1 |
/************************************************************************** |
/***************************************************************************** |
2 |
* |
* |
3 |
* Modifications: |
* XVID MPEG-4 VIDEO CODEC |
4 |
|
* - Motion Estimation related code - |
5 |
* |
* |
6 |
* 25.04.2002 partial prevMB conversion |
* Copyright(C) 2002 Christoph Lampert <gruel@web.de> |
7 |
* 22.04.2002 remove some compile warning by chenm001 <chenm001@163.com> |
* 2002 Michael Militzer <michael@xvid.org> |
8 |
* 14.04.2002 added MotionEstimationBVOP() |
* 2002-2003 Radoslaw Czyz <xvid@syskin.cjb.net> |
|
* 02.04.2002 add EPZS(^2) as ME algorithm, use PMV_USESQUARES to choose between |
|
|
* EPZS and EPZS^2 |
|
|
* 08.02.2002 split up PMVfast into three routines: PMVFast, PMVFast_MainLoop |
|
|
* PMVFast_Refine to support multiple searches with different start points |
|
|
* 07.01.2002 uv-block-based interpolation |
|
|
* 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. |
|
9 |
* |
* |
10 |
* Michael Militzer <isibaar@videocoding.de> |
* This program is free software ; you can redistribute it and/or modify |
11 |
|
* it under the terms of the GNU General Public License as published by |
12 |
|
* the Free Software Foundation ; either version 2 of the License, or |
13 |
|
* (at your option) any later version. |
14 |
* |
* |
15 |
**************************************************************************/ |
* This program is distributed in the hope that it will be useful, |
16 |
|
* but WITHOUT ANY WARRANTY ; without even the implied warranty of |
17 |
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
18 |
|
* GNU General Public License for more details. |
19 |
|
* |
20 |
|
* You should have received a copy of the GNU General Public License |
21 |
|
* along with this program ; if not, write to the Free Software |
22 |
|
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
23 |
|
* |
24 |
|
* $Id: motion_est.c,v 1.58.2.35 2003-09-06 11:24:50 Isibaar Exp $ |
25 |
|
* |
26 |
|
****************************************************************************/ |
27 |
|
|
28 |
#include <assert.h> |
#include <assert.h> |
29 |
#include <stdio.h> |
#include <stdio.h> |
30 |
#include <stdlib.h> |
#include <stdlib.h> |
31 |
|
#include <string.h> /* memcpy */ |
32 |
|
#include <math.h> /* lrint */ |
33 |
|
|
34 |
#include "../encoder.h" |
#include "../encoder.h" |
35 |
#include "../utils/mbfunctions.h" |
#include "../utils/mbfunctions.h" |
36 |
#include "../prediction/mbprediction.h" |
#include "../prediction/mbprediction.h" |
37 |
#include "../global.h" |
#include "../global.h" |
38 |
#include "../utils/timer.h" |
#include "../utils/timer.h" |
39 |
|
#include "../image/interpolate8x8.h" |
40 |
|
#include "motion_est.h" |
41 |
#include "motion.h" |
#include "motion.h" |
42 |
#include "sad.h" |
#include "sad.h" |
43 |
|
#include "gmc.h" |
44 |
|
#include "../utils/emms.h" |
45 |
|
#include "../dct/fdct.h" |
46 |
|
|
47 |
|
/***************************************************************************** |
48 |
|
* Modified rounding tables -- declared in motion.h |
49 |
|
* Original tables see ISO spec tables 7-6 -> 7-9 |
50 |
|
****************************************************************************/ |
51 |
|
|
52 |
|
const uint32_t roundtab[16] = |
53 |
|
{0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2 }; |
54 |
|
|
55 |
|
/* K = 4 */ |
56 |
|
const uint32_t roundtab_76[16] = |
57 |
|
{ 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1 }; |
58 |
|
|
59 |
|
/* K = 2 */ |
60 |
|
const uint32_t roundtab_78[8] = |
61 |
|
{ 0, 0, 1, 1, 0, 0, 0, 1 }; |
62 |
|
|
63 |
|
/* K = 1 */ |
64 |
|
const uint32_t roundtab_79[4] = |
65 |
|
{ 0, 1, 0, 0 }; |
66 |
|
|
67 |
|
#define INITIAL_SKIP_THRESH (10) |
68 |
|
#define FINAL_SKIP_THRESH (50) |
69 |
|
#define MAX_SAD00_FOR_SKIP (20) |
70 |
|
#define MAX_CHROMA_SAD_FOR_SKIP (22) |
71 |
|
|
72 |
|
#define CHECK_CANDIDATE(X,Y,D) { \ |
73 |
|
CheckCandidate((X),(Y), data, (D) ); } |
74 |
|
|
75 |
|
|
76 |
|
/***************************************************************************** |
77 |
|
* Code |
78 |
|
****************************************************************************/ |
79 |
|
|
80 |
|
static __inline uint32_t |
81 |
|
d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv) |
82 |
|
{ |
83 |
|
int bits; |
84 |
|
const int q = (1 << (iFcode - 1)) - 1; |
85 |
|
|
86 |
|
x <<= qpel; |
87 |
|
y <<= qpel; |
88 |
|
if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); } |
89 |
|
|
90 |
|
x -= pred.x; |
91 |
|
bits = (x != 0 ? iFcode:0); |
92 |
|
x = abs(x); |
93 |
|
x += q; |
94 |
|
x >>= (iFcode - 1); |
95 |
|
bits += mvtab[x]; |
96 |
|
|
97 |
|
y -= pred.y; |
98 |
|
bits += (y != 0 ? iFcode:0); |
99 |
|
y = abs(y); |
100 |
|
y += q; |
101 |
|
y >>= (iFcode - 1); |
102 |
|
bits += mvtab[y]; |
103 |
|
|
104 |
|
return bits; |
105 |
|
} |
106 |
|
|
107 |
|
static int32_t ChromaSAD2(const int fx, const int fy, const int bx, const int by, |
108 |
|
const SearchData * const data) |
109 |
|
{ |
110 |
|
int sad; |
111 |
|
const uint32_t stride = data->iEdgedWidth/2; |
112 |
|
uint8_t *f_refu, *f_refv, *b_refu, *b_refv; |
113 |
|
|
114 |
|
const INTERPOLATE8X8_PTR interpolate8x8_halfpel[] = { |
115 |
|
NULL, |
116 |
|
interpolate8x8_halfpel_v, |
117 |
|
interpolate8x8_halfpel_h, |
118 |
|
interpolate8x8_halfpel_hv |
119 |
|
}; |
120 |
|
|
121 |
// very large value |
int offset = (fx>>1) + (fy>>1)*stride; |
122 |
#define MV_MAX_ERROR (4096 * 256) |
int filter = ((fx & 1) << 1) | (fy & 1); |
123 |
|
|
124 |
// stop search if sdelta < THRESHOLD |
if (filter != 0) { |
125 |
#define MV16_THRESHOLD 192 |
f_refu = data->RefQ; |
126 |
#define MV8_THRESHOLD 56 |
f_refv = data->RefQ + 8; |
127 |
|
interpolate8x8_halfpel[filter](f_refu, data->RefP[4] + offset, stride, data->rounding); |
128 |
|
interpolate8x8_halfpel[filter](f_refv, data->RefP[5] + offset, stride, data->rounding); |
129 |
|
} else { |
130 |
|
f_refu = (uint8_t*)data->RefP[4] + offset; |
131 |
|
f_refv = (uint8_t*)data->RefP[5] + offset; |
132 |
|
} |
133 |
|
|
134 |
/* sad16(0,0) bias; mpeg4 spec suggests nb/2+1 */ |
offset = (bx>>1) + (by>>1)*stride; |
135 |
/* nb = vop pixels * 2^(bpp-8) */ |
filter = ((bx & 1) << 1) | (by & 1); |
|
#define MV16_00_BIAS (128+1) |
|
136 |
|
|
137 |
/* INTER bias for INTER/INTRA decision; mpeg4 spec suggests 2*nb */ |
if (filter != 0) { |
138 |
#define INTER_BIAS 512 |
b_refu = data->RefQ + 16; |
139 |
|
b_refv = data->RefQ + 24; |
140 |
|
interpolate8x8_halfpel[filter](b_refu, data->b_RefP[4] + offset, stride, data->rounding); |
141 |
|
interpolate8x8_halfpel[filter](b_refv, data->b_RefP[5] + offset, stride, data->rounding); |
142 |
|
} else { |
143 |
|
b_refu = (uint8_t*)data->b_RefP[4] + offset; |
144 |
|
b_refv = (uint8_t*)data->b_RefP[5] + offset; |
145 |
|
} |
146 |
|
|
147 |
/* Parameters which control inter/inter4v decision */ |
sad = sad8bi(data->CurU, b_refu, f_refu, stride); |
148 |
#define IMV16X16 5 |
sad += sad8bi(data->CurV, b_refv, f_refv, stride); |
149 |
|
|
150 |
/* vector map (vlc delta size) smoother parameters */ |
return sad; |
151 |
#define NEIGH_TEND_16X16 2 |
} |
|
#define NEIGH_TEND_8X8 2 |
|
152 |
|
|
153 |
|
static int32_t |
154 |
|
ChromaSAD(const int dx, const int dy, const SearchData * const data) |
155 |
|
{ |
156 |
|
int sad; |
157 |
|
const uint32_t stride = data->iEdgedWidth/2; |
158 |
|
int offset = (dx>>1) + (dy>>1)*stride; |
159 |
|
int next = 1; |
160 |
|
|
161 |
// fast ((A)/2)*2 |
if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; /* it has been checked recently */ |
162 |
#define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) |
data->temp[5] = dx; data->temp[6] = dy; /* backup */ |
163 |
|
|
164 |
|
switch (((dx & 1) << 1) | (dy & 1)) { |
165 |
|
case 0: |
166 |
|
sad = sad8(data->CurU, data->RefP[4] + offset, stride); |
167 |
|
sad += sad8(data->CurV, data->RefP[5] + offset, stride); |
168 |
|
break; |
169 |
|
case 1: |
170 |
|
next = stride; |
171 |
|
case 2: |
172 |
|
sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + next, stride); |
173 |
|
sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + next, stride); |
174 |
|
break; |
175 |
|
default: |
176 |
|
interpolate8x8_halfpel_hv(data->RefQ, data->RefP[4] + offset, stride, data->rounding); |
177 |
|
sad = sad8(data->CurU, data->RefQ, stride); |
178 |
|
|
179 |
int32_t PMVfastSearch16( |
interpolate8x8_halfpel_hv(data->RefQ, data->RefP[5] + offset, stride, data->rounding); |
180 |
const uint8_t * const pRef, |
sad += sad8(data->CurV, data->RefQ, stride); |
181 |
const uint8_t * const pRefH, |
break; |
182 |
const uint8_t * const pRefV, |
} |
183 |
const uint8_t * const pRefHV, |
data->temp[7] = sad; /* backup, part 2 */ |
184 |
const IMAGE * const pCur, |
return sad; |
185 |
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); |
|
186 |
|
|
187 |
int32_t EPZSSearch16( |
static __inline const uint8_t * |
188 |
const uint8_t * const pRef, |
GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data) |
189 |
const uint8_t * const pRefH, |
{ |
190 |
const uint8_t * const pRefV, |
/* dir : 0 = forward, 1 = backward */ |
191 |
const uint8_t * const pRefHV, |
const uint8_t *const *const direction = ( dir == 0 ? data->RefP : data->b_RefP ); |
192 |
const IMAGE * const pCur, |
const int picture = ((x&1)<<1) | (y&1); |
193 |
const int x, const int y, |
const int offset = (x>>1) + (y>>1)*data->iEdgedWidth; |
194 |
const uint32_t MotionFlags, |
return direction[picture] + offset; |
195 |
const uint32_t iQuant, |
} |
196 |
const uint32_t iFcode, |
|
197 |
const MBParam * const pParam, |
/* this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate */ |
198 |
const MACROBLOCK * const pMBs, |
static __inline const uint8_t * |
199 |
const MACROBLOCK * const prevMBs, |
GetReference(const int x, const int y, const SearchData * const data) |
200 |
VECTOR * const currMV, |
{ |
201 |
VECTOR * const currPMV); |
const int picture = ((x&1)<<1) | (y&1); |
202 |
|
const int offset = (x>>1) + (y>>1)*data->iEdgedWidth; |
203 |
|
return data->RefP[picture] + offset; |
204 |
|
} |
205 |
|
|
206 |
|
static uint8_t * |
207 |
|
Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data) |
208 |
|
{ |
209 |
|
/* create or find a qpel-precision reference picture; return pointer to it */ |
210 |
|
uint8_t * Reference = data->RefQ + 16*dir; |
211 |
|
const uint32_t iEdgedWidth = data->iEdgedWidth; |
212 |
|
const uint32_t rounding = data->rounding; |
213 |
|
const int halfpel_x = x/2; |
214 |
|
const int halfpel_y = y/2; |
215 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
216 |
|
|
217 |
|
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
218 |
|
ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
219 |
|
switch( ((x&1)<<1) + (y&1) ) { |
220 |
|
case 3: /* x and y in qpel resolution - the "corners" (top left/right and */ |
221 |
|
/* bottom left/right) during qpel refinement */ |
222 |
|
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
223 |
|
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
224 |
|
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
225 |
|
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
226 |
|
ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
227 |
|
ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
228 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
229 |
|
break; |
230 |
|
|
231 |
|
case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */ |
232 |
|
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
233 |
|
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
234 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
235 |
|
break; |
236 |
|
|
237 |
int32_t PMVfastSearch8( |
case 2: /* x qpel, y halfpel - left or right during qpel refinement */ |
238 |
const uint8_t * const pRef, |
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
239 |
const uint8_t * const pRefH, |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
240 |
const uint8_t * const pRefV, |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
241 |
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); |
|
242 |
|
|
243 |
int32_t EPZSSearch8( |
default: /* pure halfpel position */ |
244 |
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); |
|
245 |
|
|
246 |
|
} |
247 |
|
return Reference; |
248 |
|
} |
249 |
|
|
250 |
typedef int32_t (MainSearch16Func)( |
static uint8_t * |
251 |
const uint8_t * const pRef, |
Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data) |
252 |
const uint8_t * const pRefH, |
{ |
253 |
const uint8_t * const pRefV, |
/* create or find a qpel-precision reference picture; return pointer to it */ |
254 |
const uint8_t * const pRefHV, |
uint8_t * Reference = data->RefQ + 16*dir; |
255 |
const uint8_t * const cur, |
const uint32_t iEdgedWidth = data->iEdgedWidth; |
256 |
const int x, const int y, |
const uint32_t rounding = data->rounding; |
257 |
int32_t startx, int32_t starty, |
const int halfpel_x = x/2; |
258 |
int32_t iMinSAD, |
const int halfpel_y = y/2; |
259 |
VECTOR * const currMV, |
const uint8_t *ref1, *ref2, *ref3, *ref4; |
|
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); |
|
260 |
|
|
261 |
typedef MainSearch16Func* MainSearch16FuncPtr; |
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
262 |
|
switch( ((x&1)<<1) + (y&1) ) { |
263 |
|
case 3: |
264 |
|
/* |
265 |
|
* x and y in qpel resolution - the "corners" (top left/right and |
266 |
|
* bottom left/right) during qpel refinement |
267 |
|
*/ |
268 |
|
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
269 |
|
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
270 |
|
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
271 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
272 |
|
interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); |
273 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); |
274 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); |
275 |
|
break; |
276 |
|
|
277 |
|
case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */ |
278 |
|
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
279 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
280 |
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
281 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); |
282 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
283 |
|
break; |
284 |
|
|
285 |
typedef int32_t (MainSearch8Func)( |
case 2: /* x qpel, y halfpel - left or right during qpel refinement */ |
286 |
const uint8_t * const pRef, |
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
287 |
const uint8_t * const pRefH, |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
288 |
const uint8_t * const pRefV, |
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
289 |
const uint8_t * const pRefHV, |
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); |
290 |
const uint8_t * const cur, |
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
291 |
const int x, const int y, |
break; |
|
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 |
|
|
}; |
|
292 |
|
|
293 |
|
|
294 |
static __inline uint32_t mv_bits(int32_t component, const uint32_t iFcode) |
default: /* pure halfpel position */ |
295 |
{ |
return (uint8_t *) ref1; |
296 |
if (component == 0) |
} |
297 |
return 1; |
return Reference; |
298 |
|
} |
299 |
|
|
300 |
if (component < 0) |
/* CHECK_CANDIATE FUNCTIONS START */ |
|
component = -component; |
|
301 |
|
|
302 |
if (iFcode == 1) |
static void |
303 |
|
CheckCandidate16(const int x, const int y, const SearchData * const data, const int Direction) |
304 |
{ |
{ |
305 |
if (component > 32) |
int xc, yc; |
306 |
component = 32; |
const uint8_t * Reference; |
307 |
|
VECTOR * current; |
308 |
|
int32_t sad; uint32_t t; |
309 |
|
|
310 |
return mvtab[component] + 1; |
if ( (x > data->max_dx) || (x < data->min_dx) |
311 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
312 |
|
|
313 |
|
if (!data->qpel_precision) { |
314 |
|
Reference = GetReference(x, y, data); |
315 |
|
current = data->currentMV; |
316 |
|
xc = x; yc = y; |
317 |
|
} else { /* x and y are in 1/4 precision */ |
318 |
|
Reference = Interpolate16x16qpel(x, y, 0, data); |
319 |
|
xc = x/2; yc = y/2; /* for chroma sad */ |
320 |
|
current = data->currentQMV; |
321 |
} |
} |
322 |
|
|
323 |
component += (1 << (iFcode - 1)) - 1; |
sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp); |
324 |
component >>= (iFcode - 1); |
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
325 |
|
|
326 |
|
sad += (data->lambda16 * t * sad)>>10; |
327 |
|
data->temp[0] += (data->lambda8 * t * (data->temp[0] + NEIGH_8X8_BIAS))>>10; |
328 |
|
|
329 |
if (component > 32) |
if (data->chroma && sad < data->iMinSAD[0]) |
330 |
component = 32; |
sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
331 |
|
(yc >> 1) + roundtab_79[yc & 0x3], data); |
332 |
|
|
333 |
return mvtab[component] + 1 + iFcode - 1; |
if (sad < data->iMinSAD[0]) { |
334 |
|
data->iMinSAD[0] = sad; |
335 |
|
current[0].x = x; current[0].y = y; |
336 |
|
*data->dir = Direction; |
337 |
} |
} |
338 |
|
|
339 |
|
if (data->temp[0] < data->iMinSAD[1]) { |
340 |
|
data->iMinSAD[1] = data->temp[0]; current[1].x = x; current[1].y = y; } |
341 |
|
if (data->temp[1] < data->iMinSAD[2]) { |
342 |
|
data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; } |
343 |
|
if (data->temp[2] < data->iMinSAD[3]) { |
344 |
|
data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; } |
345 |
|
if (data->temp[3] < data->iMinSAD[4]) { |
346 |
|
data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; } |
347 |
|
} |
348 |
|
|
349 |
static __inline uint32_t calc_delta_16(const int32_t dx, const int32_t dy, const uint32_t iFcode) |
static void |
350 |
|
CheckCandidate16_subpel(const int x, const int y, const SearchData * const data, const int Direction) |
351 |
{ |
{ |
352 |
return NEIGH_TEND_16X16 * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); |
int xc, yc; |
353 |
} |
const uint8_t *Reference; |
354 |
|
VECTOR *current, *current2; |
355 |
|
int32_t sad; uint32_t t; |
356 |
|
|
357 |
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) |
358 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
359 |
|
|
360 |
{ |
if (!data->qpel_precision) { |
361 |
return NEIGH_TEND_8X8 * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); |
Reference = GetReference(x, y, data); |
362 |
|
current = data->currentMV; |
363 |
|
current2 = data->currentMV2; |
364 |
|
xc = x; yc = y; |
365 |
|
} else { /* x and y are in 1/4 precision */ |
366 |
|
Reference = Interpolate16x16qpel(x, y, 0, data); |
367 |
|
xc = x/2; yc = y/2; /* for chroma sad */ |
368 |
|
current = data->currentQMV; |
369 |
|
current2 = data->currentQMV2; |
370 |
} |
} |
371 |
|
|
372 |
|
sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp); |
373 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
374 |
|
|
375 |
|
sad += (data->lambda16 * t * sad)>>10; |
376 |
|
data->temp[0] += (data->lambda8 * t * (data->temp[0] + NEIGH_8X8_BIAS))>>10; |
377 |
|
|
378 |
|
if (data->chroma && sad < data->iMinSAD[0]) |
379 |
|
sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
380 |
|
(yc >> 1) + roundtab_79[yc & 0x3], data); |
381 |
|
|
382 |
|
if (data->temp[0] < data->iMinSAD[1]) { |
383 |
|
data->iMinSAD[1] = data->temp[0]; current[1].x = x; current[1].y = y; } |
384 |
|
if (data->temp[1] < data->iMinSAD[2]) { |
385 |
|
data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; } |
386 |
|
if (data->temp[2] < data->iMinSAD[3]) { |
387 |
|
data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; } |
388 |
|
if (data->temp[3] < data->iMinSAD[4]) { |
389 |
|
data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; } |
390 |
|
|
391 |
#ifndef SEARCH16 |
if (sad < data->iMinSAD[0]) { |
392 |
#define SEARCH16 PMVfastSearch16 |
*(data->iMinSAD2) = *(data->iMinSAD); |
393 |
//#define SEARCH16 FullSearch16 |
current2->x = current->x; current2->y = current->y; |
|
//#define SEARCH16 EPZSSearch16 |
|
|
#endif |
|
394 |
|
|
395 |
#ifndef SEARCH8 |
data->iMinSAD[0] = sad; |
396 |
#define SEARCH8 PMVfastSearch8 |
current[0].x = x; current[0].y = y; |
397 |
//#define SEARCH8 EPZSSearch8 |
*data->dir = Direction; |
398 |
#endif |
return; |
399 |
|
} |
400 |
|
|
401 |
bool MotionEstimation( |
if (sad < *(data->iMinSAD2)) { |
402 |
MBParam * const pParam, |
*(data->iMinSAD2) = sad; |
403 |
FRAMEINFO * const current, |
current2->x = x; current2->y = y; |
404 |
FRAMEINFO * const reference, |
*data->dir = Direction; |
405 |
const IMAGE * const pRefH, |
} |
406 |
const IMAGE * const pRefV, |
} |
|
const IMAGE * const pRefHV, |
|
|
const uint32_t iLimit) |
|
407 |
|
|
408 |
|
static void |
409 |
|
CheckCandidate8(const int x, const int y, const SearchData * const data, const int Direction) |
410 |
{ |
{ |
411 |
const uint32_t iWcount = pParam->mb_width; |
int32_t sad; uint32_t t; |
412 |
const uint32_t iHcount = pParam->mb_height; |
const uint8_t * Reference; |
413 |
MACROBLOCK * pMBs = current->mbs; |
VECTOR * current; |
|
IMAGE * pCurrent = ¤t->image; |
|
|
|
|
|
MACROBLOCK * prevMBs = reference->mbs; // previous frame |
|
|
IMAGE * pRef = &reference->image; |
|
|
|
|
414 |
|
|
415 |
uint32_t i, j, iIntra = 0; |
if ( (x > data->max_dx) || (x < data->min_dx) |
416 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
417 |
|
|
418 |
VECTOR mv16; |
if (!data->qpel_precision) { |
419 |
VECTOR pmv16; |
Reference = GetReference(x, y, data); |
420 |
|
current = data->currentMV; |
421 |
int32_t sad8 = 0; |
} else { /* x and y are in 1/4 precision */ |
422 |
int32_t sad16; |
Reference = Interpolate8x8qpel(x, y, 0, 0, data); |
423 |
int32_t deviation; |
current = data->currentQMV; |
424 |
|
} |
|
if (sadInit) |
|
|
(*sadInit)(); |
|
425 |
|
|
426 |
|
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
427 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
428 |
|
|
429 |
/* eventhough we have a seperate prevMBs, |
sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10; |
|
pmvfast/epsz does something "funny" with the previous frames data */ |
|
430 |
|
|
431 |
for (i = 0; i < iHcount; i++) |
if (sad < *(data->iMinSAD)) { |
432 |
for (j = 0; j < iWcount; j++) |
*(data->iMinSAD) = sad; |
433 |
{ |
current->x = x; current->y = y; |
434 |
pMBs[j + i * iWcount].mvs[0] = prevMBs[j + i * iWcount].mvs[0]; |
*data->dir = Direction; |
435 |
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]; |
|
|
} |
|
|
|
|
|
/*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); |
|
436 |
} |
} |
|
*/ |
|
437 |
|
|
438 |
// note: i==horizontal, j==vertical |
static void |
439 |
for (i = 0; i < iHcount; i++) |
CheckCandidate32(const int x, const int y, const SearchData * const data, const int Direction) |
|
for (j = 0; j < iWcount; j++) |
|
440 |
{ |
{ |
441 |
MACROBLOCK *pMB = &pMBs[j + i * iWcount]; |
uint32_t t; |
442 |
MACROBLOCK *prevMB = &prevMBs[j + i * iWcount]; |
const uint8_t * Reference; |
443 |
|
int sad; |
|
sad16 = SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
|
|
j, i, current->motion_flags, current->quant, current->fcode, |
|
|
pParam, pMBs, prevMBs, &mv16, &pmv16); |
|
|
pMB->sad16=sad16; |
|
|
|
|
444 |
|
|
445 |
/* decide: MODE_INTER or MODE_INTRA |
if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || /* non-zero even value */ |
446 |
if (dev_intra < sad_inter - 2 * nb) use_intra |
(x > data->max_dx) || (x < data->min_dx) |
447 |
*/ |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
448 |
|
|
449 |
deviation = dev16(pCurrent->y + j*16 + i*16*pParam->edged_width, pParam->edged_width); |
Reference = GetReference(x, y, data); |
450 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1); |
451 |
|
|
452 |
if (deviation < (sad16 - INTER_BIAS)) |
sad = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp); |
|
{ |
|
|
pMB->mode = MODE_INTRA; |
|
|
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
|
|
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; |
|
453 |
|
|
454 |
pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = 0; |
sad += (data->lambda16 * t * sad) >> 10; |
455 |
|
data->temp[0] += (data->lambda8 * t * (data->temp[0] + NEIGH_8X8_BIAS))>>10; |
456 |
|
|
457 |
iIntra++; |
if (sad < data->iMinSAD[0]) { |
458 |
if(iIntra >= iLimit) |
data->iMinSAD[0] = sad; |
459 |
return 1; |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
460 |
|
*data->dir = Direction; |
461 |
|
} |
462 |
|
|
463 |
continue; |
if (data->temp[0] < data->iMinSAD[1]) { |
464 |
|
data->iMinSAD[1] = data->temp[0]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
465 |
|
if (data->temp[1] < data->iMinSAD[2]) { |
466 |
|
data->iMinSAD[2] = data->temp[1]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
467 |
|
if (data->temp[2] < data->iMinSAD[3]) { |
468 |
|
data->iMinSAD[3] = data->temp[2]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
469 |
|
if (data->temp[3] < data->iMinSAD[4]) { |
470 |
|
data->iMinSAD[4] = data->temp[3]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
471 |
} |
} |
472 |
|
|
473 |
if (current->global_flags & XVID_INTER4V) |
static void |
474 |
|
CheckCandidate16no4v(const int x, const int y, const SearchData * const data, const int Direction) |
475 |
{ |
{ |
476 |
pMB->sad8[0] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
int32_t sad, xc, yc; |
477 |
2 * j, 2 * i, mv16.x, mv16.y, |
const uint8_t * Reference; |
478 |
current->motion_flags, current->quant, current->fcode, |
uint32_t t; |
479 |
pParam, pMBs, prevMBs, &pMB->mvs[0], &pMB->pmvs[0]); |
VECTOR * current; |
480 |
|
|
481 |
pMB->sad8[1] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
if ( (x > data->max_dx) || ( x < data->min_dx) |
482 |
2 * j + 1, 2 * i, mv16.x, mv16.y, |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
|
current->motion_flags, current->quant, current->fcode, |
|
|
pParam, pMBs, prevMBs, &pMB->mvs[1], &pMB->pmvs[1]); |
|
483 |
|
|
484 |
pMB->sad8[2] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; /* non-zero even value */ |
|
2 * j, 2 * i + 1, mv16.x, mv16.y, |
|
|
current->motion_flags, current->quant, current->fcode, |
|
|
pParam, pMBs, prevMBs, &pMB->mvs[2], &pMB->pmvs[2]); |
|
485 |
|
|
486 |
pMB->sad8[3] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
if (data->qpel_precision) { /* x and y are in 1/4 precision */ |
487 |
2 * j + 1, 2 * i + 1, mv16.x, mv16.y, |
Reference = Interpolate16x16qpel(x, y, 0, data); |
488 |
current->motion_flags, current->quant, current->fcode, |
current = data->currentQMV; |
489 |
pParam, pMBs, prevMBs, &pMB->mvs[3], &pMB->pmvs[3]); |
xc = x/2; yc = y/2; |
490 |
|
} else { |
491 |
sad8 = pMB->sad8[0] + pMB->sad8[1] + pMB->sad8[2] + pMB->sad8[3]; |
Reference = GetReference(x, y, data); |
492 |
|
current = data->currentMV; |
493 |
|
xc = x; yc = y; |
494 |
} |
} |
495 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, |
496 |
|
data->qpel^data->qpel_precision, data->rrv); |
497 |
|
|
498 |
|
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
499 |
|
sad += (data->lambda16 * t * sad)>>10; |
500 |
|
|
501 |
/* decide: MODE_INTER or MODE_INTER4V |
if (data->chroma && sad < *data->iMinSAD) |
502 |
mpeg4: if (sad8 < sad16 - nb/2+1) use_inter4v |
sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
503 |
*/ |
(yc >> 1) + roundtab_79[yc & 0x3], data); |
|
|
|
|
if (!(current->global_flags & XVID_LUMIMASKING) || pMB->dquant == NO_CHANGE) |
|
|
{ |
|
|
if (((current->global_flags & XVID_INTER4V)==0) || |
|
|
(sad16 < (sad8 + (int32_t)(IMV16X16 * current->quant)))) |
|
|
{ |
|
504 |
|
|
505 |
sad8 = sad16; |
if (sad < *(data->iMinSAD)) { |
506 |
pMB->mode = MODE_INTER; |
*(data->iMinSAD) = sad; |
507 |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; |
current->x = x; current->y = y; |
508 |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; |
*data->dir = Direction; |
|
pMB->pmvs[0].x = pmv16.x; |
|
|
pMB->pmvs[0].y = pmv16.y; |
|
|
} |
|
|
else |
|
|
pMB->mode = MODE_INTER4V; |
|
|
} |
|
|
else |
|
|
{ |
|
|
sad8 = sad16; |
|
|
pMB->mode = MODE_INTER; |
|
|
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; |
|
|
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; |
|
|
pMB->pmvs[0].x = pmv16.x; |
|
|
pMB->pmvs[0].y = pmv16.y; |
|
509 |
} |
} |
510 |
} |
} |
511 |
|
|
512 |
/* dprintf("*** AFTER ***", pMBs[0].b_mvs[0].x); |
static void |
513 |
for (i = 0; i < iHcount; i++) |
CheckCandidate16I(const int x, const int y, const SearchData * const data, const int Direction) |
514 |
for (j = 0; j < iWcount; j++) |
{ |
515 |
{ |
int sad; |
516 |
dprintf(" [%i,%i] mode=%i dquant=%i mvs=(%i %i %i %i) sad8=(%i %i %i %i) sad16=(%i)", j,i, |
// int xc, yc; |
517 |
pMBs[j + i * iWcount].mode, |
const uint8_t * Reference; |
518 |
pMBs[j + i * iWcount].dquant, |
// VECTOR * current; |
|
pMBs[j + i * iWcount].mvs[0], |
|
|
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); |
|
|
} |
|
|
*/ |
|
519 |
|
|
520 |
return 0; |
if ( (x > data->max_dx) || ( x < data->min_dx) |
521 |
} |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
522 |
|
|
523 |
#define MVzero(A) ( ((A).x)==(0) && ((A).y)==(0) ) |
Reference = GetReference(x, y, data); |
524 |
|
// xc = x; yc = y; |
525 |
|
|
526 |
#define MVequal(A,B) ( ((A).x)==((B).x) && ((A).y)==((B).y) ) |
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
527 |
|
// sad += d_mv_bits(x, y, data->predMV, data->iFcode, 0, 0); |
528 |
|
|
529 |
|
/* if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
530 |
|
(yc >> 1) + roundtab_79[yc & 0x3], data); |
531 |
|
*/ |
532 |
|
|
533 |
#define CHECK_MV16_ZERO {\ |
if (sad < data->iMinSAD[0]) { |
534 |
if ( (0 <= max_dx) && (0 >= min_dx) \ |
data->iMinSAD[0] = sad; |
535 |
&& (0 <= max_dy) && (0 >= min_dy) ) \ |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
536 |
{ \ |
*data->dir = Direction; |
537 |
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0, 0 , iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); \ |
} |
|
iSAD += calc_delta_16(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD <= iQuant * 96) \ |
|
|
iSAD -= MV16_00_BIAS; \ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } } \ |
|
|
} |
|
|
|
|
|
#define NOCHECK_MV16_CANDIDATE(X,Y) { \ |
|
|
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(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; } } \ |
|
538 |
} |
} |
539 |
|
|
540 |
/* too slow and not fully functional at the moment */ |
static void |
541 |
/* |
CheckCandidate32I(const int x, const int y, const SearchData * const data, const int Direction) |
|
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) |
|
542 |
{ |
{ |
543 |
const int32_t iEdgedWidth = pParam->edged_width; |
/* maximum speed - for P/B/I decision */ |
544 |
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
int32_t sad; |
|
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; |
|
545 |
|
|
546 |
return iSAD; |
if ( (x > data->max_dx) || (x < data->min_dx) |
547 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
548 |
|
|
549 |
|
sad = sad32v_c(data->Cur, data->RefP[0] + x + y*((int)data->iEdgedWidth), |
550 |
|
data->iEdgedWidth, data->temp); |
551 |
|
|
552 |
|
if (sad < *(data->iMinSAD)) { |
553 |
|
*(data->iMinSAD) = sad; |
554 |
|
data->currentMV[0].x = x; data->currentMV[0].y = y; |
555 |
|
*data->dir = Direction; |
556 |
} |
} |
557 |
*/ |
if (data->temp[0] < data->iMinSAD[1]) { |
558 |
|
data->iMinSAD[1] = data->temp[0]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
559 |
|
if (data->temp[1] < data->iMinSAD[2]) { |
560 |
|
data->iMinSAD[2] = data->temp[1]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
561 |
|
if (data->temp[2] < data->iMinSAD[3]) { |
562 |
|
data->iMinSAD[3] = data->temp[2]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
563 |
|
if (data->temp[3] < data->iMinSAD[4]) { |
564 |
|
data->iMinSAD[4] = data->temp[3]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
565 |
|
|
|
int32_t Diamond16_MainSearch( |
|
|
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, |
|
|
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_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); |
|
566 |
} |
} |
567 |
else |
|
568 |
|
static void |
569 |
|
CheckCandidateInt(const int xf, const int yf, const SearchData * const data, const int Direction) |
570 |
{ |
{ |
571 |
currMV->x = startx; |
int32_t sad, xb, yb, xcf, ycf, xcb, ycb; |
572 |
currMV->y = starty; |
uint32_t t; |
573 |
} |
const uint8_t *ReferenceF, *ReferenceB; |
574 |
return iMinSAD; |
VECTOR *current; |
|
} |
|
575 |
|
|
576 |
int32_t Square16_MainSearch( |
if ((xf > data->max_dx) || (xf < data->min_dx) || |
577 |
const uint8_t * const pRef, |
(yf > data->max_dy) || (yf < data->min_dy)) |
578 |
const uint8_t * const pRefH, |
return; |
|
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) |
|
|
{ |
|
|
/* 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 |
|
|
*/ |
|
579 |
|
|
580 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
if (!data->qpel_precision) { |
581 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
ReferenceF = GetReference(xf, yf, data); |
582 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
xb = data->currentMV[1].x; yb = data->currentMV[1].y; |
583 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
ReferenceB = GetReferenceB(xb, yb, 1, data); |
584 |
|
current = data->currentMV; |
585 |
|
xcf = xf; ycf = yf; |
586 |
|
xcb = xb; ycb = yb; |
587 |
|
} else { |
588 |
|
ReferenceF = Interpolate16x16qpel(xf, yf, 0, data); |
589 |
|
xb = data->currentQMV[1].x; yb = data->currentQMV[1].y; |
590 |
|
current = data->currentQMV; |
591 |
|
ReferenceB = Interpolate16x16qpel(xb, yb, 1, data); |
592 |
|
xcf = xf/2; ycf = yf/2; |
593 |
|
xcb = xb/2; ycb = yb/2; |
594 |
|
} |
595 |
|
|
596 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0) |
597 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
+ d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
|
598 |
|
|
599 |
|
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
600 |
|
sad += (data->lambda16 * t * sad)>>10; |
601 |
|
|
602 |
if (iDirection) |
if (data->chroma && sad < *data->iMinSAD) |
603 |
while (!iFound) |
sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3], |
604 |
{ |
(ycf >> 1) + roundtab_79[ycf & 0x3], |
605 |
iFound = 1; |
(xcb >> 1) + roundtab_79[xcb & 0x3], |
606 |
backupMV=*currMV; |
(ycb >> 1) + roundtab_79[ycb & 0x3], data); |
607 |
|
|
608 |
switch (iDirection) |
if (sad < *(data->iMinSAD)) { |
609 |
{ |
*(data->iMinSAD) = sad; |
610 |
case 1: |
current->x = xf; current->y = yf; |
611 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
*data->dir = Direction; |
612 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
} |
613 |
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; |
|
614 |
|
|
615 |
case 3: |
static void |
616 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
CheckCandidateDirect(const int x, const int y, const SearchData * const data, const int Direction) |
617 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
{ |
618 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0; |
619 |
break; |
uint32_t k; |
620 |
|
const uint8_t *ReferenceF; |
621 |
|
const uint8_t *ReferenceB; |
622 |
|
VECTOR mvs, b_mvs; |
623 |
|
|
624 |
case 4: |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
|
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; |
|
625 |
|
|
626 |
case 5: |
for (k = 0; k < 4; k++) { |
627 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
mvs.x = data->directmvF[k].x + x; |
628 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
b_mvs.x = ((x == 0) ? |
629 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
data->directmvB[k].x |
630 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
: mvs.x - data->referencemv[k].x); |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
|
|
break; |
|
631 |
|
|
632 |
case 6: |
mvs.y = data->directmvF[k].y + y; |
633 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
b_mvs.y = ((y == 0) ? |
634 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
data->directmvB[k].y |
635 |
|
: mvs.y - data->referencemv[k].y); |
636 |
|
|
637 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
if ((mvs.x > data->max_dx) || (mvs.x < data->min_dx) || |
638 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
(mvs.y > data->max_dy) || (mvs.y < data->min_dy) || |
639 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
(b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx) || |
640 |
|
(b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) ) |
641 |
|
return; |
642 |
|
|
643 |
break; |
if (data->qpel) { |
644 |
|
xcf += mvs.x/2; ycf += mvs.y/2; |
645 |
|
xcb += b_mvs.x/2; ycb += b_mvs.y/2; |
646 |
|
} else { |
647 |
|
xcf += mvs.x; ycf += mvs.y; |
648 |
|
xcb += b_mvs.x; ycb += b_mvs.y; |
649 |
|
mvs.x *= 2; mvs.y *= 2; /* we move to qpel precision anyway */ |
650 |
|
b_mvs.x *= 2; b_mvs.y *= 2; |
651 |
|
} |
652 |
|
|
653 |
case 7: |
ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); |
654 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
|
|
break; |
|
655 |
|
|
656 |
case 8: |
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
657 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
ReferenceF, ReferenceB, data->iEdgedWidth); |
658 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
if (sad > *(data->iMinSAD)) return; |
|
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); |
|
|
break; |
|
|
default: |
|
|
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); |
|
|
|
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
|
|
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); |
|
|
break; |
|
|
} |
|
|
} |
|
|
else |
|
|
{ |
|
|
currMV->x = startx; |
|
|
currMV->y = starty; |
|
|
} |
|
|
return iMinSAD; |
|
659 |
} |
} |
660 |
|
|
661 |
|
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
662 |
|
|
663 |
int32_t Full16_MainSearch( |
if (data->chroma && sad < *data->iMinSAD) |
664 |
const uint8_t * const pRef, |
sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
665 |
const uint8_t * const pRefH, |
(ycf >> 3) + roundtab_76[ycf & 0xf], |
666 |
const uint8_t * const pRefV, |
(xcb >> 3) + roundtab_76[xcb & 0xf], |
667 |
const uint8_t * const pRefHV, |
(ycb >> 3) + roundtab_76[ycb & 0xf], data); |
|
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_MV16_CANDIDATE(dx,dy); |
|
668 |
|
|
669 |
return iMinSAD; |
if (sad < *(data->iMinSAD)) { |
670 |
|
*(data->iMinSAD) = sad; |
671 |
|
data->currentMV->x = x; data->currentMV->y = y; |
672 |
|
*data->dir = Direction; |
673 |
} |
} |
|
|
|
|
int32_t Full8_MainSearch( |
|
|
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, |
|
|
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); |
|
|
|
|
|
return iMinSAD; |
|
674 |
} |
} |
675 |
|
|
676 |
|
static void |
677 |
|
CheckCandidateDirectno4v(const int x, const int y, const SearchData * const data, const int Direction) |
678 |
|
{ |
679 |
|
int32_t sad, xcf, ycf, xcb, ycb; |
680 |
|
const uint8_t *ReferenceF; |
681 |
|
const uint8_t *ReferenceB; |
682 |
|
VECTOR mvs, b_mvs; |
683 |
|
|
684 |
|
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
685 |
|
|
686 |
int32_t Halfpel16_Refine( |
mvs.x = data->directmvF[0].x + x; |
687 |
const uint8_t * const pRef, |
b_mvs.x = ((x == 0) ? |
688 |
const uint8_t * const pRefH, |
data->directmvB[0].x |
689 |
const uint8_t * const pRefV, |
: mvs.x - data->referencemv[0].x); |
|
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) */ |
|
690 |
|
|
691 |
int32_t iSAD; |
mvs.y = data->directmvF[0].y + y; |
692 |
VECTOR backupMV = *currMV; |
b_mvs.y = ((y == 0) ? |
693 |
|
data->directmvB[0].y |
694 |
|
: mvs.y - data->referencemv[0].y); |
695 |
|
|
696 |
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y-1); |
if ( (mvs.x > data->max_dx) || (mvs.x < data->min_dx) |
697 |
CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y-1); |
|| (mvs.y > data->max_dy) || (mvs.y < data->min_dy) |
698 |
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y-1); |
|| (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx) |
699 |
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y); |
|| (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) ) return; |
|
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); |
|
700 |
|
|
701 |
return iMinSAD; |
if (data->qpel) { |
702 |
|
xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2); |
703 |
|
xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2); |
704 |
|
ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data); |
705 |
|
ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); |
706 |
|
} else { |
707 |
|
xcf = 4*mvs.x; ycf = 4*mvs.y; |
708 |
|
xcb = 4*b_mvs.x; ycb = 4*b_mvs.y; |
709 |
|
ReferenceF = GetReference(mvs.x, mvs.y, data); |
710 |
|
ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data); |
711 |
} |
} |
712 |
|
|
713 |
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
714 |
|
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
715 |
|
|
716 |
|
if (data->chroma && sad < *data->iMinSAD) |
717 |
|
sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
718 |
|
(ycf >> 3) + roundtab_76[ycf & 0xf], |
719 |
|
(xcb >> 3) + roundtab_76[xcb & 0xf], |
720 |
|
(ycb >> 3) + roundtab_76[ycb & 0xf], data); |
721 |
|
|
722 |
int32_t PMVfastSearch16( |
if (sad < *(data->iMinSAD)) { |
723 |
const uint8_t * const pRef, |
*(data->iMinSAD) = sad; |
724 |
const uint8_t * const pRefH, |
data->currentMV->x = x; data->currentMV->y = y; |
725 |
const uint8_t * const pRefV, |
*data->dir = Direction; |
726 |
const uint8_t * const pRefHV, |
} |
727 |
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) |
|
|
{ |
|
|
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; |
|
728 |
|
|
|
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
|
729 |
|
|
730 |
int32_t iDiamondSize; |
static void |
731 |
|
CheckCandidateRD16(const int x, const int y, const SearchData * const data, const int Direction) |
732 |
|
{ |
733 |
|
|
734 |
int32_t min_dx; |
int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; |
735 |
int32_t max_dx; |
int32_t rd = 0; |
736 |
int32_t min_dy; |
VECTOR * current; |
737 |
int32_t max_dy; |
const uint8_t * ptr; |
738 |
|
int i, cbp = 0, t, xc, yc; |
739 |
|
|
740 |
int32_t iFound; |
if ( (x > data->max_dx) || (x < data->min_dx) |
741 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
742 |
|
|
743 |
VECTOR newMV; |
if (!data->qpel_precision) { |
744 |
VECTOR backupMV; /* just for PMVFAST */ |
ptr = GetReference(x, y, data); |
745 |
|
current = data->currentMV; |
746 |
|
xc = x; yc = y; |
747 |
|
} else { /* x and y are in 1/4 precision */ |
748 |
|
ptr = Interpolate16x16qpel(x, y, 0, data); |
749 |
|
current = data->currentQMV; |
750 |
|
xc = x/2; yc = y/2; |
751 |
|
} |
752 |
|
|
753 |
VECTOR pmv[4]; |
for(i = 0; i < 4; i++) { |
754 |
int32_t psad[4]; |
int s = 8*((i&1) + (i>>1)*data->iEdgedWidth); |
755 |
|
transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth); |
756 |
|
rd += data->temp[i] = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, i); |
757 |
|
} |
758 |
|
|
759 |
const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
rd += t = BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
|
const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; |
|
760 |
|
|
761 |
static int32_t threshA,threshB; |
if (data->temp[0] + t < data->iMinSAD[1]) { |
762 |
int32_t bPredEq; |
data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; data->cbp[1] = (data->cbp[1]&~32) | (cbp&32); } |
763 |
int32_t iMinSAD,iSAD; |
if (data->temp[1] < data->iMinSAD[2]) { |
764 |
|
data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; data->cbp[1] = (data->cbp[1]&~16) | (cbp&16); } |
765 |
|
if (data->temp[2] < data->iMinSAD[3]) { |
766 |
|
data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; data->cbp[1] = (data->cbp[1]&~8) | (cbp&8); } |
767 |
|
if (data->temp[3] < data->iMinSAD[4]) { |
768 |
|
data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; data->cbp[1] = (data->cbp[1]&~4) | (cbp&4); } |
769 |
|
|
770 |
/* Get maximum range */ |
rd += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; |
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
|
|
x, y, 16, iWidth, iHeight, iFcode); |
|
771 |
|
|
772 |
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
if (rd >= data->iMinSAD[0]) return; |
773 |
|
|
774 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
/* chroma */ |
775 |
{ min_dx = EVEN(min_dx); |
xc = (xc >> 1) + roundtab_79[xc & 0x3]; |
776 |
max_dx = EVEN(max_dx); |
yc = (yc >> 1) + roundtab_79[yc & 0x3]; |
|
min_dy = EVEN(min_dy); |
|
|
max_dy = EVEN(max_dy); |
|
|
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
|
777 |
|
|
778 |
|
/* chroma U */ |
779 |
|
ptr = interpolate8x8_switch2(data->RefQ, data->RefP[4], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); |
780 |
|
transfer_8to16subro(in, data->CurU, ptr, data->iEdgedWidth/2); |
781 |
|
rd += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 4); |
782 |
|
if (rd >= data->iMinSAD[0]) return; |
783 |
|
|
784 |
bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
/* chroma V */ |
785 |
|
ptr = interpolate8x8_switch2(data->RefQ, data->RefP[5], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); |
786 |
|
transfer_8to16subro(in, data->CurV, ptr, data->iEdgedWidth/2); |
787 |
|
rd += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5); |
788 |
|
|
789 |
if ((x==0) && (y==0) ) |
rd += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; |
|
{ |
|
|
threshA = 512; |
|
|
threshB = 1024; |
|
790 |
|
|
791 |
|
if (rd < data->iMinSAD[0]) { |
792 |
|
data->iMinSAD[0] = rd; |
793 |
|
current[0].x = x; current[0].y = y; |
794 |
|
*data->dir = Direction; |
795 |
|
*data->cbp = cbp; |
796 |
} |
} |
|
else |
|
|
{ |
|
|
threshA = psad[0]; |
|
|
threshB = threshA+256; |
|
|
if (threshA< 512) threshA = 512; |
|
|
if (threshA>1024) threshA = 1024; |
|
|
if (threshB>1792) threshB = 1792; |
|
797 |
} |
} |
798 |
|
|
799 |
iFound=0; |
static void |
800 |
|
CheckCandidateRD8(const int x, const int y, const SearchData * const data, const int Direction) |
801 |
/* 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 |
|
|
*/ |
|
|
|
|
|
if ((bPredEq) && (MVequal(pmv[0],prevMB->mvs[0]) ) ) |
|
|
iFound=2; |
|
|
|
|
|
/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
|
|
Otherwise select large Diamond Search. |
|
|
*/ |
|
|
|
|
|
if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536) || (bPredEq) ) |
|
|
iDiamondSize=1; // halfpel! |
|
|
else |
|
|
iDiamondSize=2; // halfpel! |
|
|
|
|
|
if (!(MotionFlags & PMV_HALFPELDIAMOND16) ) |
|
|
iDiamondSize*=2; |
|
|
|
|
|
/* 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. |
|
|
*/ |
|
802 |
|
|
803 |
|
int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; |
804 |
|
int32_t rd; |
805 |
|
VECTOR * current; |
806 |
|
const uint8_t * ptr; |
807 |
|
int cbp = 0; |
808 |
|
|
809 |
// Prepare for main loop |
if ( (x > data->max_dx) || (x < data->min_dx) |
810 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
811 |
|
|
812 |
*currMV=pmv[0]; /* current best := prediction */ |
if (!data->qpel_precision) { |
813 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
ptr = GetReference(x, y, data); |
814 |
{ /* This should NOT be necessary! */ |
current = data->currentMV; |
815 |
currMV->x = EVEN(currMV->x); |
} else { /* x and y are in 1/4 precision */ |
816 |
currMV->y = EVEN(currMV->y); |
ptr = Interpolate8x8qpel(x, y, 0, 0, data); |
817 |
|
current = data->currentQMV; |
818 |
} |
} |
819 |
|
|
820 |
if (currMV->x > max_dx) |
transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth); |
821 |
{ |
rd = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5); |
822 |
currMV->x=max_dx; |
rd += BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
823 |
} |
|
824 |
if (currMV->x < min_dx) |
if (rd < data->iMinSAD[0]) { |
825 |
{ |
*data->cbp = cbp; |
826 |
currMV->x=min_dx; |
data->iMinSAD[0] = rd; |
827 |
} |
current[0].x = x; current[0].y = y; |
828 |
if (currMV->y > max_dy) |
*data->dir = Direction; |
|
{ |
|
|
currMV->y=max_dy; |
|
829 |
} |
} |
|
if (currMV->y < min_dy) |
|
|
{ |
|
|
currMV->y=min_dy; |
|
830 |
} |
} |
831 |
|
|
832 |
iMinSAD = sad16( cur, |
/* CHECK_CANDIATE FUNCTIONS END */ |
833 |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
|
834 |
iEdgedWidth, MV_MAX_ERROR); |
/* MAINSEARCH FUNCTIONS START */ |
|
iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; |
|
835 |
|
|
836 |
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,prevMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
static void |
837 |
|
AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection, CheckFunc * const CheckCandidate) |
838 |
{ |
{ |
839 |
|
|
840 |
if (MotionFlags & PMV_QUICKSTOP16) |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
|
goto PMVfast16_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfast16_Terminate_with_Refine; |
|
|
} |
|
841 |
|
|
842 |
/* |
unsigned int * const iDirection = data->dir; |
|
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. ******** WHAT'S THIS 'OFFSET' ??? *********** |
|
|
*/ |
|
843 |
|
|
844 |
// (0,0) is always possible |
for(;;) { /* forever */ |
845 |
|
*iDirection = 0; |
846 |
|
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
847 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
848 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
849 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
850 |
|
|
851 |
CHECK_MV16_ZERO; |
/* now we're doing diagonal checks near our candidate */ |
852 |
|
|
853 |
// previous frame MV is always possible |
if (*iDirection) { /* if anything found */ |
854 |
CHECK_MV16_CANDIDATE(prevMB->mvs[0].x,prevMB->mvs[0].y); |
bDirection = *iDirection; |
855 |
|
*iDirection = 0; |
856 |
|
x = data->currentMV->x; y = data->currentMV->y; |
857 |
|
if (bDirection & 3) { /* our candidate is left or right */ |
858 |
|
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
859 |
|
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
860 |
|
} else { /* what remains here is up or down */ |
861 |
|
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
862 |
|
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
863 |
|
} |
864 |
|
|
865 |
// left neighbour, if allowed |
if (*iDirection) { |
866 |
if (x != 0) |
bDirection += *iDirection; |
867 |
|
x = data->currentMV->x; y = data->currentMV->y; |
868 |
|
} |
869 |
|
} else { /* about to quit, eh? not so fast.... */ |
870 |
|
switch (bDirection) { |
871 |
|
case 2: |
872 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
873 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
874 |
|
break; |
875 |
|
case 1: |
876 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
877 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
878 |
|
break; |
879 |
|
case 2 + 4: |
880 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
881 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
882 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
883 |
|
break; |
884 |
|
case 4: |
885 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
886 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
887 |
|
break; |
888 |
|
case 8: |
889 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
890 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
891 |
|
break; |
892 |
|
case 1 + 4: |
893 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
894 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
895 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
896 |
|
break; |
897 |
|
case 2 + 8: |
898 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
899 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
900 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
901 |
|
break; |
902 |
|
case 1 + 8: |
903 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
904 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
905 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
906 |
|
break; |
907 |
|
default: /* 1+2+4+8 == we didn't find anything at all */ |
908 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
909 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
910 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
911 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
912 |
|
break; |
913 |
|
} |
914 |
|
if (!*iDirection) break; /* ok, the end. really */ |
915 |
|
bDirection = *iDirection; |
916 |
|
x = data->currentMV->x; y = data->currentMV->y; |
917 |
|
} |
918 |
|
} |
919 |
|
} |
920 |
|
|
921 |
|
static void |
922 |
|
SquareSearch(int x, int y, const SearchData * const data, int bDirection, CheckFunc * const CheckCandidate) |
923 |
{ |
{ |
924 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
unsigned int * const iDirection = data->dir; |
925 |
{ pmv[1].x = EVEN(pmv[1].x); |
|
926 |
pmv[1].y = EVEN(pmv[1].y); |
do { |
927 |
|
*iDirection = 0; |
928 |
|
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64); |
929 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128); |
930 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32); |
931 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128); |
932 |
|
if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64); |
933 |
|
if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128); |
934 |
|
if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128); |
935 |
|
if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128); |
936 |
|
|
937 |
|
bDirection = *iDirection; |
938 |
|
x = data->currentMV->x; y = data->currentMV->y; |
939 |
|
} while (*iDirection); |
940 |
} |
} |
941 |
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
|
942 |
|
static void |
943 |
|
DiamondSearch(int x, int y, const SearchData * const data, int bDirection, CheckFunc * const CheckCandidate) |
944 |
|
{ |
945 |
|
|
946 |
|
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
947 |
|
|
948 |
|
unsigned int * const iDirection = data->dir; |
949 |
|
|
950 |
|
do { |
951 |
|
*iDirection = 0; |
952 |
|
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
953 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
954 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
955 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
956 |
|
|
957 |
|
/* now we're doing diagonal checks near our candidate */ |
958 |
|
|
959 |
|
if (*iDirection) { /* checking if anything found */ |
960 |
|
bDirection = *iDirection; |
961 |
|
*iDirection = 0; |
962 |
|
x = data->currentMV->x; y = data->currentMV->y; |
963 |
|
if (bDirection & 3) { /* our candidate is left or right */ |
964 |
|
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
965 |
|
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
966 |
|
} else { /* what remains here is up or down */ |
967 |
|
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
968 |
|
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
969 |
|
} |
970 |
|
bDirection += *iDirection; |
971 |
|
x = data->currentMV->x; y = data->currentMV->y; |
972 |
|
} |
973 |
|
} |
974 |
|
while (*iDirection); |
975 |
} |
} |
976 |
|
|
977 |
// top neighbour, if allowed |
/* MAINSEARCH FUNCTIONS END */ |
978 |
if (y != 0) |
|
979 |
|
static void |
980 |
|
SubpelRefine_Fast(SearchData * data, CheckFunc * CheckCandidate) |
981 |
{ |
{ |
982 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
/* Do a half-pel or q-pel refinement */ |
983 |
{ pmv[2].x = EVEN(pmv[2].x); |
VECTOR centerMV; |
984 |
pmv[2].y = EVEN(pmv[2].y); |
VECTOR second_best; |
985 |
|
int best_sad = *data->iMinSAD; |
986 |
|
int xo, yo, xo2, yo2; |
987 |
|
int size = 2; |
988 |
|
CheckFunc *backupFunc = CheckCandidate; |
989 |
|
|
990 |
|
if(data->qpel_precision) |
991 |
|
size = 1; |
992 |
|
|
993 |
|
centerMV = *data->currentMV; |
994 |
|
*data->iMinSAD = 256 * 4096; |
995 |
|
|
996 |
|
CHECK_CANDIDATE(centerMV.x, centerMV.y - size, 0); |
997 |
|
CHECK_CANDIDATE(centerMV.x + size, centerMV.y - size, 0); |
998 |
|
CHECK_CANDIDATE(centerMV.x + size, centerMV.y, 0); |
999 |
|
CHECK_CANDIDATE(centerMV.x + size, centerMV.y + size, 0); |
1000 |
|
|
1001 |
|
CHECK_CANDIDATE(centerMV.x, centerMV.y + size, 0); |
1002 |
|
CHECK_CANDIDATE(centerMV.x - size, centerMV.y + size, 0); |
1003 |
|
CHECK_CANDIDATE(centerMV.x - size, centerMV.y, 0); |
1004 |
|
CHECK_CANDIDATE(centerMV.x - size, centerMV.y - size, 0); |
1005 |
|
|
1006 |
|
second_best = *data->currentMV; |
1007 |
|
|
1008 |
|
if(data->qpel_precision) { |
1009 |
|
second_best.x *= 2; second_best.y *= 2; |
1010 |
} |
} |
|
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
|
1011 |
|
|
1012 |
// top right neighbour, if allowed |
data->currentMV[0] = centerMV; |
1013 |
if ((uint32_t)x != (iWcount-1)) |
*data->iMinSAD = best_sad; |
1014 |
|
|
1015 |
|
centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV; |
1016 |
|
|
1017 |
|
xo = centerMV.x; |
1018 |
|
yo = centerMV.y; |
1019 |
|
xo2 = second_best.x; |
1020 |
|
yo2 = second_best.y; |
1021 |
|
|
1022 |
|
CheckCandidate = CheckCandidate16_subpel; |
1023 |
|
*data->iMinSAD2 = 256 * 4096; |
1024 |
|
|
1025 |
|
if (yo == yo2) |
1026 |
{ |
{ |
1027 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
CHECK_CANDIDATE((xo+xo2)>>1, yo, 0); |
1028 |
{ pmv[3].x = EVEN(pmv[3].x); |
CHECK_CANDIDATE(xo, yo-1, 0); |
1029 |
pmv[3].y = EVEN(pmv[3].y); |
CHECK_CANDIDATE(xo, yo+1, 0); |
1030 |
|
|
1031 |
|
if(best_sad <= *data->iMinSAD2) |
1032 |
|
goto ende; |
1033 |
|
|
1034 |
|
if(data->currentQMV[0].x == data->currentQMV2[0].x) { |
1035 |
|
CHECK_CANDIDATE((xo+xo2)>>1, yo-1, 0); |
1036 |
|
CHECK_CANDIDATE((xo+xo2)>>1, yo+1, 0); |
1037 |
|
goto ende; |
1038 |
} |
} |
1039 |
CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); |
else { |
1040 |
|
CHECK_CANDIDATE((xo+xo2)>>1, |
1041 |
|
(data->currentQMV[0].x == xo) ? data->currentQMV[0].y : data->currentQMV2[0].y, |
1042 |
|
0); |
1043 |
|
goto ende; |
1044 |
} |
} |
1045 |
} |
} |
1046 |
|
|
1047 |
/* Step 6: If MinSAD <= thresa goto Step 10. |
if (xo == xo2) |
1048 |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
{ |
1049 |
*/ |
CHECK_CANDIDATE(xo, (yo+yo2)>>1, 0); |
1050 |
|
CHECK_CANDIDATE(xo-1, yo, 0); |
1051 |
|
CHECK_CANDIDATE(xo+1, yo, 0); |
1052 |
|
|
1053 |
|
if(best_sad < *data->iMinSAD2) |
1054 |
|
goto ende; |
1055 |
|
|
1056 |
|
if(data->currentQMV[0].y == data->currentQMV2[0].y) { |
1057 |
|
CHECK_CANDIDATE(xo-1, (yo+yo2)>>1, 0); |
1058 |
|
CHECK_CANDIDATE(xo+1, (yo+yo2)>>1, 0); |
1059 |
|
goto ende; |
1060 |
|
} |
1061 |
|
else { |
1062 |
|
CHECK_CANDIDATE((data->currentQMV[0].y == yo) ? data->currentQMV[0].x : data->currentQMV2[0].x, (yo+yo2)>>1, 0); |
1063 |
|
goto ende; |
1064 |
|
} |
1065 |
|
} |
1066 |
|
|
1067 |
|
CHECK_CANDIDATE(xo, (yo+yo2)>>1, 0); |
1068 |
|
CHECK_CANDIDATE((xo+xo2)>>1, yo, 0); |
1069 |
|
|
1070 |
|
if(best_sad <= *data->iMinSAD2) |
1071 |
|
goto ende; |
1072 |
|
|
1073 |
|
CHECK_CANDIDATE((xo+xo2)>>1, (yo+yo2)>>1, 0); |
1074 |
|
|
1075 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,prevMB->mvs[0]) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
ende: |
1076 |
|
CheckCandidate = backupFunc; |
1077 |
|
} |
1078 |
|
|
1079 |
|
static void |
1080 |
|
SubpelRefine(const SearchData * const data, CheckFunc * const CheckCandidate) |
1081 |
{ |
{ |
1082 |
if (MotionFlags & PMV_QUICKSTOP16) |
/* Do a half-pel or q-pel refinement */ |
1083 |
goto PMVfast16_Terminate_without_Refine; |
const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV; |
1084 |
if (MotionFlags & PMV_EARLYSTOP16) |
|
1085 |
goto PMVfast16_Terminate_with_Refine; |
CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0); |
1086 |
|
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0); |
1087 |
|
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0); |
1088 |
|
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0); |
1089 |
|
CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0); |
1090 |
|
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0); |
1091 |
|
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0); |
1092 |
|
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0); |
1093 |
} |
} |
1094 |
|
|
1095 |
|
static __inline int |
1096 |
|
SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
1097 |
|
const int x, const int y, |
1098 |
|
const uint32_t stride, const uint32_t iQuant, int rrv) |
1099 |
|
|
1100 |
/************ (Diamond Search) **************/ |
{ |
1101 |
/* |
int offset = (x + y*stride)*8; |
1102 |
Step 7: Perform Diamond search, with either the small or large diamond. |
if(!rrv) { |
1103 |
If Found=2 only examine one Diamond pattern, and afterwards goto step 10 |
uint32_t sadC = sad8(current->u + offset, |
1104 |
Step 8: If small diamond, iterate small diamond search pattern until motion vector lies in the center of the diamond. |
reference->u + offset, stride); |
1105 |
If center then goto step 10. |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
1106 |
Step 9: If large diamond, iterate large diamond search pattern until motion vector lies in the center. |
sadC += sad8(current->v + offset, |
1107 |
Refine by using small diamond and goto step 10. |
reference->v + offset, stride); |
1108 |
*/ |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
1109 |
|
return 1; |
1110 |
|
|
1111 |
|
} else { |
1112 |
|
uint32_t sadC = sad16(current->u + 2*offset, |
1113 |
|
reference->u + 2*offset, stride, 256*4096); |
1114 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
1115 |
|
sadC += sad16(current->v + 2*offset, |
1116 |
|
reference->v + 2*offset, stride, 256*4096); |
1117 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
1118 |
|
return 1; |
1119 |
|
} |
1120 |
|
} |
1121 |
|
|
1122 |
|
static __inline void |
1123 |
|
ZeroMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
1124 |
|
{ |
1125 |
|
pMB->mode = MODE_INTER; |
1126 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV; |
1127 |
|
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV; |
1128 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
1129 |
|
} |
1130 |
|
|
1131 |
|
static __inline void |
1132 |
|
ModeDecision_Fast(SearchData * const Data, |
1133 |
|
MACROBLOCK * const pMB, |
1134 |
|
const MACROBLOCK * const pMBs, |
1135 |
|
const int x, const int y, |
1136 |
|
const MBParam * const pParam, |
1137 |
|
const uint32_t MotionFlags, |
1138 |
|
const uint32_t VopFlags, |
1139 |
|
const uint32_t VolFlags, |
1140 |
|
const IMAGE * const pCurrent, |
1141 |
|
const IMAGE * const pRef, |
1142 |
|
const IMAGE * const vGMC, |
1143 |
|
const int coding_type) |
1144 |
|
{ |
1145 |
|
int mode = MODE_INTER; |
1146 |
|
int mcsel = 0; |
1147 |
|
int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0); |
1148 |
|
const uint32_t iQuant = pMB->quant; |
1149 |
|
const int skip_possible = (coding_type == P_VOP) && (pMB->dquant == 0); |
1150 |
|
int sad; |
1151 |
|
int min_rd = -1, intra_rd, i, cbp, c[2] = {0, 0}; |
1152 |
|
VECTOR backup[5], *v; |
1153 |
|
int sad_backup[5]; |
1154 |
|
int InterBias = MV16_INTER_BIAS; |
1155 |
|
int thresh = 0; |
1156 |
|
int count = 0; |
1157 |
|
int top = 0, top_right = 0, left = 0; |
1158 |
|
|
1159 |
|
pMB->mcsel = 0; |
1160 |
|
|
1161 |
|
/* INTER <-> INTER4V decision */ |
1162 |
|
if ((Data->iMinSAD[0] + 75 < Data->iMinSAD[1] + |
1163 |
|
Data->iMinSAD[2] + Data->iMinSAD[3] + Data->iMinSAD[4])) { /* normal, fast, SAD-based mode decision */ |
1164 |
|
if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
1165 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) { |
1166 |
|
mode = MODE_INTER; |
1167 |
|
sad = Data->iMinSAD[0]; |
1168 |
|
} else { |
1169 |
|
mode = MODE_INTER4V; |
1170 |
|
sad = Data->iMinSAD[1] + Data->iMinSAD[2] + |
1171 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant; |
1172 |
|
Data->iMinSAD[0] = sad; |
1173 |
|
} |
1174 |
|
|
1175 |
|
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
1176 |
|
if (skip_possible && (pMB->sad16 < (int)iQuant * MAX_SAD00_FOR_SKIP)) |
1177 |
|
if ( (100*sad)/(pMB->sad16+1) > FINAL_SKIP_THRESH) |
1178 |
|
if (Data->chroma || SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant, Data->rrv)) { |
1179 |
|
mode = MODE_NOT_CODED; |
1180 |
|
goto early_out; |
1181 |
|
} |
1182 |
|
|
1183 |
|
/* mcsel */ |
1184 |
|
if (coding_type == S_VOP) { |
1185 |
|
|
1186 |
|
int32_t iSAD = sad16(Data->Cur, |
1187 |
|
vGMC->y + 16*y*Data->iEdgedWidth + 16*x, Data->iEdgedWidth, 65536); |
1188 |
|
|
1189 |
|
if (Data->chroma) { |
1190 |
|
iSAD += sad8(Data->CurU, vGMC->u + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); |
1191 |
|
iSAD += sad8(Data->CurV, vGMC->v + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); |
1192 |
|
} |
1193 |
|
|
1194 |
|
if (iSAD <= sad) { /* mode decision GMC */ |
1195 |
|
mode = MODE_INTER; |
1196 |
|
mcsel = 1; |
1197 |
|
sad = iSAD; |
1198 |
|
} |
1199 |
|
|
1200 |
|
} |
1201 |
|
} else { /* Rate-Distortion INTER<->INTER4V */ |
1202 |
|
Data->iQuant = iQuant; |
1203 |
|
Data->cbp = c; |
1204 |
|
v = Data->qpel ? Data->currentQMV : Data->currentMV; |
1205 |
|
|
1206 |
|
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
1207 |
|
if (skip_possible && (pMB->sad16 < (int)iQuant * MAX_SAD00_FOR_SKIP)) |
1208 |
|
if ( (100*Data->iMinSAD[0])/(pMB->sad16+1) > FINAL_SKIP_THRESH) |
1209 |
|
if (Data->chroma || SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant, Data->rrv)) { |
1210 |
|
mode = MODE_NOT_CODED; |
1211 |
|
goto early_out; |
1212 |
|
} |
1213 |
|
|
1214 |
|
for (i = 0; i < 5; i++) { |
1215 |
|
sad_backup[i] = Data->iMinSAD[i]; |
1216 |
|
Data->iMinSAD[i] = 256*4096; |
1217 |
|
backup[i] = v[i]; |
1218 |
|
} |
1219 |
|
|
1220 |
|
min_rd = findRDinter(Data, pMBs, x, y, pParam, MotionFlags); |
1221 |
|
cbp = *Data->cbp; |
1222 |
|
sad = sad_backup[0]; |
1223 |
|
|
1224 |
|
if (coding_type == S_VOP) { |
1225 |
|
int gmc_rd; |
1226 |
|
int32_t iSAD = sad16(Data->Cur, |
1227 |
|
vGMC->y + 16*y*Data->iEdgedWidth + 16*x, Data->iEdgedWidth, 65536); |
1228 |
|
|
1229 |
|
if (Data->chroma) { |
1230 |
|
iSAD += sad8(Data->CurU, vGMC->u + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); |
1231 |
|
iSAD += sad8(Data->CurV, vGMC->v + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); |
1232 |
|
} |
1233 |
|
|
1234 |
|
*Data->iMinSAD = min_rd += BITS_MULT*1; /* mcsel */ |
1235 |
|
gmc_rd = findRDgmc(Data, vGMC, x, y); |
1236 |
|
if (gmc_rd < min_rd) { |
1237 |
|
mcsel = 1; |
1238 |
|
*Data->iMinSAD = min_rd = gmc_rd; |
1239 |
|
mode = MODE_INTER; |
1240 |
|
cbp = *Data->cbp; |
1241 |
|
sad = iSAD; |
1242 |
|
} |
1243 |
|
} |
1244 |
|
|
1245 |
|
if (inter4v) { |
1246 |
|
int v4_rd; |
1247 |
|
v4_rd = findRDinter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup); |
1248 |
|
if (v4_rd < min_rd) { |
1249 |
|
Data->iMinSAD[0] = min_rd = v4_rd; |
1250 |
|
mode = MODE_INTER4V; |
1251 |
|
cbp = *Data->cbp; |
1252 |
|
sad = sad_backup[1] + sad_backup[2] + |
1253 |
|
sad_backup[3] + sad_backup[4] + IMV16X16 * (int32_t)iQuant; |
1254 |
|
} |
1255 |
|
} |
1256 |
|
} |
1257 |
|
|
1258 |
|
left = top = top_right = -1; |
1259 |
|
thresh = 0; |
1260 |
|
|
1261 |
|
if((x > 0) && (y > 0) && (x < (int32_t) pParam->mb_width)) { |
1262 |
|
left = (&pMBs[(x-1) + y * pParam->mb_width])->sad16; // left |
1263 |
|
top = (&pMBs[x + (y-1) * pParam->mb_width])->sad16; // top |
1264 |
|
top_right = (&pMBs[(x+1) + (y-1) * pParam->mb_width])->sad16; // top right |
1265 |
|
|
1266 |
|
if(((&pMBs[(x-1) + y * pParam->mb_width])->mode != MODE_INTRA) && |
1267 |
|
((&pMBs[x + (y-1) * pParam->mb_width])->mode != MODE_INTRA) && |
1268 |
|
((&pMBs[(x+1) + (y-1) * pParam->mb_width])->mode != MODE_INTRA)) { |
1269 |
|
thresh = MAX(MAX(top, left), top_right); |
1270 |
|
} |
1271 |
|
else |
1272 |
|
thresh = MIN(MIN(top, left), top_right); |
1273 |
|
} |
1274 |
|
|
1275 |
|
/* INTRA <-> INTER decision */ |
1276 |
|
if (sad < thresh) { /* normal, fast, SAD-based mode decision */ |
1277 |
|
/* intra decision */ |
1278 |
|
|
1279 |
|
if (iQuant > 8) InterBias += 100 * (iQuant - 8); /* to make high quants work */ |
1280 |
|
if (y != 0) |
1281 |
|
if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; |
1282 |
|
if (x != 0) |
1283 |
|
if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; |
1284 |
|
|
1285 |
|
if (Data->chroma) InterBias += 50; /* dev8(chroma) ??? <-- yes, we need dev8 (no big difference though) */ |
1286 |
|
if (Data->rrv) InterBias *= 4; |
1287 |
|
|
1288 |
|
if (InterBias < sad) { |
1289 |
|
int32_t deviation; |
1290 |
|
if (!Data->rrv) |
1291 |
|
deviation = dev16(Data->Cur, Data->iEdgedWidth); |
1292 |
|
else |
1293 |
|
deviation = dev16(Data->Cur, Data->iEdgedWidth) + /* dev32() */ |
1294 |
|
dev16(Data->Cur+16, Data->iEdgedWidth) + |
1295 |
|
dev16(Data->Cur + 16*Data->iEdgedWidth, Data->iEdgedWidth) + |
1296 |
|
dev16(Data->Cur+16+16*Data->iEdgedWidth, Data->iEdgedWidth); |
1297 |
|
|
1298 |
|
if (deviation < (sad - InterBias)) mode = MODE_INTRA; |
1299 |
|
} |
1300 |
|
|
1301 |
|
pMB->cbp = 63; |
1302 |
|
} else { /* Rate-Distortion INTRA<->INTER */ |
1303 |
|
if(min_rd < 0) { |
1304 |
|
Data->iQuant = iQuant; |
1305 |
|
Data->cbp = c; |
1306 |
|
v = Data->qpel ? Data->currentQMV : Data->currentMV; |
1307 |
|
|
1308 |
|
for (i = 0; i < 5; i++) { |
1309 |
|
Data->iMinSAD[i] = 256*4096; |
1310 |
|
backup[i] = v[i]; |
1311 |
|
} |
1312 |
|
|
1313 |
|
if(mode == MODE_INTER) { |
1314 |
|
min_rd = findRDinter(Data, pMBs, x, y, pParam, MotionFlags); |
1315 |
|
cbp = *Data->cbp; |
1316 |
|
|
1317 |
|
if (coding_type == S_VOP) { |
1318 |
|
int gmc_rd; |
1319 |
|
|
1320 |
|
*Data->iMinSAD = min_rd += BITS_MULT*1; /* mcsel */ |
1321 |
|
gmc_rd = findRDgmc(Data, vGMC, x, y); |
1322 |
|
if (gmc_rd < min_rd) { |
1323 |
|
mcsel = 1; |
1324 |
|
*Data->iMinSAD = min_rd = gmc_rd; |
1325 |
|
mode = MODE_INTER; |
1326 |
|
cbp = *Data->cbp; |
1327 |
|
} |
1328 |
|
} |
1329 |
|
} |
1330 |
|
|
1331 |
|
if(mode == MODE_INTER4V) { |
1332 |
|
int v4_rd; |
1333 |
|
v4_rd = findRDinter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup); |
1334 |
|
if (v4_rd < min_rd) { |
1335 |
|
Data->iMinSAD[0] = min_rd = v4_rd; |
1336 |
|
mode = MODE_INTER4V; |
1337 |
|
cbp = *Data->cbp; |
1338 |
|
} |
1339 |
|
} |
1340 |
|
} |
1341 |
|
|
1342 |
|
intra_rd = findRDintra(Data); |
1343 |
|
if (intra_rd < min_rd) { |
1344 |
|
*Data->iMinSAD = min_rd = intra_rd; |
1345 |
|
mode = MODE_INTRA; |
1346 |
|
} |
1347 |
|
|
1348 |
|
pMB->cbp = cbp; |
1349 |
|
} |
1350 |
|
|
1351 |
|
early_out: |
1352 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
1353 |
|
|
1354 |
|
if (Data->rrv) { |
1355 |
|
Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); |
1356 |
|
Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); |
1357 |
|
} |
1358 |
|
|
1359 |
|
if (mode == MODE_INTER && mcsel == 0) { |
1360 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
1361 |
|
|
1362 |
|
if(Data->qpel) { |
1363 |
|
pMB->qmvs[0] = pMB->qmvs[1] |
1364 |
|
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
1365 |
|
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; |
1366 |
|
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; |
1367 |
|
} else { |
1368 |
|
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
1369 |
|
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
1370 |
|
} |
1371 |
|
|
1372 |
|
} else if (mode == MODE_INTER ) { // but mcsel == 1 |
1373 |
|
|
1374 |
|
pMB->mcsel = 1; |
1375 |
|
if (Data->qpel) { |
1376 |
|
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = pMB->amv; |
1377 |
|
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = pMB->amv.x/2; |
1378 |
|
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = pMB->amv.y/2; |
1379 |
|
} else |
1380 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->amv; |
1381 |
|
|
1382 |
|
} else |
1383 |
|
if (mode == MODE_INTER4V) ; /* anything here? */ |
1384 |
|
else /* INTRA, NOT_CODED */ |
1385 |
|
ZeroMacroblockP(pMB, 0); |
1386 |
|
|
1387 |
|
pMB->mode = mode; |
1388 |
|
} |
1389 |
|
|
1390 |
|
static __inline void |
1391 |
|
ModeDecision(SearchData * const Data, |
1392 |
|
MACROBLOCK * const pMB, |
1393 |
|
const MACROBLOCK * const pMBs, |
1394 |
|
const int x, const int y, |
1395 |
|
const MBParam * const pParam, |
1396 |
|
const uint32_t MotionFlags, |
1397 |
|
const uint32_t VopFlags, |
1398 |
|
const uint32_t VolFlags, |
1399 |
|
const IMAGE * const pCurrent, |
1400 |
|
const IMAGE * const pRef, |
1401 |
|
const IMAGE * const vGMC, |
1402 |
|
const int coding_type) |
1403 |
|
{ |
1404 |
|
int mode = MODE_INTER; |
1405 |
|
int mcsel = 0; |
1406 |
|
int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0); |
1407 |
|
const uint32_t iQuant = pMB->quant; |
1408 |
|
|
1409 |
|
const int skip_possible = (coding_type == P_VOP) && (pMB->dquant == 0); |
1410 |
|
|
1411 |
|
pMB->mcsel = 0; |
1412 |
|
|
1413 |
|
if (!(VopFlags & XVID_VOP_MODEDECISION_RD)) { /* normal, fast, SAD-based mode decision */ |
1414 |
|
int sad; |
1415 |
|
int InterBias = MV16_INTER_BIAS; |
1416 |
|
if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
1417 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) { |
1418 |
|
mode = MODE_INTER; |
1419 |
|
sad = Data->iMinSAD[0]; |
1420 |
|
} else { |
1421 |
|
mode = MODE_INTER4V; |
1422 |
|
sad = Data->iMinSAD[1] + Data->iMinSAD[2] + |
1423 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant; |
1424 |
|
Data->iMinSAD[0] = sad; |
1425 |
|
} |
1426 |
|
|
1427 |
|
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
1428 |
|
if (skip_possible && (pMB->sad16 < (int)iQuant * MAX_SAD00_FOR_SKIP)) |
1429 |
|
if ( (100*sad)/(pMB->sad16+1) > FINAL_SKIP_THRESH) |
1430 |
|
if (Data->chroma || SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant, Data->rrv)) { |
1431 |
|
mode = MODE_NOT_CODED; |
1432 |
|
sad = 0; |
1433 |
|
} |
1434 |
|
|
1435 |
|
/* mcsel */ |
1436 |
|
if (coding_type == S_VOP) { |
1437 |
|
|
1438 |
|
int32_t iSAD = sad16(Data->Cur, |
1439 |
|
vGMC->y + 16*y*Data->iEdgedWidth + 16*x, Data->iEdgedWidth, 65536); |
1440 |
|
|
1441 |
|
if (Data->chroma) { |
1442 |
|
iSAD += sad8(Data->CurU, vGMC->u + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); |
1443 |
|
iSAD += sad8(Data->CurV, vGMC->v + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); |
1444 |
|
} |
1445 |
|
|
1446 |
|
if (iSAD <= sad) { /* mode decision GMC */ |
1447 |
|
mode = MODE_INTER; |
1448 |
|
mcsel = 1; |
1449 |
|
sad = iSAD; |
1450 |
|
} |
1451 |
|
|
1452 |
|
} |
1453 |
|
|
1454 |
|
/* intra decision */ |
1455 |
|
|
1456 |
|
if (iQuant > 8) InterBias += 100 * (iQuant - 8); /* to make high quants work */ |
1457 |
|
if (y != 0) |
1458 |
|
if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; |
1459 |
|
if (x != 0) |
1460 |
|
if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; |
1461 |
|
|
1462 |
|
if (Data->chroma) InterBias += 50; /* dev8(chroma) ??? <-- yes, we need dev8 (no big difference though) */ |
1463 |
|
if (Data->rrv) InterBias *= 4; |
1464 |
|
|
1465 |
|
if (InterBias < sad) { |
1466 |
|
int32_t deviation; |
1467 |
|
if (!Data->rrv) |
1468 |
|
deviation = dev16(Data->Cur, Data->iEdgedWidth); |
1469 |
|
else |
1470 |
|
deviation = dev16(Data->Cur, Data->iEdgedWidth) + /* dev32() */ |
1471 |
|
dev16(Data->Cur+16, Data->iEdgedWidth) + |
1472 |
|
dev16(Data->Cur + 16*Data->iEdgedWidth, Data->iEdgedWidth) + |
1473 |
|
dev16(Data->Cur+16+16*Data->iEdgedWidth, Data->iEdgedWidth); |
1474 |
|
|
1475 |
|
if (deviation < (sad - InterBias)) mode = MODE_INTRA; |
1476 |
|
} |
1477 |
|
|
1478 |
|
pMB->cbp = 63; |
1479 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
1480 |
|
|
1481 |
|
} else { /* Rate-Distortion */ |
1482 |
|
|
1483 |
|
int min_rd, intra_rd, i, cbp, c[2] = {0, 0}; |
1484 |
|
VECTOR backup[5], *v; |
1485 |
|
Data->iQuant = iQuant; |
1486 |
|
Data->cbp = c; |
1487 |
|
|
1488 |
|
v = Data->qpel ? Data->currentQMV : Data->currentMV; |
1489 |
|
for (i = 0; i < 5; i++) { |
1490 |
|
Data->iMinSAD[i] = 256*4096; |
1491 |
|
backup[i] = v[i]; |
1492 |
|
} |
1493 |
|
|
1494 |
|
min_rd = findRDinter(Data, pMBs, x, y, pParam, MotionFlags); |
1495 |
|
cbp = *Data->cbp; |
1496 |
|
|
1497 |
|
if (coding_type == S_VOP) { |
1498 |
|
int gmc_rd; |
1499 |
|
*Data->iMinSAD = min_rd += BITS_MULT*1; /* mcsel */ |
1500 |
|
gmc_rd = findRDgmc(Data, vGMC, x, y); |
1501 |
|
if (gmc_rd < min_rd) { |
1502 |
|
mcsel = 1; |
1503 |
|
*Data->iMinSAD = min_rd = gmc_rd; |
1504 |
|
mode = MODE_INTER; |
1505 |
|
cbp = *Data->cbp; |
1506 |
|
} |
1507 |
|
} |
1508 |
|
|
1509 |
|
if (inter4v) { |
1510 |
|
int v4_rd; |
1511 |
|
v4_rd = findRDinter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup); |
1512 |
|
if (v4_rd < min_rd) { |
1513 |
|
Data->iMinSAD[0] = min_rd = v4_rd; |
1514 |
|
mode = MODE_INTER4V; |
1515 |
|
cbp = *Data->cbp; |
1516 |
|
} |
1517 |
|
} |
1518 |
|
|
1519 |
|
intra_rd = findRDintra(Data); |
1520 |
|
if (intra_rd < min_rd) { |
1521 |
|
*Data->iMinSAD = min_rd = intra_rd; |
1522 |
|
mode = MODE_INTRA; |
1523 |
|
} |
1524 |
|
|
1525 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = 0; |
1526 |
|
pMB->cbp = cbp; |
1527 |
|
} |
1528 |
|
|
1529 |
|
if (Data->rrv) { |
1530 |
|
Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); |
1531 |
|
Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); |
1532 |
|
} |
1533 |
|
|
1534 |
|
if (mode == MODE_INTER && mcsel == 0) { |
1535 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
1536 |
|
|
1537 |
|
if(Data->qpel) { |
1538 |
|
pMB->qmvs[0] = pMB->qmvs[1] |
1539 |
|
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
1540 |
|
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; |
1541 |
|
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; |
1542 |
|
} else { |
1543 |
|
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
1544 |
|
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
1545 |
|
} |
1546 |
|
|
1547 |
|
} else if (mode == MODE_INTER ) { // but mcsel == 1 |
1548 |
|
|
1549 |
|
pMB->mcsel = 1; |
1550 |
|
if (Data->qpel) { |
1551 |
|
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = pMB->amv; |
1552 |
|
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = pMB->amv.x/2; |
1553 |
|
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = pMB->amv.y/2; |
1554 |
|
} else |
1555 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->amv; |
1556 |
|
|
1557 |
|
} else |
1558 |
|
if (mode == MODE_INTER4V) ; /* anything here? */ |
1559 |
|
else /* INTRA, NOT_CODED */ |
1560 |
|
ZeroMacroblockP(pMB, 0); |
1561 |
|
|
1562 |
|
pMB->mode = mode; |
1563 |
|
} |
1564 |
|
|
1565 |
|
bool |
1566 |
|
MotionEstimation(MBParam * const pParam, |
1567 |
|
FRAMEINFO * const current, |
1568 |
|
FRAMEINFO * const reference, |
1569 |
|
const IMAGE * const pRefH, |
1570 |
|
const IMAGE * const pRefV, |
1571 |
|
const IMAGE * const pRefHV, |
1572 |
|
const IMAGE * const pGMC, |
1573 |
|
const uint32_t iLimit) |
1574 |
|
{ |
1575 |
|
MACROBLOCK *const pMBs = current->mbs; |
1576 |
|
const IMAGE *const pCurrent = ¤t->image; |
1577 |
|
const IMAGE *const pRef = &reference->image; |
1578 |
|
|
1579 |
|
uint32_t mb_width = pParam->mb_width; |
1580 |
|
uint32_t mb_height = pParam->mb_height; |
1581 |
|
const uint32_t iEdgedWidth = pParam->edged_width; |
1582 |
|
const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->vop_flags, current->vol_flags); |
1583 |
|
int stat_thresh = 0; |
1584 |
|
|
1585 |
|
uint32_t x, y; |
1586 |
|
uint32_t iIntra = 0; |
1587 |
|
int32_t sad00; |
1588 |
|
int skip_thresh = INITIAL_SKIP_THRESH * \ |
1589 |
|
(current->vop_flags & XVID_VOP_REDUCED ? 4:1) * \ |
1590 |
|
(current->vop_flags & XVID_VOP_MODEDECISION_RD ? 2:1); |
1591 |
|
|
1592 |
|
/* some pre-initialized thingies for SearchP */ |
1593 |
|
int32_t temp[8]; uint32_t dir; |
1594 |
|
VECTOR currentMV[5]; |
1595 |
|
VECTOR currentQMV[5]; |
1596 |
|
VECTOR currentMV2[5]; |
1597 |
|
VECTOR currentQMV2[5]; |
1598 |
|
int32_t iMinSAD[5]; |
1599 |
|
int32_t iMinSAD2[5]; |
1600 |
|
DECLARE_ALIGNED_MATRIX(dct_space, 3, 64, int16_t, CACHE_LINE); |
1601 |
|
SearchData Data; |
1602 |
|
memset(&Data, 0, sizeof(SearchData)); |
1603 |
|
Data.iEdgedWidth = iEdgedWidth; |
1604 |
|
Data.currentMV = currentMV; |
1605 |
|
Data.currentQMV = currentQMV; |
1606 |
|
Data.currentMV2 = currentMV2; |
1607 |
|
Data.currentQMV2 = currentQMV2; |
1608 |
|
Data.iMinSAD = iMinSAD; |
1609 |
|
Data.iMinSAD2 = iMinSAD2; |
1610 |
|
Data.temp = temp; |
1611 |
|
Data.dir = &dir; |
1612 |
|
Data.iFcode = current->fcode; |
1613 |
|
Data.rounding = pParam->m_rounding_type; |
1614 |
|
Data.qpel = (current->vol_flags & XVID_VOL_QUARTERPEL ? 1:0); |
1615 |
|
Data.chroma = MotionFlags & XVID_ME_CHROMA_PVOP; |
1616 |
|
Data.rrv = (current->vop_flags & XVID_VOP_REDUCED) ? 1:0; |
1617 |
|
Data.dctSpace = dct_space; |
1618 |
|
Data.quant_type = !(pParam->vol_flags & XVID_VOL_MPEGQUANT); |
1619 |
|
|
1620 |
|
if ((current->vop_flags & XVID_VOP_REDUCED)) { |
1621 |
|
mb_width = (pParam->width + 31) / 32; |
1622 |
|
mb_height = (pParam->height + 31) / 32; |
1623 |
|
Data.qpel = 0; |
1624 |
|
} |
1625 |
|
|
1626 |
|
Data.RefQ = pRefV->u; /* a good place, also used in MC (for similar purpose) */ |
1627 |
|
if (sadInit) (*sadInit) (); |
1628 |
|
|
1629 |
|
for (y = 0; y < mb_height; y++) { |
1630 |
|
for (x = 0; x < mb_width; x++) { |
1631 |
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
1632 |
|
MACROBLOCK *prevMB = &reference->mbs[x + y * pParam->mb_width]; |
1633 |
|
|
1634 |
|
if (!Data.rrv) pMB->sad16 = |
1635 |
|
sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16, |
1636 |
|
pRef->y + (x + y * iEdgedWidth) * 16, |
1637 |
|
pParam->edged_width, pMB->sad8 ); |
1638 |
|
|
1639 |
|
else pMB->sad16 = |
1640 |
|
sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32, |
1641 |
|
pRef->y + (x + y * iEdgedWidth) * 32, |
1642 |
|
pParam->edged_width, pMB->sad8 ); |
1643 |
|
|
1644 |
|
if (Data.chroma) { |
1645 |
|
Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8, |
1646 |
|
pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2) |
1647 |
|
+ sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8, |
1648 |
|
pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); |
1649 |
|
pMB->sad16 += Data.temp[7]; |
1650 |
|
} |
1651 |
|
|
1652 |
|
sad00 = pMB->sad16; |
1653 |
|
|
1654 |
|
/* initial skip decision */ |
1655 |
|
/* no early skip for GMC (global vector = skip vector is unknown!) */ |
1656 |
|
if (current->coding_type != S_VOP) { /* no fast SKIP for S(GMC)-VOPs */ |
1657 |
|
if (pMB->dquant == 0 && sad00 < pMB->quant * skip_thresh) |
1658 |
|
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) { |
1659 |
|
ZeroMacroblockP(pMB, sad00); |
1660 |
|
pMB->mode = MODE_NOT_CODED; |
1661 |
|
continue; |
1662 |
|
} |
1663 |
|
} |
1664 |
|
|
1665 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
if(MotionFlags & XVID_ME_DETECT_STATIC_MOTION) { |
1666 |
|
if(x > 0 && y > 0 && x < pParam->mb_width) |
1667 |
|
if(MVequal((&pMBs[(x-1) + y * pParam->mb_width])->mvs[0], zeroMV) && |
1668 |
|
MVequal((&pMBs[x + (y-1) * pParam->mb_width])->mvs[0], zeroMV) && |
1669 |
|
MVequal((&pMBs[(x+1) + (y-1) * pParam->mb_width])->mvs[0], zeroMV) && |
1670 |
|
MVequal(prevMB->mvs[0], zeroMV)) { |
1671 |
|
stat_thresh = MAX((&pMBs[(x-1) + y * pParam->mb_width])->sad16, |
1672 |
|
MAX((&pMBs[x + (y-1) * pParam->mb_width])->sad16, |
1673 |
|
MAX((&pMBs[(x+1) + (y-1) * pParam->mb_width])->sad16, |
1674 |
|
prevMB->sad16))); |
1675 |
|
} |
1676 |
|
else |
1677 |
|
stat_thresh = MIN((&pMBs[(x-1) + y * pParam->mb_width])->sad16, |
1678 |
|
MIN((&pMBs[x + (y-1) * pParam->mb_width])->sad16, |
1679 |
|
MIN((&pMBs[(x+1) + (y-1) * pParam->mb_width])->sad16, |
1680 |
|
prevMB->sad16))); |
1681 |
|
} |
1682 |
|
|
1683 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
if ((current->vop_flags & XVID_VOP_CARTOON) && |
1684 |
iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
(sad00 < pMB->quant * 4 * skip_thresh) || (sad00 < stat_thresh)) { /* favorize (0,0) vector for cartoons */ |
1685 |
x, y, |
ZeroMacroblockP(pMB, sad00); |
1686 |
currMV->x, currMV->y, iMinSAD, &newMV, |
continue; |
1687 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
} |
1688 |
|
|
1689 |
|
SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
1690 |
|
y, MotionFlags, current->vop_flags, current->vol_flags, |
1691 |
|
&Data, pParam, pMBs, reference->mbs, pMB); |
1692 |
|
|
1693 |
|
if(current->vop_flags & XVID_VOP_FAST_MODEDECISION_RD) { |
1694 |
|
ModeDecision_Fast(&Data, pMB, pMBs, x, y, pParam, |
1695 |
|
MotionFlags, current->vop_flags, current->vol_flags, |
1696 |
|
pCurrent, pRef, pGMC, current->coding_type); |
1697 |
|
} |
1698 |
|
else { |
1699 |
|
ModeDecision(&Data, pMB, pMBs, x, y, pParam, |
1700 |
|
MotionFlags, current->vop_flags, current->vol_flags, |
1701 |
|
pCurrent, pRef, pGMC, current->coding_type); |
1702 |
|
} |
1703 |
|
|
1704 |
|
if (pMB->mode == MODE_INTRA) |
1705 |
|
if (++iIntra > iLimit) return 1; |
1706 |
|
} |
1707 |
|
} |
1708 |
|
|
1709 |
if (iSAD < iMinSAD) |
return 0; |
1710 |
|
} |
1711 |
|
|
1712 |
|
/* check if given vector is equal to any vector checked before */ |
1713 |
|
static __inline int |
1714 |
|
vector_repeats(const VECTOR * const pmv, const int i) |
1715 |
|
{ |
1716 |
|
unsigned int j; |
1717 |
|
for (j = 0; j < i; j++) |
1718 |
|
if (MVequal(pmv[i], pmv[j])) return 1; /* same vector has been checked already */ |
1719 |
|
return 0; |
1720 |
|
} |
1721 |
|
|
1722 |
|
/* make a binary mask that prevents diamonds/squares |
1723 |
|
from checking a vector which has been checked as a prediction */ |
1724 |
|
static __inline int |
1725 |
|
make_mask(const VECTOR * const pmv, const int i, const int current) |
1726 |
{ |
{ |
1727 |
*currMV = newMV; |
unsigned int mask = 255, j; |
1728 |
iMinSAD = iSAD; |
for (j = 0; j < i; j++) { |
1729 |
|
if (pmv[current].x == pmv[j].x) { |
1730 |
|
if (pmv[current].y == pmv[j].y + iDiamondSize) mask &= ~4; |
1731 |
|
else if (pmv[current].y == pmv[j].y - iDiamondSize) mask &= ~8; |
1732 |
|
} else |
1733 |
|
if (pmv[current].y == pmv[j].y) { |
1734 |
|
if (pmv[current].x == pmv[j].x + iDiamondSize) mask &= ~1; |
1735 |
|
else if (pmv[current].x == pmv[j].x - iDiamondSize) mask &= ~2; |
1736 |
|
} |
1737 |
|
} |
1738 |
|
return mask; |
1739 |
} |
} |
1740 |
|
|
1741 |
if (MotionFlags & PMV_EXTSEARCH16) |
static __inline void |
1742 |
|
PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, |
1743 |
|
int iHcount, const MACROBLOCK * const prevMB, int rrv) |
1744 |
{ |
{ |
1745 |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
/* this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself */ |
1746 |
|
if (rrv) { iWcount /= 2; iHcount /= 2; } |
1747 |
|
|
1748 |
|
if ( (y != 0) && (x < (iWcount-1)) ) { /* [5] top-right neighbour */ |
1749 |
|
pmv[5].x = EVEN(pmv[3].x); |
1750 |
|
pmv[5].y = EVEN(pmv[3].y); |
1751 |
|
} else pmv[5].x = pmv[5].y = 0; |
1752 |
|
|
1753 |
|
if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }/* pmv[3] is left neighbour */ |
1754 |
|
else pmv[3].x = pmv[3].y = 0; |
1755 |
|
|
1756 |
|
if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }/* [4] top neighbour */ |
1757 |
|
else pmv[4].x = pmv[4].y = 0; |
1758 |
|
|
1759 |
if (!(MVequal(pmv[0],backupMV)) ) |
/* [1] median prediction */ |
1760 |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); |
1761 |
x, y, |
|
1762 |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
pmv[0].x = pmv[0].y = 0; /* [0] is zero; not used in the loop (checked before) but needed here for make_mask */ |
1763 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
1764 |
|
pmv[2].x = EVEN(prevMB->mvs[0].x); /* [2] is last frame */ |
1765 |
|
pmv[2].y = EVEN(prevMB->mvs[0].y); |
1766 |
|
|
1767 |
|
if ((x < iWcount-1) && (y < iHcount-1)) { |
1768 |
|
pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); /* [6] right-down neighbour in last frame */ |
1769 |
|
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
1770 |
|
} else pmv[6].x = pmv[6].y = 0; |
1771 |
|
|
1772 |
|
if (rrv) { |
1773 |
|
int i; |
1774 |
|
for (i = 0; i < 7; i++) { |
1775 |
|
pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); |
1776 |
|
pmv[i].y = RRV_MV_SCALEUP(pmv[i].y); |
1777 |
|
} |
1778 |
|
} |
1779 |
|
} |
1780 |
|
|
1781 |
if (iSAD < iMinSAD) |
static void |
1782 |
|
SearchP(const IMAGE * const pRef, |
1783 |
|
const uint8_t * const pRefH, |
1784 |
|
const uint8_t * const pRefV, |
1785 |
|
const uint8_t * const pRefHV, |
1786 |
|
const IMAGE * const pCur, |
1787 |
|
const int x, |
1788 |
|
const int y, |
1789 |
|
const uint32_t MotionFlags, |
1790 |
|
const uint32_t VopFlags, |
1791 |
|
const uint32_t VolFlags, |
1792 |
|
SearchData * const Data, |
1793 |
|
const MBParam * const pParam, |
1794 |
|
const MACROBLOCK * const pMBs, |
1795 |
|
const MACROBLOCK * const prevMBs, |
1796 |
|
MACROBLOCK * const pMB) |
1797 |
{ |
{ |
1798 |
*currMV = newMV; |
|
1799 |
iMinSAD = iSAD; |
int i, threshA; |
1800 |
|
VECTOR pmv[7]; |
1801 |
|
int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0); |
1802 |
|
CheckFunc * CheckCandidate; |
1803 |
|
|
1804 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
1805 |
|
pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, Data->rrv); |
1806 |
|
|
1807 |
|
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); |
1808 |
|
|
1809 |
|
Data->temp[5] = Data->temp[6] = 0; /* chroma-sad cache */ |
1810 |
|
i = Data->rrv ? 2 : 1; |
1811 |
|
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i; |
1812 |
|
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1813 |
|
Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1814 |
|
|
1815 |
|
Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16*i; |
1816 |
|
Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16*i; |
1817 |
|
Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16*i; |
1818 |
|
Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16*i; |
1819 |
|
Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1820 |
|
Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1821 |
|
|
1822 |
|
Data->lambda16 = lambda_vec16[pMB->quant]; |
1823 |
|
Data->lambda8 = lambda_vec8[pMB->quant]; |
1824 |
|
Data->qpel_precision = 0; |
1825 |
|
*Data->dir = 0; |
1826 |
|
|
1827 |
|
memset(Data->currentMV, 0, 5*sizeof(VECTOR)); |
1828 |
|
|
1829 |
|
if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
1830 |
|
else Data->predMV = pmv[0]; |
1831 |
|
|
1832 |
|
i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0); |
1833 |
|
Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10); |
1834 |
|
Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10); |
1835 |
|
Data->iMinSAD[2] = pMB->sad8[1]; |
1836 |
|
Data->iMinSAD[3] = pMB->sad8[2]; |
1837 |
|
Data->iMinSAD[4] = pMB->sad8[3]; |
1838 |
|
|
1839 |
|
if ((!(VopFlags & XVID_VOP_MODEDECISION_RD)) && (x | y)) { |
1840 |
|
threshA = Data->temp[0]; /* that's where we keep this SAD atm */ |
1841 |
|
if (threshA < 512) threshA = 512; |
1842 |
|
else if (threshA > 1024) threshA = 1024; |
1843 |
|
} else |
1844 |
|
threshA = 512; |
1845 |
|
|
1846 |
|
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
1847 |
|
prevMBs + x + y * pParam->mb_width, Data->rrv); |
1848 |
|
|
1849 |
|
if (!Data->rrv) { |
1850 |
|
if (inter4v) CheckCandidate = CheckCandidate16; |
1851 |
|
else CheckCandidate = CheckCandidate16no4v; /* for extra speed */ |
1852 |
|
} else CheckCandidate = CheckCandidate32; |
1853 |
|
|
1854 |
|
/* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/ |
1855 |
|
|
1856 |
|
for (i = 1; i < 7; i++) |
1857 |
|
if (!vector_repeats(pmv, i)) { |
1858 |
|
CheckCandidate(pmv[i].x, pmv[i].y, Data, i); |
1859 |
|
if (Data->iMinSAD[0] <= threshA) { i++; break; } |
1860 |
|
} |
1861 |
|
|
1862 |
|
if ((Data->iMinSAD[0] <= threshA) || |
1863 |
|
(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
1864 |
|
(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) |
1865 |
|
inter4v = 0; |
1866 |
|
else { |
1867 |
|
|
1868 |
|
MainSearchFunc * MainSearchPtr; |
1869 |
|
int mask = make_mask(pmv, i, *Data->dir); // all vectors pmv[0..i-1] have been checked |
1870 |
|
|
1871 |
|
if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch; |
1872 |
|
else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1873 |
|
else MainSearchPtr = DiamondSearch; |
1874 |
|
|
1875 |
|
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate); |
1876 |
|
|
1877 |
|
/* extended search, diamond starting in 0,0 and in prediction. |
1878 |
|
note that this search is/might be done in halfpel positions, |
1879 |
|
which makes it more different than the diamond above */ |
1880 |
|
|
1881 |
|
if (MotionFlags & XVID_ME_EXTSEARCH16) { |
1882 |
|
int32_t bSAD; |
1883 |
|
VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; |
1884 |
|
if (Data->rrv) { |
1885 |
|
startMV.x = RRV_MV_SCALEUP(startMV.x); |
1886 |
|
startMV.y = RRV_MV_SCALEUP(startMV.y); |
1887 |
|
} |
1888 |
|
if (!(MVequal(startMV, backupMV))) { |
1889 |
|
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1890 |
|
|
1891 |
|
CheckCandidate(startMV.x, startMV.y, Data, 255); |
1892 |
|
MainSearchPtr(startMV.x, startMV.y, Data, 255, CheckCandidate); |
1893 |
|
if (bSAD < Data->iMinSAD[0]) { |
1894 |
|
Data->currentMV[0] = backupMV; |
1895 |
|
Data->iMinSAD[0] = bSAD; } |
1896 |
|
} |
1897 |
|
|
1898 |
|
backupMV = Data->currentMV[0]; |
1899 |
|
startMV.x = startMV.y = 1; |
1900 |
|
if (!(MVequal(startMV, backupMV))) { |
1901 |
|
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1902 |
|
|
1903 |
|
CheckCandidate(startMV.x, startMV.y, Data, 255); |
1904 |
|
MainSearchPtr(startMV.x, startMV.y, Data, 255, CheckCandidate); |
1905 |
|
if (bSAD < Data->iMinSAD[0]) { |
1906 |
|
Data->currentMV[0] = backupMV; |
1907 |
|
Data->iMinSAD[0] = bSAD; |
1908 |
|
} |
1909 |
|
} |
1910 |
|
} |
1911 |
} |
} |
1912 |
|
|
1913 |
|
if (MotionFlags & XVID_ME_HALFPELREFINE16) |
1914 |
|
SubpelRefine(Data, CheckCandidate); |
1915 |
|
|
1916 |
|
for(i = 0; i < 5; i++) { |
1917 |
|
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* initialize qpel vectors */ |
1918 |
|
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
1919 |
|
} |
1920 |
|
|
1921 |
|
if (Data->qpel) { |
1922 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
1923 |
|
pParam->width, pParam->height, Data->iFcode, 2, 0); |
1924 |
|
Data->qpel_precision = 1; |
1925 |
|
if (MotionFlags & XVID_ME_QUARTERPELREFINE16) |
1926 |
|
if(MotionFlags & XVID_ME_FASTREFINE16) |
1927 |
|
SubpelRefine_Fast(Data, CheckCandidate); |
1928 |
|
else |
1929 |
|
SubpelRefine(Data, CheckCandidate); |
1930 |
|
} |
1931 |
|
|
1932 |
|
if (Data->iMinSAD[0] < (int32_t)pMB->quant * 30) |
1933 |
|
inter4v = 0; |
1934 |
|
|
1935 |
|
if (inter4v) { |
1936 |
|
SearchData Data8; |
1937 |
|
memcpy(&Data8, Data, sizeof(SearchData)); /* quick copy of common data */ |
1938 |
|
|
1939 |
|
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
1940 |
|
Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); |
1941 |
|
Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); |
1942 |
|
Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); |
1943 |
|
|
1944 |
|
if ((Data->chroma) && (!(VopFlags & XVID_VOP_MODEDECISION_RD))) { |
1945 |
|
/* chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, it will not be used */ |
1946 |
|
int sumx = 0, sumy = 0; |
1947 |
|
|
1948 |
|
if (Data->qpel) |
1949 |
|
for (i = 1; i < 5; i++) { |
1950 |
|
sumx += Data->currentQMV[i].x/2; |
1951 |
|
sumy += Data->currentQMV[i].y/2; |
1952 |
|
} |
1953 |
|
else |
1954 |
|
for (i = 1; i < 5; i++) { |
1955 |
|
sumx += Data->currentMV[i].x; |
1956 |
|
sumy += Data->currentMV[i].y; |
1957 |
} |
} |
1958 |
|
|
1959 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf], |
1960 |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
(sumy >> 3) + roundtab_76[sumy & 0xf], Data); |
1961 |
x, y, |
} |
1962 |
0, 0, iMinSAD, &newMV, |
} else Data->iMinSAD[1] = 4096*256; |
1963 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
} |
1964 |
|
|
1965 |
if (iSAD < iMinSAD) |
static void |
1966 |
|
Search8(const SearchData * const OldData, |
1967 |
|
const int x, const int y, |
1968 |
|
const uint32_t MotionFlags, |
1969 |
|
const MBParam * const pParam, |
1970 |
|
MACROBLOCK * const pMB, |
1971 |
|
const MACROBLOCK * const pMBs, |
1972 |
|
const int block, |
1973 |
|
SearchData * const Data) |
1974 |
{ |
{ |
1975 |
*currMV = newMV; |
int i = 0; |
1976 |
iMinSAD = iSAD; |
CheckFunc * CheckCandidate; |
1977 |
|
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
1978 |
|
Data->currentMV = OldData->currentMV + 1 + block; |
1979 |
|
Data->currentQMV = OldData->currentQMV + 1 + block; |
1980 |
|
|
1981 |
|
if(Data->qpel) { |
1982 |
|
Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
1983 |
|
if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y, |
1984 |
|
Data->predMV, Data->iFcode, 0, 0); |
1985 |
|
} else { |
1986 |
|
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
1987 |
|
if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y, |
1988 |
|
Data->predMV, Data->iFcode, 0, Data->rrv); |
1989 |
} |
} |
1990 |
|
|
1991 |
|
*(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10; |
1992 |
|
|
1993 |
|
if (MotionFlags & (XVID_ME_EXTSEARCH8|XVID_ME_HALFPELREFINE8|XVID_ME_QUARTERPELREFINE8)) { |
1994 |
|
|
1995 |
|
if (Data->rrv) i = 16; else i = 8; |
1996 |
|
|
1997 |
|
Data->RefP[0] = OldData->RefP[0] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1998 |
|
Data->RefP[1] = OldData->RefP[1] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1999 |
|
Data->RefP[2] = OldData->RefP[2] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
2000 |
|
Data->RefP[3] = OldData->RefP[3] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
2001 |
|
|
2002 |
|
Data->Cur = OldData->Cur + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
2003 |
|
Data->qpel_precision = 0; |
2004 |
|
|
2005 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 3, |
2006 |
|
pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, Data->rrv); |
2007 |
|
|
2008 |
|
if (!Data->rrv) CheckCandidate = CheckCandidate8; |
2009 |
|
else CheckCandidate = CheckCandidate16no4v; |
2010 |
|
|
2011 |
|
if (MotionFlags & XVID_ME_EXTSEARCH8 && (!(MotionFlags & XVID_ME_EXTSEARCH_RD))) { |
2012 |
|
int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */ |
2013 |
|
|
2014 |
|
MainSearchFunc *MainSearchPtr; |
2015 |
|
if (MotionFlags & XVID_ME_USESQUARES8) MainSearchPtr = SquareSearch; |
2016 |
|
else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
2017 |
|
else MainSearchPtr = DiamondSearch; |
2018 |
|
|
2019 |
|
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255, CheckCandidate); |
2020 |
|
|
2021 |
|
if(*(Data->iMinSAD) < temp_sad) { |
2022 |
|
Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */ |
2023 |
|
Data->currentQMV->y = 2 * Data->currentMV->y; |
2024 |
} |
} |
2025 |
} |
} |
2026 |
|
|
2027 |
/* |
if (MotionFlags & XVID_ME_HALFPELREFINE8) { |
2028 |
Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */ |
|
*/ |
|
2029 |
|
|
2030 |
PMVfast16_Terminate_with_Refine: |
SubpelRefine(Data, CheckCandidate); /* perform halfpel refine of current best vector */ |
|
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); |
|
2031 |
|
|
2032 |
PMVfast16_Terminate_without_Refine: |
if(*(Data->iMinSAD) < temp_sad) { /* we have found a better match */ |
2033 |
currPMV->x = currMV->x - pmv[0].x; |
Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */ |
2034 |
currPMV->y = currMV->y - pmv[0].y; |
Data->currentQMV->y = 2 * Data->currentMV->y; |
2035 |
return iMinSAD; |
} |
2036 |
} |
} |
2037 |
|
|
2038 |
|
if (Data->qpel && MotionFlags & XVID_ME_QUARTERPELREFINE8) { |
2039 |
|
Data->qpel_precision = 1; |
2040 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 3, |
2041 |
|
pParam->width, pParam->height, Data->iFcode, 2, 0); |
2042 |
|
SubpelRefine(Data, CheckCandidate); |
2043 |
|
} |
2044 |
|
} |
2045 |
|
|
2046 |
|
if (Data->rrv) { |
2047 |
|
Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x); |
2048 |
|
Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y); |
2049 |
|
} |
2050 |
|
|
2051 |
|
if(Data->qpel) { |
2052 |
|
pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x; |
2053 |
|
pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y; |
2054 |
|
pMB->qmvs[block] = *Data->currentQMV; |
2055 |
|
} else { |
2056 |
|
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
2057 |
|
pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; |
2058 |
|
} |
2059 |
|
|
2060 |
|
pMB->mvs[block] = *Data->currentMV; |
2061 |
|
pMB->sad8[block] = 4 * *Data->iMinSAD; |
2062 |
|
} |
2063 |
|
|
2064 |
|
/* motion estimation for B-frames */ |
2065 |
|
|
2066 |
int32_t Diamond8_MainSearch( |
static __inline VECTOR |
2067 |
const uint8_t * const pRef, |
ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) |
2068 |
|
{ |
2069 |
|
/* the stupidiest function ever */ |
2070 |
|
return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]); |
2071 |
|
} |
2072 |
|
|
2073 |
|
static void __inline |
2074 |
|
PreparePredictionsBF(VECTOR * const pmv, const int x, const int y, |
2075 |
|
const uint32_t iWcount, |
2076 |
|
const MACROBLOCK * const pMB, |
2077 |
|
const uint32_t mode_curr) |
2078 |
|
{ |
2079 |
|
|
2080 |
|
/* [0] is prediction */ |
2081 |
|
pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); |
2082 |
|
|
2083 |
|
pmv[1].x = pmv[1].y = 0; /* [1] is zero */ |
2084 |
|
|
2085 |
|
pmv[2] = ChoosePred(pMB, mode_curr); |
2086 |
|
pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); |
2087 |
|
|
2088 |
|
if ((y != 0)&&(x != (int)(iWcount+1))) { /* [3] top-right neighbour */ |
2089 |
|
pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); |
2090 |
|
pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); |
2091 |
|
} else pmv[3].x = pmv[3].y = 0; |
2092 |
|
|
2093 |
|
if (y != 0) { |
2094 |
|
pmv[4] = ChoosePred(pMB-iWcount, mode_curr); |
2095 |
|
pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y); |
2096 |
|
} else pmv[4].x = pmv[4].y = 0; |
2097 |
|
|
2098 |
|
if (x != 0) { |
2099 |
|
pmv[5] = ChoosePred(pMB-1, mode_curr); |
2100 |
|
pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
2101 |
|
} else pmv[5].x = pmv[5].y = 0; |
2102 |
|
|
2103 |
|
if (x != 0 && y != 0) { |
2104 |
|
pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); |
2105 |
|
pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y); |
2106 |
|
} else pmv[6].x = pmv[6].y = 0; |
2107 |
|
} |
2108 |
|
|
2109 |
|
|
2110 |
|
/* search backward or forward */ |
2111 |
|
static void |
2112 |
|
SearchBF( const IMAGE * const pRef, |
2113 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
2114 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
2115 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
2116 |
const uint8_t * const cur, |
const IMAGE * const pCur, |
2117 |
|
const int x, const int y, |
2118 |
|
const uint32_t MotionFlags, |
2119 |
|
const uint32_t iFcode, |
2120 |
|
const MBParam * const pParam, |
2121 |
|
MACROBLOCK * const pMB, |
2122 |
|
const VECTOR * const predMV, |
2123 |
|
int32_t * const best_sad, |
2124 |
|
const int32_t mode_current, |
2125 |
|
SearchData * const Data) |
2126 |
|
{ |
2127 |
|
|
2128 |
|
int i; |
2129 |
|
VECTOR pmv[7]; |
2130 |
|
MainSearchFunc *MainSearchPtr; |
2131 |
|
*Data->iMinSAD = MV_MAX_ERROR; |
2132 |
|
Data->iFcode = iFcode; |
2133 |
|
Data->qpel_precision = 0; |
2134 |
|
Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; /* reset chroma-sad cache */ |
2135 |
|
|
2136 |
|
Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16; |
2137 |
|
Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16; |
2138 |
|
Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16; |
2139 |
|
Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16; |
2140 |
|
Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
2141 |
|
Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
2142 |
|
|
2143 |
|
Data->predMV = *predMV; |
2144 |
|
|
2145 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
2146 |
|
pParam->width, pParam->height, iFcode - Data->qpel, 1, 0); |
2147 |
|
|
2148 |
|
pmv[0] = Data->predMV; |
2149 |
|
if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } |
2150 |
|
|
2151 |
|
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
2152 |
|
|
2153 |
|
Data->currentMV->x = Data->currentMV->y = 0; |
2154 |
|
|
2155 |
|
/* main loop. checking all predictions */ |
2156 |
|
for (i = 0; i < 7; i++) |
2157 |
|
if (!vector_repeats(pmv, i) ) |
2158 |
|
CheckCandidate16no4v(pmv[i].x, pmv[i].y, Data, i); |
2159 |
|
|
2160 |
|
if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch; |
2161 |
|
else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
2162 |
|
else MainSearchPtr = DiamondSearch; |
2163 |
|
|
2164 |
|
if (*Data->iMinSAD > 512) { |
2165 |
|
unsigned int mask = make_mask(pmv, 7, *Data->dir); |
2166 |
|
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate16no4v); |
2167 |
|
} |
2168 |
|
|
2169 |
|
SubpelRefine(Data, CheckCandidate16no4v); |
2170 |
|
|
2171 |
|
if (Data->qpel && *Data->iMinSAD < *best_sad + 300) { |
2172 |
|
Data->currentQMV->x = 2*Data->currentMV->x; |
2173 |
|
Data->currentQMV->y = 2*Data->currentMV->y; |
2174 |
|
Data->qpel_precision = 1; |
2175 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
2176 |
|
pParam->width, pParam->height, iFcode, 2, 0); |
2177 |
|
SubpelRefine(Data, CheckCandidate16no4v); |
2178 |
|
} |
2179 |
|
|
2180 |
|
/* three bits are needed to code backward mode. four for forward */ |
2181 |
|
|
2182 |
|
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
2183 |
|
else *Data->iMinSAD += 3 * Data->lambda16; |
2184 |
|
|
2185 |
|
if (*Data->iMinSAD < *best_sad) { |
2186 |
|
*best_sad = *Data->iMinSAD; |
2187 |
|
pMB->mode = mode_current; |
2188 |
|
if (Data->qpel) { |
2189 |
|
pMB->pmvs[0].x = Data->currentQMV->x - predMV->x; |
2190 |
|
pMB->pmvs[0].y = Data->currentQMV->y - predMV->y; |
2191 |
|
if (mode_current == MODE_FORWARD) |
2192 |
|
pMB->qmvs[0] = *Data->currentQMV; |
2193 |
|
else |
2194 |
|
pMB->b_qmvs[0] = *Data->currentQMV; |
2195 |
|
} else { |
2196 |
|
pMB->pmvs[0].x = Data->currentMV->x - predMV->x; |
2197 |
|
pMB->pmvs[0].y = Data->currentMV->y - predMV->y; |
2198 |
|
} |
2199 |
|
if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; |
2200 |
|
else pMB->b_mvs[0] = *Data->currentMV; |
2201 |
|
} |
2202 |
|
|
2203 |
|
if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; |
2204 |
|
else *(Data->currentMV+1) = *Data->currentMV; /* we store currmv for interpolate search */ |
2205 |
|
} |
2206 |
|
|
2207 |
|
static void |
2208 |
|
SkipDecisionB(const IMAGE * const pCur, |
2209 |
|
const IMAGE * const f_Ref, |
2210 |
|
const IMAGE * const b_Ref, |
2211 |
|
MACROBLOCK * const pMB, |
2212 |
|
const uint32_t x, const uint32_t y, |
2213 |
|
const SearchData * const Data) |
2214 |
|
{ |
2215 |
|
int dx = 0, dy = 0, b_dx = 0, b_dy = 0; |
2216 |
|
int32_t sum; |
2217 |
|
int k; |
2218 |
|
const uint32_t stride = Data->iEdgedWidth/2; |
2219 |
|
/* this is not full chroma compensation, only it's fullpel approximation. should work though */ |
2220 |
|
|
2221 |
|
for (k = 0; k < 4; k++) { |
2222 |
|
dy += Data->directmvF[k].y >> Data->qpel; |
2223 |
|
dx += Data->directmvF[k].x >> Data->qpel; |
2224 |
|
b_dy += Data->directmvB[k].y >> Data->qpel; |
2225 |
|
b_dx += Data->directmvB[k].x >> Data->qpel; |
2226 |
|
} |
2227 |
|
|
2228 |
|
dy = (dy >> 3) + roundtab_76[dy & 0xf]; |
2229 |
|
dx = (dx >> 3) + roundtab_76[dx & 0xf]; |
2230 |
|
b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf]; |
2231 |
|
b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf]; |
2232 |
|
|
2233 |
|
sum = sad8bi(pCur->u + 8 * x + 8 * y * stride, |
2234 |
|
f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2, |
2235 |
|
b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, |
2236 |
|
stride); |
2237 |
|
|
2238 |
|
if (sum >= MAX_CHROMA_SAD_FOR_SKIP * Data->iQuant) return; /* no skip */ |
2239 |
|
|
2240 |
|
sum += sad8bi(pCur->v + 8*x + 8 * y * stride, |
2241 |
|
f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2, |
2242 |
|
b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, |
2243 |
|
stride); |
2244 |
|
|
2245 |
|
if (sum < MAX_CHROMA_SAD_FOR_SKIP * Data->iQuant) { |
2246 |
|
pMB->mode = MODE_DIRECT_NONE_MV; /* skipped */ |
2247 |
|
for (k = 0; k < 4; k++) { |
2248 |
|
pMB->qmvs[k] = pMB->mvs[k]; |
2249 |
|
pMB->b_qmvs[k] = pMB->b_mvs[k]; |
2250 |
|
} |
2251 |
|
} |
2252 |
|
} |
2253 |
|
|
2254 |
|
static __inline uint32_t |
2255 |
|
SearchDirect(const IMAGE * const f_Ref, |
2256 |
|
const uint8_t * const f_RefH, |
2257 |
|
const uint8_t * const f_RefV, |
2258 |
|
const uint8_t * const f_RefHV, |
2259 |
|
const IMAGE * const b_Ref, |
2260 |
|
const uint8_t * const b_RefH, |
2261 |
|
const uint8_t * const b_RefV, |
2262 |
|
const uint8_t * const b_RefHV, |
2263 |
|
const IMAGE * const pCur, |
2264 |
|
const int x, const int y, |
2265 |
|
const uint32_t MotionFlags, |
2266 |
|
const int32_t TRB, const int32_t TRD, |
2267 |
|
const MBParam * const pParam, |
2268 |
|
MACROBLOCK * const pMB, |
2269 |
|
const MACROBLOCK * const b_mb, |
2270 |
|
int32_t * const best_sad, |
2271 |
|
SearchData * const Data) |
2272 |
|
|
2273 |
|
{ |
2274 |
|
int32_t skip_sad; |
2275 |
|
int k = (x + Data->iEdgedWidth*y) * 16; |
2276 |
|
MainSearchFunc *MainSearchPtr; |
2277 |
|
CheckFunc * CheckCandidate; |
2278 |
|
|
2279 |
|
*Data->iMinSAD = 256*4096; |
2280 |
|
Data->RefP[0] = f_Ref->y + k; |
2281 |
|
Data->RefP[2] = f_RefH + k; |
2282 |
|
Data->RefP[1] = f_RefV + k; |
2283 |
|
Data->RefP[3] = f_RefHV + k; |
2284 |
|
Data->b_RefP[0] = b_Ref->y + k; |
2285 |
|
Data->b_RefP[2] = b_RefH + k; |
2286 |
|
Data->b_RefP[1] = b_RefV + k; |
2287 |
|
Data->b_RefP[3] = b_RefHV + k; |
2288 |
|
Data->RefP[4] = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; |
2289 |
|
Data->RefP[5] = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; |
2290 |
|
Data->b_RefP[4] = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; |
2291 |
|
Data->b_RefP[5] = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; |
2292 |
|
|
2293 |
|
k = Data->qpel ? 4 : 2; |
2294 |
|
Data->max_dx = k * (pParam->width - x * 16); |
2295 |
|
Data->max_dy = k * (pParam->height - y * 16); |
2296 |
|
Data->min_dx = -k * (16 + x * 16); |
2297 |
|
Data->min_dy = -k * (16 + y * 16); |
2298 |
|
|
2299 |
|
Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs; |
2300 |
|
Data->qpel_precision = 0; |
2301 |
|
|
2302 |
|
for (k = 0; k < 4; k++) { |
2303 |
|
pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); |
2304 |
|
pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD; |
2305 |
|
pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); |
2306 |
|
pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; |
2307 |
|
|
2308 |
|
if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx) |
2309 |
|
| (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) { |
2310 |
|
|
2311 |
|
*best_sad = 256*4096; /* in that case, we won't use direct mode */ |
2312 |
|
pMB->mode = MODE_DIRECT; /* just to make sure it doesn't say "MODE_DIRECT_NONE_MV" */ |
2313 |
|
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
2314 |
|
return 256*4096; |
2315 |
|
} |
2316 |
|
if (b_mb->mode != MODE_INTER4V) { |
2317 |
|
pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; |
2318 |
|
pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0]; |
2319 |
|
Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0]; |
2320 |
|
Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0]; |
2321 |
|
break; |
2322 |
|
} |
2323 |
|
} |
2324 |
|
|
2325 |
|
CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v; |
2326 |
|
|
2327 |
|
CheckCandidate(0, 0, Data, 255); |
2328 |
|
|
2329 |
|
/* initial (fast) skip decision */ |
2330 |
|
if (*Data->iMinSAD < Data->iQuant * INITIAL_SKIP_THRESH * (Data->chroma?3:2)) { |
2331 |
|
/* possible skip */ |
2332 |
|
if (Data->chroma) { |
2333 |
|
pMB->mode = MODE_DIRECT_NONE_MV; |
2334 |
|
return *Data->iMinSAD; /* skip. */ |
2335 |
|
} else { |
2336 |
|
SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data); |
2337 |
|
if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; /* skip. */ |
2338 |
|
} |
2339 |
|
} |
2340 |
|
|
2341 |
|
*Data->iMinSAD += Data->lambda16; |
2342 |
|
skip_sad = *Data->iMinSAD; |
2343 |
|
|
2344 |
|
/* |
2345 |
|
* DIRECT MODE DELTA VECTOR SEARCH. |
2346 |
|
* This has to be made more effective, but at the moment I'm happy it's running at all |
2347 |
|
*/ |
2348 |
|
|
2349 |
|
if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch; |
2350 |
|
else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
2351 |
|
else MainSearchPtr = DiamondSearch; |
2352 |
|
|
2353 |
|
MainSearchPtr(0, 0, Data, 255, CheckCandidate); |
2354 |
|
|
2355 |
|
SubpelRefine(Data, CheckCandidate); |
2356 |
|
|
2357 |
|
*best_sad = *Data->iMinSAD; |
2358 |
|
|
2359 |
|
if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT; |
2360 |
|
else pMB->mode = MODE_DIRECT_NO4V; /* for faster compensation */ |
2361 |
|
|
2362 |
|
pMB->pmvs[3] = *Data->currentMV; |
2363 |
|
|
2364 |
|
for (k = 0; k < 4; k++) { |
2365 |
|
pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x; |
2366 |
|
pMB->b_mvs[k].x = ( (Data->currentMV->x == 0) |
2367 |
|
? Data->directmvB[k].x |
2368 |
|
:pMB->mvs[k].x - Data->referencemv[k].x); |
2369 |
|
pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y); |
2370 |
|
pMB->b_mvs[k].y = ((Data->currentMV->y == 0) |
2371 |
|
? Data->directmvB[k].y |
2372 |
|
: pMB->mvs[k].y - Data->referencemv[k].y); |
2373 |
|
if (Data->qpel) { |
2374 |
|
pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2; |
2375 |
|
pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2; |
2376 |
|
pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2; |
2377 |
|
pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2; |
2378 |
|
} |
2379 |
|
|
2380 |
|
if (b_mb->mode != MODE_INTER4V) { |
2381 |
|
pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; |
2382 |
|
pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; |
2383 |
|
pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0]; |
2384 |
|
pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0]; |
2385 |
|
break; |
2386 |
|
} |
2387 |
|
} |
2388 |
|
return skip_sad; |
2389 |
|
} |
2390 |
|
|
2391 |
|
static void |
2392 |
|
SearchInterpolate(const IMAGE * const f_Ref, |
2393 |
|
const uint8_t * const f_RefH, |
2394 |
|
const uint8_t * const f_RefV, |
2395 |
|
const uint8_t * const f_RefHV, |
2396 |
|
const IMAGE * const b_Ref, |
2397 |
|
const uint8_t * const b_RefH, |
2398 |
|
const uint8_t * const b_RefV, |
2399 |
|
const uint8_t * const b_RefHV, |
2400 |
|
const IMAGE * const pCur, |
2401 |
const int x, const int y, |
const int x, const int y, |
2402 |
int32_t startx, int32_t starty, |
const uint32_t fcode, |
2403 |
int32_t iMinSAD, |
const uint32_t bcode, |
2404 |
VECTOR * const currMV, |
const uint32_t MotionFlags, |
2405 |
const VECTOR * const pmv, |
const MBParam * const pParam, |
2406 |
const int32_t min_dx, const int32_t max_dx, |
const VECTOR * const f_predMV, |
2407 |
const int32_t min_dy, const int32_t max_dy, |
const VECTOR * const b_predMV, |
2408 |
