_pos << 4); |
|
|
pU_Cur = pCurrent->u + (y_pos << 3) * stride2 + (x_pos << 3); |
|
|
pV_Cur = pCurrent->v + (y_pos << 3) * stride2 + (x_pos << 3); |
|
|
|
|
|
start_timer(); |
|
|
transfer_8to16copy(&data[0 * 64], pY_Cur, stride); |
|
|
transfer_8to16copy(&data[1 * 64], pY_Cur + 8, stride); |
|
|
transfer_8to16copy(&data[2 * 64], pY_Cur + next_block, stride); |
|
|
transfer_8to16copy(&data[3 * 64], pY_Cur + next_block + 8, stride); |
|
|
transfer_8to16copy(&data[4 * 64], pU_Cur, stride2); |
|
|
transfer_8to16copy(&data[5 * 64], pV_Cur, stride2); |
|
|
stop_transfer_timer(); |
|
|
} |
|
|
|
|
|
void |
|
|
MBTransAdd(const MBParam * pParam, |
|
483 |
FRAMEINFO * frame, |
FRAMEINFO * frame, |
484 |
MACROBLOCK * pMB, |
MACROBLOCK * pMB, |
485 |
const uint32_t x_pos, |
const uint32_t x_pos, |
486 |
const uint32_t y_pos, |
const uint32_t y_pos, |
487 |
int16_t data[6 * 64], |
int16_t data[6 * 64], |
488 |
const uint8_t cbp) |
int16_t qcoeff[6 * 64]) |
489 |
{ |
{ |
490 |
uint8_t *pY_Cur, *pU_Cur, *pV_Cur; |
uint8_t cbp; |
491 |
uint32_t stride = pParam->edged_width; |
uint32_t limit; |
|
uint32_t stride2 = stride / 2; |
|
|
uint32_t next_block = stride * 8; |
|
|
IMAGE *pCurrent = &frame->image; |
|
492 |
|
|
493 |
pY_Cur = pCurrent->y + (y_pos << 4) * stride + (x_pos << 4); |
/* |
494 |
pU_Cur = pCurrent->u + (y_pos << 3) * stride2 + (x_pos << 3); |
* There is no MBTrans8to16 for Inter block, that's done in motion compensation |
495 |
pV_Cur = pCurrent->v + (y_pos << 3) * stride2 + (x_pos << 3); |
* already |
496 |
|
*/ |
497 |
|
|
498 |
if (pMB->field_dct) { |
/* Perform DCT (and field decision) */ |
499 |
next_block = stride; |
MBfDCT(pParam, frame, pMB, x_pos, y_pos, data); |
|
stride *= 2; |
|
|
} |
|
500 |
|
|
501 |
start_timer(); |
/* Set the limit threshold */ |
502 |
if (cbp & 32) |
limit = BVOP_TOOSMALL_LIMIT; |
|
transfer_16to8add(pY_Cur, &data[0 * 64], stride); |
|
|
if (cbp & 16) |
|
|
transfer_16to8add(pY_Cur + 8, &data[1 * 64], stride); |
|
|
if (cbp & 8) |
|
|
transfer_16to8add(pY_Cur + next_block, &data[2 * 64], stride); |
|
|
if (cbp & 4) |
|
|
transfer_16to8add(pY_Cur + next_block + 8, &data[3 * 64], stride); |
|
|
if (cbp & 2) |
|
|
transfer_16to8add(pU_Cur, &data[4 * 64], stride2); |
|
|
if (cbp & 1) |
|
|
transfer_16to8add(pV_Cur, &data[5 * 64], stride2); |
|
|
stop_transfer_timer(); |
|
|
} |
|
503 |
|
|
504 |
|
if (frame->vop_flags & XVID_VOP_CARTOON) |
505 |
|
limit *= 2; |
506 |
|
|
507 |
|
/* Quantize the block */ |
508 |
|
cbp = MBQuantInter(pParam, frame, pMB, data, qcoeff, 1, limit); |
509 |
|
|
510 |
/* permute block and return field dct choice */ |
/* |
511 |
|
* History comment: |
512 |
|
* We don't have to DeQuant, iDCT and Transfer back data for B-frames. |
513 |
|
* |
514 |
|
* BUT some plugins require the original frame to be passed so we have |
515 |
|
* to take care of that here |
516 |
|
*/ |
517 |
|
if((pParam->plugin_flags & XVID_REQORIGINAL)) { |
518 |
|
|
519 |
|
/* DeQuantize the block */ |
520 |
|
MBDeQuantInter(pParam, pMB->quant, data, qcoeff, cbp); |
521 |
|
|
522 |
uint32_t |
/* Perform inverse DCT*/ |
523 |
MBDecideFieldDCT(int16_t data[6 * 64]) |
MBiDCT(data, cbp); |
|
{ |
|
|
uint32_t field = MBFieldTest(data); |
|
524 |
|
|
525 |
if (field) { |
/* Transfer back the data -- Add the data */ |
526 |
MBFrameToField(data); |
MBTrans16to8(pParam, frame, pMB, x_pos, y_pos, data, 1, cbp); |
527 |
} |
} |
528 |
|
|
529 |
return field; |
return(cbp); |
530 |
} |
} |
531 |
|
|
|
|
|
532 |
/* if sum(diff between field lines) < sum(diff between frame lines), use field dct */ |
/* if sum(diff between field lines) < sum(diff between frame lines), use field dct */ |
|
|
|
533 |
uint32_t |
uint32_t |
534 |
MBFieldTest_c(int16_t data[6 * 64]) |
MBFieldTest_c(int16_t data[6 * 64]) |
535 |
{ |
{ |
543 |
for (i = 0; i < 7; ++i) { |
for (i = 0; i < 7; ++i) { |
544 |
for (j = 0; j < 8; ++j) { |
for (j = 0; j < 8; ++j) { |
545 |
frame += |
frame += |
546 |
ABS(data[0 * 64 + (i + 1) * 8 + j] - data[0 * 64 + i * 8 + j]); |
abs(data[0 * 64 + (i + 1) * 8 + j] - data[0 * 64 + i * 8 + j]); |
547 |
frame += |
frame += |
548 |
ABS(data[1 * 64 + (i + 1) * 8 + j] - data[1 * 64 + i * 8 + j]); |
abs(data[1 * 64 + (i + 1) * 8 + j] - data[1 * 64 + i * 8 + j]); |
549 |
frame += |
frame += |
550 |
ABS(data[2 * 64 + (i + 1) * 8 + j] - data[2 * 64 + i * 8 + j]); |
abs(data[2 * 64 + (i + 1) * 8 + j] - data[2 * 64 + i * 8 + j]); |
551 |
frame += |
frame += |
552 |
ABS(data[3 * 64 + (i + 1) * 8 + j] - data[3 * 64 + i * 8 + j]); |
abs(data[3 * 64 + (i + 1) * 8 + j] - data[3 * 64 + i * 8 + j]); |
553 |
|
|
554 |
field += |
field += |
555 |
ABS(data[blocks[i + 1] + lines[i + 1] + j] - |
abs(data[blocks[i + 1] + lines[i + 1] + j] - |
556 |
data[blocks[i] + lines[i] + j]); |
data[blocks[i] + lines[i] + j]); |
557 |
field += |
field += |
558 |
ABS(data[blocks[i + 1] + lines[i + 1] + 8 + j] - |
abs(data[blocks[i + 1] + lines[i + 1] + 8 + j] - |
559 |
data[blocks[i] + lines[i] + 8 + j]); |
data[blocks[i] + lines[i] + 8 + j]); |
560 |
field += |
field += |
561 |
ABS(data[blocks[i + 1] + 64 + lines[i + 1] + j] - |
abs(data[blocks[i + 1] + 64 + lines[i + 1] + j] - |
562 |
data[blocks[i] + 64 + lines[i] + j]); |
data[blocks[i] + 64 + lines[i] + j]); |
563 |
field += |
field += |
564 |
ABS(data[blocks[i + 1] + 64 + lines[i + 1] + 8 + j] - |
abs(data[blocks[i + 1] + 64 + lines[i + 1] + 8 + j] - |
565 |
data[blocks[i] + 64 + lines[i] + 8 + j]); |
data[blocks[i] + 64 + lines[i] + 8 + j]); |
566 |
} |
} |
567 |
} |
} |
582 |
|
|
583 |
/* left blocks */ |
/* left blocks */ |
584 |
|
|
585 |
// 1=2, 2=4, 4=8, 8=1 |
/* 1=2, 2=4, 4=8, 8=1 */ |
586 |
MOVLINE(tmp, LINE(0, 1)); |
MOVLINE(tmp, LINE(0, 1)); |
587 |
MOVLINE(LINE(0, 1), LINE(0, 2)); |
MOVLINE(LINE(0, 1), LINE(0, 2)); |
588 |
MOVLINE(LINE(0, 2), LINE(0, 4)); |
MOVLINE(LINE(0, 2), LINE(0, 4)); |
589 |
MOVLINE(LINE(0, 4), LINE(2, 0)); |
MOVLINE(LINE(0, 4), LINE(2, 0)); |
590 |
MOVLINE(LINE(2, 0), tmp); |
MOVLINE(LINE(2, 0), tmp); |
591 |
|
|
592 |
// 3=6, 6=12, 12=9, 9=3 |
/* 3=6, 6=12, 12=9, 9=3 */ |
593 |
MOVLINE(tmp, LINE(0, 3)); |
MOVLINE(tmp, LINE(0, 3)); |
594 |
MOVLINE(LINE(0, 3), LINE(0, 6)); |
MOVLINE(LINE(0, 3), LINE(0, 6)); |
595 |
MOVLINE(LINE(0, 6), LINE(2, 4)); |
MOVLINE(LINE(0, 6), LINE(2, 4)); |
596 |
MOVLINE(LINE(2, 4), LINE(2, 1)); |
MOVLINE(LINE(2, 4), LINE(2, 1)); |
597 |
MOVLINE(LINE(2, 1), tmp); |
MOVLINE(LINE(2, 1), tmp); |
598 |
|
|
599 |
// 5=10, 10=5 |
/* 5=10, 10=5 */ |
600 |
MOVLINE(tmp, LINE(0, 5)); |
MOVLINE(tmp, LINE(0, 5)); |
601 |
MOVLINE(LINE(0, 5), LINE(2, 2)); |
MOVLINE(LINE(0, 5), LINE(2, 2)); |
602 |
MOVLINE(LINE(2, 2), tmp); |
MOVLINE(LINE(2, 2), tmp); |
603 |
|
|
604 |
// 7=14, 14=13, 13=11, 11=7 |
/* 7=14, 14=13, 13=11, 11=7 */ |
605 |
MOVLINE(tmp, LINE(0, 7)); |
MOVLINE(tmp, LINE(0, 7)); |
606 |
MOVLINE(LINE(0, 7), LINE(2, 6)); |
MOVLINE(LINE(0, 7), LINE(2, 6)); |
607 |
MOVLINE(LINE(2, 6), LINE(2, 5)); |
MOVLINE(LINE(2, 6), LINE(2, 5)); |
610 |
|
|
611 |
/* right blocks */ |
/* right blocks */ |
612 |
|
|
613 |
// 1=2, 2=4, 4=8, 8=1 |
/* 1=2, 2=4, 4=8, 8=1 */ |
614 |
MOVLINE(tmp, LINE(1, 1)); |
MOVLINE(tmp, LINE(1, 1)); |
615 |
MOVLINE(LINE(1, 1), LINE(1, 2)); |
MOVLINE(LINE(1, 1), LINE(1, 2)); |
616 |
MOVLINE(LINE(1, 2), LINE(1, 4)); |
MOVLINE(LINE(1, 2), LINE(1, 4)); |
617 |
MOVLINE(LINE(1, 4), LINE(3, 0)); |
MOVLINE(LINE(1, 4), LINE(3, 0)); |
618 |
MOVLINE(LINE(3, 0), tmp); |
MOVLINE(LINE(3, 0), tmp); |
619 |
|
|
620 |
// 3=6, 6=12, 12=9, 9=3 |
/* 3=6, 6=12, 12=9, 9=3 */ |
621 |
MOVLINE(tmp, LINE(1, 3)); |
MOVLINE(tmp, LINE(1, 3)); |
622 |
MOVLINE(LINE(1, 3), LINE(1, 6)); |
MOVLINE(LINE(1, 3), LINE(1, 6)); |
623 |
MOVLINE(LINE(1, 6), LINE(3, 4)); |
MOVLINE(LINE(1, 6), LINE(3, 4)); |
624 |
MOVLINE(LINE(3, 4), LINE(3, 1)); |
MOVLINE(LINE(3, 4), LINE(3, 1)); |
625 |
MOVLINE(LINE(3, 1), tmp); |
MOVLINE(LINE(3, 1), tmp); |
626 |
|
|
627 |
// 5=10, 10=5 |
/* 5=10, 10=5 */ |
628 |
MOVLINE(tmp, LINE(1, 5)); |
MOVLINE(tmp, LINE(1, 5)); |
629 |
MOVLINE(LINE(1, 5), LINE(3, 2)); |
MOVLINE(LINE(1, 5), LINE(3, 2)); |
630 |
MOVLINE(LINE(3, 2), tmp); |
MOVLINE(LINE(3, 2), tmp); |
631 |
|
|
632 |
// 7=14, 14=13, 13=11, 11=7 |
/* 7=14, 14=13, 13=11, 11=7 */ |
633 |
MOVLINE(tmp, LINE(1, 7)); |
MOVLINE(tmp, LINE(1, 7)); |
634 |
MOVLINE(LINE(1, 7), LINE(3, 6)); |
MOVLINE(LINE(1, 7), LINE(3, 6)); |
635 |
MOVLINE(LINE(3, 6), LINE(3, 5)); |
MOVLINE(LINE(3, 6), LINE(3, 5)); |
636 |
MOVLINE(LINE(3, 5), LINE(3, 3)); |
MOVLINE(LINE(3, 5), LINE(3, 3)); |
637 |
MOVLINE(LINE(3, 3), tmp); |
MOVLINE(LINE(3, 3), tmp); |
638 |
} |
} |
639 |
|
|
640 |
|
|
641 |
|
|
642 |
|
|
643 |
|
|
644 |
|
/***************************************************************************** |
645 |
|
* Trellis based R-D optimal quantization |
646 |
|
* |
647 |
|
* Trellis Quant code (C) 2003 Pascal Massimino skal(at)planet-d.net |
648 |
|
* |
649 |
|
****************************************************************************/ |
650 |
|
|
651 |
|
|
652 |
|
#if 0 |
653 |
|
static int |
654 |
|
dct_quantize_trellis_mpeg_c(int16_t *const Out, |
655 |
|
const int16_t *const In, |
656 |
|
int Q, |
657 |
|
const uint16_t * const Zigzag, |
658 |
|
int Non_Zero) |
659 |
|
{ |
660 |
|
return 63; |
661 |
|
} |
662 |
|
#endif |
663 |
|
|
664 |
|
/*---------------------------------------------------------------------------- |
665 |
|
* |
666 |
|
* Trellis-Based quantization |
667 |
|
* |
668 |
|
* So far I understand this paper: |
669 |
|
* |
670 |
|
* "Trellis-Based R-D Optimal Quantization in H.263+" |
671 |
|
* J.Wen, M.Luttrell, J.Villasenor |
672 |
|
* IEEE Transactions on Image Processing, Vol.9, No.8, Aug. 2000. |
673 |
|
* |
674 |
|
* we are at stake with a simplified Bellmand-Ford / Dijkstra Single |
675 |
|
* Source Shorted Path algo. But due to the underlying graph structure |
676 |
|
* ("Trellis"), it can be turned into a dynamic programming algo, |
677 |
|
* partially saving the explicit graph's nodes representation. And |
678 |
|
* without using a heap, since the open frontier of the DAG is always |
679 |
|
* known, and of fixed sized. |
680 |
|
*--------------------------------------------------------------------------*/ |
681 |
|
|
682 |
|
|
683 |
|
|
684 |
|
/* Codes lengths for relevant levels. */ |
685 |
|
|
686 |
|
/* let's factorize: */ |
687 |
|
static const uint8_t Code_Len0[64] = { |
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 |
|
30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 }; |
690 |
|
static const uint8_t Code_Len1[64] = { |
691 |
|
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, |
692 |
|
30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 }; |
693 |
|
static const uint8_t Code_Len2[64] = { |
694 |
|
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, |
695 |
|
30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 }; |
696 |
|
static const uint8_t Code_Len3[64] = { |
697 |
|
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, |
698 |
|
30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 }; |
699 |
|
static const uint8_t Code_Len4[64] = { |
700 |
|
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, |
701 |
|
30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 }; |
702 |
|
static const uint8_t Code_Len5[64] = { |
703 |
|
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, |
704 |
|
30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 }; |
705 |
|
static const uint8_t Code_Len6[64] = { |
706 |
|
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, |
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_Len7[64] = { |
709 |
|
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, |
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_Len8[64] = { |
712 |
|
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, |
713 |
|
30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 }; |
714 |
|
static const uint8_t Code_Len9[64] = { |
715 |
|
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, |
716 |
|
30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 }; |
717 |
|
static const uint8_t Code_Len10[64] = { |
718 |
|
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, |
719 |
|
30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 }; |
720 |
|
static const uint8_t Code_Len11[64] = { |
721 |
|
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, |
722 |
|
30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 }; |
723 |
|
static const uint8_t Code_Len12[64] = { |
724 |
|
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, |
725 |
|
30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 }; |
726 |
|
static const uint8_t Code_Len13[64] = { |
727 |
|
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, |
728 |
|
30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 }; |
729 |
|
static const uint8_t Code_Len14[64] = { |
730 |
|
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, |
731 |
|
30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 }; |
732 |
|
static const uint8_t Code_Len15[64] = { |
733 |
|
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, |
734 |
|
30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 }; |
735 |
|
static const uint8_t Code_Len16[64] = { |
736 |
|
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, |
737 |
|
30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30}; |
738 |
|
static const uint8_t Code_Len17[64] = { |
739 |
|
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, |
740 |
|
30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 }; |
741 |
|
static const uint8_t Code_Len18[64] = { |
742 |
|
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, |
743 |
|
30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 }; |
744 |
|
static const uint8_t Code_Len19[64] = { |
745 |
|
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, |
746 |
|
30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30 }; |
747 |
|
static const uint8_t Code_Len20[64] = { |
748 |
|
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, |
749 |
|
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 }; |
750 |
|
|
751 |
|
/* a few more table for LAST table: */ |
752 |
|
static const uint8_t Code_Len21[64] = { |
753 |
|
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, |
754 |
|
30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30}; |
755 |
|
static const uint8_t Code_Len22[64] = { |
756 |
|
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, |
757 |
|
30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30,30}; |
758 |
|
static const uint8_t Code_Len23[64] = { |
759 |
|
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, |
760 |
|
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}; |
761 |
|
static const uint8_t Code_Len24[64] = { |
762 |
|
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, |
763 |
|
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}; |
764 |
|
|
765 |
|
|
766 |
|
static const uint8_t * const B16_17_Code_Len[24] = { /* levels [1..24] */ |
767 |
|
Code_Len20,Code_Len19,Code_Len18,Code_Len17, |
768 |
|
Code_Len16,Code_Len15,Code_Len14,Code_Len13, |
769 |
|
Code_Len12,Code_Len11,Code_Len10,Code_Len9, |
770 |
|
Code_Len8, Code_Len7 ,Code_Len6 ,Code_Len5, |
771 |
|
Code_Len4, Code_Len3, Code_Len3 ,Code_Len2, |
772 |
|
Code_Len2, Code_Len1, Code_Len1, Code_Len1, |
773 |
|
}; |
774 |
|
|
775 |
|
static const uint8_t * const B16_17_Code_Len_Last[6] = { /* levels [1..6] */ |
776 |
|
Code_Len24,Code_Len23,Code_Len22,Code_Len21, Code_Len3, Code_Len1, |
777 |
|
}; |
778 |
|
|
779 |
|
#define TL(q) 0xfe00/(q*q) |
780 |
|
|
781 |
|
static const int Trellis_Lambda_Tabs[31] = { |
782 |
|
TL( 1),TL( 2),TL( 3),TL( 4),TL( 5),TL( 6), TL( 7), |
783 |
|
TL( 8),TL( 9),TL(10),TL(11),TL(12),TL(13),TL(14), TL(15), |
784 |
|
TL(16),TL(17),TL(18),TL(19),TL(20),TL(21),TL(22), TL(23), |
785 |
|
TL(24),TL(25),TL(26),TL(27),TL(28),TL(29),TL(30), TL(31) |
786 |
|
}; |
787 |
|
#undef TL |
788 |
|
|
789 |
|
static int __inline |
790 |
|
Find_Last(const int16_t *C, const uint16_t *Zigzag, int i) |
791 |
|
{ |
792 |
|
while(i>=0) |
793 |
|
if (C[Zigzag[i]]) |
794 |
|
return i; |
795 |
|
else i--; |
796 |
|
return -1; |
797 |
|
} |
798 |
|
|
799 |
|
static int __inline |
800 |
|
Compute_Sum(const int16_t *C, int last) |
801 |
|
{ |
802 |
|
int sum = 0; |
803 |
|
|
804 |
|
while(last--) |
805 |
|
sum += abs(C[last]); |
806 |
|
|
807 |
|
return(sum); |
808 |
|
} |
809 |
|
/* this routine has been strippen of all debug code */ |
810 |
|
|
811 |
|
static int |
812 |
|
dct_quantize_trellis_h263_c(int16_t *const Out, const int16_t *const In, int Q, const uint16_t * const Zigzag, int Non_Zero) |
813 |
|
{ |
814 |
|
|
815 |
|
/* |
816 |
|
* Note: We should search last non-zero coeffs on *real* DCT input coeffs (In[]), |
817 |
|
* not quantized one (Out[]). However, it only improves the result *very* |
818 |
|
* slightly (~0.01dB), whereas speed drops to crawling level :) |
819 |
|
* Well, actually, taking 1 more coeff past Non_Zero into account sometimes helps. |
820 |
|
*/ |
821 |
|
typedef struct { int16_t Run, Level; } NODE; |
822 |
|
|
823 |
|
NODE Nodes[65], Last; |
824 |
|
uint32_t Run_Costs0[64+1]; |
825 |
|
uint32_t * const Run_Costs = Run_Costs0 + 1; |
826 |
|
const int Mult = 2*Q; |
827 |
|
const int Bias = (Q-1) | 1; |
828 |
|
const int Lev0 = Mult + Bias; |
829 |
|
const int Lambda = Trellis_Lambda_Tabs[Q-1]; /* it's 1/lambda, actually */ |
