1 |
/************************************************************************** |
/************************************************************************** |
2 |
* |
* |
3 |
* Modifications: |
* XVID MPEG-4 VIDEO CODEC |
4 |
|
* motion estimation |
5 |
* |
* |
6 |
* 25.04.2002 partial prevMB conversion |
* This program is an implementation of a part of one or more MPEG-4 |
7 |
* 22.04.2002 remove some compile warning by chenm001 <chenm001@163.com> |
* Video tools as specified in ISO/IEC 14496-2 standard. Those intending |
8 |
* 14.04.2002 added MotionEstimationBVOP() |
* to use this software module in hardware or software products are |
9 |
* 02.04.2002 add EPZS(^2) as ME algorithm, use PMV_USESQUARES to choose between |
* advised that its use may infringe existing patents or copyrights, and |
10 |
* EPZS and EPZS^2 |
* any such use would be at such party's own risk. The original |
11 |
* 08.02.2002 split up PMVfast into three routines: PMVFast, PMVFast_MainLoop |
* developer of this software module and his/her company, and subsequent |
12 |
* PMVFast_Refine to support multiple searches with different start points |
* editors and their companies, will have no liability for use of this |
13 |
* 07.01.2002 uv-block-based interpolation |
* software or modifications or derivatives thereof. |
|
* 06.01.2002 INTER/INTRA-decision is now done before any SEARCH8 (speedup) |
|
|
* changed INTER_BIAS to 150 (as suggested by suxen_drol) |
|
|
* removed halfpel refinement step in PMVfastSearch8 + quality=5 |
|
|
* added new quality mode = 6 which performs halfpel refinement |
|
|
* filesize difference between quality 5 and 6 is smaller than 1% |
|
|
* (Isibaar) |
|
|
* 31.12.2001 PMVfastSearch16 and PMVfastSearch8 (gruel) |
|
|
* 30.12.2001 get_range/MotionSearchX simplified; blue/green bug fix |
|
|
* 22.12.2001 commented best_point==99 check |
|
|
* 19.12.2001 modified get_range (purple bug fix) |
|
|
* 15.12.2001 moved pmv displacement from mbprediction |
|
|
* 02.12.2001 motion estimation/compensation split (Isibaar) |
|
|
* 16.11.2001 rewrote/tweaked search algorithms; pross@cs.rmit.edu.au |
|
|
* 10.11.2001 support for sad16/sad8 functions |
|
|
* 28.08.2001 reactivated MODE_INTER4V for EXT_MODE |
|
|
* 24.08.2001 removed MODE_INTER4V_Q, disabled MODE_INTER4V for EXT_MODE |
|
|
* 22.08.2001 added MODE_INTER4V_Q |
|
|
* 20.08.2001 added pragma to get rid of internal compiler error with VC6 |
|
|
* idea by Cyril. Thanks. |
|
14 |
* |
* |
15 |
* Michael Militzer <isibaar@videocoding.de> |
* This program is free software; you can redistribute it and/or modify |
16 |
|
* it under the terms of the GNU General Public License as published by |
17 |
|
* the Free Software Foundation; either version 2 of the License, or |
18 |
|
* (at your option) any later version. |
19 |
* |
* |
20 |
**************************************************************************/ |
* This program is distributed in the hope that it will be useful, |
21 |
|
* but WITHOUT ANY WARRANTY; without even the implied warranty of |
22 |
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
23 |
|
* GNU General Public License for more details. |
24 |
|
* |
25 |
|
* You should have received a copy of the GNU General Public License |
26 |
|
* along with this program; if not, write to the Free Software |
27 |
|
* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
28 |
|
* |
29 |
|
*************************************************************************/ |
30 |
|
|
31 |
#include <assert.h> |
#include <assert.h> |
32 |
#include <stdio.h> |
#include <stdio.h> |
33 |
#include <stdlib.h> |
#include <stdlib.h> |
34 |
|
#include <string.h> // memcpy |
35 |
|
#include <math.h> // lrint |
36 |
|
|
37 |
#include "../encoder.h" |
#include "../encoder.h" |
38 |
#include "../utils/mbfunctions.h" |
#include "../utils/mbfunctions.h" |
39 |
#include "../prediction/mbprediction.h" |
#include "../prediction/mbprediction.h" |
40 |
#include "../global.h" |
#include "../global.h" |
41 |
#include "../utils/timer.h" |
#include "../utils/timer.h" |
42 |
|
#include "../image/interpolate8x8.h" |
43 |
|
#include "motion_est.h" |
44 |
#include "motion.h" |
#include "motion.h" |
45 |
#include "sad.h" |
#include "sad.h" |
46 |
|
#include "../utils/emms.h" |
47 |
|
#include "../dct/fdct.h" |
48 |
|
|
49 |
// very large value |
/***************************************************************************** |
50 |
#define MV_MAX_ERROR (4096 * 256) |
* Modified rounding tables -- declared in motion.h |
51 |
|
* Original tables see ISO spec tables 7-6 -> 7-9 |
52 |
// stop search if sdelta < THRESHOLD |
****************************************************************************/ |
53 |
#define MV16_THRESHOLD 192 |
|
54 |
#define MV8_THRESHOLD 56 |
const uint32_t roundtab[16] = |
55 |
|
{0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2 }; |
56 |
/* sad16(0,0) bias; mpeg4 spec suggests nb/2+1 */ |
|
57 |
/* nb = vop pixels * 2^(bpp-8) */ |
/* K = 4 */ |
58 |
#define MV16_00_BIAS (128+1) |
const uint32_t roundtab_76[16] = |
59 |
#define MV8_00_BIAS (0) |
{ 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1 }; |
60 |
|
|
61 |
|
/* K = 2 */ |
62 |
|
const uint32_t roundtab_78[8] = |
63 |
|
{ 0, 0, 1, 1, 0, 0, 0, 1 }; |
64 |
|
|
65 |
|
/* K = 1 */ |
66 |
|
const uint32_t roundtab_79[4] = |
67 |
|
{ 0, 1, 0, 0 }; |
68 |
|
|
69 |
|
#define INITIAL_SKIP_THRESH (10) |
70 |
|
#define FINAL_SKIP_THRESH (50) |
71 |
|
#define MAX_SAD00_FOR_SKIP (20) |
72 |
|
#define MAX_CHROMA_SAD_FOR_SKIP (22) |
73 |
|
|
74 |
|
#define CHECK_CANDIDATE(X,Y,D) { \ |
75 |
|
CheckCandidate((X),(Y), (D), &iDirection, data ); } |
76 |
|
|
77 |
|
/***************************************************************************** |
78 |
|
* Code |
79 |
|
****************************************************************************/ |
80 |
|
|
81 |
|
static __inline uint32_t |
82 |
|
d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv) |
83 |
|
{ |
84 |
|
int bits; |
85 |
|
const int q = (1 << (iFcode - 1)) - 1; |
86 |
|
|
87 |
|
x <<= qpel; |
88 |
|
y <<= qpel; |
89 |
|
if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); } |
90 |
|
|
91 |
|
x -= pred.x; |
92 |
|
bits = (x != 0 ? iFcode:0); |
93 |
|
x = abs(x); |
94 |
|
x += q; |
95 |
|
x >>= (iFcode - 1); |
96 |
|
bits += mvtab[x]; |
97 |
|
|
98 |
|
y -= pred.y; |
99 |
|
bits += (y != 0 ? iFcode:0); |
100 |
|
y = abs(y); |
101 |
|
y += q; |
102 |
|
y >>= (iFcode - 1); |
103 |
|
bits += mvtab[y]; |
104 |
|
|
105 |
|
return bits; |
106 |
|
} |
107 |
|
|
108 |
|
static int32_t ChromaSAD2(const int fx, const int fy, const int bx, const int by, |
109 |
|
const SearchData * const data) |
110 |
|
{ |
111 |
|
int sad; |
112 |
|
const uint32_t stride = data->iEdgedWidth/2; |
113 |
|
uint8_t * f_refu = data->RefQ, |
114 |
|
* f_refv = data->RefQ + 8, |
115 |
|
* b_refu = data->RefQ + 16, |
116 |
|
* b_refv = data->RefQ + 24; |
117 |
|
int offset = (fx>>1) + (fy>>1)*stride; |
118 |
|
|
119 |
|
switch (((fx & 1) << 1) | (fy & 1)) { |
120 |
|
case 0: |
121 |
|
f_refu = (uint8_t*)data->RefP[4] + offset; |
122 |
|
f_refv = (uint8_t*)data->RefP[5] + offset; |
123 |
|
break; |
124 |
|
case 1: |
125 |
|
interpolate8x8_halfpel_v(f_refu, data->RefP[4] + offset, stride, data->rounding); |
126 |
|
interpolate8x8_halfpel_v(f_refv, data->RefP[5] + offset, stride, data->rounding); |
127 |
|
break; |
128 |
|
case 2: |
129 |
|
interpolate8x8_halfpel_h(f_refu, data->RefP[4] + offset, stride, data->rounding); |
130 |
|
interpolate8x8_halfpel_h(f_refv, data->RefP[5] + offset, stride, data->rounding); |
131 |
|
break; |
132 |
|
default: |
133 |
|
interpolate8x8_halfpel_hv(f_refu, data->RefP[4] + offset, stride, data->rounding); |
134 |
|
interpolate8x8_halfpel_hv(f_refv, data->RefP[5] + offset, stride, data->rounding); |
135 |
|
break; |
136 |
|
} |
137 |
|
|
138 |
/* INTER bias for INTER/INTRA decision; mpeg4 spec suggests 2*nb */ |
offset = (bx>>1) + (by>>1)*stride; |
139 |
#define INTER_BIAS 512 |
switch (((bx & 1) << 1) | (by & 1)) { |
140 |
|
case 0: |
141 |
|
b_refu = (uint8_t*)data->b_RefP[4] + offset; |
142 |
|
b_refv = (uint8_t*)data->b_RefP[5] + offset; |
143 |
|
break; |
144 |
|
case 1: |
145 |
|
interpolate8x8_halfpel_v(b_refu, data->b_RefP[4] + offset, stride, data->rounding); |
146 |
|
interpolate8x8_halfpel_v(b_refv, data->b_RefP[5] + offset, stride, data->rounding); |
147 |
|
break; |
148 |
|
case 2: |
149 |
|
interpolate8x8_halfpel_h(b_refu, data->b_RefP[4] + offset, stride, data->rounding); |
150 |
|
interpolate8x8_halfpel_h(b_refv, data->b_RefP[5] + offset, stride, data->rounding); |
151 |
|
break; |
152 |
|
default: |
153 |
|
interpolate8x8_halfpel_hv(b_refu, data->b_RefP[4] + offset, stride, data->rounding); |
154 |
|
interpolate8x8_halfpel_hv(b_refv, data->b_RefP[5] + offset, stride, data->rounding); |
155 |
|
break; |
156 |
|
} |
157 |
|
|
158 |
/* Parameters which control inter/inter4v decision */ |
sad = sad8bi(data->CurU, b_refu, f_refu, stride); |
159 |
#define IMV16X16 5 |
sad += sad8bi(data->CurV, b_refv, f_refv, stride); |
160 |
|
|
161 |
/* vector map (vlc delta size) smoother parameters */ |
return sad; |
162 |
#define NEIGH_TEND_16X16 2 |
} |
|
#define NEIGH_TEND_8X8 2 |
|
163 |
|
|
164 |
|
static int32_t |
165 |
|
ChromaSAD(const int dx, const int dy, const SearchData * const data) |
166 |
|
{ |
167 |
|
int sad; |
168 |
|
const uint32_t stride = data->iEdgedWidth/2; |
169 |
|
int offset = (dx>>1) + (dy>>1)*stride; |
170 |
|
|
171 |
// fast ((A)/2)*2 |
if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently |
172 |
#define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) |
data->temp[5] = dx; data->temp[6] = dy; // backup |
173 |
|
|
174 |
|
switch (((dx & 1) << 1) | (dy & 1)) { |
175 |
|
case 0: |
176 |
|
sad = sad8(data->CurU, data->RefP[4] + offset, stride); |
177 |
|
sad += sad8(data->CurV, data->RefP[5] + offset, stride); |
178 |
|
break; |
179 |
|
case 1: |
180 |
|
sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + stride, stride); |
181 |
|
sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + stride, stride); |
182 |
|
break; |
183 |
|
case 2: |
184 |
|
sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + 1, stride); |
185 |
|
sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + 1, stride); |
186 |
|
break; |
187 |
|
default: |
188 |
|
interpolate8x8_halfpel_hv(data->RefQ, data->RefP[4] + offset, stride, data->rounding); |
189 |
|
sad = sad8(data->CurU, data->RefQ, stride); |
190 |
|
|
191 |
int32_t PMVfastSearch16( |
interpolate8x8_halfpel_hv(data->RefQ, data->RefP[5] + offset, stride, data->rounding); |
192 |
const uint8_t * const pRef, |
sad += sad8(data->CurV, data->RefQ, stride); |
193 |
const uint8_t * const pRefH, |
break; |
194 |
const uint8_t * const pRefV, |
} |
195 |
const uint8_t * const pRefHV, |
data->temp[7] = sad; //backup, part 2 |
196 |
const IMAGE * const pCur, |
return sad; |
197 |
const int x, const int y, |
} |
|
const uint32_t MotionFlags, |
|
|
const uint32_t iQuant, |
|
|
const uint32_t iFcode, |
|
|
const MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const MACROBLOCK * const prevMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV); |
|
198 |
|
|
199 |
int32_t EPZSSearch16( |
static __inline const uint8_t * |
200 |
const uint8_t * const pRef, |
GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data) |
201 |
const uint8_t * const pRefH, |
{ |
202 |
const uint8_t * const pRefV, |
// dir : 0 = forward, 1 = backward |
203 |
const uint8_t * const pRefHV, |
const uint8_t *const *const direction = ( dir == 0 ? data->RefP : data->b_RefP ); |
204 |
const IMAGE * const pCur, |
const int picture = ((x&1)<<1) | (y&1); |
205 |
const int x, const int y, |
const int offset = (x>>1) + (y>>1)*data->iEdgedWidth; |
206 |
const uint32_t MotionFlags, |
return direction[picture] + offset; |
207 |
const uint32_t iQuant, |
} |
208 |
const uint32_t iFcode, |
|
209 |
const MBParam * const pParam, |
// this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate |
210 |
const MACROBLOCK * const pMBs, |
static __inline const uint8_t * |
211 |
const MACROBLOCK * const prevMBs, |
GetReference(const int x, const int y, const SearchData * const data) |
212 |
VECTOR * const currMV, |
{ |
213 |
VECTOR * const currPMV); |
const int picture = ((x&1)<<1) | (y&1); |
214 |
|
const int offset = (x>>1) + (y>>1)*data->iEdgedWidth; |
215 |
|
return data->RefP[picture] + offset; |
216 |
|
} |
217 |
|
|
218 |
|
static uint8_t * |
219 |
|
Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data) |
220 |
|
{ |
221 |
|
// create or find a qpel-precision reference picture; return pointer to it |
222 |
|
uint8_t * Reference = data->RefQ + 16*dir; |
223 |
|
const uint32_t iEdgedWidth = data->iEdgedWidth; |
224 |
|
const uint32_t rounding = data->rounding; |
225 |
|
const int halfpel_x = x/2; |
226 |
|
const int halfpel_y = y/2; |
227 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
228 |
|
|
229 |
|
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
230 |
|
ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
231 |
|
switch( ((x&1)<<1) + (y&1) ) { |
232 |
|
case 3: // x and y in qpel resolution - the "corners" (top left/right and |
233 |
|
// bottom left/right) during qpel refinement |
234 |
|
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
235 |
|
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
236 |
|
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
237 |
|
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
238 |
|
ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
239 |
|
ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
240 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
241 |
|
break; |
242 |
|
|
243 |
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
244 |
|
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
245 |
|
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
246 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
247 |
|
break; |
248 |
|
|
249 |
int32_t PMVfastSearch8( |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
250 |
const uint8_t * const pRef, |
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
251 |
const uint8_t * const pRefH, |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
252 |
const uint8_t * const pRefV, |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
253 |
const uint8_t * const pRefHV, |
break; |
|
const IMAGE * const pCur, |
|
|
const int x, const int y, |
|
|
const int start_x, const int start_y, |
|
|
const uint32_t MotionFlags, |
|
|
const uint32_t iQuant, |
|
|
const uint32_t iFcode, |
|
|
const MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const MACROBLOCK * const prevMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV); |
|
254 |
|
|
255 |
int32_t EPZSSearch8( |
default: // pure halfpel position |
256 |
const uint8_t * const pRef, |
return (uint8_t *) ref1; |
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const IMAGE * const pCur, |
|
|
const int x, const int y, |
|
|
const int start_x, const int start_y, |
|
|
const uint32_t MotionFlags, |
|
|
const uint32_t iQuant, |
|
|
const uint32_t iFcode, |
|
|
const MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const MACROBLOCK * const prevMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV); |
|
257 |
|
|
258 |
|
} |
259 |
|
return Reference; |
260 |
|
} |
261 |
|
|
262 |
typedef int32_t (MainSearch16Func)( |
static uint8_t * |
263 |
const uint8_t * const pRef, |
Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data) |
264 |
const uint8_t * const pRefH, |
{ |
265 |
const uint8_t * const pRefV, |
// create or find a qpel-precision reference picture; return pointer to it |
266 |
const uint8_t * const pRefHV, |
uint8_t * Reference = data->RefQ + 16*dir; |
267 |
const uint8_t * const cur, |
const uint32_t iEdgedWidth = data->iEdgedWidth; |
268 |
const int x, const int y, |
const uint32_t rounding = data->rounding; |
269 |
int32_t startx, int32_t starty, |
const int halfpel_x = x/2; |
270 |
int32_t iMinSAD, |
const int halfpel_y = y/2; |
271 |
VECTOR * const currMV, |
const uint8_t *ref1, *ref2, *ref3, *ref4; |
272 |
const VECTOR * const pmv, |
|
273 |
const int32_t min_dx, const int32_t max_dx, |
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
274 |
const int32_t min_dy, const int32_t max_dy, |
switch( ((x&1)<<1) + (y&1) ) { |
275 |
const int32_t iEdgedWidth, |
case 3: // x and y in qpel resolution - the "corners" (top left/right and |
276 |
const int32_t iDiamondSize, |
// bottom left/right) during qpel refinement |
277 |
const int32_t iFcode, |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
278 |
const int32_t iQuant, |
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
279 |
int iFound); |
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
280 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
281 |
|
interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); |
282 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); |
283 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); |
284 |
|
break; |
285 |
|
|
286 |
typedef MainSearch16Func* MainSearch16FuncPtr; |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
287 |
|
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
288 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
289 |
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
290 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); |
291 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
292 |
|
break; |
293 |
|
|
294 |
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
295 |
|
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
296 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
297 |
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
298 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); |
299 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
300 |
|
break; |
301 |
|
|
302 |
typedef int32_t (MainSearch8Func)( |
default: // pure halfpel position |
303 |
const uint8_t * const pRef, |
return (uint8_t *) ref1; |
304 |
const uint8_t * const pRefH, |
} |
305 |
const uint8_t * const pRefV, |
return Reference; |
306 |
const uint8_t * const pRefHV, |
} |
|
const uint8_t * const cur, |
|
|
const int x, const int y, |
|
|
int32_t startx, int32_t starty, |
|
|
int32_t iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const VECTOR * const pmv, |
|
|
const int32_t min_dx, const int32_t max_dx, |
|
|
const int32_t min_dy, const int32_t max_dy, |
|
|
const int32_t iEdgedWidth, |
|
|
const int32_t iDiamondSize, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
int iFound); |
|
|
|
|
|
typedef MainSearch8Func* MainSearch8FuncPtr; |
|
|
|
|
|
// mv.length table |
|
|
static const uint32_t mvtab[33] = { |
|
|
1, 2, 3, 4, 6, 7, 7, 7, |
|
|
9, 9, 9, 10, 10, 10, 10, 10, |
|
|
10, 10, 10, 10, 10, 10, 10, 10, |
|
|
10, 11, 11, 11, 11, 11, 11, 12, 12 |
|
|
}; |
|
307 |
|
|
308 |
|
/* CHECK_CANDIATE FUNCTIONS START */ |
309 |
|
|
310 |
static __inline uint32_t mv_bits(int32_t component, const uint32_t iFcode) |
static void |
311 |
|
CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
312 |
{ |
{ |
313 |
if (component == 0) |
int xc, yc; |
314 |
return 1; |
const uint8_t * Reference; |
315 |
|
VECTOR * current; |
316 |
if (component < 0) |
int32_t sad; uint32_t t; |
|
component = -component; |
|
317 |
|
|
318 |
if (iFcode == 1) |
if ( (x > data->max_dx) || (x < data->min_dx) |
319 |
{ |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
|
if (component > 32) |
|
|
component = 32; |
|
320 |
|
|
321 |
return mvtab[component] + 1; |
if (!data->qpel_precision) { |
322 |
|
Reference = GetReference(x, y, data); |
323 |
|
current = data->currentMV; |
324 |
|
xc = x; yc = y; |
325 |
|
} else { // x and y are in 1/4 precision |
326 |
|
Reference = Interpolate16x16qpel(x, y, 0, data); |
327 |
|
xc = x/2; yc = y/2; //for chroma sad |
328 |
|
current = data->currentQMV; |
329 |
} |
} |
330 |
|
|
331 |
component += (1 << (iFcode - 1)) - 1; |
sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
332 |
component >>= (iFcode - 1); |
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
