28 |
* |
* |
29 |
*************************************************************************/ |
*************************************************************************/ |
30 |
|
|
|
/************************************************************************** |
|
|
* |
|
|
* Modifications: |
|
|
* |
|
|
* 01.05.2002 updated MotionEstimationBVOP |
|
|
* 25.04.2002 partial prevMB conversion |
|
|
* 22.04.2002 remove some compile warning by chenm001 <chenm001@163.com> |
|
|
* 14.04.2002 added MotionEstimationBVOP() |
|
|
* 02.04.2002 add EPZS(^2) as ME algorithm, use PMV_USESQUARES to choose between |
|
|
* EPZS and EPZS^2 |
|
|
* 08.02.2002 split up PMVfast into three routines: PMVFast, PMVFast_MainLoop |
|
|
* PMVFast_Refine to support multiple searches with different start points |
|
|
* 07.01.2002 uv-block-based interpolation |
|
|
* 06.01.2002 INTER/INTRA-decision is now done before any SEARCH8 (speedup) |
|
|
* changed INTER_BIAS to 150 (as suggested by suxen_drol) |
|
|
* removed halfpel refinement step in PMVfastSearch8 + quality=5 |
|
|
* added new quality mode = 6 which performs halfpel refinement |
|
|
* filesize difference between quality 5 and 6 is smaller than 1% |
|
|
* (Isibaar) |
|
|
* 31.12.2001 PMVfastSearch16 and PMVfastSearch8 (gruel) |
|
|
* 30.12.2001 get_range/MotionSearchX simplified; blue/green bug fix |
|
|
* 22.12.2001 commented best_point==99 check |
|
|
* 19.12.2001 modified get_range (purple bug fix) |
|
|
* 15.12.2001 moved pmv displacement from mbprediction |
|
|
* 02.12.2001 motion estimation/compensation split (Isibaar) |
|
|
* 16.11.2001 rewrote/tweaked search algorithms; pross@cs.rmit.edu.au |
|
|
* 10.11.2001 support for sad16/sad8 functions |
|
|
* 28.08.2001 reactivated MODE_INTER4V for EXT_MODE |
|
|
* 24.08.2001 removed MODE_INTER4V_Q, disabled MODE_INTER4V for EXT_MODE |
|
|
* 22.08.2001 added MODE_INTER4V_Q |
|
|
* 20.08.2001 added pragma to get rid of internal compiler error with VC6 |
|
|
* idea by Cyril. Thanks. |
|
|
* |
|
|
* Michael Militzer <isibaar@videocoding.de> |
|
|
* |
|
|
**************************************************************************/ |
|
|
|
|
31 |
#include <assert.h> |
#include <assert.h> |
32 |
#include <stdio.h> |
#include <stdio.h> |
33 |
#include <stdlib.h> |
#include <stdlib.h> |
37 |
#include "../prediction/mbprediction.h" |
#include "../prediction/mbprediction.h" |
38 |
#include "../global.h" |
#include "../global.h" |
39 |
#include "../utils/timer.h" |
#include "../utils/timer.h" |
40 |
|
#include "motion_est.h" |
41 |
#include "motion.h" |
#include "motion.h" |
42 |
#include "sad.h" |
#include "sad.h" |
43 |
|
|
44 |
|
#define INITIAL_SKIP_THRESH (10) |
45 |
|
#define FINAL_SKIP_THRESH (50) |
46 |
|
#define MAX_SAD00_FOR_SKIP (20) |
47 |
|
#define MAX_CHROMA_SAD_FOR_SKIP (18) |
48 |
|
#define SKIP_THRESH_B (10) |
49 |
|
|
50 |
|
#define CHECK_CANDIDATE(X,Y,D) { \ |
51 |
|
(*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } |
52 |
|
|
53 |
static int32_t lambda_vec16[32] = /* rounded values for lambda param for weight of motion bits as in modified H.26L */ |
#define iDiamondSize 2 |
|
{ 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 */ |
|
|
|
|
54 |
|
|
55 |
|
//FILE * debug; |
56 |
|
|
57 |
// mv.length table |
static __inline int |
58 |
static const uint32_t mvtab[33] = { |
d_mv_bits(int x, int y, const uint32_t iFcode) |
|
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) |
|
59 |
{ |
{ |
60 |
if (component == 0) |
int xb, yb; |
|
return 1; |
|
|
|
|
|
if (component < 0) |
|
|
component = -component; |
|
61 |
|
|
62 |
if (iFcode == 1) { |
if (x == 0) xb = 1; |
63 |
if (component > 32) |
else { |
64 |
component = 32; |
if (x < 0) x = -x; |
65 |
|
x += (1 << (iFcode - 1)) - 1; |
66 |
return mvtab[component] + 1; |
x >>= (iFcode - 1); |
67 |
|
if (x > 32) x = 32; |
68 |
|
xb = mvtab[x] + iFcode; |
69 |
} |
} |
70 |
|
|
71 |
component += (1 << (iFcode - 1)) - 1; |
if (y == 0) yb = 1; |
72 |
component >>= (iFcode - 1); |
else { |
73 |
|
if (y < 0) y = -y; |
74 |
if (component > 32) |
y += (1 << (iFcode - 1)) - 1; |
75 |
component = 32; |
y >>= (iFcode - 1); |
76 |
|
if (y > 32) y = 32; |
77 |
return mvtab[component] + 1 + iFcode - 1; |
yb = mvtab[y] + iFcode; |
78 |
} |
} |
79 |
|
return xb + yb; |
|
|
|
|
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)); |
|
80 |
} |
} |
81 |
|
|
82 |
static __inline uint32_t |
/* CHACK_CANDIATE FUNCTIONS START */ |
|
calc_delta_8(const int32_t dx, |
|
|
const int32_t dy, |
|
|
const uint32_t iFcode, |
|
|
const uint32_t iQuant) |
|
|
{ |
|
|
return NEIGH_TEND_8X8 * lambda_vec8[iQuant] * (mv_bits(dx, iFcode) + |
|
|
mv_bits(dy, iFcode)); |
|
|
} |
|
83 |
|
|
84 |
bool |
static void |
85 |
MotionEstimation(MBParam * const pParam, |
CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
FRAMEINFO * const current, |
|
|
FRAMEINFO * const reference, |
|
|
const IMAGE * const pRefH, |
|
|
const IMAGE * const pRefV, |
|
|
const IMAGE * const pRefHV, |
|
|
const uint32_t iLimit) |
|
86 |
{ |
{ |
87 |
const uint32_t iWcount = pParam->mb_width; |
int32_t * const sad = data->temp; |
88 |
const uint32_t iHcount = pParam->mb_height; |
// static int32_t sad[5]; |
89 |
MACROBLOCK *const pMBs = current->mbs; |
int t; |
90 |
MACROBLOCK *const prevMBs = reference->mbs; |
const uint8_t * Reference; |
|
const IMAGE *const pCurrent = ¤t->image; |
|
|
const IMAGE *const pRef = &reference->image; |
|
|
|
|
|
static const VECTOR zeroMV = { 0, 0 }; |
|
|
VECTOR predMV; |
|
|
|
|
|
int32_t x, y; |
|
|
int32_t iIntra = 0; |
|
|
VECTOR pmv; |
|
91 |
|
|
92 |
if (sadInit) |
if (( x > data->max_dx) || ( x < data->min_dx) |
93 |
(*sadInit) (); |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
94 |
|
|
95 |
for (y = 0; y < iHcount; y++) { |
switch ( ((x&1)<<1) + (y&1) ) { |
96 |
for (x = 0; x < iWcount; x ++) { |
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
97 |
|
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
98 |
MACROBLOCK *const pMB = &pMBs[x + y * iWcount]; |
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
99 |
|
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
100 |
|
} |
101 |
|
|
102 |
if (pMB->mode == MODE_NOT_CODED) |
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, sad+1); |
|
continue; |
|
103 |
|
|
104 |
predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
105 |
|
data->temp[0] += lambda_vec16[data->iQuant] * t; |
106 |
|
data->temp[1] += lambda_vec8[data->iQuant] * t; |
107 |
|
|
108 |
pMB->sad16 = |
if (data->temp[0] < data->iMinSAD[0]) { |
109 |
SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
data->iMinSAD[0] = data->temp[0]; |
110 |
x, y, predMV.x, predMV.y, predMV.x, predMV.y, |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
111 |
current->motion_flags, current->quant, |
*dir = Direction; } |
|
current->fcode, pParam, pMBs, prevMBs, &pMB->mv16, |
|
|
&pMB->pmvs[0]); |
|
112 |
|
|
113 |
if (0 < (pMB->sad16 - MV16_INTER_BIAS)) { |
if (data->temp[1] < data->iMinSAD[1]) { |
114 |
int32_t deviation; |
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
115 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
116 |
|
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
117 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
118 |
|
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
119 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
120 |
|
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
121 |
|
|
122 |
deviation = |
} |
|
dev16(pCurrent->y + x * 16 + y * 16 * pParam->edged_width, |
|
|
pParam->edged_width); |
|
123 |
|
|
124 |
if (deviation < (pMB->sad16 - MV16_INTER_BIAS)) { |
static void |
125 |
pMB->mode = MODE_INTRA; |
CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
126 |
pMB->mv16 = pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = |
{ |
127 |
pMB->mvs[3] = zeroMV; |
int32_t sad; |
128 |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = |
const uint8_t * Reference; |
|
pMB->sad8[3] = 0; |
|
129 |
|
|
130 |
iIntra++; |
if (( x > data->max_dx) || ( x < data->min_dx) |
131 |
if (iIntra >= iLimit) |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|
return 1; |
|
132 |
|
|
133 |
continue; |
switch ( ((x&1)<<1) + (y&1) ) |
134 |
} |
{ |
135 |
|
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
136 |
|
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
137 |
|
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
138 |
|
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
139 |
} |
} |
140 |
|
|
141 |
pmv = pMB->pmvs[0]; |
sad = lambda_vec16[data->iQuant] * |
142 |
if (current->global_flags & XVID_INTER4V) |
d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
143 |
if ((!(current->global_flags & XVID_LUMIMASKING) || |
sad += sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
144 |
pMB->dquant == NO_CHANGE)) { |
|
145 |
int32_t sad8 = IMV16X16 * current->quant; |
if (sad < *(data->iMinSAD)) { |
146 |
|
*(data->iMinSAD) = sad; |
147 |
if (sad8 < pMB->sad16) { |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
148 |
sad8 += pMB->sad8[0] = |
*dir = Direction; } |
|
SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, |
|
|
pCurrent, 2 * x, 2 * y, |
|
|
pMB->mv16.x, pMB->mv16.y, predMV.x, predMV.y, |
|
|
current->motion_flags, |
|
|
current->quant, current->fcode, pParam, |
|
|
pMBs, prevMBs, &pMB->mvs[0], |
|
|
&pMB->pmvs[0]); |
|
|
} |
|
|
if (sad8 < pMB->sad16) { |
|
|
|
|
|
predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 1); |
|
|
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, predMV.x, predMV.y, |
|
|
current->motion_flags, |
|
|
current->quant, current->fcode, pParam, |
|
|
pMBs, prevMBs, &pMB->mvs[1], |
|
|
&pMB->pmvs[1]); |
|
|
} |
|
|
if (sad8 < pMB->sad16) { |
|
|
predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 2); |
|
|
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, predMV.x, predMV.y, |
|
|
current->motion_flags, |
|
|
current->quant, current->fcode, pParam, |
|
|
pMBs, prevMBs, &pMB->mvs[2], |
|
|
&pMB->pmvs[2]); |
|
|
} |
|
|
if (sad8 < pMB->sad16) { |
|
|
predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 3); |
|
|
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, predMV.x, predMV.y, |
|
|
current->motion_flags, |
|
|
current->quant, current->fcode, pParam, |
|
|
pMBs, prevMBs, |
|
|
&pMB->mvs[3], |
|
|
&pMB->pmvs[3]); |
|
149 |
} |
} |
150 |
|
|
151 |
/* decide: MODE_INTER or MODE_INTER4V |
static void |
152 |
mpeg4: if (sad8 < pMB->sad16 - nb/2+1) use_inter4v |
CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) |
153 |
*/ |
{ |
154 |
|
int32_t sad; |
155 |
|
const int xb = data->currentMV[1].x; |
156 |
|
const int yb = data->currentMV[1].y; |
157 |
|
const uint8_t *ReferenceF, *ReferenceB; |
158 |
|
|
159 |
if (sad8 < pMB->sad16) { |
if (( xf > data->max_dx) || ( xf < data->min_dx) |
160 |
pMB->mode = MODE_INTER4V; |
|| ( yf > data->max_dy) || (yf < data->min_dy)) return; |
|
pMB->sad8[0] *= 4; |
|
|
pMB->sad8[1] *= 4; |
|
|
pMB->sad8[2] *= 4; |
|
|
pMB->sad8[3] *= 4; |
|
|
continue; |
|
|
} |
|
161 |
|
|
162 |
|
switch ( ((xf&1)<<1) + (yf&1) ) { |
163 |
|
case 0 : ReferenceF = data->Ref + xf/2 + (yf/2)*(data->iEdgedWidth); break; |
164 |
|
case 1 : ReferenceF = data->RefV + xf/2 + ((yf-1)/2)*(data->iEdgedWidth); break; |
165 |
|
case 2 : ReferenceF = data->RefH + (xf-1)/2 + (yf/2)*(data->iEdgedWidth); break; |
166 |
|
default : ReferenceF = data->RefHV + (xf-1)/2 + ((yf-1)/2)*(data->iEdgedWidth); break; |
167 |
} |
} |
168 |
|
|
169 |
pMB->mode = MODE_INTER; |
switch ( ((xb&1)<<1) + (yb&1) ) { |
170 |
pMB->pmvs[0] = pmv; /* pMB->pmvs[1] = pMB->pmvs[2] = pMB->pmvs[3] are not needed for INTER */ |
case 0 : ReferenceB = data->bRef + xb/2 + (yb/2)*(data->iEdgedWidth); break; |
171 |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mv16; |
case 1 : ReferenceB = data->bRefV + xb/2 + ((yb-1)/2)*(data->iEdgedWidth); break; |
172 |
pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = |
case 2 : ReferenceB = data->bRefH + (xb-1)/2 + (yb/2)*(data->iEdgedWidth); break; |
173 |
pMB->sad16; |
default : ReferenceB = data->bRefHV + (xb-1)/2 + ((yb-1)/2)*(data->iEdgedWidth); break; |
|
} |
|
174 |
} |
} |
175 |
|
|
176 |
return 0; |
sad = lambda_vec16[data->iQuant] * |
177 |
} |
( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) + |
178 |
|
d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode) ); |
179 |
|
|
180 |
|
sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
181 |
|
|
182 |
#define CHECK_MV16_ZERO {\ |
if (sad < *(data->iMinSAD)) { |
183 |
if ( (0 <= max_dx) && (0 >= min_dx) \ |
*(data->iMinSAD) = sad; |
184 |
&& (0 <= max_dy) && (0 >= min_dy) ) \ |
data->currentMV->x = xf; data->currentMV->y = yf; |
185 |
{ \ |
*dir = Direction; } |
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0, 0 , iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); \ |
|
|
iSAD += calc_delta_16(-center_x, -center_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) - center_x, (Y) - center_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) - center_x, (Y) - center_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) - center_x, (Y) - center_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) - center_x, (Y) - center_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(-center_x, -center_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)-center_x, (Y)-center_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)-center_x, (Y)-center_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)-center_x, (Y)-center_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)-center_x, (Y)-center_y, (uint8_t)iFcode, iQuant);\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
|
186 |
} |
} |
187 |
|
|
188 |
/* too slow and not fully functional at the moment */ |
static void |
189 |
/* |
CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
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) |
|
190 |
{ |
{ |
191 |
const int32_t iEdgedWidth = pParam->edged_width; |
int32_t sad; |
192 |
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
int k; |
193 |
int32_t iSAD; |
const uint8_t *ReferenceF; |
194 |
VECTOR pred; |
const uint8_t *ReferenceB; |
195 |
|
VECTOR mvs, b_mvs; |
196 |
|
|
197 |
|
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
198 |
|
|
199 |
pred = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1); |
200 |
|
|
201 |
iSAD = sad16( cur, |
for (k = 0; k < 4; k++) { |
202 |
get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0,0, iEdgedWidth), |
mvs.x = data->directmvF[k].x + x; |
203 |
iEdgedWidth, MV_MAX_ERROR); |
b_mvs.x = ((x == 0) ? |
204 |
if (iSAD <= iQuant * 96) |
data->directmvB[k].x |
205 |
iSAD -= MV16_00_BIAS; |
: mvs.x - data->referencemv[k].x); |
206 |
|
|
207 |
currMV->x = 0; |
mvs.y = data->directmvF[k].y + y; |
208 |
currMV->y = 0; |
b_mvs.y = ((y == 0) ? |
209 |
currPMV->x = -pred.x; |
data->directmvB[k].y |
210 |
currPMV->y = -pred.y; |
: mvs.y - data->referencemv[k].y); |
211 |
|
|
212 |
return iSAD; |
if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
213 |
|
|| ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) |
214 |
|
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
215 |
|
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
216 |
|
|
217 |
|
switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
218 |
|
case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
219 |
|
case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
220 |
|
case 2 : ReferenceF = data->RefH + (mvs.x-1)/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
221 |
|
default : ReferenceF = data->RefHV + (mvs.x-1)/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
222 |
} |
} |
|
*/ |
|
|
|
|
|
int32_t |
|
|
Diamond16_MainSearch(const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, |
|
|
const int y, |
|
|
const int start_x, |
|
|
const int start_y, |
|
|
int iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const int center_x, |
|
|
const int center_y, |
|
|
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 iDirectionBackup; |
|
|
int32_t iSAD; |
|
|
VECTOR backupMV; |
|
|
|
|
|
backupMV.x = start_x; |
|
|
backupMV.y = start_y; |
|
|
|
|
|
/* 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); |
|
223 |
|
|
224 |
if (iDirection) { |
switch ( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { |
225 |
while (!iFound) { |
case 0 : ReferenceB = data->bRef + b_mvs.x/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
226 |
iFound = 1; |
case 1 : ReferenceB = data->bRefV + b_mvs.x/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
227 |
backupMV = *currMV; |
case 2 : ReferenceB = data->bRefH + (b_mvs.x-1)/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
228 |
iDirectionBackup = iDirection; |
default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
|
|
|
|
if (iDirectionBackup != 2) |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y, 1); |
|
|
if (iDirectionBackup != 1) |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y, 2); |
|
|
if (iDirectionBackup != 4) |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, |
|
|
backupMV.y - iDiamondSize, 3); |
|
|
if (iDirectionBackup != 3) |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, |
|
|
backupMV.y + iDiamondSize, 4); |
|
229 |
} |
} |
230 |
} else { |
|
231 |
currMV->x = start_x; |
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
232 |
currMV->y = start_y; |
ReferenceF + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
233 |
|
ReferenceB + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
234 |
|
data->iEdgedWidth); |
235 |
|
if (sad > *(data->iMinSAD)) return; |
236 |
} |
} |
237 |
return iMinSAD; |
|
238 |
|
if (sad < *(data->iMinSAD)) { |
239 |
|
*(data->iMinSAD) = sad; |
240 |
|
data->currentMV->x = x; data->currentMV->y = y; |
241 |
|
*dir = Direction; } |
242 |
} |
} |
243 |
|
|
244 |
int32_t |
static void |
245 |
Square16_MainSearch(const uint8_t * const pRef, |
CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
246 |
const uint8_t * const pRefH, |
{ |
247 |
const uint8_t * const pRefV, |
int32_t sad; |
248 |
const uint8_t * const pRefHV, |
const uint8_t *ReferenceF; |
249 |
const uint8_t * const cur, |
const uint8_t *ReferenceB; |
250 |
const int x, |
VECTOR mvs, b_mvs; |
|
const int y, |
|
|
const int start_x, |
|
|
const int start_y, |
|
|
int iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const int center_x, |
|
|
const int center_y, |
|
|
const int32_t min_dx, |
|
|
const int32_t max_dx, |
|
|
const int32_t min_dy, |
|
|
const int32_t max_dy, |
|
|
const int32_t iEdgedWidth, |
|
|
const int32_t iDiamondSize, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
int iFound) |
|
|
{ |
|
|
/* Do a square search around given starting point, return SAD of best */ |
|
|
|
|
|
int32_t iDirection = 0; |
|
|
int32_t iSAD; |
|
|
VECTOR backupMV; |
|
|
|
|
|
backupMV.x = start_x; |
|
|
backupMV.y = start_y; |
|
|
|
|
|
/* 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 |
|
|
*/ |
|
251 |
|
|
252 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, backupMV.y, 1); |
if (( x > 31) || ( x < -31) || ( y > 31) || (y < -31)) return; |
|
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); |
|
253 |
|
|
254 |
|
sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1); |
255 |
|
|
256 |
if (iDirection) { |
mvs.x = data->directmvF[0].x + x; |
257 |
while (!iFound) { |
b_mvs.x = ((x == 0) ? |
258 |
iFound = 1; |
data->directmvB[0].x |
259 |
backupMV = *currMV; |
: mvs.x - data->referencemv[0].x); |
260 |
|
|
261 |
switch (iDirection) { |
mvs.y = data->directmvF[0].y + y; |
262 |
case 1: |
b_mvs.y = ((y == 0) ? |
263 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
data->directmvB[0].y |
264 |
backupMV.y, 1); |
: mvs.y - data->referencemv[0].y); |
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 5); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 7); |
|
|
break; |
|
|
case 2: |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, |
|
|
2); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 6); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 8); |
|
|
break; |
|
265 |
|
|
266 |
case 3: |
if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
267 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, |
|| ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) |
268 |
4); |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
269 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
|
backupMV.y - iDiamondSize, 7); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 8); |
|
|
break; |
|
270 |
|
|
271 |
case 4: |
switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
272 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, |
case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
273 |
3); |
case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
274 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
case 2 : ReferenceF = data->RefH + (mvs.x-1)/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
275 |
backupMV.y - iDiamondSize, 5); |
default : ReferenceF = data->RefHV + (mvs.x-1)/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
276 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
} |
|
backupMV.y + iDiamondSize, 6); |
|
|
break; |
|
277 |
|
|
278 |
case 5: |
switch ( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { |
279 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, |
case 0 : ReferenceB = data->bRef + b_mvs.x/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
280 |
1); |
case 1 : ReferenceB = data->bRefV + b_mvs.x/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
281 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, |
case 2 : ReferenceB = data->bRefH + (b_mvs.x-1)/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
282 |
3); |
default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
283 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
} |
|
backupMV.y - iDiamondSize, 5); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 6); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 7); |
|
|
break; |
|
284 |
|
|
285 |
case 6: |
sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, |
|
|
2); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, |
|
|
3); |
|
|
|
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 5); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 6); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 8); |
|
286 |
|
|
287 |
break; |
if (sad < *(data->iMinSAD)) { |
288 |
|
*(data->iMinSAD) = sad; |
289 |
|
data->currentMV->x = x; data->currentMV->y = y; |
290 |
|
*dir = Direction; } |
291 |
|
} |
292 |
|
|
293 |
case 7: |
static void |
294 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
295 |
backupMV.y, 1); |
{ |
296 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, |
int32_t sad; |
297 |
4); |
const uint8_t * Reference; |
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 5); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 7); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 8); |
|
|
break; |
|
298 |
|
|
299 |
case 8: |
if (( x > data->max_dx) || ( x < data->min_dx) |
300 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|
2); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, |
|
|
4); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 6); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 7); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 8); |
|
|
break; |
|
|
default: |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, |
|
|
1); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, |
|
|
2); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, |
|
|
3); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, |
|
|
4); |
|
|
|
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 5); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 6); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 7); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 8); |
|
|
break; |
|
|
} |
|
|
} |
|
|
} else { |
|
|
currMV->x = start_x; |
|
|
currMV->y = start_y; |
|
|
} |
|
|
return iMinSAD; |
|
|
} |
|
301 |
|
|
302 |
|
switch ( ((x&1)<<1) + (y&1) ) |
303 |
|
{ |
304 |
|
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
305 |
|
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
306 |
|
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
307 |
|
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
308 |
|
} |
309 |
|
|
310 |
int32_t |
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
311 |
Full16_MainSearch(const uint8_t * const pRef, |
sad += lambda_vec8[data->iQuant] * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
|
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, |
|
|
const int start_x, |
|
|
const int start_y, |
|
|
int iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const int center_x, |
|
|
const int center_y, |
|
|
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 = start_x; |
|
|
backupMV.y = start_y; |
|
|
|
|
|
for (dx = min_dx; dx <= max_dx; dx += iDiamondSize) |
|
|
for (dy = min_dy; dy <= max_dy; dy += iDiamondSize) |
|
|
NOCHECK_MV16_CANDIDATE(dx, dy); |
|
312 |
|
|
313 |
return iMinSAD; |
if (sad < *(data->iMinSAD)) { |
314 |
|
*(data->iMinSAD) = sad; |
315 |
|
data->currentMV->x = x; data->currentMV->y = y; |
316 |
|
*dir = Direction; } |
317 |
} |
} |
318 |
|
|
319 |
int32_t |
/* CHACK_CANDIATE FUNCTIONS END */ |
|
AdvDiamond16_MainSearch(const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, |
|
|
const int y, |
|
|
int start_x, |
|
|
int start_y, |
|
|
int iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const int center_x, |
|
|
const int center_y, |
|
|
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 iDirection) |
|
|
{ |
|
320 |
|
|
321 |
int32_t iSAD; |
/* MAINSEARCH FUNCTIONS START */ |
322 |
|
|
323 |
|
static void |
324 |
|
AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection) |
325 |
|
{ |
326 |
|
|
327 |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
328 |
|
|
329 |
if (iDirection) { |
int iDirection; |
|
CHECK_MV16_CANDIDATE(start_x - iDiamondSize, start_y); |
|
|
CHECK_MV16_CANDIDATE(start_x + iDiamondSize, start_y); |
|
|
CHECK_MV16_CANDIDATE(start_x, start_y - iDiamondSize); |
|
|
CHECK_MV16_CANDIDATE(start_x, start_y + iDiamondSize); |
|
|
} else { |
|
|
int bDirection = 1 + 2 + 4 + 8; |
|
330 |
|
|
331 |
do { |
do { |
332 |
iDirection = 0; |
iDirection = 0; |
333 |
if (bDirection & 1) //we only want to check left if we came from the right (our last motion was to the left, up-left or down-left) |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
334 |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
335 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
336 |
if (bDirection & 2) |
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
|
CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, start_y, 2); |
|
|
|
|
|
if (bDirection & 4) |
|
|
CHECK_MV16_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); |
|
|
|
|
|
if (bDirection & 8) |
|
|
CHECK_MV16_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); |
|
337 |
|
|
338 |
/* now we're doing diagonal checks near our candidate */ |
/* now we're doing diagonal checks near our candidate */ |
339 |
|
|
340 |
if (iDirection) //checking if anything found |
if (iDirection) { //checking if anything found |
|
{ |
|
341 |
bDirection = iDirection; |
bDirection = iDirection; |
342 |
iDirection = 0; |
iDirection = 0; |
343 |
start_x = currMV->x; |
x = data->currentMV->x; y = data->currentMV->y; |
344 |
start_y = currMV->y; |
if (bDirection & 3) { //our candidate is left or right |
345 |
if (bDirection & 3) //our candidate is left or right |
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
346 |
{ |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
347 |
CHECK_MV16_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); |
} else { // what remains here is up or down |
348 |
CHECK_MV16_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
349 |
} else // what remains here is up or down |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); } |
|
{ |
|
|
CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, start_y, 2); |
|
|
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); |
|
|
} |
|
350 |
|
|
351 |
if (iDirection) { |
if (iDirection) { |
352 |
bDirection += iDirection; |
bDirection += iDirection; |
353 |
start_x = currMV->x; |
x = data->currentMV->x; y = data->currentMV->y; } |
354 |
start_y = currMV->y; |
} else { //about to quit, eh? not so fast.... |
|
} |
|
|
} else //about to quit, eh? not so fast.... |
|
|
{ |
|
355 |
switch (bDirection) { |
switch (bDirection) { |
356 |
case 2: |
case 2: |
357 |
CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
358 |
start_y - iDiamondSize, 2 + 4); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
|
CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, |
|
|
start_y + iDiamondSize, 2 + 8); |
|
359 |
break; |
break; |
360 |
case 1: |
case 1: |
361 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
362 |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
|
start_y - iDiamondSize, 1 + 4); |
|
|
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
|
|
start_y + iDiamondSize, 1 + 8); |
|
363 |
break; |
break; |
364 |
case 2 + 4: |
case 2 + 4: |
365 |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
366 |
start_y - iDiamondSize, 1 + 4); |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
367 |
CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
|
start_y - iDiamondSize, 2 + 4); |
|
|
CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, |
|
|
start_y + iDiamondSize, 2 + 8); |
|
368 |
break; |
break; |
369 |
case 4: |
case 4: |
370 |
CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
371 |
start_y - iDiamondSize, 2 + 4); |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
|
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
|
|
start_y - iDiamondSize, 1 + 4); |
|
372 |
break; |
break; |
373 |
case 8: |
case 8: |
374 |
CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
375 |
start_y + iDiamondSize, 2 + 8); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
|
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
|
|
start_y + iDiamondSize, 1 + 8); |
|
376 |
break; |
break; |
377 |
case 1 + 4: |
case 1 + 4: |
378 |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
379 |
start_y + iDiamondSize, 1 + 8); |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
380 |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
|
start_y - iDiamondSize, 1 + 4); |
|
|
CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, |
|
|
start_y - iDiamondSize, 2 + 4); |
|
381 |
break; |
break; |
382 |
case 2 + 8: |
case 2 + 8: |
383 |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
384 |
start_y - iDiamondSize, 1 + 4); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
385 |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
|
start_y + iDiamondSize, 1 + 8); |
|
|
CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, |
|
|
start_y + iDiamondSize, 2 + 8); |
|
386 |
break; |
break; |
387 |
case 1 + 8: |
case 1 + 8: |
388 |
CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
389 |
start_y - iDiamondSize, 2 + 4); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
390 |
CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
|
start_y + iDiamondSize, 2 + 8); |
|
|
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
|
|
start_y + iDiamondSize, 1 + 8); |
|
391 |
break; |
break; |
392 |
default: //1+2+4+8 == we didn't find anything at all |
default: //1+2+4+8 == we didn't find anything at all |
393 |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
394 |
start_y - iDiamondSize, 1 + 4); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
395 |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
396 |
start_y + iDiamondSize, 1 + 8); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
|
CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, |
|
|
start_y - iDiamondSize, 2 + 4); |
|
|
CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, |
|
|
start_y + iDiamondSize, 2 + 8); |
|
397 |
break; |
break; |
398 |
} |
} |
399 |
if (!iDirection) |
if (!iDirection) break; //ok, the end. really |
|
break; //ok, the end. really |
|
|
else { |
|
400 |
bDirection = iDirection; |
bDirection = iDirection; |
401 |
start_x = currMV->x; |
x = data->currentMV->x; y = data->currentMV->y; |
|
start_y = currMV->y; |
|
|
} |
|
402 |
} |
} |
403 |
} |
} |
404 |
while (1); //forever |
while (1); //forever |
405 |
} |
} |
|
return iMinSAD; |
|
|
} |
|
406 |
|
|
407 |
#define CHECK_MV16_F_INTERPOL(X,Y) { \ |
static void |
408 |
if ( ((X) <= f_max_dx) && ((X) >= f_min_dx) \ |
SquareSearch(int x, int y, const SearchData * const data, int bDirection) |
409 |
&& ((Y) <= f_max_dy) && ((Y) >= f_min_dy) ) \ |
{ |
410 |
{ \ |
int iDirection; |
|
iSAD = sad16bi( cur, \ |
|
|
get_ref(f_pRef, f_pRefH, f_pRefV, f_pRefHV, x, y, 16, X, Y, iEdgedWidth), \ |
|
|
get_ref(b_pRef, b_pRefH, b_pRefV, b_pRefHV, x, y, 16, b_currMV->x, b_currMV->y, iEdgedWidth), \ |
|
|
iEdgedWidth); \ |
|
|
iSAD += calc_delta_16((X) - f_center_x, (Y) - f_center_y, (uint8_t)f_iFcode, iQuant);\ |
|
|
iSAD += calc_delta_16(b_currMV->x - b_center_x, b_currMV->y - b_center_y, (uint8_t)b_iFcode, iQuant);\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; f_currMV->x=(X); f_currMV->y=(Y); } } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV16_F_INTERPOL_FOUND(X,Y) { \ |
|
|
if ( ((X) <= f_max_dx) && ((X) >= f_min_dx) \ |
|
|
&& ((Y) <= f_max_dy) && ((Y) >= f_min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16bi( cur, \ |
|
|
get_ref(f_pRef, f_pRefH, f_pRefV, f_pRefHV, x, y, 16, X, Y, iEdgedWidth), \ |
|
|
get_ref(b_pRef, b_pRefH, b_pRefV, b_pRefHV, x, y, 16, b_currMV->x, b_currMV->y, iEdgedWidth), \ |
|
|
iEdgedWidth); \ |
|
|
iSAD += calc_delta_16((X) - f_center_x, (Y) - f_center_y, (uint8_t)f_iFcode, iQuant);\ |
|
|
iSAD += calc_delta_16(b_currMV->x - b_center_x, b_currMV->y - b_center_y, (uint8_t)b_iFcode, iQuant);\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; f_currMV->x=(X); f_currMV->y=(Y); iFound=0;} } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV16_B_INTERPOL(X,Y) { \ |
|
|
if ( ((X) <= b_max_dx) && ((X) >= b_min_dx) \ |
|
|
&& ((Y) <= b_max_dy) && ((Y) >= b_min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16bi( cur, \ |
|
|
get_ref(f_pRef, f_pRefH, f_pRefV, f_pRefHV, x, y, 16, f_currMV->x, f_currMV->y, iEdgedWidth), \ |
|
|
get_ref(b_pRef, b_pRefH, b_pRefV, b_pRefHV, x, y, 16, X, Y, iEdgedWidth), \ |
|
|
iEdgedWidth); \ |
|
|
iSAD += calc_delta_16(f_currMV->x - f_center_x, f_currMV->y - f_center_y, (uint8_t)f_iFcode, iQuant);\ |
|
|
iSAD += calc_delta_16((X) - b_center_x, (Y) - b_center_y, (uint8_t)b_iFcode, iQuant);\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; b_currMV->x=(X); b_currMV->y=(Y); } } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV16_B_INTERPOL_FOUND(X,Y) { \ |
|
|
if ( ((X) <= b_max_dx) && ((X) >= b_min_dx) \ |
|
|
&& ((Y) <= b_max_dy) && ((Y) >= b_min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16bi( cur, \ |
|
|
get_ref(f_pRef, f_pRefH, f_pRefV, f_pRefHV, x, y, 16, f_currMV->x, f_currMV->y, iEdgedWidth), \ |
|
|
get_ref(b_pRef, b_pRefH, b_pRefV, b_pRefHV, x, y, 16, X, Y, iEdgedWidth), \ |
|
|
iEdgedWidth); \ |
|
|
iSAD += calc_delta_16(f_currMV->x - f_center_x, f_currMV->y - f_center_y, (uint8_t)f_iFcode, iQuant);\ |
|
|
iSAD += calc_delta_16((X) - b_center_x, (Y) - b_center_y, (uint8_t)b_iFcode, iQuant);\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; b_currMV->x=(X); b_currMV->y=(Y); iFound=0;} } \ |
|
|
} |
|
|
|
|
|
int32_t |
|
|
Diamond16_InterpolMainSearch( |
|
|
const uint8_t * const f_pRef, |
|
|
const uint8_t * const f_pRefH, |
|
|
const uint8_t * const f_pRefV, |
|
|
const uint8_t * const f_pRefHV, |
|
|
|
|
|
const uint8_t * const cur, |
|
|
|
|
|
const uint8_t * const b_pRef, |
|
|
const uint8_t * const b_pRefH, |
|
|
const uint8_t * const b_pRefV, |
|
|
const uint8_t * const b_pRefHV, |
|
|
|
|
|
const int x, |
|
|
const int y, |
|
|
|
|
|
const int f_start_x, |
|
|
const int f_start_y, |
|
|
const int b_start_x, |
|
|
const int b_start_y, |
|
|
|
|
|
int iMinSAD, |
|
|
VECTOR * const f_currMV, |
|
|
VECTOR * const b_currMV, |
|
|
|
|
|
const int f_center_x, |
|
|
const int f_center_y, |
|
|
const int b_center_x, |
|
|
const int b_center_y, |
|
|
|
|
|
const int32_t f_min_dx, |
|
|
const int32_t f_max_dx, |
|
|
const int32_t f_min_dy, |
|
|
const int32_t f_max_dy, |
|
|
|
|
|
const int32_t b_min_dx, |
|
|
const int32_t b_max_dx, |
|
|
const int32_t b_min_dy, |
|
|
const int32_t b_max_dy, |
|
|
|
|
|
const int32_t iEdgedWidth, |
|
|
const int32_t iDiamondSize, |
|
|
|
|
|
const int32_t f_iFcode, |
|
|
const int32_t b_iFcode, |
|
|
|
|
|
const int32_t iQuant, |
|
|
int iFound) |
|
|
{ |
|
|
/* Do a diamond search around given starting point, return SAD of best */ |
|
|
|
|
|
int32_t iSAD; |
|
|
|
|
|
VECTOR f_backupMV; |
|
|
VECTOR b_backupMV; |
|
|
|
|
|
f_currMV->x = f_start_x; |
|
|
f_currMV->y = f_start_y; |
|
|
b_currMV->x = b_start_x; |
|
|
b_currMV->y = b_start_y; |
|
|
|
|
|
do |
|
|
{ |
|
|
iFound = 1; |
|
|
|
|
|
f_backupMV = *f_currMV; |
|
|
|
|
|
CHECK_MV16_F_INTERPOL_FOUND(f_backupMV.x - iDiamondSize, f_backupMV.y); |
|
|
CHECK_MV16_F_INTERPOL_FOUND(f_backupMV.x + iDiamondSize, f_backupMV.y); |
|
|
CHECK_MV16_F_INTERPOL_FOUND(f_backupMV.x, f_backupMV.y - iDiamondSize); |
|
|
CHECK_MV16_F_INTERPOL_FOUND(f_backupMV.x, f_backupMV.y + iDiamondSize); |
|
|
|
|
|
b_backupMV = *b_currMV; |
|
|
|
|
|
CHECK_MV16_B_INTERPOL_FOUND(b_backupMV.x - iDiamondSize, b_backupMV.y); |
|
|
CHECK_MV16_B_INTERPOL_FOUND(b_backupMV.x + iDiamondSize, b_backupMV.y); |
|
|
CHECK_MV16_B_INTERPOL_FOUND(b_backupMV.x, b_backupMV.y - iDiamondSize); |
|
|
CHECK_MV16_B_INTERPOL_FOUND(b_backupMV.x, b_backupMV.y + iDiamondSize); |
|
|
|
|
|
} while (!iFound); |
|
|
|
|
|
return iMinSAD; |
|
|
} |
|
|
|
|
|
/* Sorry, these MACROS really got too large... I'll turn them into function soon! */ |
|
|
|
|
|
#define CHECK_MV16_DIRECT_FOUND(X,Y) \ |
|
|
if ( (X)>=(-32) && (X)<=(31) && ((Y)>=-32) && ((Y)<=31) ) \ |
|
|
{ int k;\ |
|
|
VECTOR mvs,b_mvs; \ |
|
|
iSAD = 0;\ |
|
|
for (k = 0; k < 4; k++) { \ |
|
|
mvs.x = (int32_t) ((TRB * directmv[k].x) / TRD + (X)); \ |
|
|
b_mvs.x = (int32_t) (((X) == 0) \ |
|
|
? ((TRB - TRD) * directmv[k].x) / TRD \ |
|
|
: mvs.x - directmv[k].x); \ |
|
|
\ |
|
|
mvs.y = (int32_t) ((TRB * directmv[k].y) / TRD + (Y)); \ |
|
|
b_mvs.y = (int32_t) (((Y) == 0) \ |
|
|
? ((TRB - TRD) * directmv[k].y) / TRD \ |
|
|
: mvs.y - directmv[k].y); \ |
|
|
\ |
|
|
if ( (mvs.x <= max_dx) && (mvs.x >= min_dx) \ |
|
|
&& (mvs.y <= max_dy) && (mvs.y >= min_dy) \ |
|
|
&& (b_mvs.x <= max_dx) && (b_mvs.x >= min_dx) \ |
|
|
&& (b_mvs.y <= max_dy) && (b_mvs.y >= min_dy) ) { \ |
|
|
iSAD += sad8bi( cur + 8*(k&1) + 8*(k>>1)*iEdgedWidth, \ |
|
|
get_ref(f_pRef, f_pRefH, f_pRefV, f_pRefHV, 2*x+(k&1), 2*y+(k>>1), 8, \ |
|
|
mvs.x, mvs.y, iEdgedWidth), \ |
|
|
get_ref(b_pRef, b_pRefH, b_pRefV, b_pRefHV, 2*x+(k&1), 2*y+(k>>1), 8, \ |
|
|
b_mvs.x, b_mvs.y, iEdgedWidth), \ |
|
|
iEdgedWidth); \ |
|
|
} \ |
|
|
else \ |
|
|
iSAD = 65535; \ |
|
|
} \ |
|
|
iSAD += calc_delta_16((X),(Y), 1, iQuant);\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iFound=0; } \ |
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
int32_t |
|
|
Diamond16_DirectMainSearch( |
|
|
const uint8_t * const f_pRef, |
|
|
const uint8_t * const f_pRefH, |
|
|
const uint8_t * const f_pRefV, |
|
|
const uint8_t * const f_pRefHV, |
|
|
|
|
|
const uint8_t * const cur, |
|
|
|
|
|
const uint8_t * const b_pRef, |
|
|
const uint8_t * const b_pRefH, |
|
|
const uint8_t * const b_pRefV, |
|
|
const uint8_t * const b_pRefHV, |
|
411 |
|
|
412 |
const int x, |
do { |
413 |
const int y, |
iDirection = 0; |
414 |
|
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64); |
415 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128); |
416 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32); |
417 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128); |
418 |
|
if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64); |
419 |
|
if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128); |
420 |
|
if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128); |
421 |
|
if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128); |
422 |
|
|
423 |
const int TRB, |
bDirection = iDirection; |
424 |
const int TRD, |
x = data->currentMV->x; y = data->currentMV->y; |
425 |
|
} while (iDirection); |
426 |
|
} |
427 |
|
|
428 |
const int start_x, |
static void |
429 |
const int start_y, |
DiamondSearch(int x, int y, const SearchData * const data, int bDirection) |
430 |
|
{ |
431 |
|
|
432 |
int iMinSAD, |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
|
VECTOR * const currMV, |
|
|
const VECTOR * const directmv, |
|
433 |
|
|
434 |
const int32_t min_dx, |
int iDirection; |
|
const int32_t max_dx, |
|
|
const int32_t min_dy, |
|
|
const int32_t max_dy, |
|
435 |
|
|
436 |
const int32_t iEdgedWidth, |
do { |
437 |
const int32_t iDiamondSize, |
iDirection = 0; |
438 |
|
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
439 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
440 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
441 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
442 |
|
|
443 |
const int32_t iQuant, |
/* now we're doing diagonal checks near our candidate */ |
|
int iFound) |
|
|
{ |
|
|
/* Do a diamond search around given starting point, return SAD of best */ |
|
444 |
|
|
445 |
int32_t iSAD; |
if (iDirection) { //checking if anything found |
446 |
|
bDirection = iDirection; |
447 |
|
iDirection = 0; |
448 |
|
x = data->currentMV->x; y = data->currentMV->y; |
449 |
|
if (bDirection & 3) { //our candidate is left or right |
450 |
|
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
451 |
|
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
452 |
|
} else { // what remains here is up or down |
453 |
|
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
454 |
|
CHECK_CANDIDATE(x - iDiamondSize, y, 1); } |
455 |
|
|
456 |
VECTOR backupMV; |
bDirection += iDirection; |
457 |
|
x = data->currentMV->x; y = data->currentMV->y; |
458 |
|
} |
459 |
|
} |
460 |
|
while (iDirection); |
461 |
|
} |
462 |
|
|
463 |
currMV->x = start_x; |
/* MAINSEARCH FUNCTIONS END */ |
|
currMV->y = start_y; |
|
464 |
|
|
465 |
/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ |
/* HALFPELREFINE COULD BE A MAINSEARCH FUNCTION, BUT THERE IS NO NEED FOR IT */ |
466 |
|
|
467 |
do |
static void |
468 |
|
HalfpelRefine(const SearchData * const data) |
469 |
{ |
{ |
470 |
iFound = 1; |
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
471 |
|
|
472 |
backupMV = *currMV; |
VECTOR backupMV = *(data->currentMV); |
473 |
|
int iDirection; //not needed |
474 |
|
|
475 |
CHECK_MV16_DIRECT_FOUND(backupMV.x - iDiamondSize, backupMV.y); |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); |
476 |
CHECK_MV16_DIRECT_FOUND(backupMV.x + iDiamondSize, backupMV.y); |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); |
477 |
CHECK_MV16_DIRECT_FOUND(backupMV.x, backupMV.y - iDiamondSize); |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); |
478 |
CHECK_MV16_DIRECT_FOUND(backupMV.x, backupMV.y + iDiamondSize); |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0); |
479 |
|
|
480 |
} while (!iFound); |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); |
481 |
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); |
482 |
|
|
483 |
return iMinSAD; |
CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); |
484 |
|
CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); |
485 |
} |
} |
486 |
|
|
487 |
|
static __inline int |
488 |
|
SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
489 |
|
const int x, const int y, |
490 |
|
const uint32_t iEdgedWidth, const uint32_t iQuant) |
491 |
|
|
|
int32_t |
|
|
AdvDiamond8_MainSearch(const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, |
|
|
const int y, |
|
|
int start_x, |
|
|
int start_y, |
|
|
int iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const int center_x, |
|
|
const int center_y, |
|
|
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 iDirection) |
|
492 |
{ |
{ |
493 |
|
/* keep repeating checks for all b-frames before this P frame, |
494 |
|
to make sure that SKIP is possible (todo) |
495 |
|
how: if skip is not possible set sad00 to a very high value */ |
496 |
|
|
497 |
|
uint32_t sadC = sad8(current->u + x*8 + y*(iEdgedWidth/2)*8, |
498 |
|
reference->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); |
499 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
500 |
|
sadC += sad8(current->v + x*8 + y*(iEdgedWidth/2)*8, |
501 |
|
reference->v + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); |
502 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
503 |
|
|
504 |
int32_t iSAD; |
return 1; |
505 |
|
} |
506 |
|
|
507 |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
static __inline void |
508 |
|
SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
509 |
|
{ |
510 |
|
pMB->mode = MODE_NOT_CODED; |
511 |
|
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = pMB->mv16.x = 0; |
512 |
|
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = pMB->mv16.y = 0; |
513 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
514 |
|
} |
515 |
|
|
516 |
if (iDirection) { |
bool |
517 |
CHECK_MV8_CANDIDATE(start_x - iDiamondSize, start_y); |
MotionEstimation(MBParam * const pParam, |
518 |
CHECK_MV8_CANDIDATE(start_x + iDiamondSize, start_y); |
FRAMEINFO * const current, |
519 |
CHECK_MV8_CANDIDATE(start_x, start_y - iDiamondSize); |
FRAMEINFO * const reference, |
520 |
CHECK_MV8_CANDIDATE(start_x, start_y + iDiamondSize); |
const IMAGE * const pRefH, |
521 |
} else { |
const IMAGE * const pRefV, |
522 |
int bDirection = 1 + 2 + 4 + 8; |
const IMAGE * const pRefHV, |
523 |
|
const uint32_t iLimit) |
524 |
|
{ |
525 |
|
MACROBLOCK *const pMBs = current->mbs; |
526 |
|
const IMAGE *const pCurrent = ¤t->image; |
527 |
|
const IMAGE *const pRef = &reference->image; |
528 |
|
|
529 |
do { |
const VECTOR zeroMV = { 0, 0 }; |
|
iDirection = 0; |
|
|
if (bDirection & 1) //we only want to check left if we came from the right (our last motion was to the left, up-left or down-left) |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); |
|
530 |
|
|
531 |
if (bDirection & 2) |
uint32_t x, y; |
532 |
CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, start_y, 2); |
uint32_t iIntra = 0; |
533 |
|
int32_t InterBias; |
534 |
|
|
535 |
|
if (sadInit) (*sadInit) (); |
536 |
|
|
537 |
|
for (y = 0; y < pParam->mb_height; y++) { |
538 |
|
for (x = 0; x < pParam->mb_width; x++) { |
539 |
|
|
540 |
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
541 |
|
int32_t sad00 = pMB->sad16 |
542 |
|
= sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, |
543 |
|
pRef->y + (x + y * pParam->edged_width) * 16, |
544 |
|
pParam->edged_width, pMB->sad8 ); |
545 |
|
|
546 |
|
if (!(current->global_flags & XVID_LUMIMASKING)) { |
547 |
|
pMB->dquant = NO_CHANGE; |
548 |
|
pMB->quant = current->quant; } |
549 |
|
|
550 |
|
//initial skip decision |
551 |
|
|
552 |
|
if ((pMB->dquant == NO_CHANGE) && (sad00 <= MAX_SAD00_FOR_SKIP * pMB->quant) |
553 |
|
&& (SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) ) { |
554 |
|
if (pMB->sad16 < pMB->quant * INITIAL_SKIP_THRESH) { |
555 |
|
SkipMacroblockP(pMB, sad00); |
556 |
|
continue; |
557 |
|
} |
558 |
|
} else sad00 = 256*4096; // skip not allowed - for final skip decision |
559 |
|
|
560 |
if (bDirection & 4) |
SearchP(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
561 |
CHECK_MV8_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); |
y, current->motion_flags, pMB->quant, |
562 |
|
current->fcode, pParam, pMBs, reference->mbs, |
563 |
|
current->global_flags & XVID_INTER4V, pMB); |
564 |
|
|
565 |
|
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
566 |
|
if (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
567 |
|
if ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) |
568 |
|
{ SkipMacroblockP(pMB, sad00); continue; } |
569 |
|
|
570 |
if (bDirection & 8) |
/* finally, intra decision */ |
|
CHECK_MV8_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); |
|
571 |
|
|
572 |
/* now we're doing diagonal checks near our candidate */ |
InterBias = MV16_INTER_BIAS; |
573 |
|
if (pMB->quant > 8) InterBias += 50 * (pMB->quant - 8); // to make high quants work |
574 |
|
if (y != 0) |
575 |
|
if ((pMB - pParam->mb_width)->mode == MODE_INTER ) InterBias -= 50; |
576 |
|
if (x != 0) |
577 |
|
if ((pMB - 1)->mode == MODE_INTER ) InterBias -= 50; |
578 |
|
|
579 |
if (iDirection) //checking if anything found |
if (InterBias < pMB->sad16) { |
580 |
{ |
const int32_t deviation = |
581 |
bDirection = iDirection; |
dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
582 |
iDirection = 0; |
pParam->edged_width); |
|
start_x = currMV->x; |
|
|
start_y = currMV->y; |
|
|
if (bDirection & 3) //our candidate is left or right |
|
|
{ |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); |
|
|
} else // what remains here is up or down |
|
|
{ |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, start_y, 2); |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); |
|
|
} |
|
583 |
|
|
584 |
if (iDirection) { |
if (deviation < (pMB->sad16 - InterBias)) { |
585 |
bDirection += iDirection; |
if (++iIntra >= iLimit) return 1; |
586 |
start_x = currMV->x; |
pMB->mode = MODE_INTRA; |
587 |
start_y = currMV->y; |
pMB->mv16 = pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = |
588 |
} |
pMB->mvs[3] = zeroMV; |
589 |
} else //about to quit, eh? not so fast.... |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = |
590 |
{ |
pMB->sad8[3] = 0; |
|
switch (bDirection) { |
|
|
case 2: |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, |
|
|
start_y - iDiamondSize, 2 + 4); |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, |
|
|
start_y + iDiamondSize, 2 + 8); |
|
|
break; |
|
|
case 1: |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, |
|
|
start_y - iDiamondSize, 1 + 4); |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, |
|
|
start_y + iDiamondSize, 1 + 8); |
|
|
break; |
|
|
case 2 + 4: |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, |
|
|
start_y - iDiamondSize, 1 + 4); |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, |
|
|
start_y - iDiamondSize, 2 + 4); |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, |
|
|
start_y + iDiamondSize, 2 + 8); |
|
|
break; |
|
|
case 4: |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, |
|
|
start_y - iDiamondSize, 2 + 4); |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, |
|
|
start_y - iDiamondSize, 1 + 4); |
|
|
break; |
|
|
case 8: |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, |
|
|
start_y + iDiamondSize, 2 + 8); |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, |
|
|
start_y + iDiamondSize, 1 + 8); |
|
|
break; |
|
|
case 1 + 4: |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, |
|
|
start_y + iDiamondSize, 1 + 8); |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, |
|
|
start_y - iDiamondSize, 1 + 4); |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, |
|
|
start_y - iDiamondSize, 2 + 4); |
|
|
break; |
|
|
case 2 + 8: |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, |
|
|
start_y - iDiamondSize, 1 + 4); |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, |
|
|
start_y + iDiamondSize, 1 + 8); |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, |
|
|
start_y + iDiamondSize, 2 + 8); |
|
|
break; |
|
|
case 1 + 8: |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, |
|
|
start_y - iDiamondSize, 2 + 4); |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, |
|
|
start_y + iDiamondSize, 2 + 8); |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, |
|
|
start_y + iDiamondSize, 1 + 8); |
|
|
break; |
|
|
default: //1+2+4+8 == we didn't find anything at all |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, |
|
|
start_y - iDiamondSize, 1 + 4); |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, |
|
|
start_y + iDiamondSize, 1 + 8); |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, |
|
|
start_y - iDiamondSize, 2 + 4); |
|
|
CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, |
|
|
start_y + iDiamondSize, 2 + 8); |
|
|
break; |
|
|
} |
|
|
if (!(iDirection)) |
|
|
break; //ok, the end. really |
|
|
else { |
|
|
bDirection = iDirection; |
|
|
start_x = currMV->x; |
|
|
start_y = currMV->y; |
|
591 |
} |
} |
592 |
} |
} |
593 |
} |
} |
|
while (1); //forever |
|
594 |
} |
} |
595 |
return iMinSAD; |
return 0; |
596 |
} |
} |
597 |
|
|
598 |
|
|
599 |
int32_t |
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
|
Full8_MainSearch(const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, |
|
|
const int y, |
|
|
const int start_x, |
|
|
const int start_y, |
|
|
int iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const int center_x, |
|
|
const int center_y, |
|
|
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 = start_x; |
|
|
backupMV.y = start_y; |
|
|
|
|
|
for (dx = min_dx; dx <= max_dx; dx += iDiamondSize) |
|
|
for (dy = min_dy; dy <= max_dy; dy += iDiamondSize) |
|
|
NOCHECK_MV8_CANDIDATE(dx, dy); |
|
600 |
|
|
601 |
return iMinSAD; |
static __inline int |
602 |
|
make_mask(const VECTOR * const pmv, const int i) |
603 |
|
{ |
604 |
|
int mask = 0xFF, j; |
605 |
|
for (j = 0; j < i; j++) { |
606 |
|
if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already |
607 |
|
if (pmv[i].x == pmv[j].x) { |
608 |
|
if (pmv[i].y == pmv[j].y + iDiamondSize) { mask &= ~4; continue; } |
609 |
|
if (pmv[i].y == pmv[j].y - iDiamondSize) { mask &= ~8; continue; } |
610 |
|
} else |
611 |
|
if (pmv[i].y == pmv[j].y) { |
612 |
|
if (pmv[i].x == pmv[j].x + iDiamondSize) { mask &= ~1; continue; } |
613 |
|
if (pmv[i].x == pmv[j].x - iDiamondSize) { mask &= ~2; continue; } |
614 |
|
} |
615 |
|
} |
616 |
|
return mask; |
617 |
} |
} |
618 |
|
|
619 |
Halfpel8_RefineFuncPtr Halfpel8_Refine; |
static __inline void |
620 |
|
PreparePredictