1 |
/************************************************************************** |
/************************************************************************** |
2 |
* |
* |
3 |
* Modifications: |
* XVID MPEG-4 VIDEO CODEC |
4 |
|
* motion estimation |
5 |
* |
* |
6 |
* 01.05.2002 updated MotionEstimationBVOP |
* This program is an implementation of a part of one or more MPEG-4 |
7 |
* 25.04.2002 partial prevMB conversion |
* Video tools as specified in ISO/IEC 14496-2 standard. Those intending |
8 |
* 22.04.2002 remove some compile warning by chenm001 <chenm001@163.com> |
* to use this software module in hardware or software products are |
9 |
* 14.04.2002 added MotionEstimationBVOP() |
* advised that its use may infringe existing patents or copyrights, and |
10 |
* 02.04.2002 add EPZS(^2) as ME algorithm, use PMV_USESQUARES to choose between |
* any such use would be at such party's own risk. The original |
11 |
* EPZS and EPZS^2 |
* developer of this software module and his/her company, and subsequent |
12 |
* 08.02.2002 split up PMVfast into three routines: PMVFast, PMVFast_MainLoop |
* editors and their companies, will have no liability for use of this |
13 |
* PMVFast_Refine to support multiple searches with different start points |
* software or modifications or derivatives thereof. |
|
* 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. |
|
14 |
* |
* |
15 |
* Michael Militzer <isibaar@videocoding.de> |
* This program is free software; you can redistribute it and/or modify |
16 |
|
* it under the terms of the GNU General Public License as published by |
17 |
|
* the Free Software Foundation; either version 2 of the License, or |
18 |
|
* (at your option) any later version. |
19 |
* |
* |
20 |
**************************************************************************/ |
* This program is distributed in the hope that it will be useful, |
21 |
|
* but WITHOUT ANY WARRANTY; without even the implied warranty of |
22 |
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
23 |
|
* GNU General Public License for more details. |
24 |
|
* |
25 |
|
* You should have received a copy of the GNU General Public License |
26 |
|
* along with this program; if not, write to the Free Software |
27 |
|
* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
28 |
|
* |
29 |
|
*************************************************************************/ |
30 |
|
|
31 |
#include <assert.h> |
#include <assert.h> |
32 |
#include <stdio.h> |
#include <stdio.h> |
33 |
#include <stdlib.h> |
#include <stdlib.h> |
34 |
|
#include <string.h> // memcpy |
35 |
|
#include <math.h> // lrint |
36 |
|
|
37 |
#include "../encoder.h" |
#include "../encoder.h" |
38 |
#include "../utils/mbfunctions.h" |
#include "../utils/mbfunctions.h" |
39 |
#include "../prediction/mbprediction.h" |
#include "../prediction/mbprediction.h" |
40 |
#include "../global.h" |
#include "../global.h" |
41 |
#include "../utils/timer.h" |
#include "../utils/timer.h" |
42 |
|
#include "../image/interpolate8x8.h" |
43 |
|
#include "motion_est.h" |
44 |
#include "motion.h" |
#include "motion.h" |
45 |
#include "sad.h" |
#include "sad.h" |
46 |
|
#include "../utils/emms.h" |
47 |
|
|
48 |
// very large value |
#define INITIAL_SKIP_THRESH (10) |
49 |
#define MV_MAX_ERROR (4096 * 256) |
#define FINAL_SKIP_THRESH (50) |
50 |
|
#define MAX_SAD00_FOR_SKIP (20) |
51 |
// stop search if sdelta < THRESHOLD |
#define MAX_CHROMA_SAD_FOR_SKIP (22) |
52 |
#define MV16_THRESHOLD 192 |
|
53 |
#define MV8_THRESHOLD 56 |
#define CHECK_CANDIDATE(X,Y,D) { \ |
54 |
|
(*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } |
55 |
#define NEIGH_MOVE_THRESH 8 |
|
56 |
// how much a block's MV must differ from his neighbour |
static __inline uint32_t |
57 |
// to be search for INTER4V. The more, the faster... |
d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv) |
58 |
|
{ |
59 |
/* sad16(0,0) bias; mpeg4 spec suggests nb/2+1 */ |
int xb, yb; |
60 |
/* nb = vop pixels * 2^(bpp-8) */ |
x = qpel ? x<<1 : x; |
61 |
#define MV16_00_BIAS (128+1) |
y = qpel ? y<<1 : y; |
62 |
#define MV8_00_BIAS (0) |
if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); } |
63 |
|
|
64 |
/* INTER bias for INTER/INTRA decision; mpeg4 spec suggests 2*nb */ |
x -= pred.x; |
65 |
#define MV16_INTER_BIAS 512 |
y -= pred.y; |
66 |
|
|
67 |
/* Parameters which control inter/inter4v decision */ |
if (x) { |
68 |
#define IMV16X16 5 |
x = ABS(x); |
69 |
|
x += (1 << (iFcode - 1)) - 1; |
70 |
/* vector map (vlc delta size) smoother parameters */ |
x >>= (iFcode - 1); |
71 |
#define NEIGH_TEND_16X16 2 |
if (x > 32) x = 32; |
72 |
#define NEIGH_TEND_8X8 2 |
xb = mvtab[x] + iFcode; |
73 |
|
} else xb = 1; |
74 |
// fast ((A)/2)*2 |
|
75 |
#define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) |
if (y) { |
76 |
|
y = ABS(y); |
77 |
#define MVzero(A) ( ((A).x)==(0) && ((A).y)==(0) ) |
y += (1 << (iFcode - 1)) - 1; |
78 |
#define MVequal(A,B) ( ((A).x)==((B).x) && ((A).y)==((B).y) ) |
y >>= (iFcode - 1); |
79 |
|
if (y > 32) y = 32; |
80 |
int32_t PMVfastSearch16( |
yb = mvtab[y] + iFcode; |
81 |
const uint8_t * const pRef, |
} else yb = 1; |
82 |
const uint8_t * const pRefH, |
return xb + yb; |
83 |
const uint8_t * const pRefV, |
} |
84 |
const uint8_t * const pRefHV, |
|
85 |
const IMAGE * const pCur, |
static int32_t ChromaSAD2(int fx, int fy, int bx, int by, const SearchData * const data) |
86 |
const int x, const int y, |
{ |
87 |
const uint32_t MotionFlags, |
int sad; |
88 |
const uint32_t iQuant, |
const uint32_t stride = data->iEdgedWidth/2; |
89 |
const uint32_t iFcode, |
uint8_t * f_refu = data->RefQ, |
90 |
const MBParam * const pParam, |
* f_refv = data->RefQ + 8, |
91 |
const MACROBLOCK * const pMBs, |
* b_refu = data->RefQ + 16, |
92 |
const MACROBLOCK * const prevMBs, |
* b_refv = data->RefQ + 24; |
93 |
VECTOR * const currMV, |
|
94 |
VECTOR * const currPMV); |
switch (((fx & 1) << 1) | (fy & 1)) { |
95 |
|
case 0: |
96 |
int32_t EPZSSearch16( |
fx = fx / 2; fy = fy / 2; |
97 |
const uint8_t * const pRef, |
f_refu = (uint8_t*)data->RefCU + fy * stride + fx, stride; |
98 |
const uint8_t * const pRefH, |
f_refv = (uint8_t*)data->RefCV + fy * stride + fx, stride; |
99 |
const uint8_t * const pRefV, |
break; |
100 |
const uint8_t * const pRefHV, |
case 1: |
101 |
const IMAGE * const pCur, |
fx = fx / 2; fy = (fy - 1) / 2; |
102 |
const int x, const int y, |
interpolate8x8_halfpel_v(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding); |
103 |
const uint32_t MotionFlags, |
interpolate8x8_halfpel_v(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding); |
104 |
const uint32_t iQuant, |
break; |
105 |
const uint32_t iFcode, |
case 2: |
106 |
const MBParam * const pParam, |
fx = (fx - 1) / 2; fy = fy / 2; |
107 |
const MACROBLOCK * const pMBs, |
interpolate8x8_halfpel_h(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding); |
108 |
const MACROBLOCK * const prevMBs, |
interpolate8x8_halfpel_h(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding); |
109 |
VECTOR * const currMV, |
break; |
110 |
VECTOR * const currPMV); |
default: |
111 |
|
fx = (fx - 1) / 2; fy = (fy - 1) / 2; |
112 |
|
interpolate8x8_halfpel_hv(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding); |
113 |
int32_t PMVfastSearch8( |
interpolate8x8_halfpel_hv(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding); |
114 |
const uint8_t * const pRef, |
break; |
|
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); |
|
|
|
|
|
int32_t EPZSSearch8( |
|
|
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 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); |
|
|
|
|
|
|
|
|
typedef int32_t (MainSearch16Func)( |
|
|
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); |
|
|
|
|
|
typedef MainSearch16Func* MainSearch16FuncPtr; |
|
|
|
|
|
|
|
|
typedef int32_t (MainSearch8Func)( |
|
|
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); |
|
|
|
|
|
typedef MainSearch8Func* MainSearch8FuncPtr; |
|
|
|
|
|
static int32_t lambda_vec16[32] = /* rounded values for lambda param for weight of motion bits as in modified H.26L */ |
|
|
{ 0 ,(int)(1.00235+0.5), (int)(1.15582+0.5), (int)(1.31976+0.5), (int)(1.49591+0.5), (int)(1.68601+0.5), |
|
|
(int)(1.89187+0.5), (int)(2.11542+0.5), (int)(2.35878+0.5), (int)(2.62429+0.5), (int)(2.91455+0.5), |
|
|
(int)(3.23253+0.5), (int)(3.58158+0.5), (int)(3.96555+0.5), (int)(4.38887+0.5), (int)(4.85673+0.5), |
|
|
(int)(5.37519+0.5), (int)(5.95144+0.5), (int)(6.59408+0.5), (int)(7.31349+0.5), (int)(8.12242+0.5), |
|
|
(int)(9.03669+0.5), (int)(10.0763+0.5), (int)(11.2669+0.5), (int)(12.6426+0.5), (int)(14.2493+0.5), |
|
|
(int)(16.1512+0.5), (int)(18.442+0.5), (int)(21.2656+0.5), (int)(24.8580+0.5), (int)(29.6436+0.5), |
|
|
(int)(36.4949+0.5) }; |
|
|
|
|
|
static int32_t *lambda_vec8 = lambda_vec16; /* same table for INTER and INTER4V for now*/ |
|
|
|
|
|
|
|
|
|
|
|
// 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 |
|
|
}; |
|
|
|
|
|
|
|
|
static __inline uint32_t mv_bits(int32_t component, const uint32_t iFcode) |
|
|
{ |
|
|
if (component == 0) |
|
|
return 1; |
|
|
|
|
|
if (component < 0) |
|
|
component = -component; |
|
|
|
|
|
if (iFcode == 1) |
|
|
{ |
|
|
if (component > 32) |
|
|
component = 32; |
|
|
|
|
|
return mvtab[component] + 1; |
|
|
} |
|
|
|
|
|
component += (1 << (iFcode - 1)) - 1; |
|
|
component >>= (iFcode - 1); |
|
|
|
|
|
if (component > 32) |
|
|
component = 32; |
|
|
|
|
|
return mvtab[component] + 1 + iFcode - 1; |
|
|
} |
|
|
|
|
|
|
|
|
static __inline uint32_t calc_delta_16(const int32_t dx, const int32_t dy, const uint32_t iFcode, const uint32_t iQuant) |
|
|
{ |
|
|
return NEIGH_TEND_16X16 * lambda_vec16[iQuant] * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); |
|
115 |
} |
} |
116 |
|
|
117 |
static __inline uint32_t calc_delta_8(const int32_t dx, const int32_t dy, const uint32_t iFcode, const uint32_t iQuant) |
switch (((bx & 1) << 1) | (by & 1)) { |
118 |
|
case 0: |
119 |
{ |
bx = bx / 2; by = by / 2; |
120 |
return NEIGH_TEND_8X8 * lambda_vec8[iQuant] * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); |
b_refu = (uint8_t*)data->b_RefCU + by * stride + bx, stride; |
121 |
|
b_refv = (uint8_t*)data->b_RefCV + by * stride + bx, stride; |
122 |
|
break; |
123 |
|
case 1: |
124 |
|
bx = bx / 2; by = (by - 1) / 2; |
125 |
|
interpolate8x8_halfpel_v(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding); |
126 |
|
interpolate8x8_halfpel_v(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding); |
127 |
|
break; |
128 |
|
case 2: |
129 |
|
bx = (bx - 1) / 2; by = by / 2; |
130 |
|
interpolate8x8_halfpel_h(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding); |
131 |
|
interpolate8x8_halfpel_h(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding); |
132 |
|
break; |
133 |
|
default: |
134 |
|
bx = (bx - 1) / 2; by = (by - 1) / 2; |
135 |
|
interpolate8x8_halfpel_hv(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding); |
136 |
|
interpolate8x8_halfpel_hv(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding); |
137 |
|
break; |
138 |
} |
} |
139 |
|
|
140 |
|
sad = sad8bi(data->CurU, b_refu, f_refu, stride); |
141 |
|
sad += sad8bi(data->CurV, b_refv, f_refv, stride); |
142 |
|
|
143 |
|
return sad; |
|
|
|
|
|
|
|
#ifndef SEARCH16 |
|
|
#define SEARCH16 PMVfastSearch16 |
|
|
//#define SEARCH16 FullSearch16 |
|
|
//#define SEARCH16 EPZSSearch16 |
|
|
#endif |
|
|
|
|
|
#ifndef SEARCH8 |
|
|
#define SEARCH8 PMVfastSearch8 |
|
|
//#define SEARCH8 EPZSSearch8 |
|
|
#endif |
|
|
|
|
|
bool MotionEstimation( |
|
|
MBParam * const pParam, |
|
|
FRAMEINFO * const current, |
|
|
FRAMEINFO * const reference, |
|
|
const IMAGE * const pRefH, |
|
|
const IMAGE * const pRefV, |
|
|
const IMAGE * const pRefHV, |
|
|
const uint32_t iLimit) |
|
|
|
|
|
{ |
|
|
const uint32_t iWcount = pParam->mb_width; |
|
|
const uint32_t iHcount = pParam->mb_height; |
|
|
MACROBLOCK * const pMBs = current->mbs; |
|
|
MACROBLOCK * const prevMBs = reference->mbs; // previous frame |
|
|
|
|
|
const IMAGE * const pCurrent = ¤t->image; |
|
|
const IMAGE * const pRef = &reference->image; |
|
|
|
|
|
const VECTOR zeroMV = {0,0}; |
|
|
|
|
|
int32_t x, y; |
|
|
int32_t iIntra = 0; |
|
|
VECTOR pmv; |
|
|
|
|
|
if (sadInit) |
|
|
(*sadInit)(); |
|
|
|
|
|
for (y = 0; y < iHcount; y++) |
|
|
for (x = 0; x < iWcount; x++) |
|
|
{ |
|
|
MACROBLOCK* const pMB = &pMBs[x + y * iWcount]; |
|
|
|
|
|
pMB->sad16 = SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
|
|
x, y, current->motion_flags, current->quant, current->fcode, |
|
|
pParam, pMBs, prevMBs, &pMB->mv16, &pMB->pmvs[0]); |
|
144 |
} |
} |
145 |
|
|
|
for (y = 0; y < iHcount; y++) |
|
|
for (x = 0; x < iWcount; x++) |
|
|
{ |
|
|
MACROBLOCK* const pMB = &pMBs[x + y * iWcount]; |
|
146 |
|
|
147 |
if (0 < (pMB->sad16 - MV16_INTER_BIAS)) |
static int32_t |
148 |
|
ChromaSAD(int dx, int dy, const SearchData * const data) |
149 |
{ |
{ |
150 |
int32_t deviation; |
int sad; |
151 |
deviation = dev16(pCurrent->y + x*16 + y*16*pParam->edged_width, |
const uint32_t stride = data->iEdgedWidth/2; |
|
pParam->edged_width); |
|
152 |
|
|
153 |
if (deviation < (pMB->sad16 - MV16_INTER_BIAS)) |
if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently |
154 |
{ |
data->temp[5] = dx; data->temp[6] = dy; // backup |
|
pMB->mode = MODE_INTRA; |
|
|
pMB->mv16 = pMB->mvs[0] = pMB->mvs[1] |
|
|
= pMB->mvs[2] = pMB->mvs[3] = zeroMV; |
|
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] |
|
|
= pMB->sad8[2] = pMB->sad8[3] = 0; |
|
155 |
|
|
156 |
iIntra++; |
switch (((dx & 1) << 1) | (dy & 1)) { |
157 |
if (iIntra >= iLimit) |
case 0: |
158 |
return 1; |
dx = dx / 2; dy = dy / 2; |
159 |
|
sad = sad8(data->CurU, data->RefCU + dy * stride + dx, stride); |
160 |
|
sad += sad8(data->CurV, data->RefCV + dy * stride + dx, stride); |
161 |
|
break; |
162 |
|
case 1: |
163 |
|
dx = dx / 2; dy = (dy - 1) / 2; |
164 |
|
sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + (dy+1) * stride + dx, stride); |
165 |
|
sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + (dy+1) * stride + dx, stride); |
166 |
|
break; |
167 |
|
case 2: |
168 |
|
dx = (dx - 1) / 2; dy = dy / 2; |
169 |
|
sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + dy * stride + dx+1, stride); |
170 |
|
sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + dy * stride + dx+1, stride); |
171 |
|
break; |
172 |
|
default: |
173 |
|
dx = (dx - 1) / 2; dy = (dy - 1) / 2; |
174 |
|
interpolate8x8_halfpel_hv(data->RefQ, data->RefCU + dy * stride + dx, stride, data->rounding); |
175 |
|
sad = sad8(data->CurU, data->RefQ, stride); |
176 |
|
|
177 |
continue; |
interpolate8x8_halfpel_hv(data->RefQ, data->RefCV + dy * stride + dx, stride, data->rounding); |
178 |
|
sad += sad8(data->CurV, data->RefQ, stride); |
179 |
|
break; |
180 |
} |
} |
181 |
|
data->temp[7] = sad; //backup, part 2 |
182 |
|
return sad; |
183 |
} |
} |
184 |
|
|
185 |
if ( (current->global_flags & XVID_INTER4V) |
static __inline const uint8_t * |
186 |
&& (!(current->global_flags & XVID_LUMIMASKING) |
GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data) |
|
|| pMB->dquant == NO_CHANGE) ) |
|
|
{ |
|
|
int32_t neigh=0; |
|
|
|
|
|
if (x>0) |
|
|
{ neigh += abs((pMB->mv16.x)-((pMB-1)->mv16.x)); |
|
|
neigh += abs((pMB->mv16.y)-((pMB-1)->mv16.y)); |
|
|
} |
|
|
if (y>0) |
|
|
{ neigh += abs((pMB->mv16.x)-((pMB-iWcount)->mv16.x)); |
|
|
neigh += abs((pMB->mv16.y)-((pMB-iWcount)->mv16.y)); |
|
|
} |
|
|
if (x<(iWcount-1)) |
|
|
{ neigh += abs((pMB->mv16.x)-((pMB+1)->mv16.x)); |
|
|
neigh += abs((pMB->mv16.y)-((pMB+1)->mv16.y)); |
|
|
} |
|
|
if (y<(iHcount-1)) |
|
|
{ neigh += abs((pMB->mv16.x)-((pMB+iHcount)->mv16.x)); |
|
|
neigh += abs((pMB->mv16.y)-((pMB+iHcount)->mv16.y)); |
|
|
} |
|
|
|
|
|
if (neigh > NEIGH_MOVE_THRESH) |
|
|
{ |
|
|
int32_t sad8 = 129; //IMV16X16 * current->quant; |
|
|
|
|
|
if (sad8 < pMB->sad16) |
|
|
sad8 += pMB->sad8[0] |
|
|
= SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
|
|
2*x, 2*y, pMB->mv16.x, pMB->mv16.y, |
|
|
current->motion_flags, current->quant, current->fcode, |
|
|
pParam, pMBs, prevMBs, &pMB->mvs[0], &pMB->pmvs[0]); |
|
|
|
|
|
if (sad8 < pMB->sad16) |
|
|
sad8 += pMB->sad8[1] |
|
|
= SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
|
|
2*x+1, 2*y, pMB->mv16.x, pMB->mv16.y, |
|
|
current->motion_flags, current->quant, current->fcode, |
|
|
pParam, pMBs, prevMBs, &pMB->mvs[1], &pMB->pmvs[1]); |
|
|
|
|
|
if (sad8 < pMB->sad16) |
|
|
sad8 += pMB->sad8[2] |
|
|
= SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
|
|
2*x, 2*y+1, pMB->mv16.x, pMB->mv16.y, |
|
|