830 |
|
|
831 |
|
int Run_Start = -1; |
832 |
|
uint32_t Min_Cost = 2<<16; |
833 |
|
|
834 |
|
int Last_Node = -1; |
835 |
|
uint32_t Last_Cost = 0; |
836 |
|
|
837 |
|
int i, j, sum; |
838 |
|
Run_Costs[-1] = 2<<16; /* source (w/ CBP penalty) */ |
839 |
|
|
840 |
|
Non_Zero = Find_Last(Out, Zigzag, Non_Zero); |
841 |
|
if (Non_Zero<0) |
842 |
|
return 0; /* Sum is zero if there are only zero coeffs */ |
843 |
|
|
844 |
|
for(i=0; i<=Non_Zero; i++) { |
845 |
|
const int AC = In[Zigzag[i]]; |
846 |
|
const int Level1 = Out[Zigzag[i]]; |
847 |
|
const int Dist0 = Lambda* AC*AC; |
848 |
|
uint32_t Best_Cost = 0xf0000000; |
849 |
|
Last_Cost += Dist0; |
850 |
|
|
851 |
|
/* very specialized loop for -1,0,+1 */ |
852 |
|
if ((uint32_t)(Level1+1)<3) { |
853 |
|
int dQ; |
854 |
|
int Run; |
855 |
|
uint32_t Cost0; |
856 |
|
|
857 |
|
if (AC<0) { |
858 |
|
Nodes[i].Level = -1; |
859 |
|
dQ = Lev0 + AC; |
860 |
|
} else { |
861 |
|
Nodes[i].Level = 1; |
862 |
|
dQ = Lev0 - AC; |
863 |
|
} |
864 |
|
Cost0 = Lambda*dQ*dQ; |
865 |
|
|
866 |
|
Nodes[i].Run = 1; |
867 |
|
Best_Cost = (Code_Len20[0]<<16) + Run_Costs[i-1]+Cost0; |
868 |
|
for(Run=i-Run_Start; Run>0; --Run) { |
869 |
|
const uint32_t Cost_Base = Cost0 + Run_Costs[i-Run]; |
870 |
|
const uint32_t Cost = Cost_Base + (Code_Len20[Run-1]<<16); |
871 |
|
const uint32_t lCost = Cost_Base + (Code_Len24[Run-1]<<16); |
872 |
|
|
873 |
|
/* |
874 |
|
* TODO: what about tie-breaks? Should we favor short runs or |
875 |
|
* long runs? Although the error is the same, it would not be |
876 |
|
* spread the same way along high and low frequencies... |
877 |
|
*/ |
878 |
|
|
879 |
|
/* (I'd say: favour short runs => hifreq errors (HVS) -- gruel ) */ |
880 |
|
|
881 |
|
if (Cost<Best_Cost) { |
882 |
|
Best_Cost = Cost; |
883 |
|
Nodes[i].Run = Run; |
884 |
|
} |
885 |
|
|
886 |
|
if (lCost<Last_Cost) { |
887 |
|
Last_Cost = lCost; |
888 |
|
Last.Run = Run; |
889 |
|
Last_Node = i; |
890 |
|
} |
891 |
|
} |
892 |
|
if (Last_Node==i) |
893 |
|
Last.Level = Nodes[i].Level; |
894 |
|
} else { /* "big" levels */ |
895 |
|
const uint8_t *Tbl_L1, *Tbl_L2, *Tbl_L1_Last, *Tbl_L2_Last; |
896 |
|
int Level2; |
897 |
|
int dQ1, dQ2; |
898 |
|
int Run; |
899 |
|
uint32_t Dist1,Dist2; |
900 |
|
int dDist21; |
901 |
|
|
902 |
|
if (Level1>1) { |
903 |
|
dQ1 = Level1*Mult-AC + Bias; |
904 |
|
dQ2 = dQ1 - Mult; |
905 |
|
Level2 = Level1-1; |
906 |
|
Tbl_L1 = (Level1<=24) ? B16_17_Code_Len[Level1-1] : Code_Len0; |
907 |
|
Tbl_L2 = (Level2<=24) ? B16_17_Code_Len[Level2-1] : Code_Len0; |
908 |
|
Tbl_L1_Last = (Level1<=6) ? B16_17_Code_Len_Last[Level1-1] : Code_Len0; |
909 |
|
Tbl_L2_Last = (Level2<=6) ? B16_17_Code_Len_Last[Level2-1] : Code_Len0; |
910 |
|
} else { /* Level1<-1 */ |
911 |
|
dQ1 = Level1*Mult-AC - Bias; |
912 |
|
dQ2 = dQ1 + Mult; |
913 |
|
Level2 = Level1 + 1; |
914 |
|
Tbl_L1 = (Level1>=-24) ? B16_17_Code_Len[Level1^-1] : Code_Len0; |
915 |
|
Tbl_L2 = (Level2>=-24) ? B16_17_Code_Len[Level2^-1] : Code_Len0; |
916 |
|
Tbl_L1_Last = (Level1>=- 6) ? B16_17_Code_Len_Last[Level1^-1] : Code_Len0; |
917 |
|
Tbl_L2_Last = (Level2>=- 6) ? B16_17_Code_Len_Last[Level2^-1] : Code_Len0; |
918 |
|
} |
919 |
|
|
920 |
|
Dist1 = Lambda*dQ1*dQ1; |
921 |
|
Dist2 = Lambda*dQ2*dQ2; |
922 |
|
dDist21 = Dist2-Dist1; |
923 |
|
|
924 |
|
for(Run=i-Run_Start; Run>0; --Run) |
925 |
|
{ |
926 |
|
const uint32_t Cost_Base = Dist1 + Run_Costs[i-Run]; |
927 |
|
uint32_t Cost1, Cost2; |
928 |
|
int bLevel; |
929 |
|
|
930 |
|
/* |
931 |
|
* for sub-optimal (but slightly worth it, speed-wise) search, uncomment the following: |
932 |
|
* if (Cost_Base>=Best_Cost) continue; |
933 |
|
* (? doesn't seem to have any effect -- gruel ) |
934 |
|
*/ |
935 |
|
|
936 |
|
Cost1 = Cost_Base + (Tbl_L1[Run-1]<<16); |
937 |
|
Cost2 = Cost_Base + (Tbl_L2[Run-1]<<16) + dDist21; |
938 |
|
|
939 |
|
if (Cost2<Cost1) { |
940 |
|
Cost1 = Cost2; |
941 |
|
bLevel = Level2; |
942 |
|
} else { |
943 |
|
bLevel = Level1; |
944 |
|
} |
945 |
|
|
946 |
|
if (Cost1<Best_Cost) { |
947 |
|
Best_Cost = Cost1; |
948 |
|
Nodes[i].Run = Run; |
949 |
|
Nodes[i].Level = bLevel; |
950 |
|
} |
951 |
|
|
952 |
|
Cost1 = Cost_Base + (Tbl_L1_Last[Run-1]<<16); |
953 |
|
Cost2 = Cost_Base + (Tbl_L2_Last[Run-1]<<16) + dDist21; |
954 |
|
|
955 |
|
if (Cost2<Cost1) { |
956 |
|
Cost1 = Cost2; |
957 |
|
bLevel = Level2; |
958 |
|
} else { |
959 |
|
bLevel = Level1; |
960 |
|
} |
961 |
|
|
962 |
|
if (Cost1<Last_Cost) { |
963 |
|
Last_Cost = Cost1; |
964 |
|
Last.Run = Run; |
965 |
|
Last.Level = bLevel; |
966 |
|
Last_Node = i; |
967 |
|
} |
968 |
|
} /* end of "for Run" */ |
969 |
|
|
970 |
|
} |
971 |
|
|
972 |
|
Run_Costs[i] = Best_Cost; |
973 |
|
|
974 |
|
if (Best_Cost < Min_Cost + Dist0) { |
975 |
|
Min_Cost = Best_Cost; |
976 |
|
Run_Start = i; |
977 |
|
} else { |
978 |
|
/* |
979 |
|
* as noticed by Michael Niedermayer (michaelni at gmx.at), there's |
980 |
|
* a code shorter by 1 bit for a larger run (!), same level. We give |
981 |
|
* it a chance by not moving the left barrier too much. |
982 |
|
*/ |
983 |
|
|
984 |
|
while( Run_Costs[Run_Start]>Min_Cost+(1<<16) ) |
985 |
|
Run_Start++; |
986 |
|
|
987 |
|
/* spread on preceding coeffs the cost incurred by skipping this one */ |
988 |
|
for(j=Run_Start; j<i; ++j) Run_Costs[j] += Dist0; |
989 |
|
Min_Cost += Dist0; |
990 |
|
} |
991 |
|
} |
992 |
|
|
993 |
|
/* It seems trellis doesn't give good results... just compute the Out sum and |
994 |
|
* quit (even if we did not modify it, upperlayer relies on this data) */ |
995 |
|
if (Last_Node<0) |
996 |
|
return Compute_Sum(Out, Non_Zero); |
997 |
|
|
998 |
|
/* reconstruct optimal sequence backward with surviving paths */ |
999 |
|
memset(Out, 0x00, 64*sizeof(*Out)); |
1000 |
|
Out[Zigzag[Last_Node]] = Last.Level; |
1001 |
|
i = Last_Node - Last.Run; |
1002 |
|
sum = 0; |
1003 |
|
while(i>=0) { |
1004 |
|
Out[Zigzag[i]] = Nodes[i].Level; |
1005 |
|
sum += abs(Nodes[i].Level); |
1006 |
|
i -= Nodes[i].Run; |
1007 |
|
} |
1008 |
|
|
1009 |
|
return sum; |
1010 |
|
} |
1011 |
|
|
1012 |
|
static int |
1013 |
|
dct_quantize_trellis_mpeg_c(int16_t *const Out, const int16_t *const In, int Q, const uint16_t * const Zigzag, int Non_Zero) |
1014 |
|
{ |
1015 |
|
/* ToDo: Ok ok it's just a place holder for Gruel -- damn write this one :-) */ |
1016 |
|
return Compute_Sum(Out, 63); |
1017 |
|
} |
1018 |
|
|
1019 |
|
/* original version including heavy debugging info */ |
1020 |
|
|
1021 |
|
#ifdef DBGTRELL |
1022 |
|
|
1023 |
|
#define DBG 0 |
1024 |
|
|
1025 |
|
static __inline uint32_t Evaluate_Cost(const int16_t *C, int Mult, int Bias, |
1026 |
|
const uint16_t * Zigzag, int Max, int Lambda) |
1027 |
|
{ |
1028 |
|
#if (DBG>0) |
1029 |
|
const int16_t * const Ref = C + 6*64; |
1030 |
|
int Last = Max; |
1031 |
|
int Bits = 0; |
1032 |
|
int Dist = 0; |
1033 |
|
int i; |
1034 |
|
uint32_t Cost; |
1035 |
|
|
1036 |
|
while(Last>=0 && C[Zigzag[Last]]==0) |
1037 |
|
Last--; |
1038 |
|
|
1039 |
|
if (Last>=0) { |
1040 |
|
int j=0, j0=0; |
1041 |
|
int Run, Level; |
1042 |
|
|
1043 |
|
Bits = 2; /* CBP */ |
1044 |
|
while(j<Last) { |
1045 |
|
while(!C[Zigzag[j]]) |
1046 |
|
j++; |
1047 |
|
if (j==Last) |
1048 |
|
break; |
1049 |
|
Level=C[Zigzag[j]]; |
1050 |
|
Run = j - j0; |
1051 |
|
j0 = ++j; |
1052 |
|
if (Level>=-24 && Level<=24) |
1053 |
|
Bits += B16_17_Code_Len[(Level<0) ? -Level-1 : Level-1][Run]; |
1054 |
|
else |
1055 |
|
Bits += 30; |
1056 |
|
} |
1057 |
|
Level = C[Zigzag[Last]]; |
1058 |
|
Run = j - j0; |
1059 |
|
if (Level>=-6 && Level<=6) |
1060 |
|
Bits += B16_17_Code_Len_Last[(Level<0) ? -Level-1 : Level-1][Run]; |
1061 |
|
else |
1062 |
|
Bits += 30; |
1063 |
|
} |
1064 |
|
|
1065 |
|
for(i=0; i<=Last; ++i) { |
1066 |
|
int V = C[Zigzag[i]]*Mult; |
1067 |
|
if (V>0) |
1068 |
|
V += Bias; |
1069 |
|
else |
1070 |
|
if (V<0) |
1071 |
|
V -= Bias; |
1072 |
|
V -= Ref[Zigzag[i]]; |
1073 |
|
Dist += V*V; |
1074 |
|
} |
1075 |
|
Cost = Lambda*Dist + (Bits<<16); |
1076 |
|
if (DBG==1) |
1077 |
|
printf( " Last:%2d/%2d Cost = [(Bits=%5.0d) + Lambda*(Dist=%6.0d) = %d ] >>12= %d ", Last,Max, Bits, Dist, Cost, Cost>>12 ); |
1078 |
|
return Cost; |
1079 |
|
|
1080 |
|
#else |
1081 |
|
return 0; |
1082 |
|
#endif |
1083 |
|
} |
1084 |
|
|
1085 |
|
|
1086 |
|
static int |
1087 |
|
dct_quantize_trellis_h263_c(int16_t *const Out, const int16_t *const In, int Q, const uint16_t * const Zigzag, int Non_Zero) |
1088 |
|
{ |
1089 |
|
|
1090 |
|
/* |
1091 |
|
* Note: We should search last non-zero coeffs on *real* DCT input coeffs (In[]), |
1092 |
|
* not quantized one (Out[]). However, it only improves the result *very* |
1093 |
|
* slightly (~0.01dB), whereas speed drops to crawling level :) |
1094 |
|
* Well, actually, taking 1 more coeff past Non_Zero into account sometimes helps. |
1095 |
|
*/ |
1096 |
|
typedef struct { int16_t Run, Level; } NODE; |
1097 |
|
|
1098 |
|
NODE Nodes[65], Last; |
1099 |
|
uint32_t Run_Costs0[64+1]; |
1100 |
|
uint32_t * const Run_Costs = Run_Costs0 + 1; |
1101 |
|
const int Mult = 2*Q; |
1102 |
|
const int Bias = (Q-1) | 1; |
1103 |
|
const int Lev0 = Mult + Bias; |
1104 |
|
const int Lambda = Trellis_Lambda_Tabs[Q-1]; /* it's 1/lambda, actually */ |
1105 |
|
|
1106 |
|
int Run_Start = -1; |
1107 |
|
Run_Costs[-1] = 2<<16; /* source (w/ CBP penalty) */ |
1108 |
|
uint32_t Min_Cost = 2<<16; |
1109 |
|
|
1110 |
|
int Last_Node = -1; |
1111 |
|
uint32_t Last_Cost = 0; |
1112 |
|
|
1113 |
|
int i, j; |
1114 |
|
|
1115 |
|
#if (DBG>0) |
1116 |
|
Last.Level = 0; Last.Run = -1; /* just initialize to smthg */ |
1117 |
|
#endif |
1118 |
|
|
1119 |
|
Non_Zero = Find_Last(Out, Zigzag, Non_Zero); |
1120 |
|
if (Non_Zero<0) |
1121 |
|
return -1; |
1122 |
|
|
1123 |
|
for(i=0; i<=Non_Zero; i++) |
1124 |
|
{ |
1125 |
|
const int AC = In[Zigzag[i]]; |
1126 |
|
const int Level1 = Out[Zigzag[i]]; |
1127 |
|
const int Dist0 = Lambda* AC*AC; |
1128 |
|
uint32_t Best_Cost = 0xf0000000; |
1129 |
|
Last_Cost += Dist0; |
1130 |
|
|
1131 |
|
if ((uint32_t)(Level1+1)<3) /* very specialized loop for -1,0,+1 */ |
1132 |
|
{ |
1133 |
|
int dQ; |
1134 |
|
int Run; |
1135 |
|
uint32_t Cost0; |
1136 |
|
|
1137 |
|
if (AC<0) { |
1138 |
|
Nodes[i].Level = -1; |
1139 |
|
dQ = Lev0 + AC; |
1140 |
|
} else { |
1141 |
|
Nodes[i].Level = 1; |
1142 |
|
dQ = Lev0 - AC; |
1143 |
|
} |
1144 |
|
Cost0 = Lambda*dQ*dQ; |
1145 |
|
|
1146 |
|
Nodes[i].Run = 1; |
1147 |
|
Best_Cost = (Code_Len20[0]<<16) + Run_Costs[i-1]+Cost0; |
1148 |
|
for(Run=i-Run_Start; Run>0; --Run) |
1149 |
|
{ |
1150 |
|
const uint32_t Cost_Base = Cost0 + Run_Costs[i-Run]; |
1151 |
|
const uint32_t Cost = Cost_Base + (Code_Len20[Run-1]<<16); |
1152 |
|
const uint32_t lCost = Cost_Base + (Code_Len24[Run-1]<<16); |
1153 |
|
|
1154 |
|
/* |
1155 |
|
* TODO: what about tie-breaks? Should we favor short runs or |
1156 |
|
* long runs? Although the error is the same, it would not be |
1157 |
|
* spread the same way along high and low frequencies... |
1158 |
|
*/ |
1159 |
|
if (Cost<Best_Cost) { |
1160 |
|
Best_Cost = Cost; |
1161 |
|
Nodes[i].Run = Run; |
1162 |
|
} |
1163 |
|
|
1164 |
|
if (lCost<Last_Cost) { |
1165 |
|
Last_Cost = lCost; |
1166 |
|
Last.Run = Run; |
1167 |
|
Last_Node = i; |
1168 |
|
} |
1169 |
|
} |
1170 |
|
if (Last_Node==i) |
1171 |
|
Last.Level = Nodes[i].Level; |
1172 |
|
|
1173 |
|
if (DBG==1) { |
1174 |
|
Run_Costs[i] = Best_Cost; |
1175 |
|
printf( "Costs #%2d: ", i); |
1176 |
|
for(j=-1;j<=Non_Zero;++j) { |
1177 |
|
if (j==Run_Start) printf( " %3.0d|", Run_Costs[j]>>12 ); |
1178 |
|
else if (j>Run_Start && j<i) printf( " %3.0d|", Run_Costs[j]>>12 ); |
1179 |
|
else if (j==i) printf( "(%3.0d)", Run_Costs[j]>>12 ); |
1180 |
|
else printf( " - |" ); |
1181 |
|
} |
1182 |
|
printf( "<%3.0d %2d %d>", Min_Cost>>12, Nodes[i].Level, Nodes[i].Run ); |
1183 |
|
printf( " Last:#%2d {%3.0d %2d %d}", Last_Node, Last_Cost>>12, Last.Level, Last.Run ); |
1184 |
|
printf( " AC:%3.0d Dist0:%3d Dist(%d)=%d", AC, Dist0>>12, Nodes[i].Level, Cost0>>12 ); |
1185 |
|
printf( "\n" ); |
1186 |
|
} |
1187 |
|
} |
1188 |
|
else /* "big" levels */ |
1189 |
|
{ |
1190 |
|
const uint8_t *Tbl_L1, *Tbl_L2, *Tbl_L1_Last, *Tbl_L2_Last; |
1191 |
|
int Level2; |
1192 |
|
int dQ1, dQ2; |
1193 |
|
int Run; |
1194 |
|
uint32_t Dist1,Dist2; |
1195 |
|
int dDist21; |
1196 |
|
|
1197 |
|
if (Level1>1) { |
1198 |
|
dQ1 = Level1*Mult-AC + Bias; |
1199 |
|
dQ2 = dQ1 - Mult; |
1200 |
|
Level2 = Level1-1; |
1201 |
|
Tbl_L1 = (Level1<=24) ? B16_17_Code_Len[Level1-1] : Code_Len0; |
1202 |
|
Tbl_L2 = (Level2<=24) ? B16_17_Code_Len[Level2-1] : Code_Len0; |
1203 |
|
Tbl_L1_Last = (Level1<=6) ? B16_17_Code_Len_Last[Level1-1] : Code_Len0; |
1204 |
|
Tbl_L2_Last = (Level2<=6) ? B16_17_Code_Len_Last[Level2-1] : Code_Len0; |
1205 |
|
} else { /* Level1<-1 */ |
1206 |
|
dQ1 = Level1*Mult-AC - Bias; |
1207 |
|
dQ2 = dQ1 + Mult; |
1208 |
|
Level2 = Level1 + 1; |
1209 |
|
Tbl_L1 = (Level1>=-24) ? B16_17_Code_Len[Level1^-1] : Code_Len0; |
1210 |
|
Tbl_L2 = (Level2>=-24) ? B16_17_Code_Len[Level2^-1] : Code_Len0; |
1211 |
|
Tbl_L1_Last = (Level1>=- 6) ? B16_17_Code_Len_Last[Level1^-1] : Code_Len0; |
1212 |
|
Tbl_L2_Last = (Level2>=- 6) ? B16_17_Code_Len_Last[Level2^-1] : Code_Len0; |
1213 |
|
} |
1214 |
|
Dist1 = Lambda*dQ1*dQ1; |
1215 |
|
Dist2 = Lambda*dQ2*dQ2; |
1216 |
|
dDist21 = Dist2-Dist1; |
1217 |
|
|
1218 |
|
for(Run=i-Run_Start; Run>0; --Run) |
1219 |
|
{ |
1220 |
|
const uint32_t Cost_Base = Dist1 + Run_Costs[i-Run]; |
1221 |
|
uint32_t Cost1, Cost2; |
1222 |
|
int bLevel; |
1223 |
|
|
1224 |
|
/* |
1225 |
|
* for sub-optimal (but slightly worth it, speed-wise) search, uncomment the following: |
1226 |
|
* if (Cost_Base>=Best_Cost) continue; |
1227 |
|
*/ |
1228 |
|
Cost1 = Cost_Base + (Tbl_L1[Run-1]<<16); |
1229 |
|
Cost2 = Cost_Base + (Tbl_L2[Run-1]<<16) + dDist21; |
1230 |
|
|
1231 |
|
if (Cost2<Cost1) { |
1232 |
|
Cost1 = Cost2; |
1233 |
|
bLevel = Level2; |
1234 |
|
} else |
1235 |
|
bLevel = Level1; |
1236 |
|
|
1237 |
|
if (Cost1<Best_Cost) { |
1238 |
|
Best_Cost = Cost1; |
1239 |
|
Nodes[i].Run = Run; |
1240 |
|
Nodes[i].Level = bLevel; |
1241 |
|
} |
1242 |
|
|
1243 |
|
Cost1 = Cost_Base + (Tbl_L1_Last[Run-1]<<16); |
1244 |
|
Cost2 = Cost_Base + (Tbl_L2_Last[Run-1]<<16) + dDist21; |
1245 |
|
|
1246 |
|
if (Cost2<Cost1) { |
1247 |
|
Cost1 = Cost2; |
1248 |
|
bLevel = Level2; |
1249 |
|
} else |
1250 |
|
bLevel = Level1; |
1251 |
|
|
1252 |
|
if (Cost1<Last_Cost) { |
1253 |
|
Last_Cost = Cost1; |
1254 |
|
Last.Run = Run; |
1255 |
|
Last.Level = bLevel; |
1256 |
|
Last_Node = i; |
1257 |
|
} |
1258 |
|
} /* end of "for Run" */ |
1259 |
|
|
1260 |
|
if (DBG==1) { |
1261 |
|
Run_Costs[i] = Best_Cost; |
1262 |
|
printf( "Costs #%2d: ", i); |
1263 |
|
for(j=-1;j<=Non_Zero;++j) { |
1264 |
|
if (j==Run_Start) printf( " %3.0d|", Run_Costs[j]>>12 ); |
1265 |
|
else if (j>Run_Start && j<i) printf( " %3.0d|", Run_Costs[j]>>12 ); |
1266 |
|
else if (j==i) printf( "(%3.0d)", Run_Costs[j]>>12 ); |
1267 |
|
else printf( " - |" ); |
1268 |
|
} |
1269 |
|
printf( "<%3.0d %2d %d>", Min_Cost>>12, Nodes[i].Level, Nodes[i].Run ); |
1270 |
|
printf( " Last:#%2d {%3.0d %2d %d}", Last_Node, Last_Cost>>12, Last.Level, Last.Run ); |
1271 |
|
printf( " AC:%3.0d Dist0:%3d Dist(%2d):%3d Dist(%2d):%3d", AC, Dist0>>12, Level1, Dist1>>12, Level2, Dist2>>12 ); |
1272 |
|
printf( "\n" ); |
1273 |
|
} |
1274 |
|
} |
1275 |
|
|
1276 |
|
Run_Costs[i] = Best_Cost; |
1277 |
|
|
1278 |
|
if (Best_Cost < Min_Cost + Dist0) { |
1279 |
|
Min_Cost = Best_Cost; |
1280 |
|
Run_Start = i; |
1281 |
|
} |
1282 |
|
else |
1283 |
|
{ |
1284 |
|
/* |
1285 |
|
* as noticed by Michael Niedermayer (michaelni at gmx.at), there's |
1286 |
|
* a code shorter by 1 bit for a larger run (!), same level. We give |
1287 |
|
* it a chance by not moving the left barrier too much. |
1288 |
|
*/ |
1289 |
|
|
1290 |
|
while( Run_Costs[Run_Start]>Min_Cost+(1<<16) ) |
1291 |
|
Run_Start++; |
1292 |
|
|
1293 |
|
/* spread on preceding coeffs the cost incurred by skipping this one */ |
1294 |
|
for(j=Run_Start; j<i; ++j) Run_Costs[j] += Dist0; |
1295 |
|
Min_Cost += Dist0; |
1296 |
|
} |
1297 |
|
} |
1298 |
|
|
1299 |
|
if (DBG) { |
1300 |
|
Last_Cost = Evaluate_Cost(Out,Mult,Bias, Zigzag,Non_Zero, Lambda); |
1301 |
|
if (DBG==1) { |
1302 |
|
printf( "=> " ); |
1303 |
|
for(i=0; i<=Non_Zero; ++i) printf( "[%3.0d] ", Out[Zigzag[i]] ); |
1304 |
|
printf( "\n" ); |
1305 |
|
} |
1306 |
|
} |
1307 |
|
|
1308 |
|
if (Last_Node<0) |
1309 |
|
return -1; |
1310 |
|
|
1311 |
|
/* reconstruct optimal sequence backward with surviving paths */ |
1312 |
|
memset(Out, 0x00, 64*sizeof(*Out)); |
1313 |
|
Out[Zigzag[Last_Node]] = Last.Level; |
1314 |
|
i = Last_Node - Last.Run; |
1315 |
|
while(i>=0) { |
1316 |
|
Out[Zigzag[i]] = Nodes[i].Level; |
1317 |
|
i -= Nodes[i].Run; |
1318 |
|
} |
1319 |
|
|
1320 |
|
if (DBG) { |
1321 |
|
uint32_t Cost = Evaluate_Cost(Out,Mult,Bias, Zigzag,Non_Zero, Lambda); |
1322 |
|
if (DBG==1) { |
1323 |
|
printf( "<= " ); |
1324 |
|
for(i=0; i<=Last_Node; ++i) printf( "[%3.0d] ", Out[Zigzag[i]] ); |
1325 |
|
printf( "\n--------------------------------\n" ); |
1326 |
|
} |
1327 |
|
if (Cost>Last_Cost) printf( "!!! %u > %u\n", Cost, Last_Cost ); |
1328 |
|
} |
1329 |
|
return Last_Node; |
1330 |
|
} |
1331 |
|
|
1332 |
|
#undef DBG |
1333 |
|
|
1334 |
|
#endif |