333 |
|
|
334 |
if (component > 32) |
sad += (data->lambda16 * t * sad)>>10; |
335 |
component = 32; |
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; |
336 |
|
|
337 |
return mvtab[component] + 1 + iFcode - 1; |
if (data->chroma) sad += ChromaSAD( (xc >> 1) + roundtab_79[xc & 0x3], |
338 |
|
(yc >> 1) + roundtab_79[yc & 0x3], data); |
339 |
|
|
340 |
|
if (sad < data->iMinSAD[0]) { |
341 |
|
data->iMinSAD[0] = sad; |
342 |
|
current[0].x = x; current[0].y = y; |
343 |
|
*dir = Direction; |
344 |
} |
} |
345 |
|
|
346 |
|
if (data->temp[1] < data->iMinSAD[1]) { |
347 |
|
data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y = y; } |
348 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
349 |
|
data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; } |
350 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
351 |
|
data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; } |
352 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
353 |
|
data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; } |
354 |
|
} |
355 |
|
|
356 |
static __inline uint32_t calc_delta_16(const int32_t dx, const int32_t dy, const uint32_t iFcode) |
static void |
357 |
|
CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
358 |
{ |
{ |
359 |
return NEIGH_TEND_16X16 * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); |
int32_t sad; uint32_t t; |
360 |
} |
const uint8_t * Reference; |
361 |
|
VECTOR * current; |
362 |
|
|
363 |
static __inline uint32_t calc_delta_8(const int32_t dx, const int32_t dy, const uint32_t iFcode) |
if ( (x > data->max_dx) || (x < data->min_dx) |
364 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
365 |
|
|
366 |
{ |
if (!data->qpel_precision) { |
367 |
return NEIGH_TEND_8X8 * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); |
Reference = GetReference(x, y, data); |
368 |
|
current = data->currentMV; |
369 |
|
} else { // x and y are in 1/4 precision |
370 |
|
Reference = Interpolate8x8qpel(x, y, 0, 0, data); |
371 |
|
current = data->currentQMV; |
372 |
} |
} |
373 |
|
|
374 |
|
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
375 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
376 |
|
|
377 |
|
sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10; |
378 |
|
|
379 |
|
if (sad < *(data->iMinSAD)) { |
380 |
|
*(data->iMinSAD) = sad; |
381 |
|
current->x = x; current->y = y; |
382 |
|
*dir = Direction; |
383 |
|
} |
384 |
|
} |
385 |
|
|
386 |
|
static void |
387 |
|
CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
388 |
|
{ |
389 |
|
uint32_t t; |
390 |
|
const uint8_t * Reference; |
391 |
|
|
392 |
#ifndef SEARCH16 |
if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero even value |
393 |
#define SEARCH16 PMVfastSearch16 |
(x > data->max_dx) || (x < data->min_dx) |
394 |
//#define SEARCH16 FullSearch16 |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
|
//#define SEARCH16 EPZSSearch16 |
|
|
#endif |
|
395 |
|
|
396 |
#ifndef SEARCH8 |
Reference = GetReference(x, y, data); |
397 |
#define SEARCH8 PMVfastSearch8 |
t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1); |
|
//#define SEARCH8 EPZSSearch8 |
|
|
#endif |
|
398 |
|
|
399 |
bool MotionEstimation( |
data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
|
MBParam * const pParam, |
|
|
FRAMEINFO * const current, |
|
|
FRAMEINFO * const reference, |
|
|
const IMAGE * const pRefH, |
|
|
const IMAGE * const pRefV, |
|
|
const IMAGE * const pRefHV, |
|
|
const uint32_t iLimit) |
|
400 |
|
|
401 |
{ |
data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10; |
402 |
const uint32_t iWcount = pParam->mb_width; |
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; |
|
const uint32_t iHcount = pParam->mb_height; |
|
|
MACROBLOCK * pMBs = current->mbs; |
|
|
IMAGE * pCurrent = ¤t->image; |
|
403 |
|
|
404 |
MACROBLOCK * prevMBs = reference->mbs; // previous frame |
if (data->temp[0] < data->iMinSAD[0]) { |
405 |
IMAGE * pRef = &reference->image; |
data->iMinSAD[0] = data->temp[0]; |
406 |
|
data->currentMV[0].x = x; data->currentMV[0].y = y; |
407 |
|
*dir = Direction; } |
408 |
|
|
409 |
|
if (data->temp[1] < data->iMinSAD[1]) { |
410 |
|
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
411 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
412 |
|
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
413 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
414 |
|
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
415 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
416 |
|
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
417 |
|
} |
418 |
|
|
419 |
uint32_t i, j, iIntra = 0; |
static void |
420 |
|
CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
421 |
|
{ |
422 |
|
int32_t sad, xc, yc; |
423 |
|
const uint8_t * Reference; |
424 |
|
uint32_t t; |
425 |
|
VECTOR * current; |
426 |
|
|
427 |
VECTOR mv16; |
if ( (x > data->max_dx) || ( x < data->min_dx) |
428 |
VECTOR pmv16; |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
429 |
|
|
430 |
int32_t sad8 = 0; |
if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; //non-zero even value |
|
int32_t sad16; |
|
|
int32_t deviation; |
|
431 |
|
|
432 |
if (sadInit) |
if (data->qpel_precision) { // x and y are in 1/4 precision |
433 |
(*sadInit)(); |
Reference = Interpolate16x16qpel(x, y, 0, data); |
434 |
|
current = data->currentQMV; |
435 |
|
xc = x/2; yc = y/2; |
436 |
|
} else { |
437 |
|
Reference = GetReference(x, y, data); |
438 |
|
current = data->currentMV; |
439 |
|
xc = x; yc = y; |
440 |
|
} |
441 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, |
442 |
|
data->qpel^data->qpel_precision, data->rrv); |
443 |
|
|
444 |
|
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
445 |
|
sad += (data->lambda16 * t * sad)>>10; |
446 |
|
|
447 |
/* eventhough we have a seperate prevMBs, |
if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
448 |
pmvfast/epsz does something "funny" with the previous frames data */ |
(yc >> 1) + roundtab_79[yc & 0x3], data); |
449 |
|
|
450 |
/* for (i = 0; i < iHcount; i++) |
if (sad < *(data->iMinSAD)) { |
451 |
for (j = 0; j < iWcount; j++) |
*(data->iMinSAD) = sad; |
452 |
{ |
current->x = x; current->y = y; |
453 |
pMBs[j + i * iWcount].mvs[0] = prevMBs[j + i * iWcount].mvs[0]; |
*dir = Direction; |
|
pMBs[j + i * iWcount].mvs[1] = prevMBs[j + i * iWcount].mvs[1]; |
|
|
pMBs[j + i * iWcount].mvs[2] = prevMBs[j + i * iWcount].mvs[2]; |
|
|
pMBs[j + i * iWcount].mvs[3] = prevMBs[j + i * iWcount].mvs[3]; |
|
454 |
} |
} |
|
*/ |
|
|
/*dprintf("*** BEFORE ***"); |
|
|
for (i = 0; i < iHcount; i++) |
|
|
for (j = 0; j < iWcount; j++) |
|
|
{ |
|
|
dprintf(" [%i,%i] mode=%i dquant=%i mvs=(%i %i %i %i) sad8=(%i %i %i %i) sad16=(%i)", j,i, |
|
|
pMBs[j + i * iWcount].mode, |
|
|
pMBs[j + i * iWcount].dquant, |
|
|
pMBs[j + i * iWcount].mvs[0], |
|
|
pMBs[j + i * iWcount].mvs[1], |
|
|
pMBs[j + i * iWcount].mvs[2], |
|
|
pMBs[j + i * iWcount].mvs[3], |
|
|
prevMBs[j + i * iWcount].sad8[0], |
|
|
prevMBs[j + i * iWcount].sad8[1], |
|
|
prevMBs[j + i * iWcount].sad8[2], |
|
|
prevMBs[j + i * iWcount].sad8[3], |
|
|
prevMBs[j + i * iWcount].sad16); |
|
455 |
} |
} |
|
*/ |
|
456 |
|
|
457 |
// note: i==horizontal, j==vertical |
static void |
458 |
for (i = 0; i < iHcount; i++) |
CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
for (j = 0; j < iWcount; j++) |
|
459 |
{ |
{ |
460 |
MACROBLOCK *pMB = &pMBs[j + i * iWcount]; |
// maximum speed - for P/B/I decision |
461 |
MACROBLOCK *prevMB = &prevMBs[j + i * iWcount]; |
int32_t sad; |
462 |
|
|
463 |
sad16 = SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
if ( (x > data->max_dx) || (x < data->min_dx) |
464 |
j, i, current->motion_flags, current->quant, current->fcode, |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
|
pParam, pMBs, prevMBs, &mv16, &pmv16); |
|
|
pMB->sad16=sad16; |
|
465 |
|
|
466 |
|
sad = sad32v_c(data->Cur, data->RefP[0] + (x>>1) + (y>>1)*(data->iEdgedWidth), |
467 |
|
data->iEdgedWidth, data->temp+1); |
468 |
|
|
469 |
/* decide: MODE_INTER or MODE_INTRA |
if (sad < *(data->iMinSAD)) { |
470 |
if (dev_intra < sad_inter - 2 * nb) use_intra |
*(data->iMinSAD) = sad; |
471 |
*/ |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
472 |
|
*dir = Direction; |
473 |
|
} |
474 |
|
if (data->temp[1] < data->iMinSAD[1]) { |
475 |
|
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
476 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
477 |
|
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
478 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
479 |
|
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
480 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
481 |
|
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
482 |
|
|
483 |
deviation = dev16(pCurrent->y + j*16 + i*16*pParam->edged_width, pParam->edged_width); |
} |
484 |
|
|
485 |
if (deviation < (sad16 - INTER_BIAS)) |
static void |
486 |
|
CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) |
487 |
{ |
{ |
488 |
pMB->mode = MODE_INTRA; |
int32_t sad, xb, yb, xcf, ycf, xcb, ycb; |
489 |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
uint32_t t; |
490 |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; |
const uint8_t *ReferenceF, *ReferenceB; |
491 |
|
VECTOR *current; |
492 |
|
|
493 |
pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = 0; |
if ((xf > data->max_dx) || (xf < data->min_dx) || |
494 |
|
(yf > data->max_dy) || (yf < data->min_dy)) |
495 |
|
return; |
496 |
|
|
497 |
iIntra++; |
if (!data->qpel_precision) { |
498 |
if(iIntra >= iLimit) |
ReferenceF = GetReference(xf, yf, data); |
499 |
return 1; |
xb = data->currentMV[1].x; yb = data->currentMV[1].y; |
500 |
|
ReferenceB = GetReferenceB(xb, yb, 1, data); |
501 |
|
current = data->currentMV; |
502 |
|
xcf = xf; ycf = yf; |
503 |
|
xcb = xb; ycb = yb; |
504 |
|
} else { |
505 |
|
ReferenceF = Interpolate16x16qpel(xf, yf, 0, data); |
506 |
|
xb = data->currentQMV[1].x; yb = data->currentQMV[1].y; |
507 |
|
current = data->currentQMV; |
508 |
|
ReferenceB = Interpolate16x16qpel(xb, yb, 1, data); |
509 |
|
xcf = xf/2; ycf = yf/2; |
510 |
|
xcb = xb/2; ycb = yb/2; |
511 |
|
} |
512 |
|
|
513 |
continue; |
t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0) |
514 |
|
+ d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
515 |
|
|
516 |
|
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
517 |
|
sad += (data->lambda16 * t * sad)>>10; |
518 |
|
|
519 |
|
if (data->chroma) sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3], |
520 |
|
(ycf >> 1) + roundtab_79[ycf & 0x3], |
521 |
|
(xcb >> 1) + roundtab_79[xcb & 0x3], |
522 |
|
(ycb >> 1) + roundtab_79[ycb & 0x3], data); |
523 |
|
|
524 |
|
if (sad < *(data->iMinSAD)) { |
525 |
|
*(data->iMinSAD) = sad; |
526 |
|
current->x = xf; current->y = yf; |
527 |
|
*dir = Direction; |
528 |
|
} |
529 |
} |
} |
530 |
|
|
531 |
if (current->global_flags & XVID_INTER4V) |
static void |
532 |
|
CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
533 |
{ |
{ |
534 |
pMB->sad8[0] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0; |
535 |
2 * j, 2 * i, mv16.x, mv16.y, |
uint32_t k; |
536 |
current->motion_flags, current->quant, current->fcode, |
const uint8_t *ReferenceF; |
537 |
pParam, pMBs, prevMBs, &pMB->mvs[0], &pMB->pmvs[0]); |
const uint8_t *ReferenceB; |
538 |
|
VECTOR mvs, b_mvs; |
539 |
|
|
540 |
pMB->sad8[1] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
|
2 * j + 1, 2 * i, mv16.x, mv16.y, |
|
|
current->motion_flags, current->quant, current->fcode, |
|
|
pParam, pMBs, prevMBs, &pMB->mvs[1], &pMB->pmvs[1]); |
|
541 |
|
|
542 |
pMB->sad8[2] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
for (k = 0; k < 4; k++) { |
543 |
2 * j, 2 * i + 1, mv16.x, mv16.y, |
mvs.x = data->directmvF[k].x + x; |
544 |
current->motion_flags, current->quant, current->fcode, |
b_mvs.x = ((x == 0) ? |
545 |
pParam, pMBs, prevMBs, &pMB->mvs[2], &pMB->pmvs[2]); |
data->directmvB[k].x |
546 |
|
: mvs.x - data->referencemv[k].x); |
547 |
|
|
548 |
pMB->sad8[3] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
mvs.y = data->directmvF[k].y + y; |
549 |
2 * j + 1, 2 * i + 1, mv16.x, mv16.y, |
b_mvs.y = ((y == 0) ? |
550 |
current->motion_flags, current->quant, current->fcode, |
data->directmvB[k].y |
551 |
pParam, pMBs, prevMBs, &pMB->mvs[3], &pMB->pmvs[3]); |
: mvs.y - data->referencemv[k].y); |
552 |
|
|
553 |
sad8 = pMB->sad8[0] + pMB->sad8[1] + pMB->sad8[2] + pMB->sad8[3]; |
if ((mvs.x > data->max_dx) || (mvs.x < data->min_dx) || |
554 |
|
(mvs.y > data->max_dy) || (mvs.y < data->min_dy) || |
555 |
|
(b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx) || |
556 |
|
(b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) ) |
557 |
|
return; |
558 |
|
|
559 |
|
if (data->qpel) { |
560 |
|
xcf += mvs.x/2; ycf += mvs.y/2; |
561 |
|
xcb += b_mvs.x/2; ycb += b_mvs.y/2; |
562 |
|
} else { |
563 |
|
xcf += mvs.x; ycf += mvs.y; |
564 |
|
xcb += b_mvs.x; ycb += b_mvs.y; |
565 |
|
mvs.x *= 2; mvs.y *= 2; //we move to qpel precision anyway |
566 |
|
b_mvs.x *= 2; b_mvs.y *= 2; |
567 |
} |
} |
568 |
|
|
569 |
|
ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); |
570 |
|
ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); |
571 |
|
|
572 |
/* decide: MODE_INTER or MODE_INTER4V |
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
573 |
mpeg4: if (sad8 < sad16 - nb/2+1) use_inter4v |
ReferenceF, ReferenceB, data->iEdgedWidth); |
574 |
*/ |
if (sad > *(data->iMinSAD)) return; |
575 |
|
} |
576 |
|
|
577 |
if (!(current->global_flags & XVID_LUMIMASKING) || pMB->dquant == NO_CHANGE) |
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
|
{ |
|
|
if (((current->global_flags & XVID_INTER4V)==0) || |
|
|
(sad16 < (sad8 + (int32_t)(IMV16X16 * current->quant)))) |
|
|
{ |
|
578 |
|
|
579 |
sad8 = sad16; |
if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
580 |
pMB->mode = MODE_INTER; |
(ycf >> 3) + roundtab_76[ycf & 0xf], |
581 |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; |
(xcb >> 3) + roundtab_76[xcb & 0xf], |
582 |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; |
(ycb >> 3) + roundtab_76[ycb & 0xf], data); |
583 |
pMB->pmvs[0].x = pmv16.x; |
|
584 |
pMB->pmvs[0].y = pmv16.y; |
if (sad < *(data->iMinSAD)) { |
585 |
|
*(data->iMinSAD) = sad; |
586 |
|
data->currentMV->x = x; data->currentMV->y = y; |
587 |
|
*dir = Direction; |
588 |
} |
} |
|
else |
|
|
pMB->mode = MODE_INTER4V; |
|
589 |
} |
} |
590 |
else |
|
591 |
|
static void |
592 |
|
CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
593 |
{ |
{ |
594 |
sad8 = sad16; |
int32_t sad, xcf, ycf, xcb, ycb; |
595 |
pMB->mode = MODE_INTER; |
const uint8_t *ReferenceF; |
596 |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; |
const uint8_t *ReferenceB; |
597 |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; |
VECTOR mvs, b_mvs; |
598 |
pMB->pmvs[0].x = pmv16.x; |
|
599 |
pMB->pmvs[0].y = pmv16.y; |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
600 |
|
|
601 |
|
mvs.x = data->directmvF[0].x + x; |
602 |
|
b_mvs.x = ((x == 0) ? |
603 |
|
data->directmvB[0].x |
604 |
|
: mvs.x - data->referencemv[0].x); |
605 |
|
|
606 |
|
mvs.y = data->directmvF[0].y + y; |
607 |
|
b_mvs.y = ((y == 0) ? |
608 |
|
data->directmvB[0].y |
609 |
|
: mvs.y - data->referencemv[0].y); |
610 |
|
|
611 |
|
if ( (mvs.x > data->max_dx) || (mvs.x < data->min_dx) |
612 |
|
|| (mvs.y > data->max_dy) || (mvs.y < data->min_dy) |
613 |
|
|| (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx) |
614 |
|
|| (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) ) return; |
615 |
|
|
616 |
|
if (data->qpel) { |
617 |
|
xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2); |
618 |
|
xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2); |
619 |
|
ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data); |
620 |
|
ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); |
621 |
|
} else { |
622 |
|
xcf = 4*mvs.x; ycf = 4*mvs.y; |
623 |
|
xcb = 4*b_mvs.x; ycb = 4*b_mvs.y; |
624 |
|
ReferenceF = GetReference(mvs.x, mvs.y, data); |
625 |
|
ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data); |
626 |
} |
} |
627 |
|
|
628 |
|
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
629 |
|
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
630 |
|
|
631 |
|
if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
632 |
|
(ycf >> 3) + roundtab_76[ycf & 0xf], |
633 |
|
(xcb >> 3) + roundtab_76[xcb & 0xf], |
634 |
|
(ycb >> 3) + roundtab_76[ycb & 0xf], data); |
635 |
|
|
636 |
|
if (sad < *(data->iMinSAD)) { |
637 |
|
*(data->iMinSAD) = sad; |
638 |
|
data->currentMV->x = x; data->currentMV->y = y; |
639 |
|
*dir = Direction; |
640 |
} |
} |
641 |
|
} |
642 |
|
|
643 |
|
|
644 |
|
static void |
645 |
|
CheckCandidateBits16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
646 |
|
{ |
647 |
|
|
648 |
|
int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; |
649 |
|
int32_t bits = 0, sum; |
650 |
|
VECTOR * current; |
651 |
|
const uint8_t * ptr; |
652 |
|
int i, cbp = 0, t, xc, yc; |
653 |
|
|
654 |
|
if ( (x > data->max_dx) || (x < data->min_dx) |
655 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
656 |
|
|
657 |
/* dprintf("*** AFTER ***", pMBs[0].b_mvs[0].x); |
if (!data->qpel_precision) { |
658 |
for (i = 0; i < iHcount; i++) |
ptr = GetReference(x, y, data); |
659 |
for (j = 0; j < iWcount; j++) |
current = data->currentMV; |
660 |
{ |
xc = x; yc = y; |
661 |
dprintf(" [%i,%i] mode=%i dquant=%i mvs=(%i %i %i %i) sad8=(%i %i %i %i) sad16=(%i)", j,i, |
} else { // x and y are in 1/4 precision |
662 |
pMBs[j + i * iWcount].mode, |
ptr = Interpolate16x16qpel(x, y, 0, data); |
663 |
pMBs[j + i * iWcount].dquant, |
current = data->currentQMV; |
664 |
pMBs[j + i * iWcount].mvs[0], |
xc = x/2; yc = y/2; |
|
pMBs[j + i * iWcount].mvs[1], |
|
|
pMBs[j + i * iWcount].mvs[2], |
|
|
pMBs[j + i * iWcount].mvs[3], |
|
|
pMBs[j + i * iWcount].sad8[0], |
|
|
pMBs[j + i * iWcount].sad8[1], |
|
|
pMBs[j + i * iWcount].sad8[2], |
|
|
pMBs[j + i * iWcount].sad8[3], |
|
|
pMBs[j + i * iWcount].sad16); |
|
665 |
} |
} |
|
*/ |
|
666 |
|
|
667 |
return 0; |
for(i = 0; i < 4; i++) { |
668 |
|
int s = 8*((i&1) + (i>>1)*data->iEdgedWidth); |
669 |
|
transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth); |
670 |
|
fdct(in); |
671 |
|
if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); |
672 |
|
else sum = quant4_inter(coeff, in, data->lambda16); |
673 |
|
if (sum > 0) { |
674 |
|
cbp |= 1 << (5 - i); |
675 |
|
bits += data->temp[i] = CodeCoeffInter_CalcBits(coeff, scan_tables[0]); |
676 |
|
} else data->temp[i] = 0; |
677 |
} |
} |
678 |
|
|
679 |
#define MVzero(A) ( ((A).x)==(0) && ((A).y)==(0) ) |
bits += t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
680 |
|
|
681 |
#define MVequal(A,B) ( ((A).x)==((B).x) && ((A).y)==((B).y) ) |
if (bits < data->iMinSAD[0]) { // there is still a chance, adding chroma |
682 |
|
xc = (xc >> 1) + roundtab_79[xc & 0x3]; |
683 |
|
yc = (yc >> 1) + roundtab_79[yc & 0x3]; |
684 |
|
|
685 |
|
//chroma U |
686 |
|
ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefP[4], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); |
687 |
|
transfer_8to16subro(in, ptr, data->CurU, data->iEdgedWidth/2); |
688 |
|
fdct(in); |
689 |
|
if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); |
690 |
|
else sum = quant4_inter(coeff, in, data->lambda16); |
691 |
|
if (sum > 0) { |
692 |
|
cbp |= 1 << (5 - 4); |
693 |
|
bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); |
694 |
|
} |
695 |
|
|
696 |
#define CHECK_MV16_ZERO {\ |
if (bits < data->iMinSAD[0]) { |
697 |
if ( (0 <= max_dx) && (0 >= min_dx) \ |
//chroma V |
698 |
&& (0 <= max_dy) && (0 >= min_dy) ) \ |
ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefP[5], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); |
699 |
{ \ |
transfer_8to16subro(in, ptr, data->CurV, data->iEdgedWidth/2); |
700 |
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0, 0 , iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); \ |
fdct(in); |
701 |
iSAD += calc_delta_16(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); |
702 |
if (iSAD < iMinSAD) \ |
else sum = quant4_inter(coeff, in, data->lambda16); |
703 |
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } } \ |
if (sum > 0) { |
704 |
} |
cbp |= 1 << (5 - 5); |
705 |
|
bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); |
706 |
#define NOCHECK_MV16_CANDIDATE(X,Y) { \ |
} |
707 |
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
} |
708 |
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
} |
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV16_CANDIDATE(X,Y) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
|
|
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV16_CANDIDATE_DIR(X,Y,D) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
|
|
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV16_CANDIDATE_FOUND(X,Y,D) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
|
|
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
|
|
} |
|
|
|
|
|
|
|
|
#define CHECK_MV8_ZERO {\ |
|
|
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, 0, 0 , iEdgedWidth), iEdgedWidth); \ |
|
|
iSAD += calc_delta_8(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } \ |
|
|
} |
|
|
|
|
|
#define NOCHECK_MV8_CANDIDATE(X,Y) \ |
|
|
{ \ |
|
|
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
|
|
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV8_CANDIDATE(X,Y) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
|
|
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV8_CANDIDATE_DIR(X,Y,D) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
|
|
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV8_CANDIDATE_FOUND(X,Y,D) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
|
|
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
|
|
} |
|
|
|
|
|
/* too slow and not fully functional at the moment */ |
|
|
/* |
|
|
int32_t ZeroSearch16( |
|
|
const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const IMAGE * const pCur, |
|
|
const int x, const int y, |
|
|
const uint32_t MotionFlags, |
|
|
const uint32_t iQuant, |
|
|
const uint32_t iFcode, |
|
|
MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const MACROBLOCK * const prevMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV) |
|
|
{ |
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
|
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
|
|
int32_t iSAD; |
|
|
int32_t pred_x,pred_y; |
|
|
|
|
|
get_pmv(pMBs, x, y, pParam->mb_width, 0, &pred_x, &pred_y); |
|
|
|
|
|
iSAD = sad16( cur, |
|
|
get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0,0, iEdgedWidth), |
|
|
iEdgedWidth, MV_MAX_ERROR); |
|
|
if (iSAD <= iQuant * 96) |
|
|
iSAD -= MV16_00_BIAS; |
|
|
|
|
|
currMV->x = 0; |
|
|
currMV->y = 0; |
|
|
currPMV->x = -pred_x; |
|
|
currPMV->y = -pred_y; |
|
709 |
|
|
710 |
return iSAD; |
bits += xvid_cbpy_tab[15-(cbp>>2)].len; |
711 |
|
bits += mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; |
712 |
|
|
713 |
|
if (bits < data->iMinSAD[0]) { |
714 |
|
data->iMinSAD[0] = bits; |
715 |
|
current[0].x = x; current[0].y = y; |
716 |
|
*dir = Direction; |
717 |
} |
} |
|
*/ |
|
718 |
|
|
719 |
int32_t Diamond16_MainSearch( |
if (data->temp[0] + t < data->iMinSAD[1]) { |
720 |
const uint8_t * const pRef, |
data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; } |
721 |
const uint8_t * const pRefH, |
if (data->temp[1] < data->iMinSAD[2]) { |
722 |
const uint8_t * const pRefV, |
data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; } |
723 |
const uint8_t * const pRefHV, |
if (data->temp[2] < data->iMinSAD[3]) { |
724 |
const uint8_t * const cur, |
data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; } |
725 |
const int x, const int y, |
if (data->temp[3] < data->iMinSAD[4]) { |
726 |
int32_t startx, int32_t starty, |
data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; } |
727 |
int32_t iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const VECTOR * const pmv, |
|
|
const int32_t min_dx, const int32_t max_dx, |
|
|
const int32_t min_dy, const int32_t max_dy, |
|
|
const int32_t iEdgedWidth, |
|
|
const int32_t iDiamondSize, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
int iFound) |
|
|
{ |
|
|
/* Do a diamond search around given starting point, return SAD of best */ |
|
|
|
|
|
int32_t iDirection=0; |
|
|
int32_t iSAD; |
|
|
VECTOR backupMV; |
|
|
backupMV.x = startx; |
|
|
backupMV.y = starty; |
|
|
|
|
|
/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ |
|
|
|
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
|
|
|
|
|
if (iDirection) |
|
|
while (!iFound) |
|
|
{ |
|
|
iFound = 1; |
|
|
backupMV=*currMV; |
|
|
|
|
|
if ( iDirection != 2) |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
|
|
if ( iDirection != 1) |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x+iDiamondSize,backupMV.y,2); |
|
|
if ( iDirection != 4) |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x,backupMV.y-iDiamondSize,3); |
|
|
if ( iDirection != 3) |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x,backupMV.y+iDiamondSize,4); |
|
728 |
} |
} |
729 |
else |
static void |
730 |
|
CheckCandidateBits8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
731 |
{ |
{ |
732 |
currMV->x = startx; |
|
733 |
currMV->y = starty; |
int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; |
734 |
} |
int32_t sum, bits; |
735 |
return iMinSAD; |
VECTOR * current; |
736 |
|
const uint8_t * ptr; |
737 |
|
int cbp; |
738 |
|
|
739 |
|
if ( (x > data->max_dx) || (x < data->min_dx) |
740 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
741 |
|
|
742 |
|
if (!data->qpel_precision) { |
743 |
|
ptr = GetReference(x, y, data); |
744 |
|
current = data->currentMV; |
745 |
|
} else { // x and y are in 1/4 precision |
746 |
|
ptr = Interpolate8x8qpel(x, y, 0, 0, data); |
747 |
|
current = data->currentQMV; |
748 |
} |
} |
749 |
|
|
750 |
int32_t Square16_MainSearch( |
transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth); |
751 |
const uint8_t * const pRef, |
fdct(in); |
752 |
const uint8_t * const pRefH, |
if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); |
753 |
const uint8_t * const pRefV, |
else sum = quant4_inter(coeff, in, data->lambda16); |
754 |
const uint8_t * const pRefHV, |
if (sum > 0) { |
755 |
const uint8_t * const cur, |
bits = CodeCoeffInter_CalcBits(coeff, scan_tables[0]); |
756 |
const int x, const int y, |
cbp = 1; |
757 |
int32_t startx, int32_t starty, |
} else cbp = bits = 0; |
|
int32_t iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const VECTOR * const pmv, |
|
|
const int32_t min_dx, const int32_t max_dx, |
|
|
const int32_t min_dy, const int32_t max_dy, |
|
|
const int32_t iEdgedWidth, |
|
|
const int32_t iDiamondSize, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
int iFound) |
|
|
{ |
|
|
/* Do a square search around given starting point, return SAD of best */ |
|
|
|
|
|
int32_t iDirection=0; |
|
|
int32_t iSAD; |
|
|
VECTOR backupMV; |
|
|
backupMV.x = startx; |
|
|
backupMV.y = starty; |
|
|
|
|
|
/* It's one search with full square pattern, and new parts for all following diamonds */ |
|
|
|
|
|
/* new direction are extra, so 1-4 is normal diamond |
|
|
537 |
|
|
1*2 |
|
|
648 |
|
|
*/ |
|
758 |
|
|
759 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
bits += sum = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
|
760 |
|
|
761 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
if (bits < data->iMinSAD[0]) { |
762 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
data->temp[0] = cbp; |
763 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
data->iMinSAD[0] = bits; |
764 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
current[0].x = x; current[0].y = y; |
765 |
|
*dir = Direction; |
766 |
|
} |
767 |
|
} |
768 |
|
|
769 |
|
/* CHECK_CANDIATE FUNCTIONS END */ |
770 |
|
|
771 |
if (iDirection) |
/* MAINSEARCH FUNCTIONS START */ |
|
while (!iFound) |
|
|
{ |
|
|
iFound = 1; |
|
|
backupMV=*currMV; |
|
772 |
|
|
773 |
switch (iDirection) |
static void |
774 |
|
AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection) |
775 |
{ |
{ |
|
case 1: |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
|
|
break; |
|
|
case 2: |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
|
|
break; |
|
776 |
|
|
777 |
case 3: |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
|
|
break; |
|
778 |
|
|
779 |
case 4: |
int iDirection; |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
|
|
break; |
|
780 |
|
|
781 |
case 5: |
for(;;) { //forever |
782 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
iDirection = 0; |
783 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
784 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
785 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
786 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
|
break; |
|
787 |
|
|
788 |
case 6: |
/* now we're doing diagonal checks near our candidate */ |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
|
789 |
|
|
790 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
if (iDirection) { //if anything found |
791 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
bDirection = iDirection; |
792 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
iDirection = 0; |
793 |
|
x = data->currentMV->x; y = data->currentMV->y; |
794 |
|
if (bDirection & 3) { //our candidate is left or right |
795 |
|
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
796 |
|
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
797 |
|
} else { // what remains here is up or down |
798 |
|
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
799 |
|
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
800 |
|
} |
801 |
|
|
802 |
|
if (iDirection) { |
803 |
|
bDirection += iDirection; |
804 |
|
x = data->currentMV->x; y = data->currentMV->y; |
805 |
|
} |
806 |
|
} else { //about to quit, eh? not so fast.... |
807 |
|
switch (bDirection) { |
808 |
|
case 2: |
809 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
810 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
811 |
break; |
break; |
812 |
|
case 1: |
813 |
case 7: |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
814 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
815 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
break; |
816 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
case 2 + 4: |
817 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
818 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
819 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
820 |
|
break; |
821 |
|
case 4: |
822 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
823 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
824 |
break; |
break; |
|
|
|
825 |
case 8: |
case 8: |
826 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
827 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
|
828 |
break; |
break; |
829 |
default: |
case 1 + 4: |
830 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
831 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
832 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
833 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
break; |
834 |
|
case 2 + 8: |
835 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
836 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
837 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
838 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
break; |
839 |
|
case 1 + 8: |
840 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
841 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
842 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
843 |
|
break; |
844 |
|
default: //1+2+4+8 == we didn't find anything at all |
845 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
846 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
847 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
848 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
849 |
break; |
break; |
850 |
} |
} |
851 |
|
if (!iDirection) break; //ok, the end. really |
852 |
|
bDirection = iDirection; |
853 |
|
x = data->currentMV->x; y = data->currentMV->y; |
854 |
} |
} |
|
else |
|
|
{ |
|
|
currMV->x = startx; |
|
|
currMV->y = starty; |
|
855 |
} |
} |
|
return iMinSAD; |
|
856 |
} |
} |
857 |
|
|
858 |
|
static void |
859 |
|
SquareSearch(int x, int y, const SearchData * const data, int bDirection) |
860 |
|
{ |
861 |
|
int iDirection; |
862 |
|
|
863 |
int32_t Full16_MainSearch( |
do { |
864 |
const uint8_t * const pRef, |
iDirection = 0; |
865 |
const uint8_t * const pRefH, |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64); |
866 |
const uint8_t * const pRefV, |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128); |
867 |
const uint8_t * const pRefHV, |
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32); |
868 |
const uint8_t * const cur, |
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128); |
869 |
const int x, const int y, |
if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64); |
870 |
int32_t startx, int32_t starty, |
if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128); |
871 |
int32_t iMinSAD, |
if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128); |
872 |
VECTOR * const currMV, |
if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128); |
|
const VECTOR * const pmv, |
|
|
const int32_t min_dx, const int32_t max_dx, |
|
|
const int32_t min_dy, const int32_t max_dy, |
|
|
const int32_t iEdgedWidth, |
|
|
const int32_t iDiamondSize, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
int iFound) |
|
|
{ |
|
|
int32_t iSAD; |
|
|
int32_t dx,dy; |
|
|
VECTOR backupMV; |
|
|
backupMV.x = startx; |
|
|
backupMV.y = starty; |
|
|
|
|
|
for (dx = min_dx; dx<=max_dx; dx+=iDiamondSize) |
|
|
for (dy = min_dy; dy<= max_dy; dy+=iDiamondSize) |
|
|
NOCHECK_MV16_CANDIDATE(dx,dy); |
|
873 |
|
|
874 |
return iMinSAD; |
bDirection = iDirection; |
875 |
|
x = data->currentMV->x; y = data->currentMV->y; |
876 |
|
} while (iDirection); |
877 |
} |
} |
878 |
|
|
879 |
int32_t Full8_MainSearch( |
static void |
880 |
const uint8_t * const pRef, |
DiamondSearch(int x, int y, const SearchData * const data, int bDirection) |
881 |
const uint8_t * const pRefH, |
{ |
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, const int y, |
|
|
int32_t startx, int32_t starty, |
|
|
int32_t iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const VECTOR * const pmv, |
|
|
const int32_t min_dx, const int32_t max_dx, |
|
|
const int32_t min_dy, const int32_t max_dy, |
|
|
const int32_t iEdgedWidth, |
|
|
const int32_t iDiamondSize, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
int iFound) |
|
|
{ |
|
|
int32_t iSAD; |
|
|
int32_t dx,dy; |
|
|
VECTOR backupMV; |
|
|
backupMV.x = startx; |
|
|
backupMV.y = starty; |
|
|
|
|
|
for (dx = min_dx; dx<=max_dx; dx+=iDiamondSize) |
|
|
for (dy = min_dy; dy<= max_dy; dy+=iDiamondSize) |
|
|
NOCHECK_MV8_CANDIDATE(dx,dy); |
|
882 |
|
|
883 |
return iMinSAD; |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
|
} |
|
884 |
|
|
885 |
|
int iDirection; |
886 |
|
|
887 |
|
do { |
888 |
|
iDirection = 0; |
889 |
|
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
890 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
891 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
892 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
893 |
|
|
894 |
int32_t Halfpel16_Refine( |
/* now we're doing diagonal checks near our candidate */ |
|
const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, const int y, |
|
|
VECTOR * const currMV, |
|
|
int32_t iMinSAD, |
|
|
const VECTOR * const pmv, |
|
|
const int32_t min_dx, const int32_t max_dx, |
|
|
const int32_t min_dy, const int32_t max_dy, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
const int32_t iEdgedWidth) |
|
|
{ |
|
|
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
|
|
|
|
|
int32_t iSAD; |
|
|
VECTOR backupMV = *currMV; |
|
|
|
|
|
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y-1); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y-1); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y-1); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y+1); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y+1); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y+1); |
|
895 |
|
|
896 |
return iMinSAD; |
if (iDirection) { //checking if anything found |
897 |
|
bDirection = iDirection; |
898 |
|
iDirection = 0; |
899 |
|
x = data->currentMV->x; y = data->currentMV->y; |
900 |
|
if (bDirection & 3) { //our candidate is left or right |
901 |
|
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
902 |
|
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
903 |
|
} else { // what remains here is up or down |
904 |
|
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
905 |
|
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
906 |
|
} |
907 |
|
bDirection += iDirection; |
908 |
|
x = data->currentMV->x; y = data->currentMV->y; |
909 |
|
} |
910 |
|
} |
911 |
|
while (iDirection); |
912 |
} |
} |
913 |
|
|
914 |
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
/* MAINSEARCH FUNCTIONS END */ |
|
|
|
915 |
|
|
916 |
int32_t PMVfastSearch16( |
static void |
917 |
const uint8_t * const pRef, |
SubpelRefine(const SearchData * const data) |
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const IMAGE * const pCur, |
|
|
const int x, const int y, |
|
|
const uint32_t MotionFlags, |
|
|
const uint32_t iQuant, |
|
|
const uint32_t iFcode, |
|
|
const MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const MACROBLOCK * const prevMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV) |
|
918 |
{ |
{ |
919 |
const uint32_t iWcount = pParam->mb_width; |
/* Do a half-pel or q-pel refinement */ |
920 |
const int32_t iWidth = pParam->width; |
const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV; |
921 |
const int32_t iHeight = pParam->height; |
int iDirection; //only needed because macro expects it |
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
|
|
|
|
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
|
|
|
|
|
int32_t iDiamondSize; |
|
922 |
|
|
923 |
int32_t min_dx; |
CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0); |
924 |
int32_t max_dx; |
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0); |
925 |
int32_t min_dy; |
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0); |
926 |
int32_t max_dy; |
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0); |
927 |
|
CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0); |
928 |
int32_t iFound; |
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0); |
929 |
|
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0); |
930 |
|
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0); |
931 |
|
} |
932 |
|
|
933 |
VECTOR newMV; |
static __inline int |
934 |
VECTOR backupMV; /* just for PMVFAST */ |
SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
935 |
|
const int x, const int y, |
936 |
|
const uint32_t stride, const uint32_t iQuant, int rrv) |
937 |
|
|
938 |
VECTOR pmv[4]; |
{ |
939 |
int32_t psad[4]; |
int offset = (x + y*stride)*8; |
940 |
|
if(!rrv) { |
941 |
|
uint32_t sadC = sad8(current->u + offset, |
942 |
|
reference->u + offset, stride); |
943 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
944 |
|
sadC += sad8(current->v + offset, |
945 |
|
reference->v + offset, stride); |
946 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
947 |
|
return 1; |
948 |
|
|
949 |
const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
} else { |
950 |
const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; |
uint32_t sadC = sad16(current->u + 2*offset, |
951 |
|
reference->u + 2*offset, stride, 256*4096); |
952 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
953 |
|
sadC += sad16(current->v + 2*offset, |
954 |
|
reference->v + 2*offset, stride, 256*4096); |
955 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
956 |
|
return 1; |
957 |
|
} |
958 |
|
} |
959 |
|
|
960 |
static int32_t threshA,threshB; |
static __inline void |
961 |
int32_t bPredEq; |
SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
962 |
int32_t iMinSAD,iSAD; |
{ |
963 |
|
pMB->mode = MODE_NOT_CODED; |
964 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV; |
965 |
|
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV; |
966 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
967 |
|
} |
968 |
|
|
969 |
/* Get maximum range */ |
bool |
970 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
MotionEstimation(MBParam * const pParam, |
971 |
x, y, 16, iWidth, iHeight, iFcode); |
FRAMEINFO * const current, |
972 |
|
FRAMEINFO * const reference, |
973 |
|
const IMAGE * const pRefH, |
974 |
|
const IMAGE * const pRefV, |
975 |
|
const IMAGE * const pRefHV, |
976 |
|
const uint32_t iLimit) |
977 |
|
{ |
978 |
|
MACROBLOCK *const pMBs = current->mbs; |
979 |
|
const IMAGE *const pCurrent = ¤t->image; |
980 |
|
const IMAGE *const pRef = &reference->image; |
981 |
|
|
982 |
|
uint32_t mb_width = pParam->mb_width; |
983 |
|
uint32_t mb_height = pParam->mb_height; |
984 |
|
const uint32_t iEdgedWidth = pParam->edged_width; |
985 |
|
const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->global_flags); |
986 |
|
|
987 |
|
uint32_t x, y; |
988 |
|
uint32_t iIntra = 0; |
989 |
|
int32_t quant = current->quant, sad00; |
990 |
|
int skip_thresh = INITIAL_SKIP_THRESH * |
991 |
|
(current->global_flags & XVID_REDUCED ? 4:1) * |
992 |
|
(current->global_flags & XVID_MODEDECISION_BITS ? 2:1); |
993 |
|
|
994 |
|
// some pre-initialized thingies for SearchP |
995 |
|
int32_t temp[8]; |
996 |
|
VECTOR currentMV[5]; |
997 |
|
VECTOR currentQMV[5]; |
998 |
|
int32_t iMinSAD[5]; |
999 |
|
DECLARE_ALIGNED_MATRIX(dct_space, 2, 64, int16_t, CACHE_LINE); |
1000 |
|
SearchData Data; |
1001 |
|
memset(&Data, 0, sizeof(SearchData)); |
1002 |
|
Data.iEdgedWidth = iEdgedWidth; |
1003 |
|
Data.currentMV = currentMV; |
1004 |
|
Data.currentQMV = currentQMV; |
1005 |
|
Data.iMinSAD = iMinSAD; |
1006 |
|
Data.temp = temp; |
1007 |
|
Data.iFcode = current->fcode; |
1008 |
|
Data.rounding = pParam->m_rounding_type; |
1009 |
|
Data.qpel = pParam->m_quarterpel; |
1010 |
|
Data.chroma = MotionFlags & PMV_CHROMA16; |
1011 |
|
Data.rrv = current->global_flags & XVID_REDUCED; |
1012 |
|
Data.dctSpace = dct_space; |
1013 |
|
|
1014 |
|
if ((current->global_flags & XVID_REDUCED)) { |
1015 |
|
mb_width = (pParam->width + 31) / 32; |
1016 |
|
mb_height = (pParam->height + 31) / 32; |
1017 |
|
Data.qpel = 0; |
1018 |
|
} |
1019 |
|
|
1020 |
|
Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose) |
1021 |
|
if (sadInit) (*sadInit) (); |
1022 |
|
|
1023 |
|
for (y = 0; y < mb_height; y++) { |
1024 |
|
for (x = 0; x < mb_width; x++) { |
1025 |
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
1026 |
|
|
1027 |
|
if (!Data.rrv) pMB->sad16 = |
1028 |
|
sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16, |
1029 |
|
pRef->y + (x + y * iEdgedWidth) * 16, |
1030 |
|
pParam->edged_width, pMB->sad8 ); |
1031 |
|
|
1032 |
|
else pMB->sad16 = |
1033 |
|
sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32, |
1034 |
|
pRef->y + (x + y * iEdgedWidth) * 32, |
1035 |
|
pParam->edged_width, pMB->sad8 ); |
1036 |
|
|
1037 |
|
if (Data.chroma) { |
1038 |
|
Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8, |
1039 |
|
pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2) |
1040 |
|
+ sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8, |
1041 |
|
pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); |
1042 |
|
pMB->sad16 += Data.temp[7]; |
1043 |
|
} |
1044 |
|
|
1045 |
|
sad00 = pMB->sad16; |
1046 |
|
|
1047 |
|
if (!(current->global_flags & XVID_LUMIMASKING)) { |
1048 |
|
pMB->dquant = NO_CHANGE; |
1049 |
|
} else { |
1050 |
|
if (pMB->dquant != NO_CHANGE) { |
1051 |
|
quant += DQtab[pMB->dquant]; |
1052 |
|
if (quant > 31) quant = 31; |
1053 |
|
else if (quant < 1) quant = 1; |
1054 |
|
} |
1055 |
|
} |
1056 |
|
pMB->quant = current->quant; |
1057 |
|
|
1058 |
|
//initial skip decision |
1059 |
|
/* no early skip for GMC (global vector = skip vector is unknown!) */ |
1060 |
|
if (!(current->global_flags & XVID_GMC)) { /* no fast SKIP for S(GMC)-VOPs */ |
1061 |
|
if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * skip_thresh) |
1062 |
|
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) { |
1063 |
|
SkipMacroblockP(pMB, sad00); |
1064 |
|
continue; |
1065 |
|
} |
1066 |
|
} |
1067 |
|
|
1068 |
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
1069 |
|
y, MotionFlags, current->global_flags, pMB->quant, |
1070 |
|
&Data, pParam, pMBs, reference->mbs, |
1071 |
|
current->global_flags & XVID_INTER4V, pMB); |
1072 |
|
|
1073 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
1074 |
{ min_dx = EVEN(min_dx); |
if (!(current->global_flags & XVID_GMC || current->global_flags & XVID_MODEDECISION_BITS)) { |
1075 |
max_dx = EVEN(max_dx); |
if ( pMB->dquant == NO_CHANGE && sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) { |
1076 |
min_dy = EVEN(min_dy); |
if ( (100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1) ) |
1077 |
max_dy = EVEN(max_dy); |
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) |
1078 |
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
SkipMacroblockP(pMB, sad00); |
1079 |
|
} |
1080 |
|
} |
1081 |
|
if (pMB->mode == MODE_INTRA) |
1082 |
|
if (++iIntra > iLimit) return 1; |
1083 |
|
} |
1084 |
|
} |
1085 |
|
|
1086 |
|
if (current->global_flags & XVID_GMC ) /* GMC only for S(GMC)-VOPs */ |
1087 |
|
{ |
1088 |
|
current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV); |
1089 |
|
} |
1090 |
|
return 0; |
1091 |
|
} |
1092 |
|
|
|
bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
|
1093 |
|
|
1094 |
if ((x==0) && (y==0) ) |
static __inline int |
1095 |
|
make_mask(const VECTOR * const pmv, const int i) |
1096 |
{ |
{ |
1097 |
threshA = 512; |
int mask = 255, j; |
1098 |
threshB = 1024; |
for (j = 0; j < i; j++) { |
1099 |
|
if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already |
1100 |
|
if (pmv[i].x == pmv[j].x) { |
1101 |
|
if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4; |
1102 |
|
else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8; |
1103 |
|
} else |
1104 |
|
if (pmv[i].y == pmv[j].y) { |
1105 |
|
if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1; |
1106 |
|
else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2; |
1107 |
} |
} |
1108 |
else |
} |
1109 |
{ |
return mask; |
|
threshA = psad[0]; |
|
|
threshB = threshA+256; |
|
|
if (threshA< 512) threshA = 512; |
|
|
if (threshA>1024) threshA = 1024; |
|
|
if (threshB>1792) threshB = 1792; |
|
1110 |
} |
} |
1111 |
|
|
1112 |
iFound=0; |
static __inline void |
1113 |
|
PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, |
1114 |
|
int iHcount, const MACROBLOCK * const prevMB, int rrv) |
1115 |
|
{ |
1116 |
|
|
1117 |
/* Step 2: Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself |
1118 |
vector of the median. |
if (rrv) { iWcount /= 2; iHcount /= 2; } |
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
|
|
*/ |
|
1119 |
|
|
1120 |
if ((bPredEq) && (MVequal(pmv[0],prevMB->mvs[0]) ) ) |
if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour |
1121 |
iFound=2; |
pmv[5].x = EVEN(pmv[3].x); |
1122 |
|
pmv[5].y = EVEN(pmv[3].y); |
1123 |
|
} else pmv[5].x = pmv[5].y = 0; |
1124 |
|
|
1125 |
/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour |
1126 |
Otherwise select large Diamond Search. |
else pmv[3].x = pmv[3].y = 0; |
|
*/ |
|
1127 |
|
|
1128 |
if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536) || (bPredEq) ) |
if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour |
1129 |
iDiamondSize=1; // halfpel! |
else pmv[4].x = pmv[4].y = 0; |
|
else |
|
|
iDiamondSize=2; // halfpel! |
|
1130 |
|
|
1131 |
if (!(MotionFlags & PMV_HALFPELDIAMOND16) ) |
// [1] median prediction |
1132 |
iDiamondSize*=2; |
pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); |
1133 |
|
|
1134 |
/* Step 4: Calculate SAD around the Median prediction. |
pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask |
|
MinSAD=SAD |
|
|
If Motion Vector equal to Previous frame motion vector |
|
|
and MinSAD<PrevFrmSAD goto Step 10. |
|
|
If SAD<=256 goto Step 10. |
|
|
*/ |
|
1135 |
|
|
1136 |
|
pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
1137 |
|
pmv[2].y = EVEN(prevMB->mvs[0].y); |
1138 |
|
|
1139 |
// Prepare for main loop |
if ((x < iWcount-1) && (y < iHcount-1)) { |
1140 |
|
pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame |
1141 |
|
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
1142 |
|
} else pmv[6].x = pmv[6].y = 0; |
1143 |
|
|
1144 |
*currMV=pmv[0]; /* current best := prediction */ |
if (rrv) { |
1145 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
int i; |
1146 |
{ /* This should NOT be necessary! */ |
for (i = 0; i < 7; i++) { |
1147 |
currMV->x = EVEN(currMV->x); |
pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); |
1148 |
currMV->y = EVEN(currMV->y); |
pmv[i].y = RRV_MV_SCALEUP(pmv[i].y); |
1149 |
} |
} |
|
|
|
|
if (currMV->x > max_dx) |
|
|
{ |
|
|
currMV->x=max_dx; |
|
1150 |
} |
} |
|
if (currMV->x < min_dx) |
|
|
{ |
|
|
currMV->x=min_dx; |
|
1151 |
} |
} |
1152 |
if (currMV->y > max_dy) |
|
1153 |
|
static int |
1154 |
|
ModeDecision(const uint32_t iQuant, SearchData * const Data, |
1155 |
|
int inter4v, |
1156 |
|
MACROBLOCK * const pMB, |
1157 |
|
const MACROBLOCK * const pMBs, |
1158 |
|
const int x, const int y, |
1159 |
|
const MBParam * const pParam, |
1160 |
|
const uint32_t MotionFlags, |
1161 |
|
const uint32_t GlobalFlags) |
1162 |
{ |
{ |
1163 |
currMV->y=max_dy; |
|
1164 |
|
int mode = MODE_INTER; |
1165 |
|
|
1166 |
|
if (!(GlobalFlags & XVID_MODEDECISION_BITS)) { //normal, fast, SAD-based mode decision |
1167 |
|
int sad; |
1168 |
|
int InterBias = MV16_INTER_BIAS; |
1169 |
|
if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
1170 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) { |
1171 |
|
mode = MODE_INTER; |
1172 |
|
sad = Data->iMinSAD[0]; |
1173 |
|
} else { |
1174 |
|
mode = MODE_INTER4V; |
1175 |
|
sad = Data->iMinSAD[1] + Data->iMinSAD[2] + |
1176 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant; |
1177 |
|
Data->iMinSAD[0] = sad; |
1178 |
} |
} |
1179 |
if (currMV->y < min_dy) |
|
1180 |
{ |
/* intra decision */ |
1181 |
currMV->y=min_dy; |
|
1182 |
|
if (iQuant > 8) InterBias += 100 * (iQuant - 8); // to make high quants work |
1183 |
|
if (y != 0) |
1184 |
|
if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; |
1185 |
|
if (x != 0) |
1186 |
|
if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; |
1187 |
|
|
1188 |
|
if (Data->chroma) InterBias += 50; // to compensate bigger SAD |
1189 |
|
if (Data->rrv) InterBias *= 4; |
1190 |
|
|
1191 |
|
if (InterBias < pMB->sad16) { |
1192 |
|
int32_t deviation; |
1193 |
|
if (!Data->rrv) deviation = dev16(Data->Cur, Data->iEdgedWidth); |
1194 |
|
else deviation = dev16(Data->Cur, Data->iEdgedWidth) + |
1195 |
|
dev16(Data->Cur+8, Data->iEdgedWidth) + |
1196 |
|
dev16(Data->Cur + 8*Data->iEdgedWidth, Data->iEdgedWidth) + |
1197 |
|
dev16(Data->Cur+8+8*Data->iEdgedWidth, Data->iEdgedWidth); |
1198 |
|
|
1199 |
|
if (deviation < (sad - InterBias)) return MODE_INTRA; |
1200 |
} |
} |
1201 |
|
return mode; |
1202 |
|
|
1203 |
iMinSAD = sad16( cur, |
} else { |
|
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
|
|
iEdgedWidth, MV_MAX_ERROR); |
|
|
iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; |
|
1204 |
|
|
1205 |
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,prevMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
int bits, intra, i; |
1206 |
{ |
VECTOR backup[5], *v; |
1207 |
|
Data->lambda16 = iQuant; |
1208 |
|
Data->lambda8 = pParam->m_quant_type; |
1209 |
|
|
1210 |
if (MotionFlags & PMV_QUICKSTOP16) |
v = Data->qpel ? Data->currentQMV : Data->currentMV; |
1211 |
goto PMVfast16_Terminate_without_Refine; |
for (i = 0; i < 5; i++) { |
1212 |
if (MotionFlags & PMV_EARLYSTOP16) |
Data->iMinSAD[i] = 256*4096; |
1213 |
goto PMVfast16_Terminate_with_Refine; |
backup[i] = v[i]; |
1214 |
} |
} |
1215 |
|
|
1216 |
/* |
bits = CountMBBitsInter(Data, pMBs, x, y, pParam, MotionFlags); |
1217 |
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
if (bits == 0) return MODE_INTER; // quick stop |
|
Also calculate (0,0) but do not subtract offset. |
|
|
Let MinSAD be the smallest SAD up to this point. |
|
|
If MV is (0,0) subtract offset. |
|
|
*/ |
|
1218 |
|
|
1219 |
// (0,0) is always possible |
if (inter4v) { |
1220 |
|
int bits_inter4v = CountMBBitsInter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup); |
1221 |
|
if (bits_inter4v < bits) { Data->iMinSAD[0] = bits = bits_inter4v; mode = MODE_INTER4V; } |
1222 |
|
} |
1223 |
|
|
1224 |
CHECK_MV16_ZERO; |
intra = CountMBBitsIntra(Data); |
1225 |
|
|
1226 |
// previous frame MV is always possible |
if (intra < bits) { *Data->iMinSAD = bits = intra; return MODE_INTRA; } |
|
CHECK_MV16_CANDIDATE(prevMB->mvs[0].x,prevMB->mvs[0].y); |
|
1227 |
|
|
1228 |
// left neighbour, if allowed |
return mode; |
|
if (x != 0) |
|
|
{ |
|
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
|
|
{ pmv[1].x = EVEN(pmv[1].x); |
|
|
pmv[1].y = EVEN(pmv[1].y); |
|
1229 |
} |
} |
|
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
|
1230 |
} |
} |
1231 |
|
|
1232 |
// top neighbour, if allowed |
static void |
1233 |
if (y != 0) |
SearchP(const IMAGE * const pRef, |
1234 |
|
const uint8_t * const pRefH, |
1235 |
|
const uint8_t * const pRefV, |
1236 |
|
const uint8_t * const pRefHV, |
1237 |
|
const IMAGE * const pCur, |
1238 |
|
const int x, |
1239 |
|
const int y, |
1240 |
|
const uint32_t MotionFlags, |
1241 |
|
const uint32_t GlobalFlags, |
1242 |
|
const uint32_t iQuant, |
1243 |
|
SearchData * const Data, |
1244 |
|
const MBParam * const pParam, |
1245 |
|
const MACROBLOCK * const pMBs, |
1246 |
|
const MACROBLOCK * const prevMBs, |
1247 |
|
int inter4v, |
1248 |
|
MACROBLOCK * const pMB) |
1249 |
{ |
{ |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
|
|
{ pmv[2].x = EVEN(pmv[2].x); |
|
|
pmv[2].y = EVEN(pmv[2].y); |
|
|
} |
|
|
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
|
1250 |
|
|
1251 |
// top right neighbour, if allowed |
int i, iDirection = 255, mask, threshA; |
1252 |
if ((uint32_t)x != (iWcount-1)) |
VECTOR pmv[7]; |
|
{ |
|
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
|
|
{ pmv[3].x = EVEN(pmv[3].x); |
|
|
pmv[3].y = EVEN(pmv[3].y); |
|
|
} |
|
|
CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); |
|
|
} |
|
|
} |
|
1253 |
|
|
1254 |
if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) && (iSAD <= iQuant * 96) ) |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1255 |
iMinSAD -= MV16_00_BIAS; |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
1256 |
|
|
1257 |
|
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); |
1258 |
|
|
1259 |
/* Step 6: If MinSAD <= thresa goto Step 10. |
Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache |
1260 |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
i = Data->rrv ? 2 : 1; |
1261 |
*/ |
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i; |
1262 |
|
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1263 |
|
Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1264 |
|
|
1265 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,prevMB->mvs[0]) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16*i; |
1266 |
{ |
Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16*i; |
1267 |
if (MotionFlags & PMV_QUICKSTOP16) |
Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16*i; |
1268 |
goto PMVfast16_Terminate_without_Refine; |
Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16*i; |
1269 |
if (MotionFlags & PMV_EARLYSTOP16) |
Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1270 |
goto PMVfast16_Terminate_with_Refine; |
Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
|
} |
|
1271 |
|
|
1272 |
|
Data->lambda16 = lambda_vec16[iQuant]; |
1273 |
|
Data->lambda8 = lambda_vec8[iQuant]; |
1274 |
|
Data->qpel_precision = 0; |
1275 |
|
|
1276 |
/************ (Diamond Search) **************/ |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
|
/* |
|
|
Step 7: Perform Diamond search, with either the small or large diamond. |
|
|
If Found=2 only examine one Diamond pattern, and afterwards goto step 10 |
|
|
Step 8: If small diamond, iterate small diamond search pattern until motion vector lies in the center of the diamond. |
|
|
If center then goto step 10. |
|
|
Step 9: If large diamond, iterate large diamond search pattern until motion vector lies in the center. |
|
|
Refine by using small diamond and goto step 10. |
|
|
*/ |
|
1277 |
|
|
1278 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
memset(Data->currentMV, 0, 5*sizeof(VECTOR)); |
1279 |
|
|
1280 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
1281 |
iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
else Data->predMV = pmv[0]; |
|
x, y, |
|
|
currMV->x, currMV->y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
1282 |
|
|
1283 |
if (iSAD < iMinSAD) |
i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0); |
1284 |
{ |
Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10); |
1285 |
*currMV = newMV; |
Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10); |
1286 |
iMinSAD = iSAD; |
Data->iMinSAD[2] = pMB->sad8[1]; |
1287 |
} |
Data->iMinSAD[3] = pMB->sad8[2]; |
1288 |
|
Data->iMinSAD[4] = pMB->sad8[3]; |
1289 |
|
|
1290 |
if (MotionFlags & PMV_EXTSEARCH16) |
if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) || (x | y)) { |
1291 |
{ |
threshA = Data->temp[0]; // that's where we keep this SAD atm |
1292 |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
if (threshA < 512) threshA = 512; |
1293 |
|
else if (threshA > 1024) threshA = 1024; |
1294 |
|
} else |
1295 |
|
threshA = 512; |
1296 |
|
|
1297 |
if (!(MVequal(pmv[0],backupMV)) ) |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
1298 |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
prevMBs + x + y * pParam->mb_width, Data->rrv); |
|
x, y, |
|
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
1299 |
|
|
1300 |
if (iSAD < iMinSAD) |
if (!Data->rrv) { |
1301 |
{ |
if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16; |
1302 |
*currMV = newMV; |
else CheckCandidate = CheckCandidate16no4v; //for extra speed |
1303 |
iMinSAD = iSAD; |
} else CheckCandidate = CheckCandidate32; |
|
} |
|
|
} |
|
1304 |
|
|
1305 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
/* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/ |
|
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
0, 0, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
1306 |
|
|
1307 |
if (iSAD < iMinSAD) |
for (i = 1; i < 7; i++) { |
1308 |
{ |
if (!(mask = make_mask(pmv, i)) ) continue; |
1309 |
*currMV = newMV; |
CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1310 |
iMinSAD = iSAD; |
if (Data->iMinSAD[0] <= threshA) break; |
1311 |
} |
} |
|
} |
|
|
} |
|
|
|
|
|
/* |
|
|
Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
|
|
*/ |
|
1312 |
|
|
1313 |
PMVfast16_Terminate_with_Refine: |
if ((Data->iMinSAD[0] <= threshA) || |
1314 |
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
1315 |
iMinSAD = Halfpel16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { |
1316 |
x, y, |
if (!(GlobalFlags & XVID_MODEDECISION_BITS)) inter4v = 0; } |
1317 |
currMV, iMinSAD, |
else { |
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
|
1318 |
|
|
1319 |
PMVfast16_Terminate_without_Refine: |
MainSearchFunc * MainSearchPtr; |
1320 |
currPMV->x = currMV->x - pmv[0].x; |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1321 |
currPMV->y = currMV->y - pmv[0].y; |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1322 |
return iMinSAD; |
else MainSearchPtr = DiamondSearch; |
|
} |
|
1323 |
|
|
1324 |
|
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
1325 |
|
|
1326 |
|
/* extended search, diamond starting in 0,0 and in prediction. |
1327 |
|
note that this search is/might be done in halfpel positions, |
1328 |
|
which makes it more different than the diamond above */ |
1329 |
|
|
1330 |
|
if (MotionFlags & PMV_EXTSEARCH16) { |
1331 |
|
int32_t bSAD; |
1332 |
|
VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; |
1333 |
|
if (Data->rrv) { |
1334 |
|
startMV.x = RRV_MV_SCALEUP(startMV.x); |
1335 |
|
startMV.y = RRV_MV_SCALEUP(startMV.y); |
1336 |
|
} |
1337 |
|
if (!(MVequal(startMV, backupMV))) { |
1338 |
|
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1339 |
|
|
1340 |
|
CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data); |
1341 |
|
MainSearchPtr(startMV.x, startMV.y, Data, 255); |
1342 |
|
if (bSAD < Data->iMinSAD[0]) { |
1343 |
|
Data->currentMV[0] = backupMV; |
1344 |
|
Data->iMinSAD[0] = bSAD; } |
1345 |
|
} |
1346 |
|
|
1347 |
|
backupMV = Data->currentMV[0]; |
1348 |
|
startMV.x = startMV.y = 1; |
1349 |
|
if (!(MVequal(startMV, backupMV))) { |
1350 |
|
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1351 |
|
|
1352 |
int32_t Diamond8_MainSearch( |
CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data); |
1353 |
const uint8_t * const pRef, |
MainSearchPtr(startMV.x, startMV.y, Data, 255); |
1354 |
const uint8_t * const pRefH, |
if (bSAD < Data->iMinSAD[0]) { |
1355 |
const uint8_t * const pRefV, |
Data->currentMV[0] = backupMV; |
1356 |
const uint8_t * const pRefHV, |
Data->iMinSAD[0] = bSAD; } |
|
const uint8_t * const cur, |
|
|
const int x, const int y, |
|
|
int32_t startx, int32_t starty, |
|
|
int32_t iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const VECTOR * const pmv, |
|
|
const int32_t min_dx, const int32_t max_dx, |
|
|
const int32_t min_dy, const int32_t max_dy, |
|
|
const int32_t iEdgedWidth, |
|
|
const int32_t iDiamondSize, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
int iFound) |
|
|
{ |
|
|
/* Do a diamond search around given starting point, return SAD of best */ |
|
|
|
|
|
int32_t iDirection=0; |
|
|
int32_t iSAD; |
|
|
VECTOR backupMV; |
|
|
backupMV.x = startx; |
|
|
backupMV.y = starty; |
|
|
|
|
|
/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ |
|
|
|
|
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
|
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
|
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
|
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
|
|
|
|
|
if (iDirection) |
|
|
while (!iFound) |
|
|
{ |
|
|
iFound = 1; |
|
|
backupMV=*currMV; // since iDirection!=0, this is well defined! |
|
|
|
|
|
if ( iDirection != 2) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
|
|
if ( iDirection != 1) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x+iDiamondSize,backupMV.y,2); |
|
|
if ( iDirection != 4) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x,backupMV.y-iDiamondSize,3); |
|
|
if ( iDirection != 3) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x,backupMV.y+iDiamondSize,4); |
|
1357 |
} |
} |
|
else |
|
|
{ |
|
|
currMV->x = startx; |
|
|
currMV->y = starty; |
|
1358 |
} |
} |
|
return iMinSAD; |
|
1359 |
} |
} |
1360 |
|
|
1361 |
int32_t Halfpel8_Refine( |
if (MotionFlags & PMV_HALFPELREFINE16) |
1362 |
const uint8_t * const pRef, |
SubpelRefine(Data); |
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, const int y, |
|
|
VECTOR * const currMV, |
|
|
int32_t iMinSAD, |
|
|
const VECTOR * const pmv, |
|
|
const int32_t min_dx, const int32_t max_dx, |
|
|
const int32_t min_dy, const int32_t max_dy, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
const int32_t iEdgedWidth) |
|
|
{ |
|
|
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
|
1363 |
|
|
1364 |
int32_t iSAD; |
for(i = 0; i < 5; i++) { |
1365 |
VECTOR backupMV = *currMV; |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors |
1366 |
|
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
1367 |
|
} |
1368 |
|
|
1369 |
CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y-1); |
if (MotionFlags & PMV_QUARTERPELREFINE16) { |
|
CHECK_MV8_CANDIDATE(backupMV.x ,backupMV.y-1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y-1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y+1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x ,backupMV.y+1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y+1); |
|
1370 |
|
|
1371 |
return iMinSAD; |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1372 |
|
pParam->width, pParam->height, Data->iFcode, 1, 0); |
1373 |
|
Data->qpel_precision = 1; |
1374 |
|
SubpelRefine(Data); |
1375 |
} |
} |
1376 |
|
|
1377 |
|
if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) && (Data->iMinSAD[0] < (int32_t)iQuant * 30)) inter4v = 0; |
1378 |
|
|
1379 |
#define PMV_HALFPEL8 (PMV_HALFPELDIAMOND8|PMV_HALFPELREFINE8) |
if (inter4v) { |
1380 |
|
SearchData Data8; |
1381 |
|
memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data |
1382 |
|
|
1383 |
int32_t PMVfastSearch8( |
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
1384 |
const uint8_t * const pRef, |
Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); |
1385 |
const uint8_t * const pRefH, |
Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); |
1386 |
const uint8_t * const pRefV, |
Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); |
|
const uint8_t * const pRefHV, |
|
|
const IMAGE * const pCur, |
|
|
const int x, const int y, |
|
|
const int start_x, const int start_y, |
|
|
const uint32_t MotionFlags, |
|
|
const uint32_t iQuant, |
|
|
const uint32_t iFcode, |
|
|
const MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const MACROBLOCK * const prevMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV) |
|
|
{ |
|
|
const uint32_t iWcount = pParam->mb_width; |
|
|
const int32_t iWidth = pParam->width; |
|
|
const int32_t iHeight = pParam->height; |
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
|
|
|
|
const uint8_t * cur = pCur->y + x*8 + y*8*iEdgedWidth; |
|
1387 |
|
|
1388 |
int32_t iDiamondSize; |
if ((Data->chroma) && (!(GlobalFlags & XVID_MODEDECISION_BITS))) { |
1389 |
|
// chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, there is no reason to compute it |
1390 |
|
int sumx = 0, sumy = 0; |
1391 |
|
const int div = Data->qpel ? 2 : 0; |
1392 |
|
const VECTOR * const mv = Data->qpel ? pMB->qmvs : pMB->mvs; |
1393 |
|
|
1394 |
int32_t min_dx; |
for (i = 0; i < 4; i++) { |
1395 |
int32_t max_dx; |
sumx += mv[i].x / div; |
1396 |
int32_t min_dy; |
sumy += mv[i].y / div; |
1397 |
int32_t max_dy; |
} |
|
|
|
|
VECTOR pmv[4]; |
|
|
int32_t psad[4]; |
|
|
VECTOR newMV; |
|
|
VECTOR backupMV; |
|
1398 |
|
|
1399 |
const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf], |
1400 |
const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; |
(sumy >> 3) + roundtab_76[sumy & 0xf], Data); |
1401 |
|
} |
1402 |
|
} |
1403 |
|
|
1404 |
static int32_t threshA,threshB; |
inter4v = ModeDecision(iQuant, Data, inter4v, pMB, pMBs, x, y, pParam, MotionFlags, GlobalFlags); |
|
int32_t iFound,bPredEq; |
|
|
int32_t iMinSAD,iSAD; |
|
1405 |
|
|
1406 |
int32_t iSubBlock = ((y&1)<<1) + (x&1); |
if (Data->rrv) { |
1407 |
|
Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); |
1408 |
|
Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); |
1409 |
|
} |
1410 |
|
|
1411 |
/* Get maximum range */ |
if (inter4v == MODE_INTER) { |
1412 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
pMB->mode = MODE_INTER; |
1413 |
x, y, 8, iWidth, iHeight, iFcode); |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
1414 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
1415 |
|
|
1416 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8 )) |
if(Data->qpel) { |
1417 |
{ min_dx = EVEN(min_dx); |
pMB->qmvs[0] = pMB->qmvs[1] |
1418 |
max_dx = EVEN(max_dx); |
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
1419 |
min_dy = EVEN(min_dy); |
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; |
1420 |
max_dy = EVEN(max_dy); |
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; |
1421 |
} /* because we might use IF (dx>max_dx) THEN dx=max_dx; */ |
} else { |
1422 |
|
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
1423 |
|
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
1424 |
|
} |
1425 |
|
|
1426 |
|
} else if (inter4v == MODE_INTER4V) { |
1427 |
|
pMB->mode = MODE_INTER4V; |
1428 |
|
pMB->sad16 = Data->iMinSAD[0]; |
1429 |
|
} else { // INTRA mode |
1430 |
|
SkipMacroblockP(pMB, 0); // not skip, but similar enough |
1431 |
|
pMB->mode = MODE_INTRA; |
1432 |
|
} |
1433 |
|
|
1434 |
bPredEq = get_pmvdata(pMBs, (x>>1), (y>>1), iWcount, iSubBlock, pmv, psad); |
} |
1435 |
|
|
1436 |
if ((x==0) && (y==0) ) |
static void |
1437 |
|
Search8(const SearchData * const OldData, |
1438 |
|
const int x, const int y, |
1439 |
|
const uint32_t MotionFlags, |
1440 |
|
const MBParam * const pParam, |
1441 |
|
MACROBLOCK * const pMB, |
1442 |
|
const MACROBLOCK * const pMBs, |
1443 |
|
const int block, |
1444 |
|
SearchData * const Data) |
1445 |
{ |
{ |
1446 |
threshA = 512/4; |
int i = 0; |
1447 |
threshB = 1024/4; |
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
1448 |
|
Data->currentMV = OldData->currentMV + 1 + block; |
1449 |
|
Data->currentQMV = OldData->currentQMV + 1 + block; |
1450 |
|
|
1451 |
} |
if(Data->qpel) { |
1452 |
else |
Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
1453 |
{ |
if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y, |
1454 |
threshA = psad[0]/4; /* good estimate */ |
Data->predMV, Data->iFcode, 0, 0); |
1455 |
threshB = threshA+256/4; |
} else { |
1456 |
if (threshA< 512/4) threshA = 512/4; |
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
1457 |
if (threshA>1024/4) threshA = 1024/4; |
if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y, |
1458 |
if (threshB>1792/4) threshB = 1792/4; |
Data->predMV, Data->iFcode, 0, Data->rrv); |
1459 |
} |
} |
1460 |
|
|
1461 |
iFound=0; |
*(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10; |
|
|
|
|
/* Step 2: Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
|
|
vector of the median. |
|
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
|
|
*/ |
|
1462 |
|
|
1463 |
if ((bPredEq) && (MVequal(pmv[0],prevMB->mvs[iSubBlock]) ) ) |
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8|PMV_QUARTERPELREFINE8)) { |
|
iFound=2; |
|
1464 |
|
|
1465 |
/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
if (Data->rrv) i = 16; else i = 8; |
|
Otherwise select large Diamond Search. |
|
|
*/ |
|
1466 |
|
|
1467 |
if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536/4) || (bPredEq) ) |
Data->RefP[0] = OldData->RefP[0] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1468 |
iDiamondSize=1; // 1 halfpel! |
Data->RefP[1] = OldData->RefP[1] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1469 |
else |
Data->RefP[2] = OldData->RefP[2] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1470 |
iDiamondSize=2; // 2 halfpel = 1 full pixel! |
Data->RefP[3] = OldData->RefP[3] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1471 |
|
|
1472 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8) ) |
Data->Cur = OldData->Cur + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1473 |
iDiamondSize*=2; |
Data->qpel_precision = 0; |
1474 |
|
|
1475 |
/* Step 4: Calculate SAD around the Median prediction. |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
1476 |
MinSAD=SAD |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
|
If Motion Vector equal to Previous frame motion vector |
|
|
and MinSAD<PrevFrmSAD goto Step 10. |
|
|
If SAD<=256 goto Step 10. |
|
|
*/ |
|
1477 |
|
|
1478 |
|
if (!Data->rrv) CheckCandidate = CheckCandidate8; |
1479 |
|
else CheckCandidate = CheckCandidate16no4v; |
1480 |
|
|
1481 |
// Prepare for main loop |
if (MotionFlags & PMV_EXTSEARCH8 && (!(MotionFlags & EXTSEARCH_BITS))) { |
1482 |
|
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
1483 |
|
|
1484 |
currMV->x=start_x; /* start with mv16 */ |
MainSearchFunc *MainSearchPtr; |
1485 |
currMV->y=start_y; |
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
1486 |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
1487 |
|
else MainSearchPtr = DiamondSearch; |
1488 |
|
|
1489 |
iMinSAD = sad8( cur, |
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255); |
|
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
|
|
iEdgedWidth); |
|
|
iMinSAD += calc_delta_8(currMV->x - pmv[0].x, currMV->y - pmv[0].y, (uint8_t)iFcode) * iQuant; |
|
1490 |
|
|
1491 |
if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,prevMB->mvs[iSubBlock])) && ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
if(*(Data->iMinSAD) < temp_sad) { |
1492 |
{ |
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
1493 |
if (MotionFlags & PMV_QUICKSTOP16) |
Data->currentQMV->y = 2 * Data->currentMV->y; |
1494 |
goto PMVfast8_Terminate_without_Refine; |
} |
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfast8_Terminate_with_Refine; |
|
1495 |
} |
} |
1496 |
|
|
1497 |
|
if (MotionFlags & PMV_HALFPELREFINE8) { |
1498 |
|
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
1499 |
|
|
1500 |
/* |
SubpelRefine(Data); // perform halfpel refine of current best vector |
|
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
|
|
Also calculate (0,0) but do not subtract offset. |
|
|
Let MinSAD be the smallest SAD up to this point. |
|
|
If MV is (0,0) subtract offset. |
|
|
*/ |
|
|
|
|
|
// the prediction might be even better than mv16 |
|
|
CHECK_MV8_CANDIDATE(pmv[0].x,pmv[0].y); |
|
|
|
|
|
// (0,0) is always possible |
|
|
CHECK_MV8_ZERO; |
|
|
|
|
|
// previous frame MV is always possible |
|
|
CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x,prevMB->mvs[iSubBlock].y); |
|
1501 |
|
|
1502 |
// left neighbour, if allowed |
if(*(Data->iMinSAD) < temp_sad) { // we have found a better match |
1503 |
if (psad[1] != MV_MAX_ERROR) |
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
1504 |
{ |
Data->currentQMV->y = 2 * Data->currentMV->y; |
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
|
|
{ pmv[1].x = EVEN(pmv[1].x); |
|
|
pmv[1].y = EVEN(pmv[1].y); |
|
|
} |
|
|
CHECK_MV8_CANDIDATE(pmv[1].x,pmv[1].y); |
|
1505 |
} |
} |
|
|
|
|
// top neighbour, if allowed |
|
|
if (psad[2] != MV_MAX_ERROR) |
|
|
{ |
|
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
|
|
{ pmv[2].x = EVEN(pmv[2].x); |
|
|
pmv[2].y = EVEN(pmv[2].y); |
|
1506 |
} |
} |
|
CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); |
|
1507 |
|
|
1508 |
// top right neighbour, if allowed |
if (Data->qpel && MotionFlags & PMV_QUARTERPELREFINE8) { |
1509 |
if (psad[3] != MV_MAX_ERROR) |
Data->qpel_precision = 1; |
1510 |
{ |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
1511 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
pParam->width, pParam->height, Data->iFcode, 1, 0); |
1512 |
{ pmv[3].x = EVEN(pmv[3].x); |
SubpelRefine(Data); |
|
pmv[3].y = EVEN(pmv[3].y); |
|
|
} |
|
|
CHECK_MV8_CANDIDATE(pmv[3].x,pmv[3].y); |
|
1513 |
} |
} |
1514 |
} |
} |
1515 |
|
|
1516 |
if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) && (iSAD <= iQuant * 96) ) |
if (Data->rrv) { |
1517 |
iMinSAD -= MV8_00_BIAS; |
Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x); |
1518 |
|
Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y); |
1519 |
|
} |
|
/* Step 6: If MinSAD <= thresa goto Step 10. |
|
|
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
|
|
*/ |
|
1520 |
|
|
1521 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,prevMB->mvs[iSubBlock]) && ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
if(Data->qpel) { |
1522 |
{ |
pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x; |
1523 |
if (MotionFlags & PMV_QUICKSTOP16) |
pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y; |
1524 |
goto PMVfast8_Terminate_without_Refine; |
pMB->qmvs[block] = *Data->currentQMV; |
1525 |
if (MotionFlags & PMV_EARLYSTOP16) |
} else { |
1526 |
goto PMVfast8_Terminate_with_Refine; |
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
1527 |
} |
pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; |
1528 |
|
} |
|
/************ (Diamond Search) **************/ |
|
|
/* |
|
|
Step 7: Perform Diamond search, with either the small or large diamond. |
|
|
If Found=2 only examine one Diamond pattern, and afterwards goto step 10 |
|
|
Step 8: If small diamond, iterate small diamond search pattern until motion vector lies in the center of the diamond. |
|
|
If center then goto step 10. |
|
|
Step 9: If large diamond, iterate large diamond search pattern until motion vector lies in the center. |
|
|
Refine by using small diamond and goto step 10. |
|
|
*/ |
|
1529 |
|
|
1530 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
pMB->mvs[block] = *Data->currentMV; |
1531 |
|
pMB->sad8[block] = 4 * *Data->iMinSAD; |
1532 |
|
} |
1533 |
|
|
1534 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
/* motion estimation for B-frames */ |
|
iSAD = Diamond8_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
currMV->x, currMV->y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
1535 |
|
|
1536 |
if (iSAD < iMinSAD) |
static __inline VECTOR |
1537 |
|
ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) |
1538 |
{ |
{ |
1539 |
*currMV = newMV; |
/* the stupidiest function ever */ |
1540 |
iMinSAD = iSAD; |
return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]); |
1541 |
} |
} |
1542 |
|
|
1543 |
if (MotionFlags & PMV_EXTSEARCH8) |
static void __inline |
1544 |
|
PreparePredictionsBF(VECTOR * const pmv, const int x, const int y, |
1545 |
|
const uint32_t iWcount, |
1546 |
|
const MACROBLOCK * const pMB, |
1547 |
|
const uint32_t mode_curr) |
1548 |
{ |
{ |
|
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
|
1549 |
|
|
1550 |
if (!(MVequal(pmv[0],backupMV)) ) |
// [0] is prediction |
1551 |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); |
|
x, y, |
|
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
1552 |
|
|
1553 |
if (iSAD < iMinSAD) |
pmv[1].x = pmv[1].y = 0; // [1] is zero |
|
{ |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
|
|
|
|
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
|
|
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
0, 0, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) |
|
|
{ |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
|
} |
|
1554 |
|
|
1555 |
/* Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
pmv[2] = ChoosePred(pMB, mode_curr); |
1556 |
By performing an optional local half-pixel search, we can refine this result even further. |
pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); |
|
*/ |
|
1557 |
|
|
1558 |
PMVfast8_Terminate_with_Refine: |
if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour |
1559 |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); |
1560 |
iMinSAD = Halfpel8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); |
1561 |
x, y, |
} else pmv[3].x = pmv[3].y = 0; |
|
currMV, iMinSAD, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
|
1562 |
|
|
1563 |
|
if (y != 0) { |
1564 |
|
pmv[4] = ChoosePred(pMB-iWcount, mode_curr); |
1565 |
|
pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y); |
1566 |
|
} else pmv[4].x = pmv[4].y = 0; |
1567 |
|
|
1568 |
PMVfast8_Terminate_without_Refine: |
if (x != 0) { |
1569 |
currPMV->x = currMV->x - pmv[0].x; |
pmv[5] = ChoosePred(pMB-1, mode_curr); |
1570 |
currPMV->y = currMV->y - pmv[0].y; |
pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
1571 |
|
} else pmv[5].x = pmv[5].y = 0; |
1572 |
|
|
1573 |
return iMinSAD; |
if (x != 0 && y != 0) { |
1574 |
|
pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); |
1575 |
|
pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y); |
1576 |
|
} else pmv[6].x = pmv[6].y = 0; |
1577 |
} |
} |
1578 |
|
|
1579 |
int32_t EPZSSearch16( |
|
1580 |
const uint8_t * const pRef, |
/* search backward or forward */ |
1581 |
|
static void |
1582 |
|
SearchBF( const IMAGE * const pRef, |
1583 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1584 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
1585 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
1586 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
1587 |
const int x, const int y, |
const int x, const int y, |
1588 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
|
const uint32_t iQuant, |
|
1589 |
const uint32_t iFcode, |
const uint32_t iFcode, |
1590 |
const MBParam * const pParam, |
const MBParam * const pParam, |
1591 |
const MACROBLOCK * const pMBs, |
MACROBLOCK * const pMB, |
1592 |
const MACROBLOCK * const prevMBs, |
const VECTOR * const predMV, |
1593 |
VECTOR * const currMV, |
int32_t * const best_sad, |
1594 |
VECTOR * const currPMV) |
const int32_t mode_current, |
1595 |
|
SearchData * const Data) |
1596 |
{ |
{ |
|
const uint32_t iWcount = pParam->mb_width; |
|
|
const uint32_t iHcount = pParam->mb_height; |
|
1597 |
|
|
1598 |
const int32_t iWidth = pParam->width; |
int i, iDirection = 255, mask; |
1599 |
const int32_t iHeight = pParam->height; |
VECTOR pmv[7]; |
1600 |
const int32_t iEdgedWidth = pParam->edged_width; |
MainSearchFunc *MainSearchPtr; |
1601 |
|
*Data->iMinSAD = MV_MAX_ERROR; |
1602 |
|
Data->iFcode = iFcode; |
1603 |
|
Data->qpel_precision = 0; |
1604 |
|
Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; // reset chroma-sad cache |
1605 |
|
|
1606 |
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16; |
1607 |
|
Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16; |
1608 |
|
Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16; |
1609 |
|
Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16; |
1610 |
|
Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
1611 |
|
Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
1612 |
|
|
1613 |
int32_t min_dx; |
Data->predMV = *predMV; |
|
int32_t max_dx; |
|
|
int32_t min_dy; |
|
|
int32_t max_dy; |
|
1614 |
|
|
1615 |
VECTOR newMV; |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1616 |
VECTOR backupMV; |
pParam->width, pParam->height, iFcode - Data->qpel, 0, 0); |
1617 |
|
|
1618 |
VECTOR pmv[4]; |
pmv[0] = Data->predMV; |
1619 |
int32_t psad[8]; |
if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } |
1620 |
|
|
1621 |
static MACROBLOCK * oldMBs = NULL; |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
|
const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
|
|
const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; |
|
|
MACROBLOCK * oldMB = NULL; |
|
1622 |
|
|
1623 |
static int32_t thresh2; |
Data->currentMV->x = Data->currentMV->y = 0; |
1624 |
int32_t bPredEq; |
CheckCandidate = CheckCandidate16no4v; |
|
int32_t iMinSAD,iSAD=9999; |
|
1625 |
|
|
1626 |
MainSearch16FuncPtr EPZSMainSearchPtr; |
// main loop. checking all predictions |
1627 |
|
for (i = 0; i < 7; i++) { |
1628 |
if (oldMBs == NULL) |
if (!(mask = make_mask(pmv, i)) ) continue; |
1629 |
{ oldMBs = (MACROBLOCK*) calloc(1,iWcount*iHcount*sizeof(MACROBLOCK)); |
CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
|
// fprintf(stderr,"allocated %d bytes for oldMBs\n",iWcount*iHcount*sizeof(MACROBLOCK)); |
|
1630 |
} |
} |
|
oldMB = oldMBs + x + y * iWcount; |
|
1631 |
|
|
1632 |
/* Get maximum range */ |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1633 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1634 |
x, y, 16, iWidth, iHeight, iFcode); |
else MainSearchPtr = DiamondSearch; |
1635 |
|
|
1636 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
|
{ min_dx = EVEN(min_dx); |
|
|
max_dx = EVEN(max_dx); |
|
|
min_dy = EVEN(min_dy); |
|
|
max_dy = EVEN(max_dy); |
|
|
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
|
1637 |
|
|
1638 |
bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
SubpelRefine(Data); |
1639 |
|
|
1640 |
/* Step 4: Calculate SAD around the Median prediction. |
if (Data->qpel && *Data->iMinSAD < *best_sad + 300) { |
1641 |
MinSAD=SAD |
Data->currentQMV->x = 2*Data->currentMV->x; |
1642 |
If Motion Vector equal to Previous frame motion vector |
Data->currentQMV->y = 2*Data->currentMV->y; |
1643 |
and MinSAD<PrevFrmSAD goto Step 10. |
Data->qpel_precision = 1; |
1644 |
If SAD<=256 goto Step 10. |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1645 |
*/ |
pParam->width, pParam->height, iFcode, 1, 0); |
1646 |
|
SubpelRefine(Data); |
1647 |
|
} |
1648 |
|
|
1649 |
// Prepare for main loop |
// three bits are needed to code backward mode. four for forward |
1650 |
|
|
1651 |
*currMV=pmv[0]; /* current best := median prediction */ |
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
1652 |
if (!(MotionFlags & PMV_HALFPEL16)) |
else *Data->iMinSAD += 3 * Data->lambda16; |
|
{ |
|
|
currMV->x = EVEN(currMV->x); |
|
|
currMV->y = EVEN(currMV->y); |
|
|
} |
|
1653 |
|
|
1654 |
if (currMV->x > max_dx) |
if (*Data->iMinSAD < *best_sad) { |
1655 |
currMV->x=max_dx; |
*best_sad = *Data->iMinSAD; |
1656 |
if (currMV->x < min_dx) |
pMB->mode = mode_current; |
1657 |
currMV->x=min_dx; |
if (Data->qpel) { |
1658 |
if (currMV->y > max_dy) |
pMB->pmvs[0].x = Data->currentQMV->x - predMV->x; |
1659 |
currMV->y=max_dy; |
pMB->pmvs[0].y = Data->currentQMV->y - predMV->y; |
1660 |
if (currMV->y < min_dy) |
if (mode_current == MODE_FORWARD) |
1661 |
currMV->y=min_dy; |
pMB->qmvs[0] = *Data->currentQMV; |
1662 |
|
else |
1663 |
|
pMB->b_qmvs[0] = *Data->currentQMV; |
1664 |
|
} else { |
1665 |
|
pMB->pmvs[0].x = Data->currentMV->x - predMV->x; |
1666 |
|
pMB->pmvs[0].y = Data->currentMV->y - predMV->y; |
1667 |
|
} |
1668 |
|
if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; |
1669 |
|
else pMB->b_mvs[0] = *Data->currentMV; |
1670 |
|
} |
1671 |
|
|
1672 |
|
if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; |
1673 |
|
else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search |
1674 |
|
} |
1675 |
|
|
1676 |
|
static void |
1677 |
|
SkipDecisionB(const IMAGE * const pCur, |
1678 |
|
const IMAGE * const f_Ref, |
1679 |
|
const IMAGE * const b_Ref, |
1680 |
|
MACROBLOCK * const pMB, |
1681 |
|
const uint32_t x, const uint32_t y, |
1682 |
|
const SearchData * const Data) |
1683 |
|
{ |
1684 |
|
int dx = 0, dy = 0, b_dx = 0, b_dy = 0; |
1685 |
|
int32_t sum; |
1686 |
|
const int div = 1 + Data->qpel; |
1687 |
|
int k; |
1688 |
|
const uint32_t stride = Data->iEdgedWidth/2; |
1689 |
|
//this is not full chroma compensation, only it's fullpel approximation. should work though |
1690 |
|
|
1691 |
|
for (k = 0; k < 4; k++) { |
1692 |
|
dy += Data->directmvF[k].y / div; |
1693 |
|
dx += Data->directmvF[k].x / div; |
1694 |
|
b_dy += Data->directmvB[k].y / div; |
1695 |
|
b_dx += Data->directmvB[k].x / div; |
1696 |
|
} |
1697 |
|
|
1698 |
|
dy = (dy >> 3) + roundtab_76[dy & 0xf]; |
1699 |
|
dx = (dx >> 3) + roundtab_76[dx & 0xf]; |
1700 |
|
b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf]; |
1701 |
|
b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf]; |
1702 |
|
|
1703 |
|
sum = sad8bi(pCur->u + 8 * x + 8 * y * stride, |
1704 |
|
f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2, |
1705 |
|
b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, |
1706 |
|
stride); |
1707 |
|
|
1708 |
|
if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; //no skip |
1709 |
|
|
1710 |
|
sum += sad8bi(pCur->v + 8*x + 8 * y * stride, |
1711 |
|
f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2, |
1712 |
|
b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, |
1713 |
|
stride); |
1714 |
|
|
1715 |
|
if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) { |
1716 |
|
pMB->mode = MODE_DIRECT_NONE_MV; //skipped |
1717 |
|
for (k = 0; k < 4; k++) { |
1718 |
|
pMB->qmvs[k] = pMB->mvs[k]; |
1719 |
|
pMB->b_qmvs[k] = pMB->b_mvs[k]; |
1720 |
|
} |
1721 |
|
} |
1722 |
|
} |
1723 |
|
|
1724 |
|
static __inline uint32_t |
1725 |
|
SearchDirect(const IMAGE * const f_Ref, |
1726 |
|
const uint8_t * const f_RefH, |
1727 |
|
const uint8_t * const f_RefV, |
1728 |
|
const uint8_t * const f_RefHV, |
1729 |
|
const IMAGE * const b_Ref, |
1730 |
|
const uint8_t * const b_RefH, |
1731 |
|
const uint8_t * const b_RefV, |
1732 |
|
const uint8_t * const b_RefHV, |
1733 |
|
const IMAGE * const pCur, |
1734 |
|
const int x, const int y, |
1735 |
|
const uint32_t MotionFlags, |
1736 |
|
const int32_t TRB, const int32_t TRD, |
1737 |
|
const MBParam * const pParam, |
1738 |
|
MACROBLOCK * const pMB, |
1739 |
|
const MACROBLOCK * const b_mb, |
1740 |
|
int32_t * const best_sad, |
1741 |
|
SearchData * const Data) |
1742 |
|
|
1743 |
|
{ |
1744 |
|
int32_t skip_sad; |
1745 |
|
int k = (x + Data->iEdgedWidth*y) * 16; |
1746 |
|
MainSearchFunc *MainSearchPtr; |
1747 |
|
|
1748 |
|
*Data->iMinSAD = 256*4096; |
1749 |
|
Data->RefP[0] = f_Ref->y + k; |
1750 |
|
Data->RefP[2] = f_RefH + k; |
1751 |
|
Data->RefP[1] = f_RefV + k; |
1752 |
|
Data->RefP[3] = f_RefHV + k; |
1753 |
|
Data->b_RefP[0] = b_Ref->y + k; |
1754 |
|
Data->b_RefP[2] = b_RefH + k; |
1755 |
|
Data->b_RefP[1] = b_RefV + k; |
1756 |
|
Data->b_RefP[3] = b_RefHV + k; |
1757 |
|
Data->RefP[4] = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; |
1758 |
|
Data->RefP[5] = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; |
1759 |
|
Data->b_RefP[4] = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; |
1760 |
|
Data->b_RefP[5] = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; |
1761 |
|
|
1762 |
|
k = Data->qpel ? 4 : 2; |
1763 |
|
Data->max_dx = k * (pParam->width - x * 16); |
1764 |
|
Data->max_dy = k * (pParam->height - y * 16); |
1765 |
|
Data->min_dx = -k * (16 + x * 16); |
1766 |
|
Data->min_dy = -k * (16 + y * 16); |
1767 |
|
|
1768 |
|
Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs; |
1769 |
|
Data->qpel_precision = 0; |
1770 |
|
|
1771 |
|
for (k = 0; k < 4; k++) { |
1772 |
|
pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); |
1773 |
|
pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD; |
1774 |
|
pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); |
1775 |
|
pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; |
1776 |
|
|
1777 |
|
if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx) |
1778 |
|
| (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) { |
1779 |
|
|
1780 |
|
*best_sad = 256*4096; // in that case, we won't use direct mode |
1781 |
|
pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV" |
1782 |
|
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
1783 |
|
return 256*4096; |
1784 |
|
} |
1785 |
|
if (b_mb->mode != MODE_INTER4V) { |
1786 |
|
pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; |
1787 |
|
pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0]; |
1788 |
|
Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0]; |
1789 |
|
Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0]; |
1790 |
|
break; |
1791 |
|
} |
1792 |
|
} |
1793 |
|
|
1794 |
/***************** This is predictor SET A: only median prediction ******************/ |
CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v; |
1795 |
|
|
1796 |
iMinSAD = sad16( cur, |
CheckCandidate(0, 0, 255, &k, Data); |
|
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
|
|
iEdgedWidth, MV_MAX_ERROR); |
|
|
iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; |
|
1797 |
|
|
1798 |
// thresh1 is fixed to 256 |
// initial (fast) skip decision |
1799 |
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,pMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (2 + Data->chroma?1:0)) { |
1800 |
{ |
//possible skip |
1801 |
if (MotionFlags & PMV_QUICKSTOP16) |
if (Data->chroma) { |
1802 |
goto EPZS16_Terminate_without_Refine; |
pMB->mode = MODE_DIRECT_NONE_MV; |
1803 |
if (MotionFlags & PMV_EARLYSTOP16) |
return *Data->iMinSAD; // skip. |
1804 |
goto EPZS16_Terminate_with_Refine; |
} else { |
1805 |
|
SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data); |
1806 |
|
if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip. |
1807 |
|
} |
1808 |
} |
} |
1809 |
|
|
1810 |
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
*Data->iMinSAD += Data->lambda16; |
1811 |
|
skip_sad = *Data->iMinSAD; |
1812 |
|
|
1813 |
// previous frame MV |
// DIRECT MODE DELTA VECTOR SEARCH. |
1814 |
CHECK_MV16_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); |
// This has to be made more effective, but at the moment I'm happy it's running at all |
1815 |
|
|
1816 |
// set threshhold based on Min of Prediction and SAD of collocated block |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1817 |
// CHECK_MV16 always uses iSAD for the SAD of last vector to check, so now iSAD is what we want |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1818 |
|
else MainSearchPtr = DiamondSearch; |
1819 |
|
|
1820 |
if ((x==0) && (y==0) ) |
MainSearchPtr(0, 0, Data, 255); |
|
{ |
|
|
thresh2 = 512; |
|
|
} |
|
|
else |
|
|
{ |
|
|
/* T_k = 1.2 * MIN(SAD_top,SAD_left,SAD_topleft,SAD_coll) +128; [Tourapis, 2002] */ |
|
1821 |
|
|
1822 |
thresh2 = MIN(psad[0],iSAD)*6/5 + 128; |
SubpelRefine(Data); |
|
} |
|
1823 |
|
|
1824 |
// MV=(0,0) is often a good choice |
*best_sad = *Data->iMinSAD; |
1825 |
|
|
1826 |
CHECK_MV16_ZERO; |
if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT; |
1827 |
|
else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
1828 |
|
|
1829 |
|
pMB->pmvs[3] = *Data->currentMV; |
1830 |
|
|
1831 |
// left neighbour, if allowed |
for (k = 0; k < 4; k++) { |
1832 |
if (x != 0) |
pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x; |
1833 |
{ |
pMB->b_mvs[k].x = ( (Data->currentMV->x == 0) |
1834 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
? Data->directmvB[k].x |
1835 |
{ pmv[1].x = EVEN(pmv[1].x); |
:pMB->mvs[k].x - Data->referencemv[k].x); |
1836 |
pmv[1].y = EVEN(pmv[1].y); |
pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y); |
1837 |
} |
pMB->b_mvs[k].y = ((Data->currentMV->y == 0) |
1838 |
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
? Data->directmvB[k].y |
1839 |
|
: pMB->mvs[k].y - Data->referencemv[k].y); |
1840 |
|
if (Data->qpel) { |
1841 |
|
pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2; |
1842 |
|
pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2; |
1843 |
|
pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2; |
1844 |
|
pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2; |
1845 |
} |
} |
1846 |
|
|
1847 |
// top neighbour, if allowed |
if (b_mb->mode != MODE_INTER4V) { |
1848 |
if (y != 0) |
pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; |
1849 |
{ |
pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; |
1850 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0]; |
1851 |
{ pmv[2].x = EVEN(pmv[2].x); |
pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0]; |
1852 |
pmv[2].y = EVEN(pmv[2].y); |
break; |
1853 |
|
} |
1854 |
|
} |
1855 |
|
return skip_sad; |
1856 |
} |
} |
|
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
|
1857 |
|
|
1858 |
// top right neighbour, if allowed |
static void |
1859 |
if ((uint32_t)x != (iWcount-1)) |
SearchInterpolate(const IMAGE * const f_Ref, |
1860 |
{ |
const uint8_t * const f_RefH, |
1861 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
const uint8_t * const f_RefV, |
1862 |
{ pmv[3].x = EVEN(pmv[3].x); |
const uint8_t * const f_RefHV, |
1863 |
pmv[3].y = EVEN(pmv[3].y); |
const IMAGE * const b_Ref, |
1864 |
|
const uint8_t * const b_RefH, |
1865 |
|
const uint8_t * const b_RefV, |
1866 |
|
const uint8_t * const b_RefHV, |
1867 |
|
const IMAGE * const pCur, |
1868 |
|
const int x, const int y, |
1869 |
|
const uint32_t fcode, |
1870 |
|
const uint32_t bcode, |
1871 |
|
const uint32_t MotionFlags, |
1872 |
|
const MBParam * const pParam, |
1873 |
|
const VECTOR * const f_predMV, |
1874 |
|
const VECTOR * const b_predMV, |
1875 |
|
MACROBLOCK * const pMB, |
1876 |
|
int32_t * const best_sad, |
1877 |
|
SearchData * const fData) |
1878 |
|
|
1879 |
|
{ |
1880 |
|
|
1881 |
|
int iDirection, i, j; |
1882 |
|
SearchData bData; |
1883 |
|
|
1884 |
|
fData->qpel_precision = 0; |
1885 |
|
memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data |
1886 |
|
*fData->iMinSAD = 4096*256; |
1887 |
|
bData.currentMV++; bData.currentQMV++; |
1888 |
|
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
1889 |
|
|
1890 |
|
i = (x + y * fData->iEdgedWidth) * 16; |
1891 |
|
|
1892 |
|
bData.b_RefP[0] = fData->RefP[0] = f_Ref->y + i; |
1893 |
|
bData.b_RefP[2] = fData->RefP[2] = f_RefH + i; |
1894 |
|
bData.b_RefP[1] = fData->RefP[1] = f_RefV + i; |
1895 |
|
bData.b_RefP[3] = fData->RefP[3] = f_RefHV + i; |
1896 |
|
bData.RefP[0] = fData->b_RefP[0] = b_Ref->y + i; |
1897 |
|
bData.RefP[2] = fData->b_RefP[2] = b_RefH + i; |
1898 |
|
bData.RefP[1] = fData->b_RefP[1] = b_RefV + i; |
1899 |
|
bData.RefP[3] = fData->b_RefP[3] = b_RefHV + i; |
1900 |
|
bData.b_RefP[4] = fData->RefP[4] = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; |
1901 |
|
bData.b_RefP[5] = fData->RefP[5] = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; |
1902 |
|
bData.RefP[4] = fData->b_RefP[4] = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; |
1903 |
|
bData.RefP[5] = fData->b_RefP[5] = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; |
1904 |
|
|
1905 |
|
bData.bpredMV = fData->predMV = *f_predMV; |
1906 |
|
fData->bpredMV = bData.predMV = *b_predMV; |
1907 |
|
fData->currentMV[0] = fData->currentMV[2]; |
1908 |
|
|
1909 |
|
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode - fData->qpel, 0, 0); |
1910 |
|
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode - fData->qpel, 0, 0); |
1911 |
|
|
1912 |
|
if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; |
1913 |
|
if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx; |
1914 |
|
if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy; |
1915 |
|
if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy; |
1916 |
|
|
1917 |
|
if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx; |
1918 |
|
if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx; |
1919 |
|
if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy; |
1920 |
|
if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy; |
1921 |
|
|
1922 |
|
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
1923 |
|
|
1924 |
|
//diamond |
1925 |
|
do { |
1926 |
|
iDirection = 255; |
1927 |
|
// forward MV moves |
1928 |
|
i = fData->currentMV[0].x; j = fData->currentMV[0].y; |
1929 |
|
|
1930 |
|
CheckCandidateInt(i + 1, j, 0, &iDirection, fData); |
1931 |
|
CheckCandidateInt(i, j + 1, 0, &iDirection, fData); |
1932 |
|
CheckCandidateInt(i - 1, j, 0, &iDirection, fData); |
1933 |
|
CheckCandidateInt(i, j - 1, 0, &iDirection, fData); |
1934 |
|
|
1935 |
|
// backward MV moves |
1936 |
|
i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
1937 |
|
fData->currentMV[2] = fData->currentMV[0]; |
1938 |
|
CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); |
1939 |
|
CheckCandidateInt(i, j + 1, 0, &iDirection, &bData); |
1940 |
|
CheckCandidateInt(i - 1, j, 0, &iDirection, &bData); |
1941 |
|
CheckCandidateInt(i, j - 1, 0, &iDirection, &bData); |
1942 |
|
|
1943 |
|
} while (!(iDirection)); |
1944 |
|
|
1945 |
|
//qpel refinement |
1946 |
|
if (fData->qpel) { |
1947 |
|
if (*fData->iMinSAD > *best_sad + 500) return; |
1948 |
|
CheckCandidate = CheckCandidateInt; |
1949 |
|
fData->qpel_precision = bData.qpel_precision = 1; |
1950 |
|
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0); |
1951 |
|
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0); |
1952 |
|
fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x; |
1953 |
|
fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y; |
1954 |
|
fData->currentQMV[1].x = 2 * fData->currentMV[1].x; |
1955 |
|
fData->currentQMV[1].y = 2 * fData->currentMV[1].y; |
1956 |
|
SubpelRefine(fData); |
1957 |
|
if (*fData->iMinSAD > *best_sad + 300) return; |
1958 |
|
fData->currentQMV[2] = fData->currentQMV[0]; |
1959 |
|
SubpelRefine(&bData); |
1960 |
|
} |
1961 |
|
|
1962 |
|
*fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode. |
1963 |
|
|
1964 |
|
if (*fData->iMinSAD < *best_sad) { |
1965 |
|
*best_sad = *fData->iMinSAD; |
1966 |
|
pMB->mvs[0] = fData->currentMV[0]; |
1967 |
|
pMB->b_mvs[0] = fData->currentMV[1]; |
1968 |
|
pMB->mode = MODE_INTERPOLATE; |
1969 |
|
if (fData->qpel) { |
1970 |
|
pMB->qmvs[0] = fData->currentQMV[0]; |
1971 |
|
pMB->b_qmvs[0] = fData->currentQMV[1]; |
1972 |
|
pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x; |
1973 |
|
pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y; |
1974 |
|
pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x; |
1975 |
|
pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y; |
1976 |
|
} else { |
1977 |
|
pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; |
1978 |
|
pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; |
1979 |
|
pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; |
1980 |
|
pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; |
1981 |
} |
} |
|
CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); |
|
1982 |
} |
} |
1983 |
} |
} |
1984 |
|
|
1985 |
/* Terminate if MinSAD <= T_2 |
void |
1986 |
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
MotionEstimationBVOP(MBParam * const pParam, |
1987 |
*/ |
FRAMEINFO * const frame, |
1988 |
|
const int32_t time_bp, |
1989 |
if ( (iMinSAD <= thresh2) |
const int32_t time_pp, |
1990 |
|| ( MVequal(*currMV,pMB->mvs[0]) && ((uint32_t)iMinSAD <= prevMB->sad16) ) ) |
// forward (past) reference |
1991 |
|
const MACROBLOCK * const f_mbs, |
1992 |
|
const IMAGE * const f_ref, |
1993 |
|
const IMAGE * const f_refH, |
1994 |
|
const IMAGE * const f_refV, |
1995 |
|
const IMAGE * const f_refHV, |
1996 |
|
// backward (future) reference |
1997 |
|
const FRAMEINFO * const b_reference, |
1998 |
|
const IMAGE * const b_ref, |
1999 |
|
const IMAGE * const b_refH, |
2000 |
|
const IMAGE * const b_refV, |
2001 |
|
const IMAGE * const b_refHV) |
2002 |
{ |
{ |
2003 |
if (MotionFlags & PMV_QUICKSTOP16) |
uint32_t i, j; |
2004 |
goto EPZS16_Terminate_without_Refine; |
int32_t best_sad; |
2005 |
if (MotionFlags & PMV_EARLYSTOP16) |
uint32_t skip_sad; |
2006 |
goto EPZS16_Terminate_with_Refine; |
int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; |
2007 |
} |
const MACROBLOCK * const b_mbs = b_reference->mbs; |
2008 |
|
|
2009 |
/***** predictor SET C: acceleration MV (new!), neighbours in prev. frame(new!) ****/ |
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
2010 |
|
|
2011 |
backupMV = pMB->mvs[0]; // last MV |
const int32_t TRB = time_pp - time_bp; |
2012 |
backupMV.x += (pMB->mvs[0].x - oldMB->mvs[0].x ); // acceleration X |
const int32_t TRD = time_pp; |
2013 |
backupMV.y += (pMB->mvs[0].y - oldMB->mvs[0].y ); // acceleration Y |
|
2014 |
|
// some pre-inintialized data for the rest of the search |
2015 |
|
|
2016 |
|
SearchData Data; |
2017 |
|
int32_t iMinSAD; |
2018 |
|
VECTOR currentMV[3]; |
2019 |
|
VECTOR currentQMV[3]; |
2020 |
|
int32_t temp[8]; |
2021 |
|
memset(&Data, 0, sizeof(SearchData)); |
2022 |
|
Data.iEdgedWidth = pParam->edged_width; |
2023 |
|
Data.currentMV = currentMV; Data.currentQMV = currentQMV; |
2024 |
|
Data.iMinSAD = &iMinSAD; |
2025 |
|
Data.lambda16 = lambda_vec16[frame->quant]; |
2026 |
|
Data.qpel = pParam->m_quarterpel; |
2027 |
|
Data.rounding = 0; |
2028 |
|
Data.chroma = frame->motion_flags & PMV_CHROMA8; |
2029 |
|
Data.temp = temp; |
2030 |
|
|
2031 |
CHECK_MV16_CANDIDATE(backupMV.x,backupMV.y); |
Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose) |
2032 |
|
// note: i==horizontal, j==vertical |
2033 |
|
for (j = 0; j < pParam->mb_height; j++) { |
2034 |
|
|
2035 |
// left neighbour |
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
|
if (x != 0) |
|
|
CHECK_MV16_CANDIDATE((oldMB-1)->mvs[0].x,oldMB->mvs[0].y); |
|
2036 |
|
|
2037 |
// top neighbour |
for (i = 0; i < pParam->mb_width; i++) { |
2038 |
if (y != 0) |
MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; |
2039 |
CHECK_MV16_CANDIDATE((oldMB-iWcount)->mvs[0].x,oldMB->mvs[0].y); |
const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; |
2040 |
|
|
2041 |
|
/* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */ |
2042 |
|
if (b_reference->coding_type != S_VOP) |
2043 |
|
if (b_mb->mode == MODE_NOT_CODED) { |
2044 |
|
pMB->mode = MODE_NOT_CODED; |
2045 |
|
continue; |
2046 |
|
} |
2047 |
|
|
2048 |
// right neighbour, if allowed (this value is not written yet, so take it from pMB->mvs |
Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
2049 |
|
Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8; |
2050 |
|
Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8; |
2051 |
|
pMB->quant = frame->quant; |
2052 |
|
|
2053 |
|
/* direct search comes first, because it (1) checks for SKIP-mode |
2054 |
|
and (2) sets very good predictions for forward and backward search */ |
2055 |
|
skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
2056 |
|
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
2057 |
|
&frame->image, |
2058 |
|
i, j, |
2059 |
|
frame->motion_flags, |
2060 |
|
TRB, TRD, |
2061 |
|
pParam, |
2062 |
|
pMB, b_mb, |
2063 |
|
&best_sad, |
2064 |
|
&Data); |
2065 |
|
|
2066 |
if ((uint32_t)x != iWcount-1) |
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
|
CHECK_MV16_CANDIDATE((pMB+1)->mvs[0].x,oldMB->mvs[0].y); |
|
2067 |
|
|
2068 |
// bottom neighbour, dito |
// forward search |
2069 |
if ((uint32_t)y != iHcount-1) |
SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
2070 |
CHECK_MV16_CANDIDATE((pMB+iWcount)->mvs[0].x,oldMB->mvs[0].y); |
&frame->image, i, j, |
2071 |
|
frame->motion_flags, |
2072 |
|
frame->fcode, pParam, |
2073 |
|
pMB, &f_predMV, &best_sad, |
2074 |
|
MODE_FORWARD, &Data); |
2075 |
|
|
2076 |
/* Terminate if MinSAD <= T_3 (here T_3 = T_2) */ |
// backward search |
2077 |
if (iMinSAD <= thresh2) |
SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y, |
2078 |
{ |
&frame->image, i, j, |
2079 |
if (MotionFlags & PMV_QUICKSTOP16) |
frame->motion_flags, |
2080 |
goto EPZS16_Terminate_without_Refine; |
frame->bcode, pParam, |
2081 |
if (MotionFlags & PMV_EARLYSTOP16) |
pMB, &b_predMV, &best_sad, |
2082 |
goto EPZS16_Terminate_with_Refine; |
MODE_BACKWARD, &Data); |
2083 |
|
|
2084 |
|
// interpolate search comes last, because it uses data from forward and backward as prediction |
2085 |
|
SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
2086 |
|
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
2087 |
|
&frame->image, |
2088 |
|
i, j, |
2089 |
|
frame->fcode, frame->bcode, |
2090 |
|
frame->motion_flags, |
2091 |
|
pParam, |
2092 |
|
&f_predMV, &b_predMV, |
2093 |
|
pMB, &best_sad, |
2094 |
|
&Data); |
2095 |
|
|
2096 |
|
// final skip decision |
2097 |
|
if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2) |
2098 |
|
&& ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) |
2099 |
|
SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data); |
2100 |
|
|
2101 |
|
switch (pMB->mode) { |
2102 |
|
case MODE_FORWARD: |
2103 |
|
f_count++; |
2104 |
|
f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; |
2105 |
|
break; |
2106 |
|
case MODE_BACKWARD: |
2107 |
|
b_count++; |
2108 |
|
b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; |
2109 |
|
break; |
2110 |
|
case MODE_INTERPOLATE: |
2111 |
|
i_count++; |
2112 |
|
f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; |
2113 |
|
b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; |
2114 |
|
break; |
2115 |
|
case MODE_DIRECT: |
2116 |
|
case MODE_DIRECT_NO4V: |
2117 |
|
d_count++; |
2118 |
|
default: |
2119 |
|
break; |
2120 |
|
} |
2121 |
|
} |
2122 |
|
} |
2123 |
} |
} |
2124 |
|
|
2125 |
/************ (if Diamond Search) **************/ |
static __inline void |
2126 |
|
MEanalyzeMB ( const uint8_t * const pRef, |
2127 |
|
const uint8_t * const pCur, |
2128 |
|
const int x, |
2129 |
|
const int y, |
2130 |
|
const MBParam * const pParam, |
2131 |
|
MACROBLOCK * const pMBs, |
2132 |
|
SearchData * const Data) |
2133 |
|
{ |
2134 |
|
|
2135 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
int i, mask; |
2136 |
|
VECTOR pmv[3]; |
2137 |
|
MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width]; |
2138 |
|
|
2139 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
2140 |
|
|
2141 |
if (MotionFlags & PMV_USESQUARES16) |
//median is only used as prediction. it doesn't have to be real |
2142 |
EPZSMainSearchPtr = Square16_MainSearch; |
if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; |
2143 |
else |
else |
2144 |
EPZSMainSearchPtr = Diamond16_MainSearch; |
if (x == 1) //left macroblock does not have any vector now |
2145 |
|
Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median |
2146 |
|
else if (y == 1) // top macroblock doesn't have it's vector |
2147 |
|
Data->predMV = (pMB - 1)->mvs[0]; // left instead of median |
2148 |
|
else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median |
2149 |
|
|
2150 |
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2151 |
x, y, |
pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, 0); |
|
currMV->x, currMV->y, iMinSAD, &newMV, pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
|
|
2, iFcode, iQuant, 0); |
|
2152 |
|
|
2153 |
if (iSAD < iMinSAD) |
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
2154 |
{ |
Data->RefP[0] = pRef + (x + y * pParam->edged_width) * 16; |
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
2155 |
|
|
2156 |
|
pmv[1].x = EVEN(pMB->mvs[0].x); |
2157 |
|
pmv[1].y = EVEN(pMB->mvs[0].y); |
2158 |
|
pmv[2].x = EVEN(Data->predMV.x); |
2159 |
|
pmv[2].y = EVEN(Data->predMV.y); |
2160 |
|
pmv[0].x = pmv[0].y = 0; |
2161 |
|
|
2162 |
if (MotionFlags & PMV_EXTSEARCH16) |
CheckCandidate32I(0, 0, 255, &i, Data); |
|
{ |
|
|
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
|
2163 |
|
|
2164 |
if (!(MVequal(pmv[0],backupMV)) ) |
if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) { |
2165 |
{ |
|
2166 |
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
if (!(mask = make_mask(pmv, 1))) |
2167 |
x, y, |
CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data); |
2168 |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
if (!(mask = make_mask(pmv, 2))) |
2169 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); |
CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data); |
2170 |
|
|
2171 |
|
if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) // diamond only if needed |
2172 |
|
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
2173 |
} |
} |
2174 |
|
|
2175 |
if (iSAD < iMinSAD) |
for (i = 0; i < 4; i++) { |
2176 |
{ |
MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width]; |
2177 |
*currMV = newMV; |
MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i]; |
2178 |
iMinSAD = iSAD; |
MB->mode = MODE_INTER; |
2179 |
|
MB->sad16 = Data->iMinSAD[i+1]; |
2180 |
|
} |
2181 |
} |
} |
2182 |
|
|
2183 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
#define INTRA_THRESH 2050 |
2184 |
{ |
#define INTER_THRESH 1200 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
0, 0, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, /*iDiamondSize*/ 2, iFcode, iQuant, 0); |
|
2185 |
|
|
2186 |
if (iSAD < iMinSAD) |
int |
2187 |
{ |
MEanalysis( const IMAGE * const pRef, |
2188 |
*currMV = newMV; |
const FRAMEINFO * const Current, |
2189 |
iMinSAD = iSAD; |
const MBParam * const pParam, |
2190 |
} |
const int maxIntra, //maximum number if non-I frames |
2191 |
|
const int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame |
2192 |
|
const int bCount, // number of B frames in a row |
2193 |
|
const int b_thresh) |
2194 |
|
{ |
2195 |
|
uint32_t x, y, intra = 0; |
2196 |
|
int sSAD = 0; |
2197 |
|
MACROBLOCK * const pMBs = Current->mbs; |
2198 |
|
const IMAGE * const pCurrent = &Current->image; |
2199 |
|
int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + 10*b_thresh; |
2200 |
|
int s = 0, blocks = 0; |
2201 |
|
|
2202 |
|
int32_t iMinSAD[5], temp[5]; |
2203 |
|
VECTOR currentMV[5]; |
2204 |
|
SearchData Data; |
2205 |
|
Data.iEdgedWidth = pParam->edged_width; |
2206 |
|
Data.currentMV = currentMV; |
2207 |
|
Data.iMinSAD = iMinSAD; |
2208 |
|
Data.iFcode = Current->fcode; |
2209 |
|
Data.temp = temp; |
2210 |
|
CheckCandidate = CheckCandidate32I; |
2211 |
|
|
2212 |
|
if (intraCount != 0 && intraCount < 10) // we're right after an I frame |
2213 |
|
IntraThresh += 8 * (intraCount - 10) * (intraCount - 10); |
2214 |
|
else |
2215 |
|
if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec |
2216 |
|
IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; |
2217 |
|
|
2218 |
|
InterThresh -= (350 - 8*b_thresh) * bCount; |
2219 |
|
if (InterThresh < 300 + 5*b_thresh) InterThresh = 300 + 5*b_thresh; |
2220 |
|
|
2221 |
|
if (sadInit) (*sadInit) (); |
2222 |
|
|
2223 |
|
for (y = 1; y < pParam->mb_height-1; y += 2) { |
2224 |
|
for (x = 1; x < pParam->mb_width-1; x += 2) { |
2225 |
|
int i; |
2226 |
|
blocks += 4; |
2227 |
|
|
2228 |
|
if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; |
2229 |
|
else { //extrapolation of the vector found for last frame |
2230 |
|
pMBs[x + y * pParam->mb_width].mvs[0].x = |
2231 |
|
(pMBs[x + y * pParam->mb_width].mvs[0].x * (bCount+1) ) / bCount; |
2232 |
|
pMBs[x + y * pParam->mb_width].mvs[0].y = |
2233 |
|
(pMBs[x + y * pParam->mb_width].mvs[0].y * (bCount+1) ) / bCount; |
2234 |
|
} |
2235 |
|
|
2236 |
|
MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data); |
2237 |
|
|
2238 |
|
for (i = 0; i < 4; i++) { |
2239 |
|
int dev; |
2240 |
|
MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width]; |
2241 |
|
if (pMB->sad16 > IntraThresh) { |
2242 |
|
dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16, |
2243 |
|
pParam->edged_width); |
2244 |
|
if (dev + IntraThresh < pMB->sad16) { |
2245 |
|
pMB->mode = MODE_INTRA; |
2246 |
|
if (++intra > ((pParam->mb_height-2)*(pParam->mb_width-2))/2) return I_VOP; |
2247 |
} |
} |
2248 |
} |
} |
2249 |
|
if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0) s++; |
2250 |
|
|
2251 |
/*************** Choose best MV found **************/ |
sSAD += pMB->sad16; |
2252 |
|
} |
2253 |
EPZS16_Terminate_with_Refine: |
} |
2254 |
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
} |
|
iMinSAD = Halfpel16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
currMV, iMinSAD, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
|
2255 |
|
|
2256 |
EPZS16_Terminate_without_Refine: |
sSAD /= blocks; |
2257 |
|
s = (10*s) / blocks; |
2258 |
|
|
2259 |
*oldMB = *pMB; |
if (s > 4) sSAD += (s - 2) * (160 - 2*b_thresh); //static block - looks bad when in bframe... |
2260 |
|
|
2261 |
currPMV->x = currMV->x - pmv[0].x; |
if (sSAD > InterThresh ) return P_VOP; |
2262 |
currPMV->y = currMV->y - pmv[0].y; |
emms(); |
2263 |
return iMinSAD; |
return B_VOP; |
2264 |
} |
} |
2265 |
|
|
2266 |
|
|
2267 |
int32_t EPZSSearch8( |
static WARPPOINTS |
2268 |
const uint8_t * const pRef, |
GlobalMotionEst(const MACROBLOCK * const pMBs, |
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const IMAGE * const pCur, |
|
|
const int x, const int y, |
|
|
const int start_x, const int start_y, |
|
|
const uint32_t MotionFlags, |
|
|
const uint32_t iQuant, |
|
|
const uint32_t iFcode, |
|
2269 |
const MBParam * const pParam, |
const MBParam * const pParam, |
2270 |
const MACROBLOCK * const pMBs, |
const FRAMEINFO * const current, |
2271 |
const MACROBLOCK * const prevMBs, |
const FRAMEINFO * const reference, |
2272 |
VECTOR * const currMV, |
const IMAGE * const pRefH, |
2273 |
VECTOR * const currPMV) |
const IMAGE * const pRefV, |
2274 |
|
const IMAGE * const pRefHV ) |
2275 |
{ |
{ |
|
const uint32_t iWcount = pParam->mb_width; |
|
|
const int32_t iWidth = pParam->width; |
|
|
const int32_t iHeight = pParam->height; |
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
|
|
|
|
const uint8_t * cur = pCur->y + x*8 + y*8*iEdgedWidth; |
|
2276 |
|
|
2277 |
int32_t iDiamondSize=1; |
const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs |
2278 |
|
const int deltay=8; |
2279 |
|
const int grad=512; // lower bound for deviation in MB |
2280 |
|
|
2281 |
int32_t min_dx; |
WARPPOINTS gmc; |
|
int32_t max_dx; |
|
|
int32_t min_dy; |
|
|
int32_t max_dy; |
|
2282 |
|
|
2283 |
VECTOR newMV; |
uint32_t mx, my; |
|
VECTOR backupMV; |
|
2284 |
|
|
2285 |
VECTOR pmv[4]; |
int MBh = pParam->mb_height; |
2286 |
int32_t psad[8]; |
int MBw = pParam->mb_width; |
2287 |
|
|
2288 |
const int32_t iSubBlock = ((y&1)<<1) + (x&1); |
int *MBmask= calloc(MBh*MBw,sizeof(int)); |
2289 |
|
double DtimesF[4] = { 0.,0., 0., 0. }; |
2290 |
|
double sol[4] = { 0., 0., 0., 0. }; |
2291 |
|
double a,b,c,n,denom; |
2292 |
|
double meanx,meany; |
2293 |
|
int num,oldnum; |
2294 |
|
|
2295 |
const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
if (!MBmask) { fprintf(stderr,"Mem error\n"); |
2296 |
const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; |
gmc.duv[0].x= gmc.duv[0].y = |
2297 |
|
gmc.duv[1].x= gmc.duv[1].y = |
2298 |
|
gmc.duv[2].x= gmc.duv[2].y = 0; |
2299 |
|
return gmc; } |
2300 |
|
|
2301 |
int32_t bPredEq; |
// filter mask of all blocks |
|
int32_t iMinSAD,iSAD=9999; |
|
2302 |
|
|
2303 |
MainSearch8FuncPtr EPZSMainSearchPtr; |
for (my = 1; my < (uint32_t)MBh-1; my++) |
2304 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
2305 |
|
{ |
2306 |
|
const int mbnum = mx + my * MBw; |
2307 |
|
const MACROBLOCK *pMB = &pMBs[mbnum]; |
2308 |
|
const VECTOR mv = pMB->mvs[0]; |
2309 |
|
|
2310 |
/* Get maximum range */ |
if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) |
2311 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
continue; |
|
x, y, 8, iWidth, iHeight, iFcode); |
|
2312 |
|
|
2313 |
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
if ( ( (abs(mv.x - (pMB-1)->mvs[0].x) < deltax) && (abs(mv.y - (pMB-1)->mvs[0].y) < deltay) ) |
2314 |
|
&& ( (abs(mv.x - (pMB+1)->mvs[0].x) < deltax) && (abs(mv.y - (pMB+1)->mvs[0].y) < deltay) ) |
2315 |
|
&& ( (abs(mv.x - (pMB-MBw)->mvs[0].x) < deltax) && (abs(mv.y - (pMB-MBw)->mvs[0].y) < deltay) ) |
2316 |
|
&& ( (abs(mv.x - (pMB+MBw)->mvs[0].x) < deltax) && (abs(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) ) |
2317 |
|
MBmask[mbnum]=1; |
2318 |
|
} |
2319 |
|
|
2320 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
for (my = 1; my < (uint32_t)MBh-1; my++) |
2321 |
{ min_dx = EVEN(min_dx); |
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
2322 |
max_dx = EVEN(max_dx); |
{ |
2323 |
min_dy = EVEN(min_dy); |
const uint8_t *const pCur = current->image.y + 16*my*pParam->edged_width + 16*mx; |
|
max_dy = EVEN(max_dy); |
|
|
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
|
2324 |
|
|
2325 |
bPredEq = get_pmvdata(pMBs, x>>1, y>>1, iWcount, iSubBlock, pmv, psad); |
const int mbnum = mx + my * MBw; |
2326 |
|
if (!MBmask[mbnum]) |
2327 |
|
continue; |
2328 |
|
|
2329 |
|
if (sad16 ( pCur, pCur+1 , pParam->edged_width, 65536) <= (uint32_t)grad ) |
2330 |
|
MBmask[mbnum] = 0; |
2331 |
|
if (sad16 ( pCur, pCur+pParam->edged_width, pParam->edged_width, 65536) <= (uint32_t)grad ) |
2332 |
|
MBmask[mbnum] = 0; |
2333 |
|
|
2334 |
/* Step 4: Calculate SAD around the Median prediction. |
} |
|
MinSAD=SAD |
|
|
If Motion Vector equal to Previous frame motion vector |
|
|
and MinSAD<PrevFrmSAD goto Step 10. |
|
|
If SAD<=256 goto Step 10. |
|
|
*/ |
|
2335 |
|
|
2336 |
// Prepare for main loop |
emms(); |
2337 |
|
|
2338 |
|
do { /* until convergence */ |
2339 |
|
|
2340 |
if (!(MotionFlags & PMV_HALFPEL8)) |
a = b = c = n = 0; |
2341 |
|
DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.; |
2342 |
|
for (my = 0; my < (uint32_t)MBh; my++) |
2343 |
|
for (mx = 0; mx < (uint32_t)MBw; mx++) |
2344 |
{ |
{ |
2345 |
currMV->x = EVEN(currMV->x); |
const int mbnum = mx + my * MBw; |
2346 |
currMV->y = EVEN(currMV->y); |
const MACROBLOCK *pMB = &pMBs[mbnum]; |
2347 |
} |
const VECTOR mv = pMB->mvs[0]; |
2348 |
|
|
2349 |
if (currMV->x > max_dx) |
if (!MBmask[mbnum]) |
2350 |
currMV->x=max_dx; |
continue; |
2351 |
if (currMV->x < min_dx) |
|
2352 |
currMV->x=min_dx; |
n++; |
2353 |
if (currMV->y > max_dy) |
a += 16*mx+8; |
2354 |
currMV->y=max_dy; |
b += 16*my+8; |
2355 |
if (currMV->y < min_dy) |
c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8); |
2356 |
currMV->y=min_dy; |
|
2357 |
|
DtimesF[0] += (double)mv.x; |
2358 |
|
DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8); |
2359 |
|
DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8); |
2360 |
|
DtimesF[3] += (double)mv.y; |
2361 |
|
} |
2362 |
|
|
2363 |
|
denom = a*a+b*b-c*n; |
2364 |
|
|
2365 |
|
/* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */ |
2366 |
|
/* D'*E*F has been calculated in the same loop as matrix */ |
2367 |
|
|
2368 |
|
sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2]; |
2369 |
|
sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3]; |
2370 |
|
sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3]; |
2371 |
|
sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3]; |
2372 |
|
|
2373 |
|
sol[0] /= denom; |
2374 |
|
sol[1] /= denom; |
2375 |
|
sol[2] /= denom; |
2376 |
|
sol[3] /= denom; |
2377 |
|
|
2378 |
|
meanx = meany = 0.; |
2379 |
|
oldnum = 0; |
2380 |
|
for (my = 0; my < (uint32_t)MBh; my++) |
2381 |
|
for (mx = 0; mx < (uint32_t)MBw; mx++) |
2382 |
|
{ |
2383 |
|
const int mbnum = mx + my * MBw; |
2384 |
|
const MACROBLOCK *pMB = &pMBs[mbnum]; |
2385 |
|
const VECTOR mv = pMB->mvs[0]; |
2386 |
|
|
2387 |
/***************** This is predictor SET A: only median prediction ******************/ |
if (!MBmask[mbnum]) |
2388 |
|
continue; |
2389 |
|
|
2390 |
|
oldnum++; |
2391 |
|
meanx += fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ); |
2392 |
|
meany += fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ); |
2393 |
|
} |
2394 |
|
|
2395 |
iMinSAD = sad8( cur, |
if (4*meanx > oldnum) /* better fit than 0.25 is useless */ |
2396 |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
meanx /= oldnum; |
2397 |
iEdgedWidth); |
else |
2398 |
iMinSAD += calc_delta_8(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; |
meanx = 0.25; |
2399 |
|
|
2400 |
|
if (4*meany > oldnum) |
2401 |
|
meany /= oldnum; |
2402 |
|
else |
2403 |
|
meany = 0.25; |
2404 |
|
|
2405 |
// thresh1 is fixed to 256 |
/* fprintf(stderr,"sol = (%8.5f, %8.5f, %8.5f, %8.5f)\n",sol[0],sol[1],sol[2],sol[3]); |
2406 |
if (iMinSAD < 256/4 ) |
fprintf(stderr,"meanx = %8.5f meany = %8.5f %d\n",meanx,meany, oldnum); |
2407 |
|
*/ |
2408 |
|
num = 0; |
2409 |
|
for (my = 0; my < (uint32_t)MBh; my++) |
2410 |
|
for (mx = 0; mx < (uint32_t)MBw; mx++) |
2411 |
{ |
{ |
2412 |
if (MotionFlags & PMV_QUICKSTOP8) |
const int mbnum = mx + my * MBw; |
2413 |
goto EPZS8_Terminate_without_Refine; |
const MACROBLOCK *pMB = &pMBs[mbnum]; |
2414 |
if (MotionFlags & PMV_EARLYSTOP8) |
const VECTOR mv = pMB->mvs[0]; |
|
goto EPZS8_Terminate_with_Refine; |
|
|
} |
|
2415 |
|
|
2416 |
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
if (!MBmask[mbnum]) |
2417 |
|
continue; |
2418 |
|
|
2419 |
|
if ( ( fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ) > meanx ) |
2420 |
|
|| ( fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ) > meany ) ) |
2421 |
|
MBmask[mbnum]=0; |
2422 |
|
else |
2423 |
|
num++; |
2424 |
|
} |
2425 |
|
|
2426 |
// previous frame MV |
} while ( (oldnum != num) && (num>=4) ); |
|
CHECK_MV8_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); |
|
2427 |
|
|
2428 |
// MV=(0,0) is often a good choice |
if (num < 4) |
2429 |
|
{ |
2430 |
|
gmc.duv[0].x= gmc.duv[0].y= gmc.duv[1].x= gmc.duv[1].y= gmc.duv[2].x= gmc.duv[2].y=0; |
2431 |
|
} else { |
2432 |
|
|
2433 |
CHECK_MV8_ZERO; |
gmc.duv[0].x=(int)(sol[0]+0.5); |
2434 |
|
gmc.duv[0].y=(int)(sol[3]+0.5); |
2435 |
|
|
2436 |
/* Terminate if MinSAD <= T_2 |
gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5); |
2437 |
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5); |
|
*/ |
|
2438 |
|
|
2439 |
if (iMinSAD < 512/4) /* T_2 == 512/4 hardcoded */ |
gmc.duv[2].x=0; |
2440 |
{ |
gmc.duv[2].y=0; |
|
if (MotionFlags & PMV_QUICKSTOP8) |
|
|
goto EPZS8_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP8) |
|
|
goto EPZS8_Terminate_with_Refine; |
|
2441 |
} |
} |
2442 |
|
// fprintf(stderr,"wp1 = ( %4d, %4d) wp2 = ( %4d, %4d) \n", gmc.duv[0].x, gmc.duv[0].y, gmc.duv[1].x, gmc.duv[1].y); |
2443 |
|
|
2444 |
/************ (if Diamond Search) **************/ |
free(MBmask); |
2445 |
|
|
2446 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
return gmc; |
2447 |
|
} |
2448 |
|
|
2449 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8)) |
// functions which perform BITS-based search/bitcount |
|
iDiamondSize *= 2; |
|
2450 |
|
|
2451 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
static int |
2452 |
|
CountMBBitsInter(SearchData * const Data, |
2453 |
|
const MACROBLOCK * const pMBs, const int x, const int y, |
2454 |
|
const MBParam * const pParam, |
2455 |
|
const uint32_t MotionFlags) |
2456 |
|
{ |
2457 |
|
int i, iDirection; |
2458 |
|
int32_t bsad[5]; |
2459 |
|
|
2460 |
// if (MotionFlags & PMV_USESQUARES8) |
CheckCandidate = CheckCandidateBits16; |
|
// EPZSMainSearchPtr = Square8_MainSearch; |
|
|
// else |
|
|
EPZSMainSearchPtr = Diamond8_MainSearch; |
|
2461 |
|
|
2462 |
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
if (Data->qpel) { |
2463 |
x, y, |
for(i = 0; i < 5; i++) { |
2464 |
currMV->x, currMV->y, iMinSAD, &newMV, |
Data->currentMV[i].x = Data->currentQMV[i].x/2; |
2465 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
Data->currentMV[i].y = Data->currentQMV[i].y/2; |
2466 |
iDiamondSize, iFcode, iQuant, 00); |
} |
2467 |
|
Data->qpel_precision = 1; |
2468 |
|
CheckCandidateBits16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data); |
2469 |
|
|
2470 |
|
//checking if this vector is perfect. if it is, we stop. |
2471 |
|
if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0) |
2472 |
|
return 0; //quick stop |
2473 |
|
|
2474 |
if (iSAD < iMinSAD) |
if (MotionFlags & (HALFPELREFINE16_BITS | EXTSEARCH_BITS)) { //we have to prepare for halfpixel-precision search |
2475 |
{ |
for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i]; |
2476 |
*currMV = newMV; |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2477 |
iMinSAD = iSAD; |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
2478 |
|
Data->qpel_precision = 0; |
2479 |
|
if (Data->currentQMV->x & 1 || Data->currentQMV->y & 1) |
2480 |
|
CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); |
2481 |
} |
} |
2482 |
|
|
2483 |
if (MotionFlags & PMV_EXTSEARCH8) |
} else { // not qpel |
|
{ |
|
|
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
|
|
|
|
|
if (!(MVequal(pmv[0],backupMV)) ) |
|
|
{ |
|
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); |
|
2484 |
|
|
2485 |
if (iSAD < iMinSAD) |
CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); |
2486 |
{ |
//checking if this vector is perfect. if it is, we stop. |
2487 |
*currMV = newMV; |
if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0) { |
2488 |
iMinSAD = iSAD; |
return 0; //inter |
2489 |
} |
} |
2490 |
} |
} |
2491 |
|
|
2492 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
if (MotionFlags&EXTSEARCH_BITS) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
|
{ |
|
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
0, 0, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); |
|
2493 |
|
|
2494 |
if (iSAD < iMinSAD) |
if (MotionFlags&HALFPELREFINE16_BITS) SubpelRefine(Data); |
2495 |
{ |
|
2496 |
*currMV = newMV; |
if (Data->qpel) { |
2497 |
iMinSAD = iSAD; |
if (MotionFlags&(EXTSEARCH_BITS | HALFPELREFINE16_BITS)) { // there was halfpel-precision search |
2498 |
|
for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) { |
2499 |
|
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // we have found a better match |
2500 |
|
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
2501 |
|
} |
2502 |
|
|
2503 |
|
// preparing for qpel-precision search |
2504 |
|
Data->qpel_precision = 1; |
2505 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2506 |
|
pParam->width, pParam->height, Data->iFcode, 1, 0); |
2507 |
|
} |
2508 |
|
if (MotionFlags&QUARTERPELREFINE16_BITS) SubpelRefine(Data); |
2509 |
} |
} |
2510 |
|
|
2511 |
|
if (MotionFlags&CHECKPREDICTION_BITS) { //let's check vector equal to prediction |
2512 |
|
VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV; |
2513 |
|
if (!(Data->predMV.x == v->x && Data->predMV.y == v->y)) |
2514 |
|
CheckCandidateBits16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data); |
2515 |
} |
} |
2516 |
|
return Data->iMinSAD[0]; |
2517 |
} |
} |
2518 |
|
|
|
/*************** Choose best MV found **************/ |
|
2519 |
|
|
2520 |
EPZS8_Terminate_with_Refine: |
static int |
2521 |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
CountMBBitsInter4v(const SearchData * const Data, |
2522 |
iMinSAD = Halfpel8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
MACROBLOCK * const pMB, const MACROBLOCK * const pMBs, |
2523 |
x, y, |
const int x, const int y, |
2524 |
currMV, iMinSAD, |
const MBParam * const pParam, const uint32_t MotionFlags, |
2525 |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
const VECTOR * const backup) |
2526 |
|
{ |
2527 |
|
|
2528 |
|
int cbp = 0, bits = 0, t = 0, i, iDirection; |
2529 |
|
SearchData Data2, *Data8 = &Data2; |
2530 |
|
int sumx = 0, sumy = 0; |
2531 |
|
int16_t *in = Data->dctSpace, *coeff = Data->dctSpace + 64; |
2532 |
|
|
2533 |
EPZS8_Terminate_without_Refine: |
memcpy(Data8, Data, sizeof(SearchData)); |
2534 |
|
CheckCandidate = CheckCandidateBits8; |
2535 |
|
|
2536 |
currPMV->x = currMV->x - pmv[0].x; |
for (i = 0; i < 4; i++) { |
2537 |
currPMV->y = currMV->y - pmv[0].y; |
Data8->iMinSAD = Data->iMinSAD + i + 1; |
2538 |
return iMinSAD; |
Data8->currentMV = Data->currentMV + i + 1; |
2539 |
|
Data8->currentQMV = Data->currentQMV + i + 1; |
2540 |
|
Data8->Cur = Data->Cur + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2541 |
|
Data8->RefP[0] = Data->RefP[0] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2542 |
|
Data8->RefP[2] = Data->RefP[2] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2543 |
|
Data8->RefP[1] = Data->RefP[1] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2544 |
|
Data8->RefP[3] = Data->RefP[3] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2545 |
|
|
2546 |
|
if(Data->qpel) { |
2547 |
|
Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i); |
2548 |
|
if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y, |
2549 |
|
Data8->predMV, Data8->iFcode, 0, 0); |
2550 |
|
} else { |
2551 |
|
Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i); |
2552 |
|
if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y, |
2553 |
|
Data8->predMV, Data8->iFcode, 0, 0); |
2554 |
} |
} |
2555 |
|
|
2556 |
|
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, |
2557 |
|
pParam->width, pParam->height, Data8->iFcode, Data8->qpel, 0); |
2558 |
|
|
2559 |
|
*Data8->iMinSAD += t; |
2560 |
|
|
2561 |
|
Data8->qpel_precision = Data8->qpel; |
2562 |
|
// checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far) |
2563 |
|
{ |
2564 |
|
VECTOR *v = Data8->qpel ? Data8->currentQMV : Data8->currentMV; |
2565 |
|
if (!( (v->x == backup[i+1].x) && (v->y == backup[i+1].y) )) |
2566 |
|
CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8); |
2567 |
|
} |
2568 |
|
|
2569 |
|
if (Data8->qpel) { |
2570 |
|
if (MotionFlags&HALFPELREFINE8_BITS || (MotionFlags&PMV_EXTSEARCH8 && MotionFlags&EXTSEARCH_BITS)) { // halfpixel motion search follows |
2571 |
|
int32_t s = *Data8->iMinSAD; |
2572 |
|
Data8->currentMV->x = Data8->currentQMV->x/2; |
2573 |
|
Data8->currentMV->y = Data8->currentQMV->y/2; |
2574 |
|
Data8->qpel_precision = 0; |
2575 |
|
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, |
2576 |
|
pParam->width, pParam->height, Data8->iFcode - 1, 0, 0); |
2577 |
|
|
2578 |
|
if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1) |
2579 |
|
CheckCandidateBits8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8); |
2580 |
|
|
2581 |
/* *********************************************************** |
if (MotionFlags & PMV_EXTSEARCH8 && MotionFlags & EXTSEARCH_BITS) |
2582 |
bvop motion estimation |
SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255); |
|
// TODO: need to incorporate prediction here (eg. sad += calc_delta_16) |
|
|
***************************************************************/ |
|
2583 |
|
|
2584 |
/* |
if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8); |
|
void MotionEstimationBVOP( |
|
|
MBParam * const pParam, |
|
|
FRAMEINFO * const frame, |
|
2585 |
|
|
2586 |
// forward (past) reference |
if(s > *Data8->iMinSAD) { //we have found a better match |
2587 |
const MACROBLOCK * const f_mbs, |
Data8->currentQMV->x = 2*Data8->currentMV->x; |
2588 |
const IMAGE * const f_ref, |
Data8->currentQMV->y = 2*Data8->currentMV->y; |
2589 |
const IMAGE * const f_refH, |
} |
|
const IMAGE * const f_refV, |
|
|
const IMAGE * const f_refHV, |
|
|
// backward (future) reference |
|
|
const MACROBLOCK * const b_mbs, |
|
|
const IMAGE * const b_ref, |
|
|
const IMAGE * const b_refH, |
|
|
const IMAGE * const b_refV, |
|
|
const IMAGE * const b_refHV) |
|
|
{ |
|
|
const uint32_t mb_width = pParam->mb_width; |
|
|
const uint32_t mb_height = pParam->mb_height; |
|
|
const int32_t edged_width = pParam->edged_width; |
|
|
|
|
|
int32_t i,j; |
|
|
|
|
|
int32_t f_sad16; |
|
|
int32_t b_sad16; |
|
|
int32_t i_sad16; |
|
|
int32_t d_sad16; |
|
|
int32_t best_sad; |
|
2590 |
|
|
2591 |
VECTOR pmv_dontcare; |
Data8->qpel_precision = 1; |
2592 |
|
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, |
2593 |
|
pParam->width, pParam->height, Data8->iFcode, 1, 0); |
2594 |
|
|
|
// note: i==horizontal, j==vertical |
|
|
for (j = 0; j < mb_height; j++) |
|
|
{ |
|
|
for (i = 0; i < mb_width; i++) |
|
|
{ |
|
|
MACROBLOCK *mb = &frame->mbs[i + j*mb_width]; |
|
|
const MACROBLOCK *f_mb = &f_mbs[i + j*mb_width]; |
|
|
const MACROBLOCK *b_mb = &b_mbs[i + j*mb_width]; |
|
|
|
|
|
if (b_mb->mode == MODE_INTER |
|
|
&& b_mb->cbp == 0 |
|
|
&& b_mb->mvs[0].x == 0 |
|
|
&& b_mb->mvs[0].y == 0) |
|
|
{ |
|
|
mb->mode = MB_IGNORE; |
|
|
mb->mvs[0].x = 0; |
|
|
mb->mvs[0].y = 0; |
|
|
mb->b_mvs[0].x = 0; |
|
|
mb->b_mvs[0].y = 0; |
|
|
continue; |
|
2595 |
} |
} |
2596 |
|
if (MotionFlags & QUARTERPELREFINE8_BITS) SubpelRefine(Data8); |
2597 |
|
|
2598 |
|
} else // not qpel |
2599 |
|
if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8); //halfpel mode, halfpel refinement |
2600 |
|
|
2601 |
// forward search |
//checking vector equal to predicion |
2602 |
f_sad16 = SEARCH16(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
if (i != 0 && MotionFlags & CHECKPREDICTION_BITS) { |
2603 |
&frame->image, |
const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV; |
2604 |
i, j, |
if (!(Data8->predMV.x == v->x && Data8->predMV.y == v->y)) |
2605 |
frame->motion_flags, frame->quant, frame->fcode, |
CheckCandidateBits8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8); |
2606 |
pParam, |
} |
|
f_mbs, |
|
|
&mb->mvs[0], &pmv_dontcare); // ignore pmv |
|
2607 |
|
|
2608 |
// backward search |
bits += *Data8->iMinSAD; |
2609 |
b_sad16 = SEARCH16(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
if (bits >= Data->iMinSAD[0]) break; // no chances for INTER4V |
2610 |
&frame->image, |
|
2611 |
i, j, |
// MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else |
2612 |
frame->motion_flags, frame->quant, frame->bcode, |
if(Data->qpel) { |
2613 |
pParam, |
pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x; |
2614 |
b_mbs, |
pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y; |
2615 |
&mb->b_mvs[0], &pmv_dontcare); // ignore pmv |
pMB->qmvs[i] = *Data8->currentQMV; |
2616 |
|
sumx += Data8->currentQMV->x/2; |
2617 |
|
sumy += Data8->currentQMV->y/2; |
2618 |
|
} else { |
2619 |
|
pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x; |
2620 |
|
pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y; |
2621 |
|
sumx += Data8->currentMV->x; |
2622 |
|
sumy += Data8->currentMV->y; |
2623 |
|
} |
2624 |
|
pMB->mvs[i] = *Data8->currentMV; |
2625 |
|
pMB->sad8[i] = 4 * *Data8->iMinSAD; |
2626 |
|
if (Data8->temp[0]) cbp |= 1 << (5 - i); |
2627 |
|
} |
2628 |
|
|
2629 |
// interpolate search (simple, but effective) |
if (bits < *Data->iMinSAD) { // there is still a chance for inter4v mode. let's check chroma |
2630 |
i_sad16 = sad16bi_c( |
const uint8_t * ptr; |
2631 |
frame->image.y + i*16 + j*16*edged_width, |
sumx = (sumx >> 3) + roundtab_76[sumx & 0xf]; |
2632 |
get_ref(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
sumy = (sumy >> 3) + roundtab_76[sumy & 0xf]; |
|
i, j, 16, mb->mvs[0].x, mb->mvs[0].y, edged_width), |
|
|
get_ref(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, 16, mb->b_mvs[0].x, mb->b_mvs[0].x, edged_width), |
|
|
edged_width); |
|
|
|
|
|
// TODO: direct search |
|
|
// predictor + range of [-32,32] |
|
|
d_sad16 = 65535; |
|
2633 |
|
|
2634 |
|
//chroma U |
2635 |
|
ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[4], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
2636 |
|
transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2); |
2637 |
|
fdct(in); |
2638 |
|
if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16); |
2639 |
|
else i = quant4_inter(coeff, in, Data->lambda16); |
2640 |
|
if (i > 0) { |
2641 |
|
bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); |
2642 |
|
cbp |= 1 << (5 - 4); |
2643 |
|
} |
2644 |
|
|
2645 |
if (f_sad16 < b_sad16) |
if (bits < *Data->iMinSAD) { // still possible |
2646 |
{ |
//chroma V |
2647 |
best_sad = f_sad16; |
ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[5], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
2648 |
mb->mode = MB_FORWARD; |
transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2); |
2649 |
|
fdct(in); |
2650 |
|
if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16); |
2651 |
|
else i = quant4_inter(coeff, in, Data->lambda16); |
2652 |
|
if (i > 0) { |
2653 |
|
bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); |
2654 |
|
cbp |= 1 << (5 - 5); |
2655 |
|
} |
2656 |
|
bits += xvid_cbpy_tab[15-(cbp>>2)].len; |
2657 |
|
bits += mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len; |
2658 |
} |
} |
|
else |
|
|
{ |
|
|
best_sad = b_sad16; |
|
|
mb->mode = MB_BACKWARD; |
|
2659 |
} |
} |
2660 |
|
|
2661 |
if (i_sad16 < best_sad) |
return bits; |
|
{ |
|
|
best_sad = i_sad16; |
|
|
mb->mode = MB_INTERPOLATE; |
|
2662 |
} |
} |
2663 |
|
|
2664 |
if (d_sad16 < best_sad) |
|
2665 |
|
static int |
2666 |
|
CountMBBitsIntra(const SearchData * const Data) |
2667 |
{ |
{ |
2668 |
best_sad = d_sad16; |
int bits = 1; //this one is ac/dc prediction flag. always 1. |
2669 |
mb->mode = MB_DIRECT; |
int cbp = 0, i, t, dc = 1024, b_dc; |
2670 |
|
const uint32_t iQuant = Data->lambda16; |
2671 |
|
int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64; |
2672 |
|
uint32_t iDcScaler = get_dc_scaler(iQuant, 1);; |
2673 |
|
|
2674 |
|
for(i = 0; i < 4; i++) { |
2675 |
|
int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2676 |
|
transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth); |
2677 |
|
fdct(in); |
2678 |
|
b_dc = in[0]; |
2679 |
|
in[0] -= dc; |
2680 |
|
dc = b_dc; |
2681 |
|
if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler); |
2682 |
|
else quant4_intra(coeff, in, iQuant, iDcScaler); |
2683 |
|
|
2684 |
|
bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcy_tab[coeff[0] + 255].len;; |
2685 |
|
Data->temp[i] = t; |
2686 |
|
if (t != 0) cbp |= 1 << (5 - i); |
2687 |
|
if (bits >= Data->iMinSAD[0]) break; |
2688 |
} |
} |
2689 |
|
|
2690 |
|
if (bits < Data->iMinSAD[0]) { // INTRA still looks good, let's add chroma |
2691 |
|
iDcScaler = get_dc_scaler(iQuant, 0); |
2692 |
|
//chroma U |
2693 |
|
transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2); |
2694 |
|
fdct(in); |
2695 |
|
in[0] -= 1024; |
2696 |
|
if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler); |
2697 |
|
else quant4_intra(coeff, in, iQuant, iDcScaler); |
2698 |
|
|
2699 |
|
bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len; |
2700 |
|
if (t != 0) cbp |= 1 << (5 - 4); |
2701 |
|
|
2702 |
|
if (bits < Data->iMinSAD[0]) { |
2703 |
|
//chroma V |
2704 |
|
transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2); |
2705 |
|
fdct(in); |
2706 |
|
in[0] -= 1024; |
2707 |
|
if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler); |
2708 |
|
else quant4_intra(coeff, in, iQuant, iDcScaler); |
2709 |
|
|
2710 |
|
bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len; |
2711 |
|
if (t != 0) cbp |= 1 << (5 - 5); |
2712 |
|
|
2713 |
|
bits += xvid_cbpy_tab[cbp>>2].len; |
2714 |
|
bits += mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len; |
2715 |
} |
} |
2716 |
} |
} |
2717 |
|
return bits; |
2718 |
} |
} |
|
|
|
|
*/ |
|