const int32_t iEdgedWidth, |
MACROBLOCK * const pMB, |
2409 |
const int32_t iDiamondSize, |
int32_t * const best_sad, |
2410 |
const int32_t iFcode, |
SearchData * const fData) |
2411 |
const int32_t iQuant, |
|
2412 |
int iFound) |
{ |
2413 |
{ |
|
2414 |
/* Do a diamond search around given starting point, return SAD of best */ |
int i, j; |
2415 |
|
SearchData bData; |
2416 |
int32_t iDirection=0; |
|
2417 |
int32_t iSAD; |
fData->qpel_precision = 0; |
2418 |
VECTOR backupMV; |
memcpy(&bData, fData, sizeof(SearchData)); /* quick copy of common data */ |
2419 |
backupMV.x = startx; |
*fData->iMinSAD = 4096*256; |
2420 |
backupMV.y = starty; |
bData.currentMV++; bData.currentQMV++; |
2421 |
|
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
2422 |
/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ |
|
2423 |
|
i = (x + y * fData->iEdgedWidth) * 16; |
2424 |
CHECK_MV8_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
|
2425 |
CHECK_MV8_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
bData.b_RefP[0] = fData->RefP[0] = f_Ref->y + i; |
2426 |
CHECK_MV8_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
bData.b_RefP[2] = fData->RefP[2] = f_RefH + i; |
2427 |
CHECK_MV8_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
bData.b_RefP[1] = fData->RefP[1] = f_RefV + i; |
2428 |
|
bData.b_RefP[3] = fData->RefP[3] = f_RefHV + i; |
2429 |
if (iDirection) |
bData.RefP[0] = fData->b_RefP[0] = b_Ref->y + i; |
2430 |
while (!iFound) |
bData.RefP[2] = fData->b_RefP[2] = b_RefH + i; |
2431 |
{ |
bData.RefP[1] = fData->b_RefP[1] = b_RefV + i; |
2432 |
iFound = 1; |
bData.RefP[3] = fData->b_RefP[3] = b_RefHV + i; |
2433 |
backupMV=*currMV; // since iDirection!=0, this is well defined! |
bData.b_RefP[4] = fData->RefP[4] = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; |
2434 |
|
bData.b_RefP[5] = fData->RefP[5] = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; |
2435 |
if ( iDirection != 2) |
bData.RefP[4] = fData->b_RefP[4] = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; |
2436 |
CHECK_MV8_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
bData.RefP[5] = fData->b_RefP[5] = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; |
2437 |
if ( iDirection != 1) |
bData.dir = fData->dir; |
2438 |
CHECK_MV8_CANDIDATE_FOUND(backupMV.x+iDiamondSize,backupMV.y,2); |
|
2439 |
if ( iDirection != 4) |
bData.bpredMV = fData->predMV = *f_predMV; |
2440 |
CHECK_MV8_CANDIDATE_FOUND(backupMV.x,backupMV.y-iDiamondSize,3); |
fData->bpredMV = bData.predMV = *b_predMV; |
2441 |
if ( iDirection != 3) |
fData->currentMV[0] = fData->currentMV[2]; |
2442 |
CHECK_MV8_CANDIDATE_FOUND(backupMV.x,backupMV.y+iDiamondSize,4); |
|
2443 |
|
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 4, pParam->width, pParam->height, fcode - fData->qpel, 1, 0); |
2444 |
|
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 4, pParam->width, pParam->height, bcode - fData->qpel, 1, 0); |
2445 |
|
|
2446 |
|
if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; |
2447 |
|
if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx; |
2448 |
|
if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy; |
2449 |
|
if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy; |
2450 |
|
|
2451 |
|
if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx; |
2452 |
|
if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx; |
2453 |
|
if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy; |
2454 |
|
if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy; |
2455 |
|
|
2456 |
|
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, fData, 255); |
2457 |
|
|
2458 |
|
/* diamond */ |
2459 |
|
do { |
2460 |
|
*fData->dir = 255; |
2461 |
|
/* forward MV moves */ |
2462 |
|
i = fData->currentMV[0].x; j = fData->currentMV[0].y; |
2463 |
|
|
2464 |
|
CheckCandidateInt(i + 1, j, fData, 0); |
2465 |
|
CheckCandidateInt(i, j + 1, fData, 0); |
2466 |
|
CheckCandidateInt(i - 1, j, fData, 0); |
2467 |
|
CheckCandidateInt(i, j - 1, fData, 0); |
2468 |
|
|
2469 |
|
/* backward MV moves */ |
2470 |
|
i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
2471 |
|
fData->currentMV[2] = fData->currentMV[0]; |
2472 |
|
CheckCandidateInt(i + 1, j, &bData, 0); |
2473 |
|
CheckCandidateInt(i, j + 1, &bData, 0); |
2474 |
|
CheckCandidateInt(i - 1, j, &bData, 0); |
2475 |
|
CheckCandidateInt(i, j - 1, &bData, 0); |
2476 |
|
|
2477 |
|
} while (!(*fData->dir)); |
2478 |
|
|
2479 |
|
/* qpel refinement */ |
2480 |
|
if (fData->qpel) { |
2481 |
|
if (*fData->iMinSAD > *best_sad + 500) return; |
2482 |
|
fData->qpel_precision = bData.qpel_precision = 1; |
2483 |
|
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 4, pParam->width, pParam->height, fcode, 2, 0); |
2484 |
|
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 4, pParam->width, pParam->height, bcode, 2, 0); |
2485 |
|
fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x; |
2486 |
|
fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y; |
2487 |
|
fData->currentQMV[1].x = 2 * fData->currentMV[1].x; |
2488 |
|
fData->currentQMV[1].y = 2 * fData->currentMV[1].y; |
2489 |
|
SubpelRefine(fData, CheckCandidateInt); |
2490 |
|
if (*fData->iMinSAD > *best_sad + 300) return; |
2491 |
|
fData->currentQMV[2] = fData->currentQMV[0]; |
2492 |
|
SubpelRefine(&bData, CheckCandidateInt); |
2493 |
|
} |
2494 |
|
|
2495 |
|
*fData->iMinSAD += (2+3) * fData->lambda16; /* two bits are needed to code interpolate mode. */ |
2496 |
|
|
2497 |
|
if (*fData->iMinSAD < *best_sad) { |
2498 |
|
*best_sad = *fData->iMinSAD; |
2499 |
|
pMB->mvs[0] = fData->currentMV[0]; |
2500 |
|
pMB->b_mvs[0] = fData->currentMV[1]; |
2501 |
|
pMB->mode = MODE_INTERPOLATE; |
2502 |
|
if (fData->qpel) { |
2503 |
|
pMB->qmvs[0] = fData->currentQMV[0]; |
2504 |
|
pMB->b_qmvs[0] = fData->currentQMV[1]; |
2505 |
|
pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x; |
2506 |
|
pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y; |
2507 |
|
pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x; |
2508 |
|
pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y; |
2509 |
|
} else { |
2510 |
|
pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; |
2511 |
|
pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; |
2512 |
|
pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; |
2513 |
|
pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; |
2514 |
|
} |
2515 |
|
} |
2516 |
|
} |
2517 |
|
|
2518 |
|
void |
2519 |
|
MotionEstimationBVOP(MBParam * const pParam, |
2520 |
|
FRAMEINFO * const frame, |
2521 |
|
const int32_t time_bp, |
2522 |
|
const int32_t time_pp, |
2523 |
|
/* forward (past) reference */ |
2524 |
|
const MACROBLOCK * const f_mbs, |
2525 |
|
const IMAGE * const f_ref, |
2526 |
|
const IMAGE * const f_refH, |
2527 |
|
const IMAGE * const f_refV, |
2528 |
|
const IMAGE * const f_refHV, |
2529 |
|
/* backward (future) reference */ |
2530 |
|
const FRAMEINFO * const b_reference, |
2531 |
|
const IMAGE * const b_ref, |
2532 |
|
const IMAGE * const b_refH, |
2533 |
|
const IMAGE * const b_refV, |
2534 |
|
const IMAGE * const b_refHV) |
2535 |
|
{ |
2536 |
|
uint32_t i, j; |
2537 |
|
int32_t best_sad; |
2538 |
|
uint32_t skip_sad; |
2539 |
|
int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; |
2540 |
|
const MACROBLOCK * const b_mbs = b_reference->mbs; |
2541 |
|
|
2542 |
|
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
2543 |
|
|
2544 |
|
const int32_t TRB = time_pp - time_bp; |
2545 |
|
const int32_t TRD = time_pp; |
2546 |
|
|
2547 |
|
/* some pre-inintialized data for the rest of the search */ |
2548 |
|
|
2549 |
|
SearchData Data; |
2550 |
|
int32_t iMinSAD; |
2551 |
|
uint32_t dir; |
2552 |
|
VECTOR currentMV[3]; |
2553 |
|
VECTOR currentQMV[3]; |
2554 |
|
int32_t temp[8]; |
2555 |
|
memset(&Data, 0, sizeof(SearchData)); |
2556 |
|
Data.iEdgedWidth = pParam->edged_width; |
2557 |
|
Data.currentMV = currentMV; Data.currentQMV = currentQMV; |
2558 |
|
Data.iMinSAD = &iMinSAD; |
2559 |
|
Data.lambda16 = lambda_vec16[MAX(frame->quant-2, 2)]; |
2560 |
|
Data.qpel = pParam->vol_flags & XVID_VOL_QUARTERPEL ? 1 : 0; |
2561 |
|
Data.rounding = 0; |
2562 |
|
Data.chroma = frame->motion_flags & XVID_ME_CHROMA_BVOP; |
2563 |
|
Data.temp = temp; |
2564 |
|
Data.dir = &dir; |
2565 |
|
Data.iQuant = frame->quant; |
2566 |
|
|
2567 |
|
Data.RefQ = f_refV->u; /* a good place, also used in MC (for similar purpose) */ |
2568 |
|
|
2569 |
|
/* note: i==horizontal, j==vertical */ |
2570 |
|
for (j = 0; j < pParam->mb_height; j++) { |
2571 |
|
|
2572 |
|
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
2573 |
|
|
2574 |
|
for (i = 0; i < pParam->mb_width; i++) { |
2575 |
|
MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; |
2576 |
|
const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; |
2577 |
|
|
2578 |
|
/* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */ |
2579 |
|
if (b_reference->coding_type != S_VOP) |
2580 |
|
if (b_mb->mode == MODE_NOT_CODED) { |
2581 |
|
pMB->mode = MODE_NOT_CODED; |
2582 |
|
continue; |
2583 |
|
} |
2584 |
|
|
2585 |
|
Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
2586 |
|
Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8; |
2587 |
|
Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8; |
2588 |
|
|
2589 |
|
/* direct search comes first, because it (1) checks for SKIP-mode |
2590 |
|
and (2) sets very good predictions for forward and backward search */ |
2591 |
|
skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
2592 |
|
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
2593 |
|
&frame->image, |
2594 |
|
i, j, |
2595 |
|
frame->motion_flags, |
2596 |
|
TRB, TRD, |
2597 |
|
pParam, |
2598 |
|
pMB, b_mb, |
2599 |
|
&best_sad, |
2600 |
|
&Data); |
2601 |
|
|
2602 |
|
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
2603 |
|
|
2604 |
|
/* forward search */ |
2605 |
|
SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
2606 |
|
&frame->image, i, j, |
2607 |
|
frame->motion_flags, |
2608 |
|
frame->fcode, pParam, |
2609 |
|
pMB, &f_predMV, &best_sad, |
2610 |
|
MODE_FORWARD, &Data); |
2611 |
|
|
2612 |
|
/* backward search */ |
2613 |
|
SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y, |
2614 |
|
&frame->image, i, j, |
2615 |
|
frame->motion_flags, |
2616 |
|
frame->bcode, pParam, |
2617 |
|
pMB, &b_predMV, &best_sad, |
2618 |
|
MODE_BACKWARD, &Data); |
2619 |
|
|
2620 |
|
/* interpolate search comes last, because it uses data from forward and backward as prediction */ |
2621 |
|
SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
2622 |
|
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
2623 |
|
&frame->image, |
2624 |
|
i, j, |
2625 |
|
frame->fcode, frame->bcode, |
2626 |
|
frame->motion_flags, |
2627 |
|
pParam, |
2628 |
|
&f_predMV, &b_predMV, |
2629 |
|
pMB, &best_sad, |
2630 |
|
&Data); |
2631 |
|
|
2632 |
|
/* final skip decision */ |
2633 |
|
if ( (skip_sad < Data.iQuant * MAX_SAD00_FOR_SKIP * 2) |
2634 |
|
&& ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) |
2635 |
|
SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data); |
2636 |
|
|
2637 |
|
switch (pMB->mode) { |
2638 |
|
case MODE_FORWARD: |
2639 |
|
f_count++; |
2640 |
|
f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; |
2641 |
|
break; |
2642 |
|
case MODE_BACKWARD: |
2643 |
|
b_count++; |
2644 |
|
b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; |
2645 |
|
break; |
2646 |
|
case MODE_INTERPOLATE: |
2647 |
|
i_count++; |
2648 |
|
f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; |
2649 |
|
b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; |
2650 |
|
break; |
2651 |
|
case MODE_DIRECT: |
2652 |
|
case MODE_DIRECT_NO4V: |
2653 |
|
d_count++; |
2654 |
|
default: |
2655 |
|
break; |
2656 |
} |
} |
|
else |
|
|
{ |
|
|
currMV->x = startx; |
|
|
currMV->y = starty; |
|
2657 |
} |
} |
|
return iMinSAD; |
|
2658 |
} |
} |
|
|
|
|
int32_t Halfpel8_Refine( |
|
|
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_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); |
|
|
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); |
|
|
|
|
|
return iMinSAD; |
|
2659 |
} |
} |
2660 |
|
|
2661 |
|
static __inline void |
2662 |
#define PMV_HALFPEL8 (PMV_HALFPELDIAMOND8|PMV_HALFPELREFINE8) |
MEanalyzeMB ( const uint8_t * const pRef, |
2663 |
|
const uint8_t * const pCur, |
2664 |
int32_t PMVfastSearch8( |
const int x, |
2665 |
const uint8_t * const pRef, |
const int y, |
|
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, |
|
2666 |
const MBParam * const pParam, |
const MBParam * const pParam, |
2667 |
const MACROBLOCK * const pMBs, |
MACROBLOCK * const pMBs, |
2668 |
const MACROBLOCK * const prevMBs, |
SearchData * const Data) |
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV) |
|
2669 |
{ |
{ |
|
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; |
|
|
|
|
|
int32_t iDiamondSize; |
|
|
|
|
|
int32_t min_dx; |
|
|
int32_t max_dx; |
|
|
int32_t min_dy; |
|
|
int32_t max_dy; |
|
2670 |
|
|
2671 |
VECTOR pmv[4]; |
int i; |
2672 |
int32_t psad[4]; |
VECTOR pmv[3]; |
2673 |
VECTOR newMV; |
MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width]; |
|
VECTOR backupMV; |
|
2674 |
|
|
2675 |
const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
unsigned int simplicity = 0; |
|
const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; |
|
2676 |
|
|
2677 |
static int32_t threshA,threshB; |
for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
|
int32_t iFound,bPredEq; |
|
|
int32_t iMinSAD,iSAD; |
|
2678 |
|
|
2679 |
int32_t iSubBlock = ((y&1)<<1) + (x&1); |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
2680 |
|
pParam->width, pParam->height, Data->iFcode - Data->qpel - 1, 0, 0); |
2681 |
|
|
2682 |
/* Get maximum range */ |
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
2683 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
Data->RefP[0] = pRef + (x + y * pParam->edged_width) * 16; |
|
x, y, 8, iWidth, iHeight, iFcode); |
|
2684 |
|
|
2685 |
/* we work with abs. MVs, not relative to prediction, so range is relative to 0,0 */ |
pmv[0].x = pMB->mvs[0].x; |
2686 |
|
pmv[0].y = pMB->mvs[0].y; |
2687 |
|
|
2688 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8 )) |
CheckCandidate32I(pmv[0].x, pmv[0].y, Data, 0); |
|
{ min_dx = EVEN(min_dx); |
|
|
max_dx = EVEN(max_dx); |
|
|
min_dy = EVEN(min_dy); |
|
|
max_dy = EVEN(max_dy); |
|
|
} /* because we might use IF (dx>max_dx) THEN dx=max_dx; */ |
|
2689 |
|
|
2690 |
|
if (*Data->iMinSAD > 200) { |
2691 |
|
|
2692 |
bPredEq = get_pmvdata(pMBs, (x>>1), (y>>1), iWcount, iSubBlock, pmv, psad); |
pmv[1].x = pmv[1].y = 0; |
2693 |
|
|
2694 |
if ((x==0) && (y==0) ) |
/* median is only used as prediction. it doesn't have to be real */ |
2695 |
{ |
if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; |
|
threshA = 512/4; |
|
|
threshB = 1024/4; |
|
|
|
|
|
} |
|
2696 |
else |
else |
2697 |
{ |
if (x == 1) /* left macroblock does not have any vector now */ |
2698 |
threshA = psad[0]/4; /* good estimate */ |
Data->predMV = (pMB - pParam->mb_width)->mvs[0]; /* top instead of median */ |
2699 |
threshB = threshA+256/4; |
else if (y == 1) /* top macroblock doesn't have it's vector */ |
2700 |
if (threshA< 512/4) threshA = 512/4; |
Data->predMV = (pMB - 1)->mvs[0]; /* left instead of median */ |
2701 |
if (threshA>1024/4) threshA = 1024/4; |
else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); /* else median */ |
2702 |
if (threshB>1792/4) threshB = 1792/4; |
|
2703 |
|
pmv[2].x = Data->predMV.x; |
2704 |
|
pmv[2].y = Data->predMV.y; |
2705 |
|
|
2706 |
|
if (!vector_repeats(pmv, 1)) |
2707 |
|
CheckCandidate32I(pmv[1].x, pmv[1].y, Data, 1); |
2708 |
|
if (!vector_repeats(pmv, 2)) |
2709 |
|
CheckCandidate32I(pmv[2].x, pmv[2].y, Data, 2); |
2710 |
|
|
2711 |
|
if (*Data->iMinSAD > 500) { /* diamond only if needed */ |
2712 |
|
unsigned int mask = make_mask(pmv, 3, *Data->dir); |
2713 |
|
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate32I); |
2714 |
|
} else simplicity++; |
2715 |
|
|
2716 |
|
if (*Data->iMinSAD > 500) /* refinement from 2-pixel to 1-pixel */ |
2717 |
|
SubpelRefine(Data, CheckCandidate32I); |
2718 |
|
else simplicity++; |
2719 |
|
} else simplicity++; |
2720 |
|
|
2721 |
|
for (i = 0; i < 4; i++) { |
2722 |
|
MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width]; |
2723 |
|
MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i]; |
2724 |
|
MB->mode = MODE_INTER; |
2725 |
|
/* if we skipped some search steps, we have to assume that SAD would be lower with them */ |
2726 |
|
MB->sad16 = Data->iMinSAD[i+1] - (simplicity<<7); |
2727 |
|
} |
2728 |
} |
} |
2729 |
|
|
2730 |
iFound=0; |
#define INTRA_THRESH 2200 |
2731 |
|
#define INTER_THRESH 40 |
2732 |
/* Step 2: Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
#define INTRA_THRESH2 95 |
2733 |
vector of the median. |
|
2734 |
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
int |
2735 |
*/ |
MEanalysis( const IMAGE * const pRef, |
2736 |
|
const FRAMEINFO * const Current, |
2737 |
|
const MBParam * const pParam, |
2738 |
|
const int maxIntra, //maximum number if non-I frames |
2739 |
|
const int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame |
2740 |
|
const int bCount, // number of B frames in a row |
2741 |
|
const int b_thresh) |
2742 |
|
{ |
2743 |
|
uint32_t x, y, intra = 0; |
2744 |
|
int sSAD = 0; |
2745 |
|
MACROBLOCK * const pMBs = Current->mbs; |
2746 |
|
const IMAGE * const pCurrent = &Current->image; |
2747 |
|
int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + b_thresh; |
2748 |
|
int blocks = 0; |
2749 |
|
int complexity = 0; |
2750 |
|
|
2751 |
|
int32_t iMinSAD[5], temp[5]; |
2752 |
|
uint32_t dir; |
2753 |
|
VECTOR currentMV[5]; |
2754 |
|
SearchData Data; |
2755 |
|
Data.iEdgedWidth = pParam->edged_width; |
2756 |
|
Data.currentMV = currentMV; |
2757 |
|
Data.iMinSAD = iMinSAD; |
2758 |
|
Data.iFcode = Current->fcode; |
2759 |
|
Data.temp = temp; |
2760 |
|
Data.dir = &dir; |
2761 |
|
Data.qpel = (pParam->vol_flags & XVID_VOL_QUARTERPEL)? 1: 0; |
2762 |
|
Data.qpel_precision = 0; |
2763 |
|
|
2764 |
|
if (intraCount != 0) { |
2765 |
|
if (intraCount < 10) // we're right after an I frame |
2766 |
|
IntraThresh += 15* (intraCount - 10) * (intraCount - 10); |
2767 |
|
else |
2768 |
|
if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec |
2769 |
|
IntraThresh -= (IntraThresh * (maxIntra - 8*(maxIntra - intraCount)))/maxIntra; |
2770 |
|
} |
2771 |
|
|
2772 |
if ((bPredEq) && (MVequal(pmv[0],pMB->mvs[iSubBlock]) ) ) |
InterThresh -= 20 * bCount; |
2773 |
iFound=2; |
if (InterThresh < 10 + b_thresh) InterThresh = 10 + b_thresh; |
2774 |
|
|
2775 |
/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
if (sadInit) (*sadInit) (); |
|
Otherwise select large Diamond Search. |
|
|
*/ |
|
2776 |
|
|
2777 |
if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536/4) || (bPredEq) ) |
for (y = 1; y < pParam->mb_height-1; y += 2) { |
2778 |
iDiamondSize=1; // 1 halfpel! |
for (x = 1; x < pParam->mb_width-1; x += 2) { |
2779 |
else |
int i; |
2780 |
iDiamondSize=2; // 2 halfpel = 1 full pixel! |
blocks += 10; |
2781 |
|
|
2782 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8) ) |
if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; |
2783 |
iDiamondSize*=2; |
else { //extrapolation of the vector found for last frame |
2784 |
|
pMBs[x + y * pParam->mb_width].mvs[0].x = |
2785 |
|
(pMBs[x + y * pParam->mb_width].mvs[0].x * (bCount+1) ) / bCount; |
2786 |
|
pMBs[x + y * pParam->mb_width].mvs[0].y = |
2787 |
|
(pMBs[x + y * pParam->mb_width].mvs[0].y * (bCount+1) ) / bCount; |
2788 |
|
} |
2789 |
|
|
2790 |
/* Step 4: Calculate SAD around the Median prediction. |
MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data); |
|
MinSAD=SAD |
|
|
If Motion Vector equal to Previous frame motion vector |
|
|
and MinSAD<PrevFrmSAD goto Step 10. |
|
|
If SAD<=256 goto Step 10. |
|
|
*/ |
|
2791 |
|
|
2792 |
|
for (i = 0; i < 4; i++) { |
2793 |
|
int dev; |
2794 |
|
MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width]; |
2795 |
|
dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16, |
2796 |
|
pParam->edged_width); |
2797 |
|
|
2798 |
// Prepare for main loop |
complexity += MAX(dev, 300); |
2799 |
|
if (dev + IntraThresh < pMB->sad16) { |
2800 |
|
pMB->mode = MODE_INTRA; |
2801 |
|
if (++intra > ((pParam->mb_height-2)*(pParam->mb_width-2))/2) return I_VOP; |
2802 |
|
} |
2803 |
|
|
2804 |
currMV->x=start_x; /* start with mv16 */ |
if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0) |
2805 |
currMV->y=start_y; |
if (dev > 1000 && pMB->sad16 < 1000) |
2806 |
|
sSAD += 1000; |
2807 |
|
|
2808 |
iMinSAD = sad8( cur, |
sSAD += (dev < 4000) ? pMB->sad16 : pMB->sad16/2; /* blocks with big contrast differences usually have large SAD - while they look very good in b-frames */ |
|
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; |
|
|
|
|
|
if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,pMB->mvs[iSubBlock])) && ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
|
|
{ |
|
|
if (MotionFlags & PMV_QUICKSTOP16) |
|
|
goto PMVfast8_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfast8_Terminate_with_Refine; |
|
2809 |
} |
} |
2810 |
|
} |
2811 |
|
} |
2812 |
|
complexity >>= 7; |
2813 |
|
|
2814 |
|
sSAD /= complexity + 4*blocks; |
2815 |
|
|
2816 |
/* |
if (intraCount > 80 && sSAD > INTRA_THRESH2 ) return I_VOP; |
2817 |
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
if (sSAD > InterThresh ) return P_VOP; |
2818 |
Also calculate (0,0) but do not subtract offset. |
emms(); |
2819 |
Let MinSAD be the smallest SAD up to this point. |
return B_VOP; |
2820 |
If MV is (0,0) subtract offset. ******** WHAT'S THIS 'OFFSET' ??? *********** |
} |
|
*/ |
|
|
|
|
|
// the prediction might be even better than mv16 |
|
|
CHECK_MV8_CANDIDATE(pmv[0].x,pmv[0].y); |
|
2821 |
|
|
|
// (0,0) is always possible |
|
|
CHECK_MV8_ZERO; |
|
2822 |
|
|
2823 |
// previous frame MV is always possible |
/* functions which perform BITS-based search/bitcount */ |
|
CHECK_MV8_CANDIDATE(pMB->mvs[iSubBlock].x,pMB->mvs[iSubBlock].y); |
|
2824 |
|
|
2825 |
// left neighbour, if allowed |
static int |
2826 |
if (psad[1] != MV_MAX_ERROR) |
findRDinter(SearchData * const Data, |
2827 |
|
const MACROBLOCK * const pMBs, const int x, const int y, |
2828 |
|
const MBParam * const pParam, |
2829 |
|
const uint32_t MotionFlags) |
2830 |
{ |
{ |
2831 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
int i; |
2832 |
{ pmv[1].x = EVEN(pmv[1].x); |
int32_t bsad[5]; |
|
pmv[1].y = EVEN(pmv[1].y); |
|
|
} |
|
|
CHECK_MV8_CANDIDATE(pmv[1].x,pmv[1].y); |
|
|
} |
|
2833 |
|
|
2834 |
// top neighbour, if allowed |
if (Data->qpel) { |
2835 |
if (psad[2] != MV_MAX_ERROR) |
for(i = 0; i < 5; i++) { |
2836 |
{ |
Data->currentMV[i].x = Data->currentQMV[i].x/2; |
2837 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
Data->currentMV[i].y = Data->currentQMV[i].y/2; |
|
{ pmv[2].x = EVEN(pmv[2].x); |
|
|
pmv[2].y = EVEN(pmv[2].y); |
|
2838 |
} |
} |
2839 |
CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); |
Data->qpel_precision = 1; |
2840 |
|
CheckCandidateRD16(Data->currentQMV[0].x, Data->currentQMV[0].y, Data, 255); |
2841 |
|
|
2842 |
// top right neighbour, if allowed |
if (MotionFlags & (XVID_ME_HALFPELREFINE16_RD | XVID_ME_EXTSEARCH_RD)) { /* we have to prepare for halfpixel-precision search */ |
2843 |
if (psad[3] != MV_MAX_ERROR) |
for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i]; |
2844 |
{ |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
2845 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, Data->rrv); |
2846 |
{ pmv[3].x = EVEN(pmv[3].x); |
Data->qpel_precision = 0; |
2847 |
pmv[3].y = EVEN(pmv[3].y); |
if (Data->currentQMV->x & 1 || Data->currentQMV->y & 1) |
2848 |
} |
CheckCandidateRD16(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255); |
|
CHECK_MV8_CANDIDATE(pmv[3].x,pmv[3].y); |
|
|
} |
|
2849 |
} |
} |
2850 |
|
|
2851 |
/* Step 6: If MinSAD <= thresa goto Step 10. |
} else { /* not qpel */ |
|
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
|
|
*/ |
|
2852 |
|
|
2853 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,pMB->mvs[iSubBlock]) && ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
CheckCandidateRD16(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255); |
|
{ |
|
|
if (MotionFlags & PMV_QUICKSTOP16) |
|
|
goto PMVfast8_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfast8_Terminate_with_Refine; |
|
2854 |
} |
} |
2855 |
|
|
2856 |
/************ (Diamond Search) **************/ |
if (MotionFlags&XVID_ME_EXTSEARCH_RD) |
2857 |
/* |
SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, 255, CheckCandidateRD16); |
|
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. |
|
|
*/ |
|
|
|
|
|
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
|
2858 |
|
|
2859 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
if (MotionFlags&XVID_ME_HALFPELREFINE16_RD) |
2860 |
iSAD = Diamond8_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
SubpelRefine(Data, CheckCandidateRD16); |
|
x, y, |
|
|
currMV->x, currMV->y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
2861 |
|
|
2862 |
if (iSAD < iMinSAD) |
if (Data->qpel) { |
2863 |
{ |
if (MotionFlags&(XVID_ME_EXTSEARCH_RD | XVID_ME_HALFPELREFINE16_RD)) { /* there was halfpel-precision search */ |
2864 |
*currMV = newMV; |
for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) { |
2865 |
iMinSAD = iSAD; |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* we have found a better match */ |
2866 |
|
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
2867 |
} |
} |
2868 |
|
|
2869 |
if (MotionFlags & PMV_EXTSEARCH8) |
/* preparing for qpel-precision search */ |
2870 |
{ |
Data->qpel_precision = 1; |
2871 |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
2872 |
|
pParam->width, pParam->height, Data->iFcode, 2, 0); |
|
if (!(MVequal(pmv[0],backupMV)) ) |
|
|
{ iSAD = Diamond16_MainSearch(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, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) |
|
|
{ |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
2873 |
} |
} |
2874 |
|
if (MotionFlags&XVID_ME_QUARTERPELREFINE16_RD) |
2875 |
|
SubpelRefine(Data, CheckCandidateRD16); |
2876 |
} |
} |
2877 |
|
|
2878 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
if (MotionFlags&XVID_ME_CHECKPREDICTION_RD) { /* let's check vector equal to prediction */ |
2879 |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV; |
2880 |
x, y, |
if (!(Data->predMV.x == v->x && Data->predMV.y == v->y)) |
2881 |
0, 0, iMinSAD, &newMV, |
CheckCandidateRD16(Data->predMV.x, Data->predMV.y, Data, 255); |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) |
|
|
{ |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
2882 |
} |
} |
2883 |
|
return Data->iMinSAD[0]; |
2884 |
} |
} |
2885 |
|
|
2886 |
/* Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
static int |
2887 |
By performing an optional local half-pixel search, we can refine this result even further. |
findRDinter4v(const SearchData * const Data, |
2888 |
*/ |
MACROBLOCK * const pMB, const MACROBLOCK * const pMBs, |
2889 |
|
const int x, const int y, |
2890 |
|
const MBParam * const pParam, const uint32_t MotionFlags, |
2891 |
|
const VECTOR * const backup) |
2892 |
|
{ |
2893 |
|
|
2894 |
PMVfast8_Terminate_with_Refine: |
int cbp = 0, bits = 0, t = 0, i; |
2895 |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
SearchData Data2, *Data8 = &Data2; |
2896 |
iMinSAD = Halfpel8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
int sumx = 0, sumy = 0; |
2897 |
x, y, |
int16_t *in = Data->dctSpace, *coeff = Data->dctSpace + 64; |
2898 |
currMV, iMinSAD, |
uint8_t * ptr; |
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
|
2899 |
|
|
2900 |
|
memcpy(Data8, Data, sizeof(SearchData)); |
2901 |
|
|
2902 |
PMVfast8_Terminate_without_Refine: |
for (i = 0; i < 4; i++) { /* for all luma blocks */ |
|
currPMV->x = currMV->x - pmv[0].x; |
|
|
currPMV->y = currMV->y - pmv[0].y; |
|
2903 |
|
|
2904 |
return iMinSAD; |
Data8->iMinSAD = Data->iMinSAD + i + 1; |
2905 |
|
Data8->currentMV = Data->currentMV + i + 1; |
2906 |
|
Data8->currentQMV = Data->currentQMV + i + 1; |
2907 |
|
Data8->Cur = Data->Cur + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2908 |
|
Data8->RefP[0] = Data->RefP[0] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2909 |
|
Data8->RefP[2] = Data->RefP[2] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2910 |
|
Data8->RefP[1] = Data->RefP[1] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2911 |
|
Data8->RefP[3] = Data->RefP[3] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2912 |
|
*Data8->cbp = (Data->cbp[1] & (1<<(5-i))) ? 1:0; // copy corresponding cbp bit |
2913 |
|
|
2914 |
|
if(Data->qpel) { |
2915 |
|
Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i); |
2916 |
|
if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y, |
2917 |
|
Data8->predMV, Data8->iFcode, 0, 0); |
2918 |
|
} else { |
2919 |
|
Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i); |
2920 |
|
if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y, |
2921 |
|
Data8->predMV, Data8->iFcode, 0, 0); |
2922 |
} |
} |
2923 |
|
|
2924 |
int32_t EPZSSearch16( |
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 3, |
2925 |
const uint8_t * const pRef, |
pParam->width, pParam->height, Data8->iFcode, Data8->qpel+1, 0); |
|
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) |
|
|
{ |
|
|
const uint32_t iWcount = pParam->mb_width; |
|
|
const uint32_t iHcount = pParam->mb_height; |
|
2926 |
|
|
2927 |
const int32_t iWidth = pParam->width; |
*Data8->iMinSAD += BITS_MULT*t; |
|
const int32_t iHeight = pParam->height; |
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
2928 |
|
|
2929 |
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
Data8->qpel_precision = Data8->qpel; |
2930 |
|
/* checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far) */ |
2931 |
|
{ |
2932 |
|
VECTOR *v = Data8->qpel ? Data8->currentQMV : Data8->currentMV; |
2933 |
|
if (!MVequal (*v, backup[i+1]) ) |
2934 |
|
CheckCandidateRD8(backup[i+1].x, backup[i+1].y, Data8, 255); |
2935 |
|
} |
2936 |
|
|
2937 |
int32_t min_dx; |
if (Data8->qpel) { |
2938 |
int32_t max_dx; |
if (MotionFlags&XVID_ME_HALFPELREFINE8_RD || (MotionFlags&XVID_ME_EXTSEARCH8 && MotionFlags&XVID_ME_EXTSEARCH_RD)) { /* halfpixel motion search follows */ |
2939 |
int32_t min_dy; |
int32_t s = *Data8->iMinSAD; |
2940 |
int32_t max_dy; |
Data8->currentMV->x = Data8->currentQMV->x/2; |
2941 |
|
Data8->currentMV->y = Data8->currentQMV->y/2; |
2942 |
|
Data8->qpel_precision = 0; |
2943 |
|
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 3, |
2944 |
|
pParam->width, pParam->height, Data8->iFcode - 1, 1, 0); |
2945 |
|
|
2946 |
VECTOR newMV; |
if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1) |
2947 |
VECTOR backupMV; |
CheckCandidateRD8(Data8->currentMV->x, Data8->currentMV->y, Data8, 255); |
2948 |
|
|
2949 |
VECTOR pmv[4]; |
if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_RD) |
2950 |
int32_t psad[8]; |
SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255, CheckCandidateRD8); |
2951 |
|
|
2952 |
static MACROBLOCK * oldMBs = NULL; |
if (MotionFlags & XVID_ME_HALFPELREFINE8_RD) |
2953 |
const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
SubpelRefine(Data8, CheckCandidateRD8); |
|
const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; |
|
|
MACROBLOCK * oldMB = NULL; |
|
2954 |
|
|
2955 |
static int32_t thresh2; |
if(s > *Data8->iMinSAD) { /* we have found a better match */ |
2956 |
int32_t bPredEq; |
Data8->currentQMV->x = 2*Data8->currentMV->x; |
2957 |
int32_t iMinSAD,iSAD=9999; |
Data8->currentQMV->y = 2*Data8->currentMV->y; |
2958 |
|
} |
2959 |
|
|
2960 |
MainSearch16FuncPtr EPZSMainSearchPtr; |
Data8->qpel_precision = 1; |
2961 |
|
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 3, |
2962 |
|
pParam->width, pParam->height, Data8->iFcode, 2, 0); |
2963 |
|
|
|
if (oldMBs == NULL) |
|
|
{ oldMBs = (MACROBLOCK*) calloc(1,iWcount*iHcount*sizeof(MACROBLOCK)); |
|
|
fprintf(stderr,"allocated %d bytes for oldMBs\n",iWcount*iHcount*sizeof(MACROBLOCK)); |
|
2964 |
} |
} |
2965 |
oldMB = oldMBs + x + y * iWcount; |
if (MotionFlags & XVID_ME_QUARTERPELREFINE8_RD) |
2966 |
|
SubpelRefine(Data8, CheckCandidateRD8); |
2967 |
|
|
2968 |
/* Get maximum range */ |
} else { /* not qpel */ |
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
|
|
x, y, 16, iWidth, iHeight, iFcode); |
|
2969 |
|
|
2970 |
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_RD) /* extsearch */ |
2971 |
|
SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255, CheckCandidateRD8); |
2972 |
|
|
2973 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
if (MotionFlags & XVID_ME_HALFPELREFINE8_RD) |
2974 |
{ min_dx = EVEN(min_dx); |
SubpelRefine(Data8, CheckCandidateRD8); /* halfpel refinement */ |
2975 |
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; */ |
|
|
|
|
|
bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
|
2976 |
|
|
2977 |
/* Step 4: Calculate SAD around the Median prediction. |
/* checking vector equal to predicion */ |
2978 |
MinSAD=SAD |
if (i != 0 && MotionFlags & XVID_ME_CHECKPREDICTION_RD) { |
2979 |
If Motion Vector equal to Previous frame motion vector |
const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV; |
2980 |
and MinSAD<PrevFrmSAD goto Step 10. |
if (!MVequal(*v, Data8->predMV)) |
2981 |
If SAD<=256 goto Step 10. |
CheckCandidateRD8(Data8->predMV.x, Data8->predMV.y, Data8, 255); |
2982 |
*/ |
} |
2983 |
|
|
2984 |
// Prepare for main loop |
bits += *Data8->iMinSAD; |
2985 |
|
if (bits >= Data->iMinSAD[0]) return bits; /* no chances for INTER4V */ |
2986 |
|
|
2987 |
*currMV=pmv[0]; /* current best := median prediction */ |
/* MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else */ |
2988 |
if (!(MotionFlags & PMV_HALFPEL16)) |
if(Data->qpel) { |
2989 |
{ |
pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x; |
2990 |
currMV->x = EVEN(currMV->x); |
pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y; |
2991 |
currMV->y = EVEN(currMV->y); |
pMB->qmvs[i] = *Data8->currentQMV; |
2992 |
|
sumx += Data8->currentQMV->x/2; |
2993 |
|
sumy += Data8->currentQMV->y/2; |
2994 |
|
} else { |
2995 |
|
pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x; |
2996 |
|
pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y; |
2997 |
|
sumx += Data8->currentMV->x; |
2998 |
|
sumy += Data8->currentMV->y; |
2999 |
} |
} |
3000 |
|
pMB->mvs[i] = *Data8->currentMV; |
3001 |
|
pMB->sad8[i] = 4 * *Data8->iMinSAD; |
3002 |
|
if (Data8->cbp[0]) cbp |= 1 << (5 - i); |
3003 |
|
|
3004 |
if (currMV->x > max_dx) |
} /* end - for all luma blocks */ |
|
currMV->x=max_dx; |
|
|
if (currMV->x < min_dx) |
|
|
currMV->x=min_dx; |
|
|
if (currMV->y > max_dy) |
|
|
currMV->y=max_dy; |
|
|
if (currMV->y < min_dy) |
|
|
currMV->y=min_dy; |
|
3005 |
|
|
3006 |
/***************** This is predictor SET A: only median prediction ******************/ |
bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; |
3007 |
|
|
3008 |
iMinSAD = sad16( cur, |
/* let's check chroma */ |
3009 |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
sumx = (sumx >> 3) + roundtab_76[sumx & 0xf]; |
3010 |
iEdgedWidth, MV_MAX_ERROR); |
sumy = (sumy >> 3) + roundtab_76[sumy & 0xf]; |
|
iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; |
|
3011 |
|
|
3012 |
// thresh1 is fixed to 256 |
/* chroma U */ |
3013 |
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,pMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[4], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
3014 |
{ |
transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2); |
3015 |
if (MotionFlags & PMV_QUICKSTOP16) |
bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4); |
|
goto EPZS16_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto EPZS16_Terminate_with_Refine; |
|
|
} |
|
3016 |
|
|
3017 |
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
if (bits >= *Data->iMinSAD) return bits; |
3018 |
|
|
3019 |
// previous frame MV |
/* chroma V */ |
3020 |
CHECK_MV16_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); |
ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[5], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
3021 |
|
transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2); |
3022 |
|
bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5); |
3023 |
|
|
3024 |
// set threshhold based on Min of Prediction and SAD of collocated block |
bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len; |
|
// CHECK_MV16 always uses iSAD for the SAD of last vector to check, so now iSAD is what we want |
|
3025 |
|
|
3026 |
if ((x==0) && (y==0) ) |
*Data->cbp = cbp; |
3027 |
{ |
return bits; |
|
thresh2 = 512; |
|
3028 |
} |
} |
3029 |
else |
|
3030 |
|
static int |
3031 |
|
findRDintra(const SearchData * const Data) |
3032 |
{ |
{ |
3033 |
/* T_k = 1.2 * MIN(SAD_top,SAD_left,SAD_topleft,SAD_coll) +128; [Tourapis, 2002] */ |
int bits = BITS_MULT*1; /* this one is ac/dc prediction flag bit */ |
3034 |
|
int cbp = 0, i, dc = 0; |
3035 |
|
int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64; |
3036 |
|
|
3037 |
thresh2 = MIN(psad[0],iSAD)*6/5 + 128; |
for(i = 0; i < 4; i++) { |
3038 |
|
int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
3039 |
|
transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth); |
3040 |
|
bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i, &dc); |
3041 |
|
|
3042 |
|
if (bits >= Data->iMinSAD[0]) return bits; |
3043 |
} |
} |
3044 |
|
|
3045 |
// MV=(0,0) is often a good choice |
bits += BITS_MULT*xvid_cbpy_tab[cbp>>2].len; |
3046 |
|
|
3047 |
CHECK_MV16_ZERO; |
/*chroma U */ |
3048 |
|
transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2); |
3049 |
|
bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4, &dc); |
3050 |
|
|
3051 |
|
if (bits >= Data->iMinSAD[0]) return bits; |
3052 |
|
|
3053 |
// left neighbour, if allowed |
/* chroma V */ |
3054 |
if (x != 0) |
transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2); |
3055 |
{ |
bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5, &dc); |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
|
|
{ pmv[1].x = EVEN(pmv[1].x); |
|
|
pmv[1].y = EVEN(pmv[1].y); |
|
|
} |
|
|
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
|
|
} |
|
3056 |
|
|
3057 |
// top neighbour, if allowed |
bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len; |
3058 |
if (y != 0) |
|
3059 |
{ |
return bits; |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
|
|
{ pmv[2].x = EVEN(pmv[2].x); |
|
|
pmv[2].y = EVEN(pmv[2].y); |
|
3060 |
} |
} |
|
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
|
3061 |
|
|
3062 |
// top right neighbour, if allowed |
static int |
3063 |
if ((uint32_t)x != (iWcount-1)) |
findRDgmc(const SearchData * const Data, const IMAGE * const vGMC, const int x, const int y) |
3064 |
{ |
{ |
3065 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
int bits = BITS_MULT*1; /* this one is mcsel */ |
3066 |
{ pmv[3].x = EVEN(pmv[3].x); |
int cbp = 0, i; |
3067 |
pmv[3].y = EVEN(pmv[3].y); |
int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64; |
3068 |
} |
|
3069 |
CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); |
for(i = 0; i < 4; i++) { |
3070 |
} |
int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
3071 |
|
transfer_8to16subro(in, Data->Cur + s, vGMC->y + s + 16*(x+y*Data->iEdgedWidth), Data->iEdgedWidth); |
3072 |
|
bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i); |
3073 |
|
if (bits >= Data->iMinSAD[0]) return bits; |
3074 |
} |
} |
3075 |
|
|
3076 |
/* Terminate if MinSAD <= T_2 |
bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; |
|
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
|
|
*/ |
|
3077 |
|
|
3078 |
if ( (iMinSAD <= thresh2) |
/*chroma U */ |
3079 |
|| ( MVequal(*currMV,pMB->mvs[0]) && ((uint32_t)iMinSAD <= prevMB->sad16) ) ) |
transfer_8to16subro(in, Data->CurU, vGMC->u + 8*(x+y*(Data->iEdgedWidth/2)), Data->iEdgedWidth/2); |
3080 |
{ |
bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4); |
|
if (MotionFlags & PMV_QUICKSTOP16) |
|
|
goto EPZS16_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto EPZS16_Terminate_with_Refine; |
|
|
} |
|
3081 |
|
|
3082 |
/***** predictor SET C: acceleration MV (new!), neighbours in prev. frame(new!) ****/ |
if (bits >= Data->iMinSAD[0]) return bits; |
3083 |
|
|
3084 |
backupMV = pMB->mvs[0]; // last MV |
/* chroma V */ |
3085 |
backupMV.x += (pMB->mvs[0].x - oldMB->mvs[0].x ); // acceleration X |
transfer_8to16subro(in, Data->CurV , vGMC->v + 8*(x+y*(Data->iEdgedWidth/2)), Data->iEdgedWidth/2); |
3086 |
backupMV.y += (pMB->mvs[0].y - oldMB->mvs[0].y ); // acceleration Y |
bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5); |
3087 |
|
|
3088 |
CHECK_MV16_CANDIDATE(backupMV.x,backupMV.y); |
bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; |
3089 |
|
|
3090 |
// left neighbour |
*Data->cbp = cbp; |
|
if (x != 0) |
|
|
CHECK_MV16_CANDIDATE((oldMB-1)->mvs[0].x,oldMB->mvs[0].y); |
|
3091 |
|
|
3092 |
// top neighbour |
return bits; |
3093 |
if (y != 0) |
} |
|
CHECK_MV16_CANDIDATE((oldMB-iWcount)->mvs[0].x,oldMB->mvs[0].y); |
|
3094 |
|
|
|
// right neighbour, if allowed (this value is not written yet, so take it from pMB->mvs |
|
3095 |
|
|
|
if ((uint32_t)x != iWcount-1) |
|
|
CHECK_MV16_CANDIDATE((pMB+1)->mvs[0].x,oldMB->mvs[0].y); |
|
3096 |
|
|
|
// bottom neighbour, dito |
|
|
if ((uint32_t)y != iHcount-1) |
|
|
CHECK_MV16_CANDIDATE((pMB+iWcount)->mvs[0].x,oldMB->mvs[0].y); |
|
3097 |
|
|
3098 |
/* Terminate if MinSAD <= T_3 (here T_3 = T_2) */ |
static __inline void |
3099 |
if (iMinSAD <= thresh2) |
GMEanalyzeMB ( const uint8_t * const pCur, |
3100 |
|
const uint8_t * const pRef, |
3101 |
|
const uint8_t * const pRefH, |
3102 |
|
const uint8_t * const pRefV, |
3103 |
|
const uint8_t * const pRefHV, |
3104 |
|
const int x, |
3105 |
|
const int y, |
3106 |
|
const MBParam * const pParam, |
3107 |
|
MACROBLOCK * const pMBs, |
3108 |
|
SearchData * const Data) |
3109 |
{ |
{ |
|
if (MotionFlags & PMV_QUICKSTOP16) |
|
|
goto EPZS16_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto EPZS16_Terminate_with_Refine; |
|
|
} |
|
3110 |
|
|
3111 |
/************ (if Diamond Search) **************/ |
int i=0; |
3112 |
|
MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width]; |
3113 |
|
|
3114 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
Data->iMinSAD[0] = MV_MAX_ERROR; |
3115 |
|
|
3116 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
3117 |
|
|
3118 |
if (MotionFlags & PMV_USESQUARES16) |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
3119 |
EPZSMainSearchPtr = Square16_MainSearch; |
pParam->width, pParam->height, 16, 1, 0); |
|
else |
|
|
EPZSMainSearchPtr = Diamond16_MainSearch; |
|
3120 |
|
|
3121 |
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
Data->Cur = pCur + 16*(x + y * pParam->edged_width); |
3122 |
x, y, |
Data->RefP[0] = pRef + 16*(x + y * pParam->edged_width); |
3123 |
currMV->x, currMV->y, iMinSAD, &newMV, pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
Data->RefP[1] = pRefV + 16*(x + y * pParam->edged_width); |
3124 |
2, iFcode, iQuant, 0); |
Data->RefP[2] = pRefH + 16*(x + y * pParam->edged_width); |
3125 |
|
Data->RefP[3] = pRefHV + 16*(x + y * pParam->edged_width); |
3126 |
|
|
3127 |
if (iSAD < iMinSAD) |
Data->currentMV[0].x = Data->currentMV[0].y = 0; |
3128 |
{ |
CheckCandidate16I(0, 0, Data, 255); |
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
3129 |
|
|
3130 |
|
if ( (Data->predMV.x !=0) || (Data->predMV.y != 0) ) |
3131 |
|
CheckCandidate16I(Data->predMV.x, Data->predMV.y, Data, 255); |
3132 |
|
|
3133 |
if (MotionFlags & PMV_EXTSEARCH16) |
DiamondSearch(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255, CheckCandidate16I); |
|
{ |
|
|
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
|
3134 |
|
|
3135 |
if (!(MVequal(pmv[0],backupMV)) ) |
SubpelRefine(Data, CheckCandidate16I); |
|
{ |
|
|
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, 2, iFcode, iQuant, 0); |
|
|
} |
|
3136 |
|
|
3137 |
if (iSAD < iMinSAD) |
|
3138 |
{ |
/* for QPel halfpel positions are worse than in halfpel mode :( */ |
3139 |
*currMV = newMV; |
/* if (Data->qpel) { |
3140 |
iMinSAD = iSAD; |
Data->currentQMV->x = 2*Data->currentMV->x; |
3141 |
|
Data->currentQMV->y = 2*Data->currentMV->y; |
3142 |
|
Data->qpel_precision = 1; |
3143 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
3144 |
|
pParam->width, pParam->height, iFcode, 2, 0); |
3145 |
|
SubpelRefine(Data); |
3146 |
} |
} |
3147 |
|
*/ |
3148 |
|
|
3149 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
3150 |
{ |
pMB->sad16 = Data->iMinSAD[0]; |
3151 |
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
pMB->mode = MODE_INTER; |
3152 |
x, y, |
pMB->sad16 += 10*d_mv_bits(pMB->mvs[0].x, pMB->mvs[0].y, Data->predMV, Data->iFcode, 0, 0); |
3153 |
0, 0, iMinSAD, &newMV, |
return; |
3154 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, /*iDiamondSize*/ 2, iFcode, iQuant, 0); |
} |
3155 |
|
|
3156 |
if (iSAD < iMinSAD) |
void |
3157 |
|
GMEanalysis(const MBParam * const pParam, |
3158 |
|
const FRAMEINFO * const current, |
3159 |
|
const FRAMEINFO * const reference, |
3160 |
|
const IMAGE * const pRefH, |
3161 |
|
const IMAGE * const pRefV, |
3162 |
|
const IMAGE * const pRefHV) |
3163 |
{ |
{ |
3164 |
*currMV = newMV; |
uint32_t x, y; |
3165 |
iMinSAD = iSAD; |
MACROBLOCK * const pMBs = current->mbs; |
3166 |
} |
const IMAGE * const pCurrent = ¤t->image; |
3167 |
} |
const IMAGE * const pReference = &reference->image; |
|
} |
|
3168 |
|
|
3169 |
/*************** Choose best MV found **************/ |
int32_t iMinSAD[5], temp[5]; |
3170 |
|
VECTOR currentMV[5]; |
3171 |
|
uint32_t dir; |
3172 |
|
SearchData Data; |
3173 |
|
memset(&Data, 0, sizeof(SearchData)); |
3174 |
|
|
3175 |
EPZS16_Terminate_with_Refine: |
Data.iEdgedWidth = pParam->edged_width; |
3176 |
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
Data.rounding = pParam->m_rounding_type; |
|
iMinSAD = Halfpel16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
currMV, iMinSAD, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
|
3177 |
|
|
3178 |
EPZS16_Terminate_without_Refine: |
Data.currentMV = ¤tMV[0]; |
3179 |
|
Data.iMinSAD = &iMinSAD[0]; |
3180 |
|
Data.iFcode = current->fcode; |
3181 |
|
Data.temp = temp; |
3182 |
|
Data.dir = &dir; |
3183 |
|
|
3184 |
*oldMB = *pMB; |
if (sadInit) (*sadInit) (); |
3185 |
|
|
3186 |
currPMV->x = currMV->x - pmv[0].x; |
for (y = 0; y < pParam->mb_height; y ++) { |
3187 |
currPMV->y = currMV->y - pmv[0].y; |
for (x = 0; x < pParam->mb_width; x ++) { |
3188 |
return iMinSAD; |
GMEanalyzeMB(pCurrent->y, pReference->y, pRefH->y, pRefV->y, pRefHV->y, x, y, pParam, pMBs, &Data); |
3189 |
|
} |
3190 |
|
} |
3191 |
|
return; |
3192 |
} |
} |
3193 |
|
|
3194 |
|
|
3195 |
int32_t EPZSSearch8( |
WARPPOINTS |
3196 |
const uint8_t * const pRef, |
GlobalMotionEst(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, |
|
3197 |
const MBParam * const pParam, |
const MBParam * const pParam, |
3198 |
const MACROBLOCK * const pMBs, |
const FRAMEINFO * const current, |
3199 |
const MACROBLOCK * const prevMBs, |
const FRAMEINFO * const reference, |
3200 |
VECTOR * const currMV, |
const IMAGE * const pRefH, |
3201 |
VECTOR * const currPMV) |
const IMAGE * const pRefV, |
3202 |
|
const IMAGE * const pRefHV) |
3203 |
{ |
{ |
|
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; |
|
3204 |
|
|
3205 |
const uint8_t * cur = pCur->y + x*8 + y*8*iEdgedWidth; |
const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs |
3206 |
|
const int deltay=8; |
3207 |
|
const unsigned int gradx=512; // lower bound for gradient in MB (ignore "flat" blocks) |
3208 |
|
const unsigned int grady=512; |
3209 |
|
|
3210 |
int32_t iDiamondSize=1; |
double sol[4] = { 0., 0., 0., 0. }; |
3211 |
|
|
3212 |
int32_t min_dx; |
WARPPOINTS gmc; |
|
int32_t max_dx; |
|
|
int32_t min_dy; |
|
|
int32_t max_dy; |
|
3213 |
|
|
3214 |
VECTOR newMV; |
uint32_t mx, my; |
|
VECTOR backupMV; |
|
3215 |
|
|
3216 |
VECTOR pmv[4]; |
int MBh = pParam->mb_height; |
3217 |
int32_t psad[8]; |
int MBw = pParam->mb_width; |
3218 |
|
const int minblocks = 9; //MBh*MBw/32+3; /* just some reasonable number 3% + 3 */ |
3219 |
|
const int maxblocks = MBh*MBw/4; /* just some reasonable number 3% + 3 */ |
3220 |
|
|
3221 |
const int32_t iSubBlock = ((y&1)<<1) + (x&1); |
int num=0; |
3222 |
|
int oldnum; |
3223 |
|
|
3224 |
const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
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; |
|
const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; |
|
3225 |
|
|
3226 |
int32_t bPredEq; |
GMEanalysis(pParam,current, reference, pRefH, pRefV, pRefHV); |
|
int32_t iMinSAD,iSAD=9999; |
|
3227 |
|
|
3228 |
MainSearch8FuncPtr EPZSMainSearchPtr; |
/* block based ME isn't done, yet, so do a quick presearch */ |
3229 |
|
|
3230 |
/* Get maximum range */ |
// filter mask of all blocks |
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
|
|
x, y, 8, iWidth, iHeight, iFcode); |
|
|
|
|
|
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
|
|
|
|
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
|
|
{ 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; */ |
|
|
|
|
|
bPredEq = get_pmvdata(pMBs, x>>1, y>>1, iWcount, iSubBlock, pmv, psad); |
|
3231 |
|
|
3232 |
|
for (my = 0; my < (uint32_t)MBh; my++) |
3233 |
|
for (mx = 0; mx < (uint32_t)MBw; mx++) |
3234 |
|
{ |
3235 |
|
const int mbnum = mx + my * MBw; |
3236 |
|
pMBs[mbnum].mcsel = 0; |
3237 |
|
} |
3238 |
|
|
|
/* 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. |
|
|
*/ |
|
3239 |
|
|
3240 |
// Prepare for main loop |
for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */ |
3241 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */ |
3242 |
|
{ |
3243 |
|
const int mbnum = mx + my * MBw; |
3244 |
|
MACROBLOCK *const pMB = &pMBs[mbnum]; |
3245 |
|
const VECTOR mv = pMB->mvs[0]; |
3246 |
|
|
3247 |
|
/* don't use object boundaries */ |
3248 |
|
if ( (abs(mv.x - (pMB-1)->mvs[0].x) < deltax) |
3249 |
|
&& (abs(mv.y - (pMB-1)->mvs[0].y) < deltay) |
3250 |
|
&& (abs(mv.x - (pMB+1)->mvs[0].x) < deltax) |
3251 |
|
&& (abs(mv.y - (pMB+1)->mvs[0].y) < deltay) |
3252 |
|
&& (abs(mv.x - (pMB-MBw)->mvs[0].x) < deltax) |
3253 |
|
&& (abs(mv.y - (pMB-MBw)->mvs[0].y) < deltay) |
3254 |
|
&& (abs(mv.x - (pMB+MBw)->mvs[0].x) < deltax) |
3255 |
|
&& (abs(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) |
3256 |
|
{ const int iEdgedWidth = pParam->edged_width; |
3257 |
|
const uint8_t *const pCur = current->image.y + 16*(my*iEdgedWidth + mx); |
3258 |
|
if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx ) |
3259 |
|
&& (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) ) |
3260 |
|
{ pMB->mcsel = 1; |
3261 |
|
num++; |
3262 |
|
} |
3263 |
|
|
3264 |
if (!(MotionFlags & PMV_HALFPEL8)) |
/* only use "structured" blocks */ |
3265 |
{ |
} |
|
currMV->x = EVEN(currMV->x); |
|
|
currMV->y = EVEN(currMV->y); |
|
3266 |
} |
} |
3267 |
|
emms(); |
3268 |
|
|
3269 |
if (currMV->x > max_dx) |
/* further filtering would be possible, but during iteration, remaining |
3270 |
currMV->x=max_dx; |
outliers usually are removed, too */ |
|
if (currMV->x < min_dx) |
|
|
currMV->x=min_dx; |
|
|
if (currMV->y > max_dy) |
|
|
currMV->y=max_dy; |
|
|
if (currMV->y < min_dy) |
|
|
currMV->y=min_dy; |
|
3271 |
|
|
3272 |
/***************** This is predictor SET A: only median prediction ******************/ |
if (num>= minblocks) |
3273 |
|
do { /* until convergence */ |
3274 |
|
double DtimesF[4]; |
3275 |
|
double a,b,c,n,invdenom; |
3276 |
|
double meanx,meany; |
3277 |
|
|
3278 |
|
a = b = c = n = 0; |
3279 |
|
DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.; |
3280 |
|
for (my = 1; my < (uint32_t)MBh-1; my++) |
3281 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
3282 |
|
{ |
3283 |
|
const int mbnum = mx + my * MBw; |
3284 |
|
const VECTOR mv = pMBs[mbnum].mvs[0]; |
3285 |
|
|
3286 |
iMinSAD = sad8( cur, |
if (!pMBs[mbnum].mcsel) |
3287 |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
continue; |
|
iEdgedWidth); |
|
|
iMinSAD += calc_delta_8(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; |
|
3288 |
|
|
3289 |
|
n++; |
3290 |
|
a += 16*mx+8; |
3291 |
|
b += 16*my+8; |
3292 |
|
c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8); |
3293 |
|
|
3294 |
// thresh1 is fixed to 256 |
DtimesF[0] += (double)mv.x; |
3295 |
if (iMinSAD < 256/4 ) |
DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8); |
3296 |
{ |
DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8); |
3297 |
if (MotionFlags & PMV_QUICKSTOP8) |
DtimesF[3] += (double)mv.y; |
|
goto EPZS8_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP8) |
|
|
goto EPZS8_Terminate_with_Refine; |
|
3298 |
} |
} |
3299 |
|
|
3300 |
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
invdenom = a*a+b*b-c*n; |
|
|
|
|
// previous frame MV |
|
|
CHECK_MV8_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); |
|
3301 |
|
|
3302 |
// MV=(0,0) is often a good choice |
/* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */ |
3303 |
|
/* D'*E*F has been calculated in the same loop as matrix */ |
3304 |
|
|
3305 |
CHECK_MV8_ZERO; |
sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2]; |
3306 |
|
sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3]; |
3307 |
|
sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3]; |
3308 |
|
sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3]; |
3309 |
|
|
3310 |
/* Terminate if MinSAD <= T_2 |
sol[0] /= invdenom; |
3311 |
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
sol[1] /= invdenom; |
3312 |
*/ |
sol[2] /= invdenom; |
3313 |
|
sol[3] /= invdenom; |
3314 |
|
|
3315 |
if (iMinSAD < 512/4) /* T_2 == 512/4 hardcoded */ |
meanx = meany = 0.; |
3316 |
|
oldnum = 0; |
3317 |
|
for (my = 1; my < (uint32_t)MBh-1; my++) |
3318 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
3319 |
{ |
{ |
3320 |
if (MotionFlags & PMV_QUICKSTOP8) |
const int mbnum = mx + my * MBw; |
3321 |
goto EPZS8_Terminate_without_Refine; |
const VECTOR mv = pMBs[mbnum].mvs[0]; |
|
if (MotionFlags & PMV_EARLYSTOP8) |
|
|
goto EPZS8_Terminate_with_Refine; |
|
|
} |
|
3322 |
|
|
3323 |
/************ (if Diamond Search) **************/ |
if (!pMBs[mbnum].mcsel) |
3324 |
|
continue; |
|
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
|
3325 |
|
|
3326 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8)) |
oldnum++; |
3327 |
iDiamondSize *= 2; |
meanx += fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - (double)mv.x ); |
3328 |
|
meany += fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - (double)mv.y ); |
3329 |
|
} |
3330 |
|
|
3331 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
if (4*meanx > oldnum) /* better fit than 0.25 (=1/4pel) is useless */ |
3332 |
|
meanx /= oldnum; |
3333 |
|
else |
3334 |
|
meanx = 0.25; |
3335 |
|
|
3336 |
// if (MotionFlags & PMV_USESQUARES8) |
if (4*meany > oldnum) |
3337 |
// EPZSMainSearchPtr = Square8_MainSearch; |
meany /= oldnum; |
3338 |
// else |
else |
3339 |
EPZSMainSearchPtr = Diamond8_MainSearch; |
meany = 0.25; |
3340 |
|
|
3341 |
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
num = 0; |
3342 |
x, y, |
for (my = 0; my < (uint32_t)MBh; my++) |
3343 |
currMV->x, currMV->y, iMinSAD, &newMV, |
for (mx = 0; mx < (uint32_t)MBw; mx++) |
3344 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
{ |
3345 |
iDiamondSize, iFcode, iQuant, 00); |
const int mbnum = mx + my * MBw; |
3346 |
|
const VECTOR mv = pMBs[mbnum].mvs[0]; |
3347 |
|
|
3348 |
|
if (!pMBs[mbnum].mcsel) |
3349 |
|
continue; |
3350 |
|
|
3351 |
if (iSAD < iMinSAD) |
if ( ( fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - (double)mv.x ) > meanx ) |
3352 |
{ |
|| ( fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - (double)mv.y ) > meany ) ) |
3353 |
*currMV = newMV; |
pMBs[mbnum].mcsel=0; |
3354 |
iMinSAD = iSAD; |
else |
3355 |
|
num++; |
3356 |
} |
} |
3357 |
|
|
3358 |
if (MotionFlags & PMV_EXTSEARCH8) |
} while ( (oldnum != num) && (num>= minblocks) ); |
|
{ |
|
|
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
|
3359 |
|
|
3360 |
if (!(MVequal(pmv[0],backupMV)) ) |
if (num < minblocks) |
3361 |
{ |
{ |
3362 |
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
const int iEdgedWidth = pParam->edged_width; |
3363 |
x, y, |
num = 0; |
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); |
|
3364 |
|
|
3365 |
if (iSAD < iMinSAD) |
/* fprintf(stderr,"Warning! Unreliable GME (%d/%d blocks), falling back to translation.\n",num,MBh*MBw); |
3366 |
{ |
*/ |
3367 |
*currMV = newMV; |
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; |
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
3368 |
|
|
3369 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
if (!(current->motion_flags & XVID_ME_GME_REFINE)) |
3370 |
{ |
return gmc; |
|
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); |
|
3371 |
|
|
3372 |
if (iSAD < iMinSAD) |
for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */ |
3373 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */ |
3374 |
{ |
{ |
3375 |
*currMV = newMV; |
const int mbnum = mx + my * MBw; |
3376 |
iMinSAD = iSAD; |
MACROBLOCK *const pMB = &pMBs[mbnum]; |
3377 |
} |
const uint8_t *const pCur = current->image.y + 16*(my*iEdgedWidth + mx); |
3378 |
|
if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx ) |
3379 |
|
&& (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) ) |
3380 |
|
{ pMB->mcsel = 1; |
3381 |
|
gmc.duv[0].x += pMB->mvs[0].x; |
3382 |
|
gmc.duv[0].y += pMB->mvs[0].y; |
3383 |
|
num++; |
3384 |
} |
} |
3385 |
} |
} |
3386 |
|
|
3387 |
/*************** Choose best MV found **************/ |
if (gmc.duv[0].x) |
3388 |
|
gmc.duv[0].x /= num; |
3389 |
|
if (gmc.duv[0].y) |
3390 |
|
gmc.duv[0].y /= num; |
3391 |
|
} else { |
3392 |
|
|
3393 |
EPZS8_Terminate_with_Refine: |
gmc.duv[0].x=(int)(sol[0]+0.5); |
3394 |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
gmc.duv[0].y=(int)(sol[3]+0.5); |
|
iMinSAD = Halfpel8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
currMV, iMinSAD, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
|
3395 |
|
|
3396 |
EPZS8_Terminate_without_Refine: |
gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5); |
3397 |
|
gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5); |
3398 |
|
|
3399 |
currPMV->x = currMV->x - pmv[0].x; |
gmc.duv[2].x=-gmc.duv[1].y; /* two warp points only */ |
3400 |
currPMV->y = currMV->y - pmv[0].y; |
gmc.duv[2].y=gmc.duv[1].x; |
3401 |
return iMinSAD; |
} |
3402 |
|
if (num>maxblocks) |
3403 |
|
{ for (my = 1; my < (uint32_t)MBh-1; my++) |
3404 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
3405 |
|
{ |
3406 |
|
const int mbnum = mx + my * MBw; |
3407 |
|
if (pMBs[mbnum-1].mcsel) |
3408 |
|
pMBs[mbnum].mcsel=0; |
3409 |
|
else |
3410 |
|
if (pMBs[mbnum-MBw].mcsel) |
3411 |
|
pMBs[mbnum].mcsel=0; |
3412 |
|
} |
3413 |
|
} |
3414 |
|
return gmc; |
3415 |
} |
} |
3416 |
|
|
3417 |
|
int |
3418 |
|
GlobalMotionEstRefine( |
3419 |
|
WARPPOINTS *const startwp, |
3420 |
|
MACROBLOCK * const pMBs, |
3421 |
|
const MBParam * const pParam, |
3422 |
|
const FRAMEINFO * const current, |
3423 |
|
const FRAMEINFO * const reference, |
3424 |
|
const IMAGE * const pCurr, |
3425 |
|
const IMAGE * const pRef, |
3426 |
|
const IMAGE * const pRefH, |
3427 |
|
const IMAGE * const pRefV, |
3428 |
|
const IMAGE * const pRefHV) |
3429 |
|
{ |
3430 |
|
uint8_t* GMCblock = (uint8_t*)malloc(16*pParam->edged_width); |
3431 |
|
WARPPOINTS bestwp=*startwp; |
3432 |
|
WARPPOINTS centerwp,currwp; |
3433 |
|
int gmcminSAD=0; |
3434 |
|
int gmcSAD=0; |
3435 |
|
int direction; |
3436 |
|
// int mx,my; |
3437 |
|
|
3438 |
|
/* use many blocks... */ |
3439 |
|
/* for (my = 0; my < (uint32_t)pParam->mb_height; my++) |
3440 |
|
for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++) |
3441 |
|
{ |
3442 |
|
const int mbnum = mx + my * pParam->mb_width; |
3443 |
|
pMBs[mbnum].mcsel=1; |
3444 |
|
} |
3445 |
|
*/ |
3446 |
|
|
3447 |
|
/* or rather don't use too many blocks... */ |
|
|
|
|
/* *********************************************************** |
|
|
bvop motion estimation |
|
|
// TODO: need to incorporate prediction here (eg. sad += calc_delta_16) |
|
|
***************************************************************/ |
|
|
|
|
3448 |
/* |
/* |
3449 |
void MotionEstimationBVOP( |
for (my = 1; my < (uint32_t)MBh-1; my++) |
3450 |
MBParam * const pParam, |
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
|
FRAMEINFO * const frame, |
|
|
|
|
|
// forward (past) reference |
|
|
const MACROBLOCK * const f_mbs, |
|
|
const IMAGE * const f_ref, |
|
|
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) |
|
3451 |
{ |
{ |
3452 |
const uint32_t mb_width = pParam->mb_width; |
const int mbnum = mx + my * MBw; |
3453 |
const uint32_t mb_height = pParam->mb_height; |
if (MBmask[mbnum-1]) |
3454 |
const int32_t edged_width = pParam->edged_width; |
MBmask[mbnum-1]=0; |
3455 |
|
else |
3456 |
int32_t i,j; |
if (MBmask[mbnum-MBw]) |
3457 |
|
MBmask[mbnum-1]=0; |
|
int32_t f_sad16; |
|
|
int32_t b_sad16; |
|
|
int32_t i_sad16; |
|
|
int32_t d_sad16; |
|
|
int32_t best_sad; |
|
|
|
|
|
VECTOR pmv_dontcare; |
|
3458 |
|
|
|
// 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; |
|
3459 |
} |
} |
3460 |
|
*/ |
3461 |
|
gmcminSAD = globalSAD(&bestwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3462 |
|
|
3463 |
|
if ( (reference->coding_type == S_VOP) |
3464 |
|
&& ( (reference->warp.duv[1].x != bestwp.duv[1].x) |
3465 |
|
|| (reference->warp.duv[1].y != bestwp.duv[1].y) |
3466 |
|
|| (reference->warp.duv[0].x != bestwp.duv[0].x) |
3467 |
|
|| (reference->warp.duv[0].y != bestwp.duv[0].y) |
3468 |
|
|| (reference->warp.duv[2].x != bestwp.duv[2].x) |
3469 |
|
|| (reference->warp.duv[2].y != bestwp.duv[2].y) ) ) |
3470 |
|
{ |
3471 |
|
gmcSAD = globalSAD(&reference->warp, pParam, pMBs, |
3472 |
|
current, pRef, pCurr, GMCblock); |
3473 |
|
|
3474 |
// forward search |
if (gmcSAD < gmcminSAD) |
3475 |
f_sad16 = SEARCH16(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
{ bestwp = reference->warp; |
3476 |
&frame->image, |
gmcminSAD = gmcSAD; |
3477 |
i, j, |
} |
3478 |
frame->motion_flags, frame->quant, frame->fcode, |
} |
|
pParam, |
|
|
f_mbs, |
|
|
&mb->mvs[0], &pmv_dontcare); // ignore pmv |
|
|
|
|
|
// backward search |
|
|
b_sad16 = SEARCH16(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
&frame->image, |
|
|
i, j, |
|
|
frame->motion_flags, frame->quant, frame->bcode, |
|
|
pParam, |
|
|
b_mbs, |
|
|
&mb->b_mvs[0], &pmv_dontcare); // ignore pmv |
|
3479 |
|
|
3480 |
// interpolate search (simple, but effective) |
do { |
3481 |
i_sad16 = sad16bi_c( |
direction = 0; |
3482 |
frame->image.y + i*16 + j*16*edged_width, |
centerwp = bestwp; |
|
get_ref(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
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; |
|
3483 |
|
|
3484 |
|
currwp = centerwp; |
3485 |
|
|
3486 |
if (f_sad16 < b_sad16) |
currwp.duv[0].x--; |
3487 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3488 |
|
if (gmcSAD < gmcminSAD) |
3489 |
|
{ bestwp = currwp; |
3490 |
|
gmcminSAD = gmcSAD; |
3491 |
|
direction = 1; |
3492 |
|
} |
3493 |
|
else |
3494 |
{ |
{ |
3495 |
best_sad = f_sad16; |
currwp = centerwp; currwp.duv[0].x++; |
3496 |
mb->mode = MB_FORWARD; |
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3497 |
|
if (gmcSAD < gmcminSAD) |
3498 |
|
{ bestwp = currwp; |
3499 |
|
gmcminSAD = gmcSAD; |
3500 |
|
direction = 2; |
3501 |
|
} |
3502 |
|
} |
3503 |
|
if (direction) continue; |
3504 |
|
|
3505 |
|
currwp = centerwp; currwp.duv[0].y--; |
3506 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3507 |
|
if (gmcSAD < gmcminSAD) |
3508 |
|
{ bestwp = currwp; |
3509 |
|
gmcminSAD = gmcSAD; |
3510 |
|
direction = 4; |
3511 |
} |
} |
3512 |
else |
else |
3513 |
{ |
{ |
3514 |
best_sad = b_sad16; |
currwp = centerwp; currwp.duv[0].y++; |
3515 |
mb->mode = MB_BACKWARD; |
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3516 |
|
if (gmcSAD < gmcminSAD) |
3517 |
|
{ bestwp = currwp; |
3518 |
|
gmcminSAD = gmcSAD; |
3519 |
|
direction = 8; |
3520 |
|
} |
3521 |
|
} |
3522 |
|
if (direction) continue; |
3523 |
|
|
3524 |
|
currwp = centerwp; currwp.duv[1].x++; |
3525 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3526 |
|
if (gmcSAD < gmcminSAD) |
3527 |
|
{ bestwp = currwp; |
3528 |
|
gmcminSAD = gmcSAD; |
3529 |
|
direction = 32; |
3530 |
|
} |
3531 |
|
currwp.duv[2].y++; |
3532 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3533 |
|
if (gmcSAD < gmcminSAD) |
3534 |
|
{ bestwp = currwp; |
3535 |
|
gmcminSAD = gmcSAD; |
3536 |
|
direction = 1024; |
3537 |
|
} |
3538 |
|
|
3539 |
|
currwp = centerwp; currwp.duv[1].x--; |
3540 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3541 |
|
if (gmcSAD < gmcminSAD) |
3542 |
|
{ bestwp = currwp; |
3543 |
|
gmcminSAD = gmcSAD; |
3544 |
|
direction = 16; |
3545 |
} |
} |
3546 |
|
else |
|
if (i_sad16 < best_sad) |
|
3547 |
{ |
{ |
3548 |
best_sad = i_sad16; |
currwp = centerwp; currwp.duv[1].x++; |
3549 |
mb->mode = MB_INTERPOLATE; |
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3550 |
|
if (gmcSAD < gmcminSAD) |
3551 |
|
{ bestwp = currwp; |
3552 |
|
gmcminSAD = gmcSAD; |
3553 |
|
direction = 32; |
3554 |
|
} |
3555 |
|
} |
3556 |
|
if (direction) continue; |
3557 |
|
|
3558 |
|
|
3559 |
|
currwp = centerwp; currwp.duv[1].y--; |
3560 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3561 |
|
if (gmcSAD < gmcminSAD) |
3562 |
|
{ bestwp = currwp; |
3563 |
|
gmcminSAD = gmcSAD; |
3564 |
|
direction = 64; |
3565 |
} |
} |
3566 |
|
else |
3567 |
if (d_sad16 < best_sad) |
{ |
3568 |
|
currwp = centerwp; currwp.duv[1].y++; |
3569 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3570 |
|
if (gmcSAD < gmcminSAD) |
3571 |
|
{ bestwp = currwp; |
3572 |
|
gmcminSAD = gmcSAD; |
3573 |
|
direction = 128; |
3574 |
|
} |
3575 |
|
} |
3576 |
|
if (direction) continue; |
3577 |
|
|
3578 |
|
currwp = centerwp; currwp.duv[2].x--; |
3579 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3580 |
|
if (gmcSAD < gmcminSAD) |
3581 |
|
{ bestwp = currwp; |
3582 |
|
gmcminSAD = gmcSAD; |
3583 |
|
direction = 256; |
3584 |
|
} |
3585 |
|
else |
3586 |
{ |
{ |
3587 |
best_sad = d_sad16; |
currwp = centerwp; currwp.duv[2].x++; |
3588 |
mb->mode = MB_DIRECT; |
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3589 |
|
if (gmcSAD < gmcminSAD) |
3590 |
|
{ bestwp = currwp; |
3591 |
|
gmcminSAD = gmcSAD; |
3592 |
|
direction = 512; |
3593 |
|
} |
3594 |
|
} |
3595 |
|
if (direction) continue; |
3596 |
|
|
3597 |
|
currwp = centerwp; currwp.duv[2].y--; |
3598 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3599 |
|
if (gmcSAD < gmcminSAD) |
3600 |
|
{ bestwp = currwp; |
3601 |
|
gmcminSAD = gmcSAD; |
3602 |
|
direction = 1024; |
3603 |
} |
} |
3604 |
|
else |
3605 |
|
{ |
3606 |
|
currwp = centerwp; currwp.duv[2].y++; |
3607 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3608 |
|
if (gmcSAD < gmcminSAD) |
3609 |
|
{ bestwp = currwp; |
3610 |
|
gmcminSAD = gmcSAD; |
3611 |
|
direction = 2048; |
3612 |
|
} |
3613 |
|
} |
3614 |
|
} while (direction); |
3615 |
|
free(GMCblock); |
3616 |
|
|
3617 |
|
*startwp = bestwp; |
3618 |
|
|
3619 |
|
return gmcminSAD; |
3620 |
} |
} |
3621 |
|
|
3622 |
|
int |
3623 |
|
globalSAD(const WARPPOINTS *const wp, |
3624 |
|
const MBParam * const pParam, |
3625 |
|
const MACROBLOCK * const pMBs, |
3626 |
|
const FRAMEINFO * const current, |
3627 |
|
const IMAGE * const pRef, |
3628 |
|
const IMAGE * const pCurr, |
3629 |
|
uint8_t *const GMCblock) |
3630 |
|
{ |
3631 |
|
NEW_GMC_DATA gmc_data; |
3632 |
|
int iSAD, gmcSAD=0; |
3633 |
|
int num=0; |
3634 |
|
unsigned int mx, my; |
3635 |
|
|
3636 |
|
generate_GMCparameters( 3, 3, wp, pParam->width, pParam->height, &gmc_data); |
3637 |
|
|
3638 |
|
for (my = 0; my < (uint32_t)pParam->mb_height; my++) |
3639 |
|
for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++) { |
3640 |
|
|
3641 |
|
const int mbnum = mx + my * pParam->mb_width; |
3642 |
|
const int iEdgedWidth = pParam->edged_width; |
3643 |
|
|
3644 |
|
if (!pMBs[mbnum].mcsel) |
3645 |
|
continue; |
3646 |
|
|
3647 |
|
gmc_data.predict_16x16(&gmc_data, GMCblock, |
3648 |
|
pRef->y, |
3649 |
|
iEdgedWidth, |
3650 |
|
iEdgedWidth, |
3651 |
|
mx, my, |
3652 |
|
pParam->m_rounding_type); |
3653 |
|
|
3654 |
|
iSAD = sad16 ( pCurr->y + 16*(my*iEdgedWidth + mx), |
3655 |
|
GMCblock , iEdgedWidth, 65536); |
3656 |
|
iSAD -= pMBs[mbnum].sad16; |
3657 |
|
|
3658 |
|
if (iSAD<0) |
3659 |
|
gmcSAD += iSAD; |
3660 |
|
num++; |
3661 |
} |
} |
3662 |
|
return gmcSAD; |
3663 |
} |
} |
3664 |
|
|
|
*/ |
|