current->motion_flags, current->quant, current->fcode, |
|
|
pParam, pMBs, prevMBs, &pMB->mvs[2], &pMB->pmvs[2]); |
|
|
|
|
|
if (sad8 < pMB->sad16) |
|
|
sad8 += pMB->sad8[3] |
|
|
= SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
|
|
2*x+1, 2*y+1, pMB->mv16.x, pMB->mv16.y, |
|
|
current->motion_flags, current->quant, current->fcode, |
|
|
pParam, pMBs, prevMBs, &pMB->mvs[3], &pMB->pmvs[3]); |
|
|
|
|
|
/* decide: MODE_INTER or MODE_INTER4V |
|
|
mpeg4: if (sad8 < pMB->sad16 - nb/2+1) use_inter4v |
|
|
*/ |
|
|
|
|
|
if (sad8 < pMB->sad16) |
|
187 |
{ |
{ |
188 |
pMB->mode = MODE_INTER4V; |
// dir : 0 = forward, 1 = backward |
189 |
pMB->sad8[0] *= 4; |
switch ( (dir << 2) | ((x&1)<<1) | (y&1) ) { |
190 |
pMB->sad8[1] *= 4; |
case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth); |
191 |
pMB->sad8[2] *= 4; |
case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); |
192 |
pMB->sad8[3] *= 4; |
case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); |
193 |
continue; |
case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); |
194 |
|
case 4 : return data->bRef + x/2 + (y/2)*(data->iEdgedWidth); |
195 |
|
case 5 : return data->bRefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); |
196 |
|
case 6 : return data->bRefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); |
197 |
|
default : return data->bRefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); |
198 |
} |
} |
199 |
} |
} |
|
} |
|
|
|
|
|
pMB->mode = MODE_INTER; |
|
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mv16; |
|
|
pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = pMB->sad16; |
|
|
|
|
|
pmv = get_pmv(pMBs, x, y, pParam->mb_width, 0); |
|
|
// get_pmv has to be called again. |
|
|
// intra-decision and inter4v change predictors |
|
|
|
|
|
pMB->pmvs[0].x = pMB->mv16.x - pmv.x; |
|
|
pMB->pmvs[0].y = pMB->mv16.y - pmv.y; |
|
|
} |
|
|
|
|
|
return 0; |
|
|
} |
|
|
|
|
|
#define CHECK_MV16_ZERO {\ |
|
|
if ( (0 <= max_dx) && (0 >= min_dx) \ |
|
|
&& (0 <= max_dy) && (0 >= min_dy) ) \ |
|
|
{ \ |
|
|
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 < 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; } } \ |
|
|
} |
|
|
|
|
|
/* too slow and not fully functional at the moment */ |
|
|
/* |
|
|
int32_t ZeroSearch16( |
|
|
const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const IMAGE * const pCur, |
|
|
const int x, const int y, |
|
|
const uint32_t MotionFlags, |
|
|
const uint32_t iQuant, |
|
|
const uint32_t iFcode, |
|
|
MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const MACROBLOCK * const prevMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV) |
|
|
{ |
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
|
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
|
|
int32_t iSAD; |
|
|
int32_t pred_x,pred_y; |
|
|
|
|
|
get_pmv(pMBs, x, y, pParam->mb_width, 0, &pred_x, &pred_y); |
|
|
|
|
|
iSAD = sad16( cur, |
|
|
get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0,0, iEdgedWidth), |
|
|
iEdgedWidth, MV_MAX_ERROR); |
|
|
if (iSAD <= iQuant * 96) |
|
|
iSAD -= MV16_00_BIAS; |
|
|
|
|
|
currMV->x = 0; |
|
|
currMV->y = 0; |
|
|
currPMV->x = -pred_x; |
|
|
currPMV->y = -pred_y; |
|
|
|
|
|
return iSAD; |
|
|
|
|
|
} |
|
|
*/ |
|
200 |
|
|
201 |
int32_t Diamond16_MainSearch( |
// this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate |
202 |
const uint8_t * const pRef, |
static __inline const uint8_t * |
203 |
const uint8_t * const pRefH, |
GetReference(const int x, const int y, const SearchData * const data) |
204 |
const uint8_t * const pRefV, |
{ |
205 |
const uint8_t * const pRefHV, |
switch ( ((x&1)<<1) | (y&1) ) { |
206 |
const uint8_t * const cur, |
case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth); |
207 |
const int x, const int y, |
case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); |
208 |
int32_t startx, int32_t starty, |
case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); |
209 |
int32_t iMinSAD, |
default : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); //case 2 |
|
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); |
|
210 |
} |
} |
|
else |
|
|
{ |
|
|
currMV->x = startx; |
|
|
currMV->y = starty; |
|
|
} |
|
|
return iMinSAD; |
|
211 |
} |
} |
212 |
|
|
213 |
int32_t Square16_MainSearch( |
static uint8_t * |
214 |
const uint8_t * const pRef, |
Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data) |
215 |
const uint8_t * const pRefH, |
{ |
216 |
const uint8_t * const pRefV, |
// create or find a qpel-precision reference picture; return pointer to it |
217 |
const uint8_t * const pRefHV, |
uint8_t * Reference = data->RefQ + 16*dir; |
218 |
const uint8_t * const cur, |
const uint32_t iEdgedWidth = data->iEdgedWidth; |
219 |
const int x, const int y, |
const uint32_t rounding = data->rounding; |
220 |
int32_t startx, int32_t starty, |
const int halfpel_x = x/2; |
221 |
int32_t iMinSAD, |
const int halfpel_y = y/2; |
222 |
VECTOR * const currMV, |
const uint8_t *ref1, *ref2, *ref3, *ref4; |
223 |
const VECTOR * const pmv, |
|
224 |
const int32_t min_dx, const int32_t max_dx, |
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
225 |
const int32_t min_dy, const int32_t max_dy, |
ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
226 |
const int32_t iEdgedWidth, |
switch( ((x&1)<<1) + (y&1) ) { |
227 |
const int32_t iDiamondSize, |
case 0: // pure halfpel position |
228 |
const int32_t iFcode, |
return (uint8_t *) ref1; |
|
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 |
|
|
*/ |
|
|
|
|
|
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); |
|
|
|
|
|
|
|
|
if (iDirection) |
|
|
while (!iFound) |
|
|
{ |
|
|
iFound = 1; |
|
|
backupMV=*currMV; |
|
|
|
|
|
switch (iDirection) |
|
|
{ |
|
|
case 1: |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
|
|
break; |
|
|
case 2: |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
|
229 |
break; |
break; |
230 |
|
|
231 |
case 3: |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
232 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
233 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
234 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
235 |
break; |
break; |
236 |
|
|
237 |
case 4: |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
238 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
239 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
240 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
241 |
break; |
break; |
242 |
|
|
243 |
case 5: |
default: // x and y in qpel resolution - the "corners" (top left/right and |
244 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
// bottom left/right) during qpel refinement |
245 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
246 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
247 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
248 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
249 |
|
ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
250 |
|
ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
251 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
252 |
break; |
break; |
253 |
|
} |
254 |
|
return Reference; |
255 |
|
} |
256 |
|
|
257 |
case 6: |
static uint8_t * |
258 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data) |
259 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
{ |
260 |
|
// create or find a qpel-precision reference picture; return pointer to it |
261 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
uint8_t * Reference = data->RefQ + 16*dir; |
262 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
const uint32_t iEdgedWidth = data->iEdgedWidth; |
263 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
const uint32_t rounding = data->rounding; |
264 |
|
const int halfpel_x = x/2; |
265 |
|
const int halfpel_y = y/2; |
266 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
267 |
|
|
268 |
|
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
269 |
|
switch( ((x&1)<<1) + (y&1) ) { |
270 |
|
case 3: // x and y in qpel resolution - the "corners" (top left/right and |
271 |
|
// bottom left/right) during qpel refinement |
272 |
|
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
273 |
|
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
274 |
|
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
275 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
276 |
|
interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); |
277 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); |
278 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); |
279 |
break; |
break; |
280 |
|
|
281 |
case 7: |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
282 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
283 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
284 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
285 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); |
286 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
287 |
break; |
break; |
288 |
|
|
289 |
case 8: |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
290 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
291 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
292 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
293 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); |
294 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
295 |
break; |
break; |
296 |
default: |
|
297 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
case 0: // pure halfpel position |
298 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
return (uint8_t *) ref1; |
|
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; |
|
299 |
} |
} |
300 |
return iMinSAD; |
return Reference; |
301 |
} |
} |
302 |
|
|
303 |
|
/* CHECK_CANDIATE FUNCTIONS START */ |
304 |
|
|
305 |
int32_t Full16_MainSearch( |
static void |
306 |
const uint8_t * const pRef, |
CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
307 |
const uint8_t * const pRefH, |
{ |
308 |
const uint8_t * const pRefV, |
int xc, yc; |
309 |
const uint8_t * const pRefHV, |
const uint8_t * Reference; |
310 |
const uint8_t * const cur, |
VECTOR * current; |
311 |
const int x, const int y, |
int32_t sad; uint32_t t; |
|
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); |
|
|
|
|
|
return iMinSAD; |
|
|
} |
|
312 |
|
|
313 |
int32_t Full8_MainSearch( |
if ( (x > data->max_dx) || (x < data->min_dx) |
314 |
const uint8_t * const pRef, |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
|
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); |
|
315 |
|
|
316 |
return iMinSAD; |
if (!data->qpel_precision) { |
317 |
|
Reference = GetReference(x, y, data); |
318 |
|
current = data->currentMV; |
319 |
|
xc = x; yc = y; |
320 |
|
} else { // x and y are in 1/4 precision |
321 |
|
Reference = Interpolate16x16qpel(x, y, 0, data); |
322 |
|
xc = x/2; yc = y/2; //for chroma sad |
323 |
|
current = data->currentQMV; |
324 |
} |
} |
325 |
|
|
326 |
|
sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
327 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
328 |
|
|
329 |
|
sad += (data->lambda16 * t * sad)>>10; |
330 |
|
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; |
331 |
|
|
332 |
int32_t Halfpel16_Refine( |
if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
333 |
const uint8_t * const pRef, |
(yc >> 1) + roundtab_79[yc & 0x3], data); |
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, const int y, |
|
|
VECTOR * const currMV, |
|
|
int32_t iMinSAD, |
|
|
const VECTOR * const pmv, |
|
|
const int32_t min_dx, const int32_t max_dx, |
|
|
const int32_t min_dy, const int32_t max_dy, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
const int32_t iEdgedWidth) |
|
|
{ |
|
|
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
|
|
|
|
|
int32_t iSAD; |
|
|
VECTOR backupMV = *currMV; |
|
334 |
|
|
335 |
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y-1); |
if (sad < data->iMinSAD[0]) { |
336 |
CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y-1); |
data->iMinSAD[0] = sad; |
337 |
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y-1); |
current[0].x = x; current[0].y = y; |
338 |
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y); |
*dir = Direction; |
|
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); |
|
|
|
|
|
return iMinSAD; |
|
339 |
} |
} |
340 |
|
|
341 |
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
if (data->temp[1] < data->iMinSAD[1]) { |
342 |
|
data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y = y; } |
343 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
344 |
|
data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; } |
345 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
346 |
|
data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; } |
347 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
348 |
|
data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; } |
349 |
|
|
350 |
|
} |
351 |
|
|
352 |
int32_t PMVfastSearch16( |
static void |
353 |
const uint8_t * const pRef, |
CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const IMAGE * const pCur, |
|
|
const int x, const int y, |
|
|
const uint32_t MotionFlags, |
|
|
const uint32_t iQuant, |
|
|
const uint32_t iFcode, |
|
|
const MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const MACROBLOCK * const prevMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV) |
|
354 |
{ |
{ |
355 |
const uint32_t iWcount = pParam->mb_width; |
int32_t sad; uint32_t t; |
356 |
const int32_t iWidth = pParam->width; |
const uint8_t * Reference; |
|
const int32_t iHeight = pParam->height; |
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
|
|
|
|
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
|
357 |
|
|
358 |
int32_t iDiamondSize; |
if ( (x > data->max_dx) || (x < data->min_dx) |
359 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
360 |
|
|
361 |
int32_t min_dx; |
if (!data->qpel_precision) Reference = GetReference(x, y, data); |
362 |
int32_t max_dx; |
else Reference = Interpolate8x8qpel(x, y, 0, 0, data); |
|
int32_t min_dy; |
|
|
int32_t max_dy; |
|
363 |
|
|
364 |
int32_t iFound; |
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
365 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
366 |
|
|
367 |
VECTOR newMV; |
sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10; |
|
VECTOR backupMV; /* just for PMVFAST */ |
|
368 |
|
|
369 |
VECTOR pmv[4]; |
if (sad < *(data->iMinSAD)) { |
370 |
int32_t psad[4]; |
*(data->iMinSAD) = sad; |
371 |
|
data->currentMV->x = x; data->currentMV->y = y; |
372 |
// const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
*dir = Direction; |
373 |
const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; |
} |
374 |
|
} |
375 |
|
|
|
static int32_t threshA,threshB; |
|
|
int32_t bPredEq; |
|
|
int32_t iMinSAD,iSAD; |
|
376 |
|
|
377 |
/* Get maximum range */ |
static void |
378 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
379 |
x, y, 16, iWidth, iHeight, iFcode); |
{ |
380 |
|
uint32_t t; |
381 |
|
const uint8_t * Reference; |
382 |
|
|
383 |
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero integer value |
384 |
|
(x > data->max_dx) || (x < data->min_dx) |
385 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
386 |
|
|
387 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
Reference = GetReference(x, y, data); |
388 |
{ min_dx = EVEN(min_dx); |
t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1); |
|
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; */ |
|
389 |
|
|
390 |
|
data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
391 |
|
|
392 |
bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10; |
393 |
|
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; |
394 |
|
|
395 |
if ((x==0) && (y==0) ) |
if (data->temp[0] < data->iMinSAD[0]) { |
396 |
{ |
data->iMinSAD[0] = data->temp[0]; |
397 |
threshA = 512; |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
398 |
threshB = 1024; |
*dir = Direction; } |
399 |
|
|
400 |
|
if (data->temp[1] < data->iMinSAD[1]) { |
401 |
|
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
402 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
403 |
|
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
404 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
405 |
|
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
406 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
407 |
|
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
408 |
} |
} |
409 |
else |
|
410 |
|
static void |
411 |
|
CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
412 |
{ |
{ |
413 |
threshA = psad[0]; |
int32_t sad, xc, yc; |
414 |
threshB = threshA+256; |
const uint8_t * Reference; |
415 |
if (threshA< 512) threshA = 512; |
uint32_t t; |
416 |
if (threshA>1024) threshA = 1024; |
VECTOR * current; |
|
if (threshB>1792) threshB = 1792; |
|
|
} |
|
417 |
|
|
418 |
iFound=0; |
if ( (x > data->max_dx) | ( x < data->min_dx) |
419 |
|
| (y > data->max_dy) | (y < data->min_dy) ) return; |
420 |
|
|
421 |
/* Step 4: Calculate SAD around the Median prediction. |
if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; //non-zero even value |
|
MinSAD=SAD |
|
|
If Motion Vector equal to Previous frame motion vector |
|
|
and MinSAD<PrevFrmSAD goto Step 10. |
|
|
If SAD<=256 goto Step 10. |
|
|
*/ |
|
422 |
|
|
423 |
*currMV=pmv[0]; /* current best := prediction */ |
if (data->qpel_precision) { // x and y are in 1/4 precision |
424 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
Reference = Interpolate16x16qpel(x, y, 0, data); |
425 |
{ /* This should NOT be necessary! */ |
current = data->currentQMV; |
426 |
currMV->x = EVEN(currMV->x); |
xc = x/2; yc = y/2; |
427 |
currMV->y = EVEN(currMV->y); |
} else { |
428 |
|
Reference = GetReference(x, y, data); |
429 |
|
current = data->currentMV; |
430 |
|
xc = x; yc = y; |
431 |
} |
} |
432 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, |
433 |
|
data->qpel^data->qpel_precision, data->rrv); |
434 |
|
|
435 |
if (currMV->x > max_dx) |
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
436 |
{ |
sad += (data->lambda16 * t * sad)>>10; |
|
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; |
|
|
} |
|
437 |
|
|
438 |
iMinSAD = sad16( cur, |
if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
439 |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
(yc >> 1) + roundtab_79[yc & 0x3], data); |
|
iEdgedWidth, MV_MAX_ERROR); |
|
|
iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode, iQuant); |
|
440 |
|
|
|
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,prevMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
|
|
{ |
|
|
if (iMinSAD < 2*iQuant) // high chances for SKIP-mode |
|
|
{ |
|
|
if (!MVzero(*currMV)) |
|
|
{ |
|
|
iMinSAD += MV16_00_BIAS; |
|
|
CHECK_MV16_ZERO; // (0,0) saves space for letterboxed pictures |
|
|
iMinSAD -= MV16_00_BIAS; |
|
|
} |
|
|
} |
|
441 |
|
|
442 |
if (MotionFlags & PMV_QUICKSTOP16) |
if (sad < *(data->iMinSAD)) { |
443 |
goto PMVfast16_Terminate_without_Refine; |
*(data->iMinSAD) = sad; |
444 |
if (MotionFlags & PMV_EARLYSTOP16) |
current->x = x; current->y = y; |
445 |
goto PMVfast16_Terminate_with_Refine; |
*dir = Direction; |
446 |
|
} |
447 |
} |
} |
448 |
|
|
449 |
|
static void |
450 |
|
CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
451 |
|
{ |
452 |
|
// maximum speed - for P/B/I decision |
453 |
|
int32_t sad; |
454 |
|
|
455 |
/* Step 2 (lazy eval): Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
if ( (x > data->max_dx) || (x < data->min_dx) |
456 |
vector of the median. |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
|
|
*/ |
|
|
|
|
|
if ((bPredEq) && (MVequal(pmv[0],prevMB->mvs[0]) ) ) |
|
|
iFound=2; |
|
457 |
|
|
458 |
/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
sad = sad32v_c(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), |
459 |
Otherwise select large Diamond Search. |
data->iEdgedWidth, data->temp+1); |
|
*/ |
|
460 |
|
|
461 |
if ( (!MVzero(pmv[0])) || (threshB<1536) || (bPredEq) ) |
if (sad < *(data->iMinSAD)) { |
462 |
iDiamondSize=1; // halfpel! |
*(data->iMinSAD) = sad; |
463 |
else |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
464 |
iDiamondSize=2; // halfpel! |
*dir = Direction; |
465 |
|
} |
466 |
|
if (data->temp[1] < data->iMinSAD[1]) { |
467 |
|
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
468 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
469 |
|
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
470 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
471 |
|
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
472 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
473 |
|
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
474 |
|
|
475 |
if (!(MotionFlags & PMV_HALFPELDIAMOND16) ) |
} |
|
iDiamondSize*=2; |
|
476 |
|
|
477 |
/* |
static void |
478 |
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) |
479 |
Also calculate (0,0) but do not subtract offset. |
{ |
480 |
Let MinSAD be the smallest SAD up to this point. |
int32_t sad, xb, yb, xcf, ycf, xcb, ycb; |
481 |
If MV is (0,0) subtract offset. |
uint32_t t; |
482 |
*/ |
const uint8_t *ReferenceF, *ReferenceB; |
483 |
|
VECTOR *current; |
484 |
|
|
485 |
// (0,0) is always possible |
if ( (xf > data->max_dx) | (xf < data->min_dx) |
486 |
|
| (yf > data->max_dy) | (yf < data->min_dy) ) return; |
487 |
|
|
488 |
if (!MVzero(pmv[0])) |
if (!data->qpel_precision) { |
489 |
CHECK_MV16_ZERO; |
ReferenceF = GetReference(xf, yf, data); |
490 |
|
xb = data->currentMV[1].x; yb = data->currentMV[1].y; |
491 |
|
ReferenceB = GetReferenceB(xb, yb, 1, data); |
492 |
|
current = data->currentMV; |
493 |
|
xcf = xf; ycf = yf; |
494 |
|
xcb = xb; ycb = yb; |
495 |
|
} else { |
496 |
|
ReferenceF = Interpolate16x16qpel(xf, yf, 0, data); |
497 |
|
xb = data->currentQMV[1].x; yb = data->currentQMV[1].y; |
498 |
|
current = data->currentQMV; |
499 |
|
ReferenceB = Interpolate16x16qpel(xb, yb, 1, data); |
500 |
|
xcf = xf/2; ycf = yf/2; |
501 |
|
xcb = xb/2; ycb = yb/2; |
502 |
|
} |
503 |
|
|
504 |
// previous frame MV is always possible |
t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0) |
505 |
|
+ d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
506 |
|
|
507 |
if (!MVzero(prevMB->mvs[0])) |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
508 |
if (!MVequal(prevMB->mvs[0],pmv[0])) |
sad += (data->lambda16 * t * sad)>>10; |
|
CHECK_MV16_CANDIDATE(prevMB->mvs[0].x,prevMB->mvs[0].y); |
|
509 |
|
|
510 |
// left neighbour, if allowed |
if (data->chroma) sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3], |
511 |
|
(ycf >> 1) + roundtab_79[ycf & 0x3], |
512 |
|
(xcb >> 1) + roundtab_79[xcb & 0x3], |
513 |
|
(ycb >> 1) + roundtab_79[ycb & 0x3], data); |
514 |
|
|
515 |
if (!MVzero(pmv[1])) |
if (sad < *(data->iMinSAD)) { |
516 |
if (!MVequal(pmv[1],prevMB->mvs[0])) |
*(data->iMinSAD) = sad; |
517 |
if (!MVequal(pmv[1],pmv[0])) |
current->x = xf; current->y = yf; |
518 |
{ |
*dir = Direction; |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
|
|
{ pmv[1].x = EVEN(pmv[1].x); |
|
|
pmv[1].y = EVEN(pmv[1].y); |
|
519 |
} |
} |
|
|
|
|
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
|
520 |
} |
} |
521 |
|
|
522 |
// top neighbour, if allowed |
static void |
523 |
if (!MVzero(pmv[2])) |
CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
if (!MVequal(pmv[2],prevMB->mvs[0])) |
|
|
if (!MVequal(pmv[2],pmv[0])) |
|
|
if (!MVequal(pmv[2],pmv[1])) |
|
524 |
{ |
{ |
525 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0; |
526 |
{ pmv[2].x = EVEN(pmv[2].x); |
uint32_t k; |
527 |
pmv[2].y = EVEN(pmv[2].y); |
const uint8_t *ReferenceF; |
528 |
} |
const uint8_t *ReferenceB; |
529 |
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
VECTOR mvs, b_mvs; |
530 |
|
|
531 |
// top right neighbour, if allowed |
if (( x > 31) | ( x < -32) | ( y > 31) | (y < -32)) return; |
|
if (!MVzero(pmv[3])) |
|
|
if (!MVequal(pmv[3],prevMB->mvs[0])) |
|
|
if (!MVequal(pmv[3],pmv[0])) |
|
|
if (!MVequal(pmv[3],pmv[1])) |
|
|
if (!MVequal(pmv[3],pmv[2])) |
|
|
{ |
|
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
|
|
{ pmv[3].x = EVEN(pmv[3].x); |
|
|
pmv[3].y = EVEN(pmv[3].y); |
|
|
} |
|
|
CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); |
|
|
} |
|
|
} |
|
532 |
|
|
533 |
if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96)*/ ) |
for (k = 0; k < 4; k++) { |
534 |
iMinSAD -= MV16_00_BIAS; |
mvs.x = data->directmvF[k].x + x; |
535 |
|
b_mvs.x = ((x == 0) ? |
536 |
|
data->directmvB[k].x |
537 |
|
: mvs.x - data->referencemv[k].x); |
538 |
|
|
539 |
|
mvs.y = data->directmvF[k].y + y; |
540 |
|
b_mvs.y = ((y == 0) ? |
541 |
|
data->directmvB[k].y |
542 |
|
: mvs.y - data->referencemv[k].y); |
543 |
|
|
544 |
/* Step 6: If MinSAD <= thresa goto Step 10. |
if ( (mvs.x > data->max_dx) | (mvs.x < data->min_dx) |
545 |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
| (mvs.y > data->max_dy) | (mvs.y < data->min_dy) |
546 |
*/ |
| (b_mvs.x > data->max_dx) | (b_mvs.x < data->min_dx) |
547 |
|
| (b_mvs.y > data->max_dy) | (b_mvs.y < data->min_dy) ) return; |
548 |
|
|
549 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,prevMB->mvs[0]) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
if (data->qpel) { |
550 |
{ |
xcf += mvs.x/2; ycf += mvs.y/2; |
551 |
if (MotionFlags & PMV_QUICKSTOP16) |
xcb += b_mvs.x/2; ycb += b_mvs.y/2; |
552 |
goto PMVfast16_Terminate_without_Refine; |
} else { |
553 |
if (MotionFlags & PMV_EARLYSTOP16) |
xcf += mvs.x; ycf += mvs.y; |
554 |
goto PMVfast16_Terminate_with_Refine; |
xcb += b_mvs.x; ycb += b_mvs.y; |
555 |
|
mvs.x *= 2; mvs.y *= 2; //we move to qpel precision anyway |
556 |
|
b_mvs.x *= 2; b_mvs.y *= 2; |
557 |
} |
} |
558 |
|
|
559 |
|
ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); |
560 |
|
ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); |
561 |
|
|
562 |
/************ (Diamond Search) **************/ |
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
563 |
/* |
ReferenceF, ReferenceB, data->iEdgedWidth); |
564 |
Step 7: Perform Diamond search, with either the small or large diamond. |
if (sad > *(data->iMinSAD)) return; |
|
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 */ |
|
|
|
|
|
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
|
|
iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
currMV->x, currMV->y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) |
|
|
{ |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
565 |
} |
} |
566 |
|
|
567 |
if (MotionFlags & PMV_EXTSEARCH16) |
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
|
{ |
|
|
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
|
568 |
|
|
569 |
if (!(MVequal(pmv[0],backupMV)) ) |
if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
570 |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
(ycf >> 3) + roundtab_76[ycf & 0xf], |
571 |
x, y, |
(xcb >> 3) + roundtab_76[xcb & 0xf], |
572 |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
(ycb >> 3) + roundtab_76[ycb & 0xf], data); |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
573 |
|
|
574 |
if (iSAD < iMinSAD) |
if (sad < *(data->iMinSAD)) { |
575 |
{ |
*(data->iMinSAD) = sad; |
576 |
*currMV = newMV; |
data->currentMV->x = x; data->currentMV->y = y; |
577 |
iMinSAD = iSAD; |
*dir = Direction; |
578 |
} |
} |
579 |
} |
} |
580 |
|
|
581 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
static void |
582 |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
x, y, |
|
|
0, 0, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) |
|
583 |
{ |
{ |
584 |
*currMV = newMV; |
int32_t sad, xcf, ycf, xcb, ycb; |
585 |
iMinSAD = iSAD; |
const uint8_t *ReferenceF; |
586 |
} |
const uint8_t *ReferenceB; |
587 |
} |
VECTOR mvs, b_mvs; |
|
} |
|
|
|
|
|
/* |
|
|
Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
|
|
*/ |
|
|
|
|
|
PMVfast16_Terminate_with_Refine: |
|
|
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); |
|
|
|
|
|
PMVfast16_Terminate_without_Refine: |
|
|
currPMV->x = currMV->x - pmv[0].x; |
|
|
currPMV->y = currMV->y - pmv[0].y; |
|
|
return iMinSAD; |
|
|
} |
|
|
|
|
588 |
|
|
589 |
|
if (( x > 31) | ( x < -32) | ( y > 31) | (y < -32)) return; |
590 |
|
|
591 |
|
mvs.x = data->directmvF[0].x + x; |
592 |
|
b_mvs.x = ((x == 0) ? |
593 |
|
data->directmvB[0].x |
594 |
|
: mvs.x - data->referencemv[0].x); |
595 |
|
|
596 |
|
mvs.y = data->directmvF[0].y + y; |
597 |
|
b_mvs.y = ((y == 0) ? |
598 |
|
data->directmvB[0].y |
599 |
|
: mvs.y - data->referencemv[0].y); |
600 |
|
|
601 |
|
if ( (mvs.x > data->max_dx) | (mvs.x < data->min_dx) |
602 |
|
| (mvs.y > data->max_dy) | (mvs.y < data->min_dy) |
603 |
|
| (b_mvs.x > data->max_dx) | (b_mvs.x < data->min_dx) |
604 |
|
| (b_mvs.y > data->max_dy) | (b_mvs.y < data->min_dy) ) return; |
605 |
|
|
606 |
int32_t Diamond8_MainSearch( |
if (data->qpel) { |
607 |
const uint8_t * const pRef, |
xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2); |
608 |
const uint8_t * const pRefH, |
xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2); |
609 |
const uint8_t * const pRefV, |
ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data); |
610 |
const uint8_t * const pRefHV, |
ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); |
611 |
const uint8_t * const cur, |
} else { |
612 |
const int x, const int y, |
xcf = 4*mvs.x; ycf = 4*mvs.y; |
613 |
int32_t startx, int32_t starty, |
xcb = 4*b_mvs.x; ycb = 4*b_mvs.y; |
614 |
int32_t iMinSAD, |
ReferenceF = GetReference(mvs.x, mvs.y, data); |
615 |
VECTOR * const currMV, |
ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data); |
|
const VECTOR * const pmv, |
|
|
const int32_t min_dx, const int32_t max_dx, |
|
|
const int32_t min_dy, const int32_t max_dy, |
|
|
const int32_t iEdgedWidth, |
|
|
const int32_t iDiamondSize, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
int iFound) |
|
|
{ |
|
|
/* Do a diamond search around given starting point, return SAD of best */ |
|
|
|
|
|
int32_t iDirection=0; |
|
|
int32_t iSAD; |
|
|
VECTOR backupMV; |
|
|
backupMV.x = startx; |
|
|
backupMV.y = starty; |
|
|
|
|
|
/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ |
|
|
|
|
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
|
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
|
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
|
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
|
|
|
|
|
if (iDirection) |
|
|
while (!iFound) |
|
|
{ |
|
|
iFound = 1; |
|
|
backupMV=*currMV; // since iDirection!=0, this is well defined! |
|
|
|
|
|
if ( iDirection != 2) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
|
|
if ( iDirection != 1) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x+iDiamondSize,backupMV.y,2); |
|
|
if ( iDirection != 4) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x,backupMV.y-iDiamondSize,3); |
|
|
if ( iDirection != 3) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x,backupMV.y+iDiamondSize,4); |
|
|
} |
|
|
else |
|
|
{ |
|
|
currMV->x = startx; |
|
|
currMV->y = starty; |
|
|
} |
|
|
return iMinSAD; |
|
616 |
} |
} |
617 |
|
|
618 |
int32_t Halfpel8_Refine( |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
619 |
const uint8_t * const pRef, |
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
|
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) */ |
|
620 |
|
|
621 |
int32_t iSAD; |
if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
622 |
VECTOR backupMV = *currMV; |
(ycf >> 3) + roundtab_76[ycf & 0xf], |
623 |
|
(xcb >> 3) + roundtab_76[xcb & 0xf], |
624 |
|
(ycb >> 3) + roundtab_76[ycb & 0xf], data); |
625 |
|
|
626 |
CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y-1); |
if (sad < *(data->iMinSAD)) { |
627 |
CHECK_MV8_CANDIDATE(backupMV.x ,backupMV.y-1); |
*(data->iMinSAD) = sad; |
628 |
CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y-1); |
data->currentMV->x = x; data->currentMV->y = y; |
629 |
CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y); |
*dir = Direction; |
630 |
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; |
|
631 |
} |
} |
632 |
|
|
633 |
|
/* CHECK_CANDIATE FUNCTIONS END */ |
634 |
|
|
635 |
#define PMV_HALFPEL8 (PMV_HALFPELDIAMOND8|PMV_HALFPELREFINE8) |
/* MAINSEARCH FUNCTIONS START */ |
636 |
|
|
637 |
int32_t PMVfastSearch8( |
static void |
638 |
const uint8_t * const pRef, |
AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection) |
|
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) |
|
639 |
{ |
{ |
|
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; |
|
|
|
|
|
VECTOR pmv[4]; |
|
|
int32_t psad[4]; |
|
|
VECTOR newMV; |
|
|
VECTOR backupMV; |
|
|
VECTOR startMV; |
|
640 |
|
|
641 |
// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
|
const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; |
|
642 |
|
|
643 |
static int32_t threshA,threshB; |
int iDirection; |
|
int32_t iFound,bPredEq; |
|
|
int32_t iMinSAD,iSAD; |
|
644 |
|
|
645 |
int32_t iSubBlock = (y&1)+(y&1) + (x&1); |
for(;;) { //forever |
646 |
|
iDirection = 0; |
647 |
|
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
648 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
649 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
650 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
651 |
|
|
652 |
/* Init variables */ |
/* now we're doing diagonal checks near our candidate */ |
|
startMV.x = start_x; |
|
|
startMV.y = start_y; |
|
653 |
|
|
654 |
/* Get maximum range */ |
if (iDirection) { //if anything found |
655 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
bDirection = iDirection; |
656 |
x, y, 8, iWidth, iHeight, iFcode); |
iDirection = 0; |
657 |
|
x = data->currentMV->x; y = data->currentMV->y; |
658 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8 )) |
if (bDirection & 3) { //our candidate is left or right |
659 |
{ min_dx = EVEN(min_dx); |
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
660 |
max_dx = EVEN(max_dx); |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
661 |
min_dy = EVEN(min_dy); |
} else { // what remains here is up or down |
662 |
max_dy = EVEN(max_dy); |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
663 |
} /* because we might use IF (dx>max_dx) THEN dx=max_dx; */ |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
664 |
|
} |
|
|
|
|
bPredEq = get_pmvdata(pMBs, (x>>1), (y>>1), iWcount, iSubBlock, pmv, psad); |
|
|
|
|
|
if ((x==0) && (y==0) ) |
|
|
{ |
|
|
threshA = 512/4; |
|
|
threshB = 1024/4; |
|
665 |
|
|
666 |
|
if (iDirection) { |
667 |
|
bDirection += iDirection; |
668 |
|
x = data->currentMV->x; y = data->currentMV->y; |
669 |
|
} |
670 |
|
} else { //about to quit, eh? not so fast.... |
671 |
|
switch (bDirection) { |
672 |
|
case 2: |
673 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
674 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
675 |
|
break; |
676 |
|
case 1: |
677 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
678 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
679 |
|
break; |
680 |
|
case 2 + 4: |
681 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
682 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
683 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
684 |
|
break; |
685 |
|
case 4: |
686 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
687 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
688 |
|
break; |
689 |
|
case 8: |
690 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
691 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
692 |
|
break; |
693 |
|
case 1 + 4: |
694 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
695 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
696 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
697 |
|
break; |
698 |
|
case 2 + 8: |
699 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
700 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
701 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
702 |
|
break; |
703 |
|
case 1 + 8: |
704 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
705 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
706 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
707 |
|
break; |
708 |
|
default: //1+2+4+8 == we didn't find anything at all |
709 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
710 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
711 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
712 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
713 |
|
break; |
714 |
|
} |
715 |
|
if (!iDirection) break; //ok, the end. really |
716 |
|
bDirection = iDirection; |
717 |
|
x = data->currentMV->x; y = data->currentMV->y; |
718 |
} |
} |
|
else |
|
|
{ |
|
|
threshA = psad[0]/4; /* good estimate */ |
|
|
threshB = threshA+256/4; |
|
|
if (threshA< 512/4) threshA = 512/4; |
|
|
if (threshA>1024/4) threshA = 1024/4; |
|
|
if (threshB>1792/4) threshB = 1792/4; |
|
719 |
} |
} |
|
|
|
|
iFound=0; |
|
|
|
|
|
/* 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. |
|
|
*/ |
|
|
|
|
|
|
|
|
// Prepare for main loop |
|
|
|
|
|
*currMV = startMV; |
|
|
|
|
|
iMinSAD = sad8( cur, |
|
|
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,prevMB->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; |
|
720 |
} |
} |
721 |
|
|
722 |
/* Step 2 (lazy eval): Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
static void |
723 |
vector of the median. |
SquareSearch(int x, int y, const SearchData * const data, int bDirection) |
724 |
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
{ |
725 |
*/ |
int iDirection; |
|
|
|
|
if ((bPredEq) && (MVequal(pmv[0],prevMB->mvs[iSubBlock]) ) ) |
|
|
iFound=2; |
|
726 |
|
|
727 |
/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
do { |
728 |
Otherwise select large Diamond Search. |
iDirection = 0; |
729 |
*/ |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64); |
730 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128); |
731 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32); |
732 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128); |
733 |
|
if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64); |
734 |
|
if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128); |
735 |
|
if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128); |
736 |
|
if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128); |
737 |
|
|
738 |
if ( (!MVzero(pmv[0])) || (threshB<1536/4) || (bPredEq) ) |
bDirection = iDirection; |
739 |
iDiamondSize=1; // 1 halfpel! |
x = data->currentMV->x; y = data->currentMV->y; |
740 |
else |
} while (iDirection); |
741 |
iDiamondSize=2; // 2 halfpel = 1 full pixel! |
} |
742 |
|
|
743 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8) ) |
static void |
744 |
iDiamondSize*=2; |
DiamondSearch(int x, int y, const SearchData * const data, int bDirection) |
745 |
|
{ |
746 |
|
|
747 |
|
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
748 |
|
|
749 |
/* |
int iDirection; |
|
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. |
|
|
*/ |
|
750 |
|
|
751 |
// the median prediction might be even better than mv16 |
do { |
752 |
|
iDirection = 0; |
753 |
|
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
754 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
755 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
756 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
757 |
|
|
758 |
if (!MVequal(pmv[0],startMV)) |
/* now we're doing diagonal checks near our candidate */ |
|
CHECK_MV8_CANDIDATE(pmv[0].x,pmv[0].y); |
|
759 |
|
|
760 |
// (0,0) if needed |
if (iDirection) { //checking if anything found |
761 |
if (!MVzero(pmv[0])) |
bDirection = iDirection; |
762 |
if (!MVzero(startMV)) |
iDirection = 0; |
763 |
CHECK_MV8_ZERO; |
x = data->currentMV->x; y = data->currentMV->y; |
764 |
|
if (bDirection & 3) { //our candidate is left or right |
765 |
|
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
766 |
|
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
767 |
|
} else { // what remains here is up or down |
768 |
|
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
769 |
|
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
770 |
|
} |
771 |
|
bDirection += iDirection; |
772 |
|
x = data->currentMV->x; y = data->currentMV->y; |
773 |
|
} |
774 |
|
} |
775 |
|
while (iDirection); |
776 |
|
} |
777 |
|
|
778 |
// previous frame MV if needed |
/* MAINSEARCH FUNCTIONS END */ |
|
if (!MVzero(prevMB->mvs[iSubBlock])) |
|
|
if (!MVequal(prevMB->mvs[iSubBlock],startMV)) |
|
|
if (!MVequal(prevMB->mvs[iSubBlock],pmv[0])) |
|
|
CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x,prevMB->mvs[iSubBlock].y); |
|
779 |
|
|
780 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,prevMB->mvs[iSubBlock]) && ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
static void |
781 |
|
SubpelRefine(const SearchData * const data) |
782 |
{ |
{ |
783 |
if (MotionFlags & PMV_QUICKSTOP16) |
/* Do a half-pel or q-pel refinement */ |
784 |
goto PMVfast8_Terminate_without_Refine; |
const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV; |
785 |
if (MotionFlags & PMV_EARLYSTOP16) |
int iDirection; //only needed because macro expects it |
786 |
goto PMVfast8_Terminate_with_Refine; |
|
787 |
|
CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0); |
788 |
|
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0); |
789 |
|
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0); |
790 |
|
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0); |
791 |
|
CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0); |
792 |
|
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0); |
793 |
|
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0); |
794 |
|
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0); |
795 |
} |
} |
796 |
|
|
797 |
|
static __inline int |
798 |
|
SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
799 |
|
const int x, const int y, |
800 |
|
const uint32_t stride, const uint32_t iQuant, int rrv) |
801 |
|
|
|
// left neighbour, if allowed and needed |
|
|
if (!MVzero(pmv[1])) |
|
|
if (!MVequal(pmv[1],startMV)) |
|
|
if (!MVequal(pmv[1],prevMB->mvs[iSubBlock])) |
|
|
if (!MVequal(pmv[1],pmv[0])) |
|
802 |
{ |
{ |
803 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
if(!rrv) { |
804 |
{ pmv[1].x = EVEN(pmv[1].x); |
uint32_t sadC = sad8(current->u + x*8 + y*stride*8, |
805 |
pmv[1].y = EVEN(pmv[1].y); |
reference->u + x*8 + y*stride*8, stride); |
806 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
807 |
|
sadC += sad8(current->v + (x + y*stride)*8, |
808 |
|
reference->v + (x + y*stride)*8, stride); |
809 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
810 |
|
return 1; |
811 |
|
|
812 |
|
} else { |
813 |
|
uint32_t sadC = sad16(current->u + x*16 + y*stride*16, |
814 |
|
reference->u + x*16 + y*stride*16, stride, 256*4096); |
815 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
816 |
|
sadC += sad16(current->v + (x + y*stride)*16, |
817 |
|
reference->v + (x + y*stride)*16, stride, 256*4096); |
818 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
819 |
|
return 1; |
820 |
} |
} |
|
CHECK_MV8_CANDIDATE(pmv[1].x,pmv[1].y); |
|
821 |
} |
} |
822 |
|
|
823 |
// top neighbour, if allowed and needed |
static __inline void |
824 |
if (!MVzero(pmv[2])) |
SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
|
if (!MVequal(pmv[2],startMV)) |
|
|
if (!MVequal(pmv[2],prevMB->mvs[iSubBlock])) |
|
|
if (!MVequal(pmv[2],pmv[0])) |
|
|
if (!MVequal(pmv[2],pmv[1])) |
|
825 |
{ |
{ |
826 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
pMB->mode = MODE_NOT_CODED; |
827 |
{ pmv[2].x = EVEN(pmv[2].x); |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV; |
828 |
pmv[2].y = EVEN(pmv[2].y); |
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV; |
829 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
830 |
} |
} |
|
CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); |
|
831 |
|
|
832 |
// top right neighbour, if allowed and needed |
bool |
833 |
if (!MVzero(pmv[3])) |
MotionEstimation(MBParam * const pParam, |
834 |
if (!MVequal(pmv[3],startMV)) |
FRAMEINFO * const current, |
835 |
if (!MVequal(pmv[3],prevMB->mvs[iSubBlock])) |
FRAMEINFO * const reference, |
836 |
if (!MVequal(pmv[3],pmv[0])) |
const IMAGE * const pRefH, |
837 |
if (!MVequal(pmv[3],pmv[1])) |
const IMAGE * const pRefV, |
838 |
if (!MVequal(pmv[3],pmv[2])) |
const IMAGE * const pRefHV, |
839 |
|
const uint32_t iLimit) |
840 |
{ |
{ |
841 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
MACROBLOCK *const pMBs = current->mbs; |
842 |
{ pmv[3].x = EVEN(pmv[3].x); |
const IMAGE *const pCurrent = ¤t->image; |
843 |
pmv[3].y = EVEN(pmv[3].y); |
const IMAGE *const pRef = &reference->image; |
844 |
|
|
845 |
|
uint32_t mb_width = pParam->mb_width; |
846 |
|
uint32_t mb_height = pParam->mb_height; |
847 |
|
const uint32_t iEdgedWidth = pParam->edged_width; |
848 |
|
|
849 |
|
uint32_t x, y; |
850 |
|
uint32_t iIntra = 0; |
851 |
|
int32_t InterBias, quant = current->quant, sad00; |
852 |
|
|
853 |
|
// some pre-initialized thingies for SearchP |
854 |
|
int32_t temp[8]; |
855 |
|
VECTOR currentMV[5]; |
856 |
|
VECTOR currentQMV[5]; |
857 |
|
int32_t iMinSAD[5]; |
858 |
|
SearchData Data; |
859 |
|
memset(&Data, 0, sizeof(SearchData)); |
860 |
|
Data.iEdgedWidth = iEdgedWidth; |
861 |
|
Data.currentMV = currentMV; |
862 |
|
Data.currentQMV = currentQMV; |
863 |
|
Data.iMinSAD = iMinSAD; |
864 |
|
Data.temp = temp; |
865 |
|
Data.iFcode = current->fcode; |
866 |
|
Data.rounding = pParam->m_rounding_type; |
867 |
|
Data.qpel = pParam->m_quarterpel; |
868 |
|
Data.chroma = current->motion_flags & PMV_CHROMA16; |
869 |
|
Data.rrv = current->global_flags & XVID_REDUCED; |
870 |
|
|
871 |
|
if ((current->global_flags & XVID_REDUCED)) { |
872 |
|
mb_width = (pParam->width + 31) / 32; |
873 |
|
mb_height = (pParam->height + 31) / 32; |
874 |
|
Data.qpel = Data.chroma = 0; |
875 |
|
} |
876 |
|
|
877 |
|
Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose) |
878 |
|
if (sadInit) (*sadInit) (); |
879 |
|
|
880 |
|
for (y = 0; y < mb_height; y++) { |
881 |
|
for (x = 0; x < mb_width; x++) { |
882 |
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
883 |
|
|
884 |
|
if (!Data.rrv) pMB->sad16 = |
885 |
|
sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16, |
886 |
|
pRef->y + (x + y * iEdgedWidth) * 16, |
887 |
|
pParam->edged_width, pMB->sad8 ); |
888 |
|
|
889 |
|
else pMB->sad16 = |
890 |
|
sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32, |
891 |
|
pRef->y + (x + y * iEdgedWidth) * 32, |
892 |
|
pParam->edged_width, pMB->sad8 ); |
893 |
|
|
894 |
|
if (Data.chroma) { |
895 |
|
Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8, |
896 |
|
pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2) |
897 |
|
+ sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8, |
898 |
|
pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); |
899 |
|
pMB->sad16 += Data.temp[7]; |
900 |
|
} |
901 |
|
|
902 |
|
sad00 = pMB->sad16; |
903 |
|
|
904 |
|
if (!(current->global_flags & XVID_LUMIMASKING)) { |
905 |
|
pMB->dquant = NO_CHANGE; |
906 |
|
} else { |
907 |
|
if (pMB->dquant != NO_CHANGE) { |
908 |
|
quant += DQtab[pMB->dquant]; |
909 |
|
if (quant > 31) quant = 31; |
910 |
|
else if (quant < 1) quant = 1; |
911 |
|
} |
912 |
|
} |
913 |
|
pMB->quant = current->quant; |
914 |
|
|
915 |
|
//initial skip decision |
916 |
|
/* no early skip for GMC (global vector = skip vector is unknown!) */ |
917 |
|
if (!(current->global_flags & XVID_GMC)) { /* no fast SKIP for S(GMC)-VOPs */ |
918 |
|
if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH * (Data.rrv ? 4:1) ) |
919 |
|
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) { |
920 |
|
SkipMacroblockP(pMB, sad00); |
921 |
|
continue; |
922 |
|
} |
923 |
} |
} |
924 |
CHECK_MV8_CANDIDATE(pmv[3].x,pmv[3].y); |
|
925 |
|
SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
926 |
|
y, current->motion_flags, pMB->quant, |
927 |
|
&Data, pParam, pMBs, reference->mbs, |
928 |
|
current->global_flags & XVID_INTER4V, pMB); |
929 |
|
|
930 |
|
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
931 |
|
if (!(current->global_flags & XVID_GMC)) { |
932 |
|
if ( (pMB->dquant == NO_CHANGE) && (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
933 |
|
&& ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1)) ) |
934 |
|
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) { |
935 |
|
SkipMacroblockP(pMB, sad00); |
936 |
|
continue; |
937 |
} |
} |
938 |
} |
} |
939 |
|
|
940 |
if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96) */ ) |
/* finally, intra decision */ |
941 |
iMinSAD -= MV8_00_BIAS; |
|
942 |
|
InterBias = MV16_INTER_BIAS; |
943 |
|
if (pMB->quant > 8) InterBias += 100 * (pMB->quant - 8); // to make high quants work |
944 |
|
if (y != 0) |
945 |
|
if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; |
946 |
|
if (x != 0) |
947 |
|
if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; |
948 |
|
|
949 |
|
if (Data.chroma) InterBias += 50; // to compensate bigger SAD |
950 |
|
if (Data.rrv) InterBias *= 4; |
951 |
|
|
952 |
/* Step 6: If MinSAD <= thresa goto Step 10. |
if (InterBias < pMB->sad16) { |
953 |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
int32_t deviation; |
954 |
*/ |
if (!Data.rrv) |
955 |
|
deviation = dev16(pCurrent->y + (x + y * iEdgedWidth) * 16, iEdgedWidth); |
956 |
|
else { |
957 |
|
deviation = dev16(pCurrent->y + (x + y * iEdgedWidth) * 32, iEdgedWidth) |
958 |
|
+ dev16(pCurrent->y + (x + y * iEdgedWidth) * 32 + 16, iEdgedWidth) |
959 |
|
+ dev16(pCurrent->y + (x + y * iEdgedWidth) * 32 + 16 * iEdgedWidth, iEdgedWidth) |
960 |
|
+ dev16(pCurrent->y + (x + y * iEdgedWidth) * 32 + 16 * (iEdgedWidth+1), iEdgedWidth); |
961 |
|
} |
962 |
|
if (deviation < (pMB->sad16 - InterBias)) { |
963 |
|
if (++iIntra >= iLimit) return 1; |
964 |
|
SkipMacroblockP(pMB, 0); //same thing |
965 |
|
pMB->mode = MODE_INTRA; |
966 |
|
} |
967 |
|
} |
968 |
|
} |
969 |
|
} |
970 |
|
|
971 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,prevMB->mvs[iSubBlock]) && ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
if (current->global_flags & XVID_GMC ) /* GMC only for S(GMC)-VOPs */ |
972 |
{ |
{ |
973 |
if (MotionFlags & PMV_QUICKSTOP16) |
current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV); |
974 |
goto PMVfast8_Terminate_without_Refine; |
} |
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfast8_Terminate_with_Refine; |
|
|
} |
|
|
|
|
|
/************ (Diamond Search) **************/ |
|
|
/* |
|
|
Step 7: Perform Diamond search, with either the small or large diamond. |
|
|
If Found=2 only examine one Diamond pattern, and afterwards goto step 10 |
|
|
Step 8: If small diamond, iterate small diamond search pattern until motion vector lies in the center of the diamond. |
|
|
If center then goto step 10. |
|
|
Step 9: If large diamond, iterate large diamond search pattern until motion vector lies in the center. |
|
|
Refine by using small diamond and goto step 10. |
|
|
*/ |
|
975 |
|
|
976 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
return 0; |
977 |
|
} |
978 |
|
|
|
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
|
|
iSAD = Diamond8_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
currMV->x, currMV->y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
979 |
|
|
980 |
if (iSAD < iMinSAD) |
static __inline int |
981 |
|
make_mask(const VECTOR * const pmv, const int i) |
982 |
{ |
{ |
983 |
*currMV = newMV; |
int mask = 255, j; |
984 |
iMinSAD = iSAD; |
for (j = 0; j < i; j++) { |
985 |
|
if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already |
986 |
|
if (pmv[i].x == pmv[j].x) { |
987 |
|
if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4; |
988 |
|
else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8; |
989 |
|
} else |
990 |
|
if (pmv[i].y == pmv[j].y) { |
991 |
|
if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1; |
992 |
|
else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2; |
993 |
|
} |
994 |
|
} |
995 |
|
return mask; |
996 |
} |
} |
997 |
|
|
998 |
if (MotionFlags & PMV_EXTSEARCH8) |
static __inline void |
999 |
|
PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, |
1000 |
|
int iHcount, const MACROBLOCK * const prevMB, int rrv) |
1001 |
{ |
{ |
|
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
|
1002 |
|
|
1003 |
if (!(MVequal(pmv[0],backupMV)) ) |
//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself |
1004 |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
if (rrv) { iWcount /= 2; iHcount /= 2; } |
|
x, y, |
|
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
1005 |
|
|
1006 |
if (iSAD < iMinSAD) |
if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour |
1007 |
{ |
pmv[5].x = EVEN(pmv[3].x); |
1008 |
*currMV = newMV; |
pmv[5].y = EVEN(pmv[3].y); |
1009 |
iMinSAD = iSAD; |
} else pmv[5].x = pmv[5].y = 0; |
|
} |
|
|
} |
|
1010 |
|
|
1011 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour |
1012 |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
else pmv[3].x = pmv[3].y = 0; |
|
x, y, |
|
|
0, 0, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
1013 |
|
|
1014 |
if (iSAD < iMinSAD) |
if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour |
1015 |
{ |
else pmv[4].x = pmv[4].y = 0; |
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
|
} |
|
1016 |
|
|
1017 |
/* Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
// [1] median prediction |
1018 |
By performing an optional local half-pixel search, we can refine this result even further. |
pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); |
|
*/ |
|
1019 |
|
|
1020 |
PMVfast8_Terminate_with_Refine: |
pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask |
|
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
|
|
iMinSAD = Halfpel8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
currMV, iMinSAD, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
|
1021 |
|
|
1022 |
|
pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
1023 |
|
pmv[2].y = EVEN(prevMB->mvs[0].y); |
1024 |
|
|
1025 |
PMVfast8_Terminate_without_Refine: |
if ((x < iWcount-1) && (y < iHcount-1)) { |
1026 |
currPMV->x = currMV->x - pmv[0].x; |
pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame |
1027 |
currPMV->y = currMV->y - pmv[0].y; |
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
1028 |
|
} else pmv[6].x = pmv[6].y = 0; |
1029 |
|
|
1030 |
return iMinSAD; |
if (rrv) { |
1031 |
|
int i; |
1032 |
|
for (i = 0; i < 7; i++) { |
1033 |
|
pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); |
1034 |
|
pmv[i].y = RRV_MV_SCALEUP(pmv[i].y); |
1035 |
|
} |
1036 |
|
} |
1037 |
} |
} |
1038 |
|
|
1039 |
int32_t EPZSSearch16( |
static void |
1040 |
const uint8_t * const pRef, |
SearchP(const IMAGE * const pRef, |
1041 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1042 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
1043 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
1044 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
1045 |
const int x, const int y, |
const int x, |
1046 |
|
const int y, |
1047 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
1048 |
const uint32_t iQuant, |
const uint32_t iQuant, |
1049 |
const uint32_t iFcode, |
SearchData * const Data, |
1050 |
const MBParam * const pParam, |
const MBParam * const pParam, |
1051 |
const MACROBLOCK * const pMBs, |
const MACROBLOCK * const pMBs, |
1052 |
const MACROBLOCK * const prevMBs, |
const MACROBLOCK * const prevMBs, |
1053 |
VECTOR * const currMV, |
int inter4v, |
1054 |
VECTOR * const currPMV) |
MACROBLOCK * const pMB) |
1055 |
{ |
{ |
|
const uint32_t iWcount = pParam->mb_width; |
|
|
const uint32_t iHcount = pParam->mb_height; |
|
1056 |
|
|
1057 |
const int32_t iWidth = pParam->width; |
int i, iDirection = 255, mask, threshA; |
1058 |
const int32_t iHeight = pParam->height; |
VECTOR pmv[7]; |
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
1059 |
|
|
1060 |
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1061 |
|
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
1062 |
|
|
1063 |
int32_t min_dx; |
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); |
|
int32_t max_dx; |
|
|
int32_t min_dy; |
|
|
int32_t max_dy; |
|
1064 |
|
|
1065 |
VECTOR newMV; |
Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache |
1066 |
VECTOR backupMV; |
i = Data->rrv ? 2 : 1; |
1067 |
|
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i; |
1068 |
|
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1069 |
|
Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1070 |
|
|
1071 |
VECTOR pmv[4]; |
Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16*i; |
1072 |
int32_t psad[8]; |
Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16*i; |
1073 |
|
Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16*i; |
1074 |
|
Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16*i; |
1075 |
|
Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1076 |
|
Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1077 |
|
|
1078 |
static MACROBLOCK * oldMBs = NULL; |
Data->lambda16 = lambda_vec16[iQuant]; |
1079 |
// const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
Data->lambda8 = lambda_vec8[iQuant]; |
1080 |
const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; |
Data->qpel_precision = 0; |
|
MACROBLOCK * oldMB = NULL; |
|
1081 |
|
|
1082 |
static int32_t thresh2; |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
|
int32_t bPredEq; |
|
|
int32_t iMinSAD,iSAD=9999; |
|
1083 |
|
|
1084 |
MainSearch16FuncPtr EPZSMainSearchPtr; |
for(i = 0; i < 5; i++) |
1085 |
|
Data->currentMV[i].x = Data->currentMV[i].y = 0; |
1086 |
|
|
1087 |
if (oldMBs == NULL) |
if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
1088 |
{ oldMBs = (MACROBLOCK*) calloc(iWcount*iHcount,sizeof(MACROBLOCK)); |
else Data->predMV = pmv[0]; |
|
// fprintf(stderr,"allocated %d bytes for oldMBs\n",iWcount*iHcount*sizeof(MACROBLOCK)); |
|
|
} |
|
|
oldMB = oldMBs + x + y * iWcount; |
|
1089 |
|
|
1090 |
/* Get maximum range */ |
i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0); |
1091 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10); |
1092 |
x, y, 16, iWidth, iHeight, iFcode); |
Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10); |
1093 |
|
Data->iMinSAD[2] = pMB->sad8[1]; |
1094 |
|
Data->iMinSAD[3] = pMB->sad8[2]; |
1095 |
|
Data->iMinSAD[4] = pMB->sad8[3]; |
1096 |
|
|
1097 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
if (x | y) { |
1098 |
{ min_dx = EVEN(min_dx); |
threshA = Data->temp[0]; // that's when we keep this SAD atm |
1099 |
max_dx = EVEN(max_dx); |
if (threshA < 512) threshA = 512; |
1100 |
min_dy = EVEN(min_dy); |
else if (threshA > 1024) threshA = 1024; |
1101 |
max_dy = EVEN(max_dy); |
} else threshA = 512; |
|
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
|
1102 |
|
|
1103 |
bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
1104 |
|
prevMBs + x + y * pParam->mb_width, Data->rrv); |
1105 |
|
|
1106 |
/* Step 4: Calculate SAD around the Median prediction. |
if (!Data->rrv) { |
1107 |
MinSAD=SAD |
if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16; |
1108 |
If Motion Vector equal to Previous frame motion vector |
else CheckCandidate = CheckCandidate16no4v; //for extra speed |
1109 |
and MinSAD<PrevFrmSAD goto Step 10. |
} else CheckCandidate = CheckCandidate32; |
|
If SAD<=256 goto Step 10. |
|
|
*/ |
|
1110 |
|
|
1111 |
// Prepare for main loop |
/* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/ |
1112 |
|
|
1113 |
*currMV=pmv[0]; /* current best := median prediction */ |
for (i = 1; i < 7; i++) { |
1114 |
if (!(MotionFlags & PMV_HALFPEL16)) |
if (!(mask = make_mask(pmv, i)) ) continue; |
1115 |
{ |
(*CheckCandidate)(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1116 |
currMV->x = EVEN(currMV->x); |
if (Data->iMinSAD[0] <= threshA) break; |
|
currMV->y = EVEN(currMV->y); |
|
1117 |
} |
} |
1118 |
|
|
1119 |
if (currMV->x > max_dx) |
if ((Data->iMinSAD[0] <= threshA) || |
1120 |
currMV->x=max_dx; |
(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
1121 |
if (currMV->x < min_dx) |
(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) |
1122 |
currMV->x=min_dx; |
inter4v = 0; |
1123 |
if (currMV->y > max_dy) |
else { |
|
currMV->y=max_dy; |
|
|
if (currMV->y < min_dy) |
|
|
currMV->y=min_dy; |
|
1124 |
|
|
1125 |
/***************** This is predictor SET A: only median prediction ******************/ |
MainSearchFunc * MainSearchPtr; |
1126 |
|
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1127 |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1128 |
|
else MainSearchPtr = DiamondSearch; |
1129 |
|
|
1130 |
iMinSAD = sad16( cur, |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
|
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
|
|
iEdgedWidth, MV_MAX_ERROR); |
|
|
iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode, iQuant); |
|
1131 |
|
|
1132 |
// thresh1 is fixed to 256 |
/* extended search, diamond starting in 0,0 and in prediction. |
1133 |
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV, prevMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
note that this search is/might be done in halfpel positions, |
1134 |
{ |
which makes it more different than the diamond above */ |
1135 |
if (MotionFlags & PMV_QUICKSTOP16) |
|
1136 |
goto EPZS16_Terminate_without_Refine; |
if (MotionFlags & PMV_EXTSEARCH16) { |
1137 |
if (MotionFlags & PMV_EARLYSTOP16) |
int32_t bSAD; |
1138 |
goto EPZS16_Terminate_with_Refine; |
VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; |
1139 |
|
if (Data->rrv) { |
1140 |
|
startMV.x = RRV_MV_SCALEUP(startMV.x); |
1141 |
|
startMV.y = RRV_MV_SCALEUP(startMV.y); |
1142 |
|
} else |
1143 |
|
if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? |
1144 |
|
startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); |
1145 |
|
if (!(MVequal(startMV, backupMV))) { |
1146 |
|
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1147 |
|
|
1148 |
|
(*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); |
1149 |
|
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
1150 |
|
if (bSAD < Data->iMinSAD[0]) { |
1151 |
|
Data->currentMV[0] = backupMV; |
1152 |
|
Data->iMinSAD[0] = bSAD; } |
1153 |
} |
} |
1154 |
|
|
1155 |
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
backupMV = Data->currentMV[0]; |
1156 |
|
if (MotionFlags & PMV_HALFPELREFINE16 && !Data->rrv) startMV.x = startMV.y = 1; |
1157 |
|
else startMV.x = startMV.y = 0; |
1158 |
|
if (!(MVequal(startMV, backupMV))) { |
1159 |
|
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1160 |
|
|
1161 |
// previous frame MV |
(*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); |
1162 |
CHECK_MV16_CANDIDATE(prevMB->mvs[0].x,prevMB->mvs[0].y); |
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
1163 |
|
if (bSAD < Data->iMinSAD[0]) { |
1164 |
|
Data->currentMV[0] = backupMV; |
1165 |
|
Data->iMinSAD[0] = bSAD; } |
1166 |
|
} |
1167 |
|
} |
1168 |
|
} |
1169 |
|
|
1170 |
// set threshhold based on Min of Prediction and SAD of collocated block |
if (MotionFlags & PMV_HALFPELREFINE16) SubpelRefine(Data); |
|
// CHECK_MV16 always uses iSAD for the SAD of last vector to check, so now iSAD is what we want |
|
1171 |
|
|
1172 |
if ((x==0) && (y==0) ) |
for(i = 0; i < 5; i++) { |
1173 |
{ |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors |
1174 |
thresh2 = 512; |
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
1175 |
} |
} |
|
else |
|
|
{ |
|
|
/* T_k = 1.2 * MIN(SAD_top,SAD_left,SAD_topleft,SAD_coll) +128; [Tourapis, 2002] */ |
|
1176 |
|
|
1177 |
thresh2 = MIN(psad[0],iSAD)*6/5 + 128; |
if (Data->qpel && MotionFlags & PMV_QUARTERPELREFINE16) { |
1178 |
|
Data->qpel_precision = 1; |
1179 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1180 |
|
pParam->width, pParam->height, Data->iFcode, 1, 0); |
1181 |
|
|
1182 |
|
SubpelRefine(Data); |
1183 |
} |
} |
1184 |
|
|
1185 |
// MV=(0,0) is often a good choice |
if (Data->iMinSAD[0] < (int32_t)iQuant * 30) inter4v = 0; |
1186 |
|
if (inter4v) { |
1187 |
|
SearchData Data8; |
1188 |
|
memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data |
1189 |
|
|
1190 |
CHECK_MV16_ZERO; |
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
1191 |
|
Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); |
1192 |
|
Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); |
1193 |
|
Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); |
1194 |
|
|
1195 |
|
if (Data->chroma) { |
1196 |
|
int sumx = 0, sumy = 0; |
1197 |
|
const int div = 1 + Data->qpel; |
1198 |
|
const VECTOR * const mv = Data->qpel ? pMB->qmvs : pMB->mvs; |
1199 |
|
|
1200 |
// left neighbour, if allowed |
for (i = 0; i < 4; i++) { |
1201 |
if (x != 0) |
sumx += mv[i].x / div; |
1202 |
{ |
sumy += mv[i].y / div; |
1203 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
} |
1204 |
{ pmv[1].x = EVEN(pmv[1].x); |
|
1205 |
pmv[1].y = EVEN(pmv[1].y); |
Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf], |
1206 |
|
(sumy >> 3) + roundtab_76[sumy & 0xf], Data); |
1207 |
} |
} |
|
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
|
1208 |
} |
} |
1209 |
|
|
1210 |
// top neighbour, if allowed |
if (Data->rrv) { |
1211 |
if (y != 0) |
Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); |
1212 |
{ |
Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
|
|
{ pmv[2].x = EVEN(pmv[2].x); |
|
|
pmv[2].y = EVEN(pmv[2].y); |
|
1213 |
} |
} |
|
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
|
1214 |
|
|
1215 |
// top right neighbour, if allowed |
if (!(inter4v) || |
1216 |
if ((uint32_t)x != (iWcount-1)) |
(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
1217 |
{ |
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
1218 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
// INTER MODE |
1219 |
{ pmv[3].x = EVEN(pmv[3].x); |
pMB->mode = MODE_INTER; |
1220 |
pmv[3].y = EVEN(pmv[3].y); |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
1221 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
1222 |
|
|
1223 |
|
if(Data->qpel) { |
1224 |
|
pMB->qmvs[0] = pMB->qmvs[1] |
1225 |
|
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
1226 |
|
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; |
1227 |
|
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; |
1228 |
|
} else { |
1229 |
|
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
1230 |
|
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
1231 |
} |
} |
1232 |
CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); |
} else { |
1233 |
|
// INTER4V MODE; all other things are already set in Search8 |
1234 |
|
pMB->mode = MODE_INTER4V; |
1235 |
|
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + |
1236 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant; |
1237 |
} |
} |
1238 |
} |
} |
1239 |
|
|
1240 |
/* Terminate if MinSAD <= T_2 |
static void |
1241 |
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
Search8(const SearchData * const OldData, |
1242 |
*/ |
const int x, const int y, |
1243 |
|
const uint32_t MotionFlags, |
1244 |
if ( (iMinSAD <= thresh2) |
const MBParam * const pParam, |
1245 |
|| ( MVequal(*currMV,prevMB->mvs[0]) && ((uint32_t)iMinSAD <= prevMB->sad16) ) ) |
MACROBLOCK * const pMB, |
1246 |
|
const MACROBLOCK * const pMBs, |
1247 |
|
const int block, |
1248 |
|
SearchData * const Data) |
1249 |
{ |
{ |
1250 |
if (MotionFlags & PMV_QUICKSTOP16) |
int i = 0; |
1251 |
goto EPZS16_Terminate_without_Refine; |
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
1252 |
if (MotionFlags & PMV_EARLYSTOP16) |
Data->currentMV = OldData->currentMV + 1 + block; |
1253 |
goto EPZS16_Terminate_with_Refine; |
Data->currentQMV = OldData->currentQMV + 1 + block; |
1254 |
|
|
1255 |
|
if(Data->qpel) { |
1256 |
|
Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
1257 |
|
if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y, |
1258 |
|
Data->predMV, Data->iFcode, 0, 0); |
1259 |
|
} else { |
1260 |
|
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
1261 |
|
if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y, |
1262 |
|
Data->predMV, Data->iFcode, 0, Data->rrv); |
1263 |
} |
} |
1264 |
|
|
1265 |
/***** predictor SET C: acceleration MV (new!), neighbours in prev. frame(new!) ****/ |
*(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10; |
1266 |
|
|
1267 |
backupMV = prevMB->mvs[0]; // collocated MV |
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8|PMV_QUARTERPELREFINE8)) { |
1268 |
backupMV.x += (prevMB->mvs[0].x - oldMB->mvs[0].x ); // acceleration X |
if (Data->rrv) i = 2; else i = 1; |
|
backupMV.y += (prevMB->mvs[0].y - oldMB->mvs[0].y ); // acceleration Y |
|
1269 |
|
|
1270 |
CHECK_MV16_CANDIDATE(backupMV.x,backupMV.y); |
Data->Ref = OldData->Ref + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1271 |
|
Data->RefH = OldData->RefH + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1272 |
|
Data->RefV = OldData->RefV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1273 |
|
Data->RefHV = OldData->RefHV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1274 |
|
|
1275 |
// left neighbour |
Data->Cur = OldData->Cur + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1276 |
if (x != 0) |
Data->qpel_precision = 0; |
|
CHECK_MV16_CANDIDATE((prevMB-1)->mvs[0].x,(prevMB-1)->mvs[0].y); |
|
1277 |
|
|
1278 |
// top neighbour |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
1279 |
if (y != 0) |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
|
CHECK_MV16_CANDIDATE((prevMB-iWcount)->mvs[0].x,(prevMB-iWcount)->mvs[0].y); |
|
1280 |
|
|
1281 |
// right neighbour, if allowed (this value is not written yet, so take it from pMB->mvs |
if (!Data->rrv) CheckCandidate = CheckCandidate8; |
1282 |
|
else CheckCandidate = CheckCandidate16no4v; |
1283 |
|
|
1284 |
if ((uint32_t)x != iWcount-1) |
if (MotionFlags & PMV_EXTSEARCH8) { |
1285 |
CHECK_MV16_CANDIDATE((prevMB+1)->mvs[0].x,(prevMB+1)->mvs[0].y); |
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
1286 |
|
|
1287 |
// bottom neighbour, dito |
MainSearchFunc *MainSearchPtr; |
1288 |
if ((uint32_t)y != iHcount-1) |
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
1289 |
CHECK_MV16_CANDIDATE((prevMB+iWcount)->mvs[0].x,(prevMB+iWcount)->mvs[0].y); |
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
1290 |
|
else MainSearchPtr = DiamondSearch; |
1291 |
|
|
1292 |
/* Terminate if MinSAD <= T_3 (here T_3 = T_2) */ |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1293 |
if (iMinSAD <= thresh2) |
|
1294 |
{ |
if(*(Data->iMinSAD) < temp_sad) { |
1295 |
if (MotionFlags & PMV_QUICKSTOP16) |
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
1296 |
goto EPZS16_Terminate_without_Refine; |
Data->currentQMV->y = 2 * Data->currentMV->y; |
1297 |
if (MotionFlags & PMV_EARLYSTOP16) |
} |
|
goto EPZS16_Terminate_with_Refine; |
|
1298 |
} |
} |
1299 |
|
|
1300 |
/************ (if Diamond Search) **************/ |
if (MotionFlags & PMV_HALFPELREFINE8) { |
1301 |
|
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
1302 |
|
|
1303 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
SubpelRefine(Data); // perform halfpel refine of current best vector |
1304 |
|
|
1305 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
if(*(Data->iMinSAD) < temp_sad) { // we have found a better match |
1306 |
|
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
1307 |
|
Data->currentQMV->y = 2 * Data->currentMV->y; |
1308 |
|
} |
1309 |
|
} |
1310 |
|
|
1311 |
if (MotionFlags & PMV_USESQUARES16) |
if (Data->qpel && MotionFlags & PMV_QUARTERPELREFINE8) { |
1312 |
EPZSMainSearchPtr = Square16_MainSearch; |
Data->qpel_precision = 1; |
1313 |
else |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
1314 |
EPZSMainSearchPtr = Diamond16_MainSearch; |
pParam->width, pParam->height, Data->iFcode, 1, 0); |
1315 |
|
SubpelRefine(Data); |
1316 |
|
} |
1317 |
|
} |
1318 |
|
|
1319 |
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
if (Data->rrv) { |
1320 |
x, y, |
Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x); |
1321 |
currMV->x, currMV->y, iMinSAD, &newMV, pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y); |
1322 |
2, iFcode, iQuant, 0); |
} |
1323 |
|
|
1324 |
if (iSAD < iMinSAD) |
if(Data->qpel) { |
1325 |
{ |
pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x; |
1326 |
*currMV = newMV; |
pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y; |
1327 |
iMinSAD = iSAD; |
pMB->qmvs[block] = *Data->currentQMV; |
1328 |
|
} else { |
1329 |
|
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
1330 |
|
pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; |
1331 |
} |
} |
1332 |
|
|
1333 |
|
pMB->mvs[block] = *Data->currentMV; |
1334 |
|
pMB->sad8[block] = 4 * *Data->iMinSAD; |
1335 |
|
} |
1336 |
|
|
1337 |
if (MotionFlags & PMV_EXTSEARCH16) |
/* motion estimation for B-frames */ |
|
{ |
|
|
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
|
1338 |
|
|
1339 |
if (!(MVequal(pmv[0],backupMV)) ) |
static __inline VECTOR |
1340 |
|
ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) |
1341 |
{ |
{ |
1342 |
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
/* the stupidiest function ever */ |
1343 |
x, y, |
return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]); |
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); |
|
1344 |
} |
} |
1345 |
|
|
1346 |
if (iSAD < iMinSAD) |
static void __inline |
1347 |
|
PreparePredictionsBF(VECTOR * const pmv, const int x, const int y, |
1348 |
|
const uint32_t iWcount, |
1349 |
|
const MACROBLOCK * const pMB, |
1350 |
|
const uint32_t mode_curr) |
1351 |
{ |
{ |
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
1352 |
|
|
1353 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
// [0] is prediction |
1354 |
{ |
pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
0, 0, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); |
|
1355 |
|
|
1356 |
if (iSAD < iMinSAD) |
pmv[1].x = pmv[1].y = 0; // [1] is zero |
|
{ |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
|
} |
|
1357 |
|
|
1358 |
/*************** Choose best MV found **************/ |
pmv[2] = ChoosePred(pMB, mode_curr); |
1359 |
|
pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); |
1360 |
|
|
1361 |
EPZS16_Terminate_with_Refine: |
if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour |
1362 |
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); |
1363 |
iMinSAD = Halfpel16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); |
1364 |
x, y, |
} else pmv[3].x = pmv[3].y = 0; |
|
currMV, iMinSAD, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
|
1365 |
|
|
1366 |
EPZS16_Terminate_without_Refine: |
if (y != 0) { |
1367 |
|
pmv[4] = ChoosePred(pMB-iWcount, mode_curr); |
1368 |
|
pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y); |
1369 |
|
} else pmv[4].x = pmv[4].y = 0; |
1370 |
|
|
1371 |
*oldMB = *prevMB; |
if (x != 0) { |
1372 |
|
pmv[5] = ChoosePred(pMB-1, mode_curr); |
1373 |
|
pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
1374 |
|
} else pmv[5].x = pmv[5].y = 0; |
1375 |
|
|
1376 |
currPMV->x = currMV->x - pmv[0].x; |
if (x != 0 && y != 0) { |
1377 |
currPMV->y = currMV->y - pmv[0].y; |
pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); |
1378 |
return iMinSAD; |
pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y); |
1379 |
|
} else pmv[6].x = pmv[6].y = 0; |
1380 |
} |
} |
1381 |
|
|
1382 |
|
|
1383 |
int32_t EPZSSearch8( |
/* search backward or forward */ |
1384 |
const uint8_t * const pRef, |
static void |
1385 |
|
SearchBF( const IMAGE * const pRef, |
1386 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1387 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
1388 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
1389 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
1390 |
const int x, const int y, |
const int x, const int y, |
|
const int start_x, const int start_y, |
|
1391 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
|
const uint32_t iQuant, |
|
1392 |
const uint32_t iFcode, |
const uint32_t iFcode, |
1393 |
const MBParam * const pParam, |
const MBParam * const pParam, |
1394 |
const MACROBLOCK * const pMBs, |
MACROBLOCK * const pMB, |
1395 |
const MACROBLOCK * const prevMBs, |
const VECTOR * const predMV, |
1396 |
VECTOR * const currMV, |
int32_t * const best_sad, |
1397 |
VECTOR * const currPMV) |
const int32_t mode_current, |
1398 |
|
SearchData * const Data) |
1399 |
{ |
{ |
|
/* Please not that EPZS might not be a good choice for 8x8-block motion search ! */ |
|
1400 |
|
|
1401 |
const uint32_t iWcount = pParam->mb_width; |
int i, iDirection = 255, mask; |
1402 |
const int32_t iWidth = pParam->width; |
VECTOR pmv[7]; |
1403 |
const int32_t iHeight = pParam->height; |
MainSearchFunc *MainSearchPtr; |
1404 |
const int32_t iEdgedWidth = pParam->edged_width; |
*Data->iMinSAD = MV_MAX_ERROR; |
1405 |
|
Data->iFcode = iFcode; |
1406 |
|
Data->qpel_precision = 0; |
1407 |
|
Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; // reset chroma-sad cache |
1408 |
|
|
1409 |
const uint8_t * cur = pCur->y + x*8 + y*8*iEdgedWidth; |
Data->Ref = pRef->y + (x + y * Data->iEdgedWidth) * 16; |
1410 |
|
Data->RefH = pRefH + (x + y * Data->iEdgedWidth) * 16; |
1411 |
|
Data->RefV = pRefV + (x + y * Data->iEdgedWidth) * 16; |
1412 |
|
Data->RefHV = pRefHV + (x + y * Data->iEdgedWidth) * 16; |
1413 |
|
Data->RefCU = pRef->u + (x + y * Data->iEdgedWidth/2) * 8; |
1414 |
|
Data->RefCV = pRef->v + (x + y * Data->iEdgedWidth/2) * 8; |
1415 |
|
|
1416 |
int32_t iDiamondSize=1; |
Data->predMV = *predMV; |
1417 |
|
|
1418 |
int32_t min_dx; |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1419 |
int32_t max_dx; |
pParam->width, pParam->height, iFcode - Data->qpel, 0, 0); |
|
int32_t min_dy; |
|
|
int32_t max_dy; |
|
1420 |
|
|
1421 |
VECTOR newMV; |
pmv[0] = Data->predMV; |
1422 |
VECTOR backupMV; |
if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } |
1423 |
|
|
1424 |
VECTOR pmv[4]; |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
|
int32_t psad[8]; |
|
1425 |
|
|
1426 |
const int32_t iSubBlock = ((y&1)<<1) + (x&1); |
Data->currentMV->x = Data->currentMV->y = 0; |
1427 |
|
CheckCandidate = CheckCandidate16no4v; |
1428 |
|
|
1429 |
// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
// main loop. checking all predictions |
1430 |
const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; |
for (i = 0; i < 7; i++) { |
1431 |
|
if (!(mask = make_mask(pmv, i)) ) continue; |
1432 |
int32_t bPredEq; |
CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1433 |
int32_t iMinSAD,iSAD=9999; |
} |
1434 |
|
|
1435 |
MainSearch8FuncPtr EPZSMainSearchPtr; |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1436 |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1437 |
|
else MainSearchPtr = DiamondSearch; |
1438 |
|
|
1439 |
/* Get maximum range */ |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
|
|
x, y, 8, iWidth, iHeight, iFcode); |
|
1440 |
|
|
1441 |
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
SubpelRefine(Data); |
1442 |
|
|
1443 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
if (Data->qpel && *Data->iMinSAD < *best_sad + 300) { |
1444 |
{ min_dx = EVEN(min_dx); |
Data->currentQMV->x = 2*Data->currentMV->x; |
1445 |
max_dx = EVEN(max_dx); |
Data->currentQMV->y = 2*Data->currentMV->y; |
1446 |
min_dy = EVEN(min_dy); |
Data->qpel_precision = 1; |
1447 |
max_dy = EVEN(max_dy); |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1448 |
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
pParam->width, pParam->height, iFcode, 1, 0); |
1449 |
|
SubpelRefine(Data); |
1450 |
|
} |
1451 |
|
|
1452 |
bPredEq = get_pmvdata(pMBs, x>>1, y>>1, iWcount, iSubBlock, pmv, psad); |
// three bits are needed to code backward mode. four for forward |
1453 |
|
|
1454 |
|
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
1455 |
|
else *Data->iMinSAD += 3 * Data->lambda16; |
1456 |
|
|
1457 |
/* Step 4: Calculate SAD around the Median prediction. |
if (*Data->iMinSAD < *best_sad) { |
1458 |
MinSAD=SAD |
*best_sad = *Data->iMinSAD; |
1459 |
If Motion Vector equal to Previous frame motion vector |
pMB->mode = mode_current; |
1460 |
and MinSAD<PrevFrmSAD goto Step 10. |
if (Data->qpel) { |
1461 |
If SAD<=256 goto Step 10. |
pMB->pmvs[0].x = Data->currentQMV->x - predMV->x; |
1462 |
*/ |
pMB->pmvs[0].y = Data->currentQMV->y - predMV->y; |
1463 |
|
if (mode_current == MODE_FORWARD) |
1464 |
|
pMB->qmvs[0] = *Data->currentQMV; |
1465 |
|
else |
1466 |
|
pMB->b_qmvs[0] = *Data->currentQMV; |
1467 |
|
} else { |
1468 |
|
pMB->pmvs[0].x = Data->currentMV->x - predMV->x; |
1469 |
|
pMB->pmvs[0].y = Data->currentMV->y - predMV->y; |
1470 |
|
} |
1471 |
|
if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; |
1472 |
|
else pMB->b_mvs[0] = *Data->currentMV; |
1473 |
|
} |
1474 |
|
|
1475 |
|
if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; |
1476 |
|
else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search |
1477 |
|
} |
1478 |
|
|
1479 |
|
static void |
1480 |
|
SkipDecisionB(const IMAGE * const pCur, |
1481 |
|
const IMAGE * const f_Ref, |
1482 |
|
const IMAGE * const b_Ref, |
1483 |
|
MACROBLOCK * const pMB, |
1484 |
|
const uint32_t x, const uint32_t y, |
1485 |
|
const SearchData * const Data) |
1486 |
|
{ |
1487 |
|
int dx = 0, dy = 0, b_dx = 0, b_dy = 0; |
1488 |
|
int32_t sum; |
1489 |
|
const int div = 1 + Data->qpel; |
1490 |
|
int k; |
1491 |
|
const uint32_t stride = Data->iEdgedWidth/2; |
1492 |
|
//this is not full chroma compensation, only it's fullpel approximation. should work though |
1493 |
|
|
1494 |
|
for (k = 0; k < 4; k++) { |
1495 |
|
dy += Data->directmvF[k].y / div; |
1496 |
|
dx += Data->directmvF[0].x / div; |
1497 |
|
b_dy += Data->directmvB[0].y / div; |
1498 |
|
b_dx += Data->directmvB[0].x / div; |
1499 |
|
} |
1500 |
|
|
1501 |
|
dy = (dy >> 3) + roundtab_76[dy & 0xf]; |
1502 |
|
dx = (dx >> 3) + roundtab_76[dx & 0xf]; |
1503 |
|
b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf]; |
1504 |
|
b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf]; |
1505 |
|
|
1506 |
|
sum = sad8bi(pCur->u + 8 * x + 8 * y * stride, |
1507 |
|
f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2, |
1508 |
|
b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, |
1509 |
|
stride); |
1510 |
|
|
1511 |
|
if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; //no skip |
1512 |
|
|
1513 |
|
sum += sad8bi(pCur->v + 8*x + 8 * y * stride, |
1514 |
|
f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2, |
1515 |
|
b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, |
1516 |
|
stride); |
1517 |
|
|
1518 |
|
if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) pMB->mode = MODE_DIRECT_NONE_MV; //skipped |
1519 |
|
} |
1520 |
|
|
1521 |
|
static __inline uint32_t |
1522 |
|
SearchDirect(const IMAGE * const f_Ref, |
1523 |
|
const uint8_t * const f_RefH, |
1524 |
|
const uint8_t * const f_RefV, |
1525 |
|
const uint8_t * const f_RefHV, |
1526 |
|
const IMAGE * const b_Ref, |
1527 |
|
const uint8_t * const b_RefH, |
1528 |
|
const uint8_t * const b_RefV, |
1529 |
|
const uint8_t * const b_RefHV, |
1530 |
|
const IMAGE * const pCur, |
1531 |
|
const int x, const int y, |
1532 |
|
const uint32_t MotionFlags, |
1533 |
|
const int32_t TRB, const int32_t TRD, |
1534 |
|
const MBParam * const pParam, |
1535 |
|
MACROBLOCK * const pMB, |
1536 |
|
const MACROBLOCK * const b_mb, |
1537 |
|
int32_t * const best_sad, |
1538 |
|
SearchData * const Data) |
1539 |
|
|
1540 |
|
{ |
1541 |
|
int32_t skip_sad; |
1542 |
|
int k = (x + Data->iEdgedWidth*y) * 16; |
1543 |
|
MainSearchFunc *MainSearchPtr; |
1544 |
|
|
1545 |
|
*Data->iMinSAD = 256*4096; |
1546 |
|
Data->Ref = f_Ref->y + k; |
1547 |
|
Data->RefH = f_RefH + k; |
1548 |
|
Data->RefV = f_RefV + k; |
1549 |
|
Data->RefHV = f_RefHV + k; |
1550 |
|
Data->bRef = b_Ref->y + k; |
1551 |
|
Data->bRefH = b_RefH + k; |
1552 |
|
Data->bRefV = b_RefV + k; |
1553 |
|
Data->bRefHV = b_RefHV + k; |
1554 |
|
Data->RefCU = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; |
1555 |
|
Data->RefCV = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; |
1556 |
|
Data->b_RefCU = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; |
1557 |
|
Data->b_RefCV = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; |
1558 |
|
|
1559 |
|
k = Data->qpel ? 4 : 2; |
1560 |
|
Data->max_dx = k * (pParam->width - x * 16); |
1561 |
|
Data->max_dy = k * (pParam->height - y * 16); |
1562 |
|
Data->min_dx = -k * (16 + x * 16); |
1563 |
|
Data->min_dy = -k * (16 + y * 16); |
1564 |
|
|
1565 |
|
Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs; |
1566 |
|
Data->qpel_precision = 0; |
1567 |
|
|
1568 |
|
for (k = 0; k < 4; k++) { |
1569 |
|
pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); |
1570 |
|
pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD; |
1571 |
|
pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); |
1572 |
|
pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; |
1573 |
|
|
1574 |
|
if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx) |
1575 |
|
| (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) { |
1576 |
|
|
1577 |
|
*best_sad = 256*4096; // in that case, we won't use direct mode |
1578 |
|
pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV" |
1579 |
|
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
1580 |
|
return 256*4096; |
1581 |
|
} |
1582 |
|
if (b_mb->mode != MODE_INTER4V) { |
1583 |
|
pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; |
1584 |
|
pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0]; |
1585 |
|
Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0]; |
1586 |
|
Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0]; |
1587 |
|
break; |
1588 |
|
} |
1589 |
|
} |
1590 |
|
|
1591 |
// Prepare for main loop |
CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v; |
1592 |
|
|
1593 |
|
(*CheckCandidate)(0, 0, 255, &k, Data); |
1594 |
|
|
1595 |
if (!(MotionFlags & PMV_HALFPEL8)) |
// initial (fast) skip decision |
1596 |
{ |
if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (2 + Data->chroma?1:0)) { |
1597 |
currMV->x = EVEN(currMV->x); |
//possible skip |
1598 |
currMV->y = EVEN(currMV->y); |
if (Data->chroma) { |
1599 |
|
pMB->mode = MODE_DIRECT_NONE_MV; |
1600 |
|
return *Data->iMinSAD; // skip. |
1601 |
|
} else { |
1602 |
|
SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data); |
1603 |
|
if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip. |
1604 |
|
} |
1605 |
} |
} |
1606 |
|
|
1607 |
if (currMV->x > max_dx) |
skip_sad = *Data->iMinSAD; |
|
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; |
|
|
|
|
|
/***************** This is predictor SET A: only median prediction ******************/ |
|
1608 |
|
|
1609 |
|
// DIRECT MODE DELTA VECTOR SEARCH. |
1610 |
|
// This has to be made more effective, but at the moment I'm happy it's running at all |
1611 |
|
|
1612 |
iMinSAD = sad8( cur, |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1613 |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1614 |
iEdgedWidth); |
else MainSearchPtr = DiamondSearch; |
|
iMinSAD += calc_delta_8(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode, iQuant); |
|
1615 |
|
|
1616 |
|
(*MainSearchPtr)(0, 0, Data, 255); |
1617 |
|
|
1618 |
// thresh1 is fixed to 256 |
SubpelRefine(Data); |
|
if (iMinSAD < 256/4 ) |
|
|
{ |
|
|
if (MotionFlags & PMV_QUICKSTOP8) |
|
|
goto EPZS8_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP8) |
|
|
goto EPZS8_Terminate_with_Refine; |
|
|
} |
|
1619 |
|
|
1620 |
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
*best_sad = *Data->iMinSAD; |
1621 |
|
|
1622 |
|
if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT; |
1623 |
|
else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
1624 |
|
|
1625 |
// MV=(0,0) is often a good choice |
pMB->pmvs[3] = *Data->currentMV; |
|
CHECK_MV8_ZERO; |
|
1626 |
|
|
1627 |
// previous frame MV |
for (k = 0; k < 4; k++) { |
1628 |
CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x,prevMB->mvs[iSubBlock].y); |
pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x; |
1629 |
|
pMB->b_mvs[k].x = ( (Data->currentMV->x == 0) |
1630 |
|
? Data->directmvB[k].x |
1631 |
|
:pMB->mvs[k].x - Data->referencemv[k].x); |
1632 |
|
pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y); |
1633 |
|
pMB->b_mvs[k].y = ((Data->currentMV->y == 0) |
1634 |
|
? Data->directmvB[k].y |
1635 |
|
: pMB->mvs[k].y - Data->referencemv[k].y); |
1636 |
|
if (Data->qpel) { |
1637 |
|
pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2; |
1638 |
|
pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2; |
1639 |
|
pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2; |
1640 |
|
pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2; |
1641 |
|
} |
1642 |
|
|
1643 |
// left neighbour, if allowed |
if (b_mb->mode != MODE_INTER4V) { |
1644 |
if (psad[1] != MV_MAX_ERROR) |
pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; |
1645 |
{ |
pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; |
1646 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0]; |
1647 |
{ pmv[1].x = EVEN(pmv[1].x); |
pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0]; |
1648 |
pmv[1].y = EVEN(pmv[1].y); |
break; |
1649 |
} |
} |
|
CHECK_MV8_CANDIDATE(pmv[1].x,pmv[1].y); |
|
1650 |
} |
} |
1651 |
|
return skip_sad; |
|
// top neighbour, if allowed |
|
|
if (psad[2] != MV_MAX_ERROR) |
|
|
{ |
|
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
|
|
{ pmv[2].x = EVEN(pmv[2].x); |
|
|
pmv[2].y = EVEN(pmv[2].y); |
|
1652 |
} |
} |
|
CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); |
|
1653 |
|
|
1654 |
// top right neighbour, if allowed |
static void |
1655 |
if (psad[3] != MV_MAX_ERROR) |
SearchInterpolate(const IMAGE * const f_Ref, |
1656 |
{ |
const uint8_t * const f_RefH, |
1657 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
const uint8_t * const f_RefV, |
1658 |
{ pmv[3].x = EVEN(pmv[3].x); |
const uint8_t * const f_RefHV, |
1659 |
pmv[3].y = EVEN(pmv[3].y); |
const IMAGE * const b_Ref, |
1660 |
|
const uint8_t * const b_RefH, |
1661 |
|
const uint8_t * const b_RefV, |
1662 |
|
const uint8_t * const b_RefHV, |
1663 |
|
const IMAGE * const pCur, |
1664 |
|
const int x, const int y, |
1665 |
|
const uint32_t fcode, |
1666 |
|
const uint32_t bcode, |
1667 |
|
const uint32_t MotionFlags, |
1668 |
|
const MBParam * const pParam, |
1669 |
|
const VECTOR * const f_predMV, |
1670 |
|
const VECTOR * const b_predMV, |
1671 |
|
MACROBLOCK * const pMB, |
1672 |
|
int32_t * const best_sad, |
1673 |
|
SearchData * const fData) |
1674 |
|
|
1675 |
|
{ |
1676 |
|
|
1677 |
|
int iDirection, i, j; |
1678 |
|
SearchData bData; |
1679 |
|
|
1680 |
|
fData->qpel_precision = 0; |
1681 |
|
memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data |
1682 |
|
*fData->iMinSAD = 4096*256; |
1683 |
|
bData.currentMV++; bData.currentQMV++; |
1684 |
|
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
1685 |
|
|
1686 |
|
i = (x + y * fData->iEdgedWidth) * 16; |
1687 |
|
bData.bRef = fData->Ref = f_Ref->y + i; |
1688 |
|
bData.bRefH = fData->RefH = f_RefH + i; |
1689 |
|
bData.bRefV = fData->RefV = f_RefV + i; |
1690 |
|
bData.bRefHV = fData->RefHV = f_RefHV + i; |
1691 |
|
bData.Ref = fData->bRef = b_Ref->y + i; |
1692 |
|
bData.RefH = fData->bRefH = b_RefH + i; |
1693 |
|
bData.RefV = fData->bRefV = b_RefV + i; |
1694 |
|
bData.RefHV = fData->bRefHV = b_RefHV + i; |
1695 |
|
bData.b_RefCU = fData->RefCU = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; |
1696 |
|
bData.b_RefCV = fData->RefCV = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; |
1697 |
|
bData.RefCU = fData->b_RefCU = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; |
1698 |
|
bData.RefCV = fData->b_RefCV = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; |
1699 |
|
|
1700 |
|
|
1701 |
|
bData.bpredMV = fData->predMV = *f_predMV; |
1702 |
|
fData->bpredMV = bData.predMV = *b_predMV; |
1703 |
|
fData->currentMV[0] = fData->currentMV[2]; |
1704 |
|
|
1705 |
|
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode - fData->qpel, 0, 0); |
1706 |
|
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode - fData->qpel, 0, 0); |
1707 |
|
|
1708 |
|
if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; |
1709 |
|
if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx; |
1710 |
|
if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy; |
1711 |
|
if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy; |
1712 |
|
|
1713 |
|
if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx; |
1714 |
|
if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx; |
1715 |
|
if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy; |
1716 |
|
if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy; |
1717 |
|
|
1718 |
|
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
1719 |
|
|
1720 |
|
//diamond |
1721 |
|
do { |
1722 |
|
iDirection = 255; |
1723 |
|
// forward MV moves |
1724 |
|
i = fData->currentMV[0].x; j = fData->currentMV[0].y; |
1725 |
|
|
1726 |
|
CheckCandidateInt(i + 1, j, 0, &iDirection, fData); |
1727 |
|
CheckCandidateInt(i, j + 1, 0, &iDirection, fData); |
1728 |
|
CheckCandidateInt(i - 1, j, 0, &iDirection, fData); |
1729 |
|
CheckCandidateInt(i, j - 1, 0, &iDirection, fData); |
1730 |
|
|
1731 |
|
// backward MV moves |
1732 |
|
i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
1733 |
|
fData->currentMV[2] = fData->currentMV[0]; |
1734 |
|
CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); |
1735 |
|
CheckCandidateInt(i, j + 1, 0, &iDirection, &bData); |
1736 |
|
CheckCandidateInt(i - 1, j, 0, &iDirection, &bData); |
1737 |
|
CheckCandidateInt(i, j - 1, 0, &iDirection, &bData); |
1738 |
|
|
1739 |
|
} while (!(iDirection)); |
1740 |
|
|
1741 |
|
//qpel refinement |
1742 |
|
if (fData->qpel) { |
1743 |
|
if (*fData->iMinSAD > *best_sad + 500) return; |
1744 |
|
CheckCandidate = CheckCandidateInt; |
1745 |
|
fData->qpel_precision = bData.qpel_precision = 1; |
1746 |
|
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0); |
1747 |
|
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0); |
1748 |
|
fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x; |
1749 |
|
fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y; |
1750 |
|
fData->currentQMV[1].x = 2 * fData->currentMV[1].x; |
1751 |
|
fData->currentQMV[1].y = 2 * fData->currentMV[1].y; |
1752 |
|
SubpelRefine(fData); |
1753 |
|
if (*fData->iMinSAD > *best_sad + 300) return; |
1754 |
|
fData->currentQMV[2] = fData->currentQMV[0]; |
1755 |
|
SubpelRefine(&bData); |
1756 |
|
} |
1757 |
|
|
1758 |
|
*fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode. |
1759 |
|
|
1760 |
|
if (*fData->iMinSAD < *best_sad) { |
1761 |
|
*best_sad = *fData->iMinSAD; |
1762 |
|
pMB->mvs[0] = fData->currentMV[0]; |
1763 |
|
pMB->b_mvs[0] = fData->currentMV[1]; |
1764 |
|
pMB->mode = MODE_INTERPOLATE; |
1765 |
|
if (fData->qpel) { |
1766 |
|
pMB->qmvs[0] = fData->currentQMV[0]; |
1767 |
|
pMB->b_qmvs[0] = fData->currentQMV[1]; |
1768 |
|
pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x; |
1769 |
|
pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y; |
1770 |
|
pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x; |
1771 |
|
pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y; |
1772 |
|
} else { |
1773 |
|
pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; |
1774 |
|
pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; |
1775 |
|
pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; |
1776 |
|
pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; |
1777 |
} |
} |
|
CHECK_MV8_CANDIDATE(pmv[3].x,pmv[3].y); |
|
1778 |
} |
} |
1779 |
} |
} |
1780 |
|
|
1781 |
/* // this bias is zero anyway, at the moment! |
void |
1782 |
|
MotionEstimationBVOP(MBParam * const pParam, |
1783 |
if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) ) // && (iMinSAD <= iQuant * 96) |
FRAMEINFO * const frame, |
1784 |
iMinSAD -= MV8_00_BIAS; |
const int32_t time_bp, |
1785 |
|
const int32_t time_pp, |
1786 |
|
// forward (past) reference |
1787 |
|
const MACROBLOCK * const f_mbs, |
1788 |
|
const IMAGE * const f_ref, |
1789 |
|
const IMAGE * const f_refH, |
1790 |
|
const IMAGE * const f_refV, |
1791 |
|
const IMAGE * const f_refHV, |
1792 |
|
// backward (future) reference |
1793 |
|
const FRAMEINFO * const b_reference, |
1794 |
|
const IMAGE * const b_ref, |
1795 |
|
const IMAGE * const b_refH, |
1796 |
|
const IMAGE * const b_refV, |
1797 |
|
const IMAGE * const b_refHV) |
1798 |
|
{ |
1799 |
|
uint32_t i, j; |
1800 |
|
int32_t best_sad; |
1801 |
|
uint32_t skip_sad; |
1802 |
|
int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; |
1803 |
|
const MACROBLOCK * const b_mbs = b_reference->mbs; |
1804 |
|
|
1805 |
|
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
1806 |
|
|
1807 |
|
const int32_t TRB = time_pp - time_bp; |
1808 |
|
const int32_t TRD = time_pp; |
1809 |
|
|
1810 |
|
// some pre-inintialized data for the rest of the search |
1811 |
|
|
1812 |
|
SearchData Data; |
1813 |
|
int32_t iMinSAD; |
1814 |
|
VECTOR currentMV[3]; |
1815 |
|
VECTOR currentQMV[3]; |
1816 |
|
int32_t temp[8]; |
1817 |
|
memset(&Data, 0, sizeof(SearchData)); |
1818 |
|
Data.iEdgedWidth = pParam->edged_width; |
1819 |
|
Data.currentMV = currentMV; Data.currentQMV = currentQMV; |
1820 |
|
Data.iMinSAD = &iMinSAD; |
1821 |
|
Data.lambda16 = lambda_vec16[frame->quant]; |
1822 |
|
Data.qpel = pParam->m_quarterpel; |
1823 |
|
Data.rounding = 0; |
1824 |
|
Data.chroma = frame->motion_flags & PMV_CHROMA8; |
1825 |
|
Data.temp = temp; |
1826 |
|
|
1827 |
*/ |
Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose) |
1828 |
|
// note: i==horizontal, j==vertical |
1829 |
|
for (j = 0; j < pParam->mb_height; j++) { |
1830 |
|
|
1831 |
/* Terminate if MinSAD <= T_2 |
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
|
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
|
|
*/ |
|
1832 |
|
|
1833 |
if (iMinSAD < 512/4) /* T_2 == 512/4 hardcoded */ |
for (i = 0; i < pParam->mb_width; i++) { |
1834 |
{ |
MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; |
1835 |
if (MotionFlags & PMV_QUICKSTOP8) |
const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; |
1836 |
goto EPZS8_Terminate_without_Refine; |
|
1837 |
if (MotionFlags & PMV_EARLYSTOP8) |
/* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */ |
1838 |
goto EPZS8_Terminate_with_Refine; |
if (b_reference->coding_type != S_VOP) |
1839 |
|
if (b_mb->mode == MODE_NOT_CODED) { |
1840 |
|
pMB->mode = MODE_NOT_CODED; |
1841 |
|
continue; |
1842 |
} |
} |
1843 |
|
|
1844 |
/************ (Diamond Search) **************/ |
Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
1845 |
|
Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8; |
1846 |
|
Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8; |
1847 |
|
pMB->quant = frame->quant; |
1848 |
|
|
1849 |
|
/* direct search comes first, because it (1) checks for SKIP-mode |
1850 |
|
and (2) sets very good predictions for forward and backward search */ |
1851 |
|
skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
1852 |
|
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
1853 |
|
&frame->image, |
1854 |
|
i, j, |
1855 |
|
frame->motion_flags, |
1856 |
|
TRB, TRD, |
1857 |
|
pParam, |
1858 |
|
pMB, b_mb, |
1859 |
|
&best_sad, |
1860 |
|
&Data); |
1861 |
|
|
1862 |
|
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
1863 |
|
|
1864 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
// forward search |
1865 |
|
SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
1866 |
|
&frame->image, i, j, |
1867 |
|
frame->motion_flags, |
1868 |
|
frame->fcode, pParam, |
1869 |
|
pMB, &f_predMV, &best_sad, |
1870 |
|
MODE_FORWARD, &Data); |
1871 |
|
|
1872 |
|
// backward search |
1873 |
|
SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y, |
1874 |
|
&frame->image, i, j, |
1875 |
|
frame->motion_flags, |
1876 |
|
frame->bcode, pParam, |
1877 |
|
pMB, &b_predMV, &best_sad, |
1878 |
|
MODE_BACKWARD, &Data); |
1879 |
|
|
1880 |
|
// interpolate search comes last, because it uses data from forward and backward as prediction |
1881 |
|
SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
1882 |
|
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
1883 |
|
&frame->image, |
1884 |
|
i, j, |
1885 |
|
frame->fcode, frame->bcode, |
1886 |
|
frame->motion_flags, |
1887 |
|
pParam, |
1888 |
|
&f_predMV, &b_predMV, |
1889 |
|
pMB, &best_sad, |
1890 |
|
&Data); |
1891 |
|
|
1892 |
|
// final skip decision |
1893 |
|
if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2) |
1894 |
|
&& ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) |
1895 |
|
SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data); |
1896 |
|
|
1897 |
|
switch (pMB->mode) { |
1898 |
|
case MODE_FORWARD: |
1899 |
|
f_count++; |
1900 |
|
f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; |
1901 |
|
break; |
1902 |
|
case MODE_BACKWARD: |
1903 |
|
b_count++; |
1904 |
|
b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; |
1905 |
|
break; |
1906 |
|
case MODE_INTERPOLATE: |
1907 |
|
i_count++; |
1908 |
|
f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; |
1909 |
|
b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; |
1910 |
|
break; |
1911 |
|
case MODE_DIRECT: |
1912 |
|
case MODE_DIRECT_NO4V: |
1913 |
|
d_count++; |
1914 |
|
default: |
1915 |
|
break; |
1916 |
|
} |
1917 |
|
} |
1918 |
|
} |
1919 |
|
} |
1920 |
|
|
1921 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8)) |
static __inline void |
1922 |
iDiamondSize *= 2; |
MEanalyzeMB ( const uint8_t * const pRef, |
1923 |
|
const uint8_t * const pCur, |
1924 |
|
const int x, |
1925 |
|
const int y, |
1926 |
|
const MBParam * const pParam, |
1927 |
|
MACROBLOCK * const pMBs, |
1928 |
|
SearchData * const Data) |
1929 |
|
{ |
1930 |
|
|
1931 |
/* default: use best prediction as starting point for one call of EPZS_MainSearch */ |
int i, mask; |
1932 |
|
VECTOR pmv[3]; |
1933 |
|
MACROBLOCK * pMB = &pMBs[x + y * pParam->mb_width]; |
1934 |
|
|
1935 |
/* // there is no EPZS^2 for inter4v at the moment |
for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
1936 |
|
|
1937 |
if (MotionFlags & PMV_USESQUARES8) |
//median is only used as prediction. it doesn't have to be real |
1938 |
EPZSMainSearchPtr = Square8_MainSearch; |
if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; |
1939 |
else |
else |
1940 |
*/ |
if (x == 1) //left macroblock does not have any vector now |
1941 |
|
Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median |
1942 |
|
else if (y == 1) // top macroblock doesn't have it's vector |
1943 |
|
Data->predMV = (pMB - 1)->mvs[0]; // left instead of median |
1944 |
|
else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median |
1945 |
|
|
1946 |
EPZSMainSearchPtr = Diamond8_MainSearch; |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1947 |
|
pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, Data->rrv); |
1948 |
|
|
1949 |
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
1950 |
x, y, |
Data->Ref = pRef + (x + y * pParam->edged_width) * 16; |
|
currMV->x, currMV->y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
|
|
iDiamondSize, iFcode, iQuant, 0); |
|
1951 |
|
|
1952 |
|
pmv[1].x = EVEN(pMB->mvs[0].x); |
1953 |
|
pmv[1].y = EVEN(pMB->mvs[0].y); |
1954 |
|
pmv[2].x = EVEN(Data->predMV.x); |
1955 |
|
pmv[2].y = EVEN(Data->predMV.y); |
1956 |
|
pmv[0].x = pmv[0].y = 0; |
1957 |
|
|
1958 |
if (iSAD < iMinSAD) |
CheckCandidate32I(0, 0, 255, &i, Data); |
|
{ |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
|
|
|
if (MotionFlags & PMV_EXTSEARCH8) |
|
|
{ |
|
|
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
|
1959 |
|
|
1960 |
if (!(MVequal(pmv[0],backupMV)) ) |
if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) { |
|
{ |
|
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); |
|
1961 |
|
|
1962 |
if (iSAD < iMinSAD) |
if (!(mask = make_mask(pmv, 1))) |
1963 |
{ |
CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data); |
1964 |
*currMV = newMV; |
if (!(mask = make_mask(pmv, 2))) |
1965 |
iMinSAD = iSAD; |
CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data); |
|
} |
|
|
} |
|
1966 |
|
|
1967 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) // diamond only if needed |
1968 |
{ |
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
|
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); |
|
1969 |
|
|
1970 |
if (iSAD < iMinSAD) |
for (i = 0; i < 4; i++) { |
1971 |
{ |
MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width]; |
1972 |
*currMV = newMV; |
MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i]; |
1973 |
iMinSAD = iSAD; |
MB->mode = MODE_INTER; |
1974 |
|
MB->sad16 = Data->iMinSAD[i+1]; |
1975 |
} |
} |
1976 |
} |
} |
1977 |
} |
} |
1978 |
|
|
1979 |
/*************** Choose best MV found **************/ |
#define INTRA_BIAS 2500 |
1980 |
|
#define INTRA_THRESH 1500 |
1981 |
|
#define INTER_THRESH 1400 |
1982 |
|
|
1983 |
EPZS8_Terminate_with_Refine: |
int |
1984 |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
MEanalysis( const IMAGE * const pRef, |
1985 |
iMinSAD = Halfpel8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
FRAMEINFO * const Current, |
1986 |
x, y, |
MBParam * const pParam, |
1987 |
currMV, iMinSAD, |
int maxIntra, //maximum number if non-I frames |
1988 |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame |
1989 |
|
int bCount) // number of B frames in a row |
1990 |
|
{ |
1991 |
|
uint32_t x, y, intra = 0; |
1992 |
|
int sSAD = 0; |
1993 |
|
MACROBLOCK * const pMBs = Current->mbs; |
1994 |
|
const IMAGE * const pCurrent = &Current->image; |
1995 |
|
int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH; |
1996 |
|
|
1997 |
|
int32_t iMinSAD[5], temp[5]; |
1998 |
|
VECTOR currentMV[5]; |
1999 |
|
SearchData Data; |
2000 |
|
Data.iEdgedWidth = pParam->edged_width; |
2001 |
|
Data.currentMV = currentMV; |
2002 |
|
Data.iMinSAD = iMinSAD; |
2003 |
|
Data.iFcode = Current->fcode; |
2004 |
|
Data.rrv = Current->global_flags & XVID_REDUCED; |
2005 |
|
Data.temp = temp; |
2006 |
|
CheckCandidate = CheckCandidate32I; |
2007 |
|
|
2008 |
EPZS8_Terminate_without_Refine: |
if (intraCount != 0 && intraCount < 10) // we're right after an I frame |
2009 |
|
IntraThresh += 4 * (intraCount - 10) * (intraCount - 10); |
2010 |
|
else |
2011 |
|
if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec |
2012 |
|
IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; |
2013 |
|
|
2014 |
currPMV->x = currMV->x - pmv[0].x; |
InterThresh += 400 * (1 - bCount); |
2015 |
currPMV->y = currMV->y - pmv[0].y; |
if (InterThresh < 300) InterThresh = 300; |
|
return iMinSAD; |
|
|
} |
|
2016 |
|
|
2017 |
|
if (sadInit) (*sadInit) (); |
2018 |
|
|
2019 |
|
for (y = 1; y < pParam->mb_height-1; y += 2) { |
2020 |
|
for (x = 1; x < pParam->mb_width-1; x += 2) { |
2021 |
|
int i; |
2022 |
|
|
2023 |
|
if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; |
2024 |
|
|
2025 |
|
MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data); |
2026 |
|
|
2027 |
/* *********************************************************** |
for (i = 0; i < 4; i++) { |
2028 |
bvop motion estimation |
int dev; |
2029 |
// TODO: need to incorporate prediction here (eg. sad += calc_delta_16) |
MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width]; |
2030 |
***************************************************************/ |
if (pMB->sad16 > IntraThresh) { |
2031 |
|
dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16, |
2032 |
|
pParam->edged_width); |
2033 |
|
if (dev + IntraThresh < pMB->sad16) { |
2034 |
|
pMB->mode = MODE_INTRA; |
2035 |
|
if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return I_VOP; |
2036 |
|
} |
2037 |
|
} |
2038 |
|
sSAD += pMB->sad16; |
2039 |
|
} |
2040 |
|
} |
2041 |
|
} |
2042 |
|
sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); |
2043 |
|
// if (sSAD > IntraThresh + INTRA_BIAS) return I_VOP; |
2044 |
|
if (sSAD > InterThresh ) return P_VOP; |
2045 |
|
emms(); |
2046 |
|
return B_VOP; |
2047 |
|
|
2048 |
|
} |
2049 |
|
|
|
void MotionEstimationBVOP( |
|
|
MBParam * const pParam, |
|
|
FRAMEINFO * const frame, |
|
2050 |
|
|
2051 |
// forward (past) reference |
static WARPPOINTS |
2052 |
const MACROBLOCK * const f_mbs, |
GlobalMotionEst(const MACROBLOCK * const pMBs, |
2053 |
const IMAGE * const f_ref, |
const MBParam * const pParam, |
2054 |
const IMAGE * const f_refH, |
const FRAMEINFO * const current, |
2055 |
const IMAGE * const f_refV, |
const FRAMEINFO * const reference, |
2056 |
const IMAGE * const f_refHV, |
const IMAGE * const pRefH, |
2057 |
// backward (future) reference |
const IMAGE * const pRefV, |
2058 |
const MACROBLOCK * const b_mbs, |
const IMAGE * const pRefHV ) |
|
const IMAGE * const b_ref, |
|
|
const IMAGE * const b_refH, |
|
|
const IMAGE * const b_refV, |
|
|
const IMAGE * const b_refHV) |
|
2059 |
{ |
{ |
|
const uint32_t mb_width = pParam->mb_width; |
|
|
const uint32_t mb_height = pParam->mb_height; |
|
|
const int32_t edged_width = pParam->edged_width; |
|
2060 |
|
|
2061 |
uint32_t i,j; |
const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs |
2062 |
|
const int deltay=8; |
2063 |
|
const int grad=512; // lower bound for deviation in MB |
2064 |
|
|
2065 |
int32_t f_sad16; |
WARPPOINTS gmc; |
|
int32_t b_sad16; |
|
|
int32_t i_sad16; |
|
|
int32_t d_sad16; |
|
|
int32_t best_sad; |
|
2066 |
|
|
2067 |
VECTOR pmv_dontcare; |
uint32_t mx, my; |
2068 |
|
|
2069 |
// note: i==horizontal, j==vertical |
int MBh = pParam->mb_height; |
2070 |
for (j = 0; j < mb_height; j++) |
int MBw = pParam->mb_width; |
2071 |
{ |
|
2072 |
for (i = 0; i < mb_width; i++) |
int *MBmask= calloc(MBh*MBw,sizeof(int)); |
2073 |
|
double DtimesF[4] = { 0.,0., 0., 0. }; |
2074 |
|
double sol[4] = { 0., 0., 0., 0. }; |
2075 |
|
double a,b,c,n,denom; |
2076 |
|
double meanx,meany; |
2077 |
|
int num,oldnum; |
2078 |
|
|
2079 |
|
if (!MBmask) { fprintf(stderr,"Mem error\n"); return gmc;} |
2080 |
|
|
2081 |
|
// filter mask of all blocks |
2082 |
|
|
2083 |
|
for (my = 1; my < MBh-1; my++) |
2084 |
|
for (mx = 1; mx < MBw-1; mx++) |
2085 |
{ |
{ |
2086 |
MACROBLOCK *mb = &frame->mbs[i + j*mb_width]; |
const int mbnum = mx + my * MBw; |
2087 |
const MACROBLOCK *f_mb = &f_mbs[i + j*mb_width]; |
const MACROBLOCK *pMB = &pMBs[mbnum]; |
2088 |
const MACROBLOCK *b_mb = &b_mbs[i + j*mb_width]; |
const VECTOR mv = pMB->mvs[0]; |
2089 |
|
|
2090 |
if (b_mb->mode == MODE_INTER |
if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) |
|
&& b_mb->cbp == 0 |
|
|
&& b_mb->mvs[0].x == 0 |
|
|
&& b_mb->mvs[0].y == 0) |
|
|
{ |
|
|
mb->mode = MODE_NOT_CODED; |
|
|
mb->mvs[0].x = 0; |
|
|
mb->mvs[0].y = 0; |
|
|
mb->b_mvs[0].x = 0; |
|
|
mb->b_mvs[0].y = 0; |
|
2091 |
continue; |
continue; |
2092 |
|
|
2093 |
|
if ( ( (ABS(mv.x - (pMB-1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-1)->mvs[0].y) < deltay) ) |
2094 |
|
&& ( (ABS(mv.x - (pMB+1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+1)->mvs[0].y) < deltay) ) |
2095 |
|
&& ( (ABS(mv.x - (pMB-MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-MBw)->mvs[0].y) < deltay) ) |
2096 |
|
&& ( (ABS(mv.x - (pMB+MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) ) |
2097 |
|
MBmask[mbnum]=1; |
2098 |
} |
} |
2099 |
|
|
2100 |
|
for (my = 1; my < MBh-1; my++) |
2101 |
|
for (mx = 1; mx < MBw-1; mx++) |
2102 |
|
{ |
2103 |
|
const uint8_t *const pCur = current->image.y + 16*my*pParam->edged_width + 16*mx; |
2104 |
|
|
2105 |
// forward search |
const int mbnum = mx + my * MBw; |
2106 |
f_sad16 = SEARCH16(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
if (!MBmask[mbnum]) |
2107 |
&frame->image, |
continue; |
|
i, j, |
|
|
frame->motion_flags, frame->quant, frame->fcode, |
|
|
pParam, |
|
|
f_mbs, f_mbs /* todo */, |
|
|
&mb->mvs[0], &pmv_dontcare); // ignore pmv |
|
2108 |
|
|
2109 |
// backward search |
if (sad16 ( pCur, pCur+1 , pParam->edged_width, 65536) <= grad ) |
2110 |
b_sad16 = SEARCH16(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
MBmask[mbnum] = 0; |
2111 |
&frame->image, |
if (sad16 ( pCur, pCur+pParam->edged_width, pParam->edged_width, 65536) <= grad ) |
2112 |
i, j, |
MBmask[mbnum] = 0; |
2113 |
frame->motion_flags, frame->quant, frame->bcode, |
|
2114 |
pParam, |
} |
|
b_mbs, b_mbs, /* todo */ |
|
|
&mb->b_mvs[0], &pmv_dontcare); // ignore pmv |
|
2115 |
|
|
2116 |
// interpolate search (simple, but effective) |
emms(); |
|
i_sad16 = sad16bi_c( |
|
|
frame->image.y + i*16 + j*16*edged_width, |
|
|
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; |
|
2117 |
|
|
2118 |
|
do { /* until convergence */ |
2119 |
|
|
2120 |
if (f_sad16 < b_sad16) |
a = b = c = n = 0; |
2121 |
|
DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.; |
2122 |
|
for (my = 0; my < MBh; my++) |
2123 |
|
for (mx = 0; mx < MBw; mx++) |
2124 |
{ |
{ |
2125 |
best_sad = f_sad16; |
const int mbnum = mx + my * MBw; |
2126 |
mb->mode = MODE_FORWARD; |
const MACROBLOCK *pMB = &pMBs[mbnum]; |
2127 |
|
const VECTOR mv = pMB->mvs[0]; |
2128 |
|
|
2129 |
|
if (!MBmask[mbnum]) |
2130 |
|
continue; |
2131 |
|
|
2132 |
|
n++; |
2133 |
|
a += 16*mx+8; |
2134 |
|
b += 16*my+8; |
2135 |
|
c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8); |
2136 |
|
|
2137 |
|
DtimesF[0] += (double)mv.x; |
2138 |
|
DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8); |
2139 |
|
DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8); |
2140 |
|
DtimesF[3] += (double)mv.y; |
2141 |
|
} |
2142 |
|
|
2143 |
|
denom = a*a+b*b-c*n; |
2144 |
|
|
2145 |
|
/* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */ |
2146 |
|
/* D'*E*F has been calculated in the same loop as matrix */ |
2147 |
|
|
2148 |
|
sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2]; |
2149 |
|
sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3]; |
2150 |
|
sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3]; |
2151 |
|
sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3]; |
2152 |
|
|
2153 |
|
sol[0] /= denom; |
2154 |
|
sol[1] /= denom; |
2155 |
|
sol[2] /= denom; |
2156 |
|
sol[3] /= denom; |
2157 |
|
|
2158 |
|
meanx = meany = 0.; |
2159 |
|
oldnum = 0; |
2160 |
|
for (my = 0; my < MBh; my++) |
2161 |
|
for (mx = 0; mx < MBw; mx++) |
2162 |
|
{ |
2163 |
|
const int mbnum = mx + my * MBw; |
2164 |
|
const MACROBLOCK *pMB = &pMBs[mbnum]; |
2165 |
|
const VECTOR mv = pMB->mvs[0]; |
2166 |
|
|
2167 |
|
if (!MBmask[mbnum]) |
2168 |
|
continue; |
2169 |
|
|
2170 |
|
oldnum++; |
2171 |
|
meanx += ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ); |
2172 |
|
meany += ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ); |
2173 |
} |
} |
2174 |
|
|
2175 |
|
if (4*meanx > oldnum) /* better fit than 0.25 is useless */ |
2176 |
|
meanx /= oldnum; |
2177 |
else |
else |
2178 |
{ |
meanx = 0.25; |
2179 |
best_sad = b_sad16; |
|
2180 |
mb->mode = MODE_BACKWARD; |
if (4*meany > oldnum) |
2181 |
} |
meany /= oldnum; |
2182 |
|
else |
2183 |
|
meany = 0.25; |
2184 |
|
|
2185 |
if (i_sad16 < best_sad) |
/* fprintf(stderr,"sol = (%8.5f, %8.5f, %8.5f, %8.5f)\n",sol[0],sol[1],sol[2],sol[3]); |
2186 |
|
fprintf(stderr,"meanx = %8.5f meany = %8.5f %d\n",meanx,meany, oldnum); |
2187 |
|
*/ |
2188 |
|
num = 0; |
2189 |
|
for (my = 0; my < MBh; my++) |
2190 |
|
for (mx = 0; mx < MBw; mx++) |
2191 |
{ |
{ |
2192 |
best_sad = i_sad16; |
const int mbnum = mx + my * MBw; |
2193 |
mb->mode = MODE_INTERPOLATE; |
const MACROBLOCK *pMB = &pMBs[mbnum]; |
2194 |
|
const VECTOR mv = pMB->mvs[0]; |
2195 |
|
|
2196 |
|
if (!MBmask[mbnum]) |
2197 |
|
continue; |
2198 |
|
|
2199 |
|
if ( ( ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ) > meanx ) |
2200 |
|
|| ( ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ) > meany ) ) |
2201 |
|
MBmask[mbnum]=0; |
2202 |
|
else |
2203 |
|
num++; |
2204 |
} |
} |
2205 |
|
|
2206 |
if (d_sad16 < best_sad) |
} while ( (oldnum != num) && (num>=4) ); |
2207 |
|
|
2208 |
|
if (num < 4) |
2209 |
{ |
{ |
2210 |
best_sad = d_sad16; |
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; |
2211 |
mb->mode = MODE_DIRECT; |
} else { |
|
} |
|
2212 |
|
|
2213 |
|
gmc.duv[0].x=(int)(sol[0]+0.5); |
2214 |
|
gmc.duv[0].y=(int)(sol[3]+0.5); |
2215 |
|
|
2216 |
|
gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5); |
2217 |
|
gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5); |
2218 |
|
|
2219 |
|
gmc.duv[2].x=0; |
2220 |
|
gmc.duv[2].y=0; |
2221 |
} |
} |
2222 |
} |
// fprintf(stderr,"wp1 = ( %4d, %4d) wp2 = ( %4d, %4d) \n", gmc.duv[0].x, gmc.duv[0].y, gmc.duv[1].x, gmc.duv[1].y); |
2223 |
|
|
2224 |
|
free(MBmask); |
2225 |
|
|
2226 |
|
return gmc; |
2227 |
} |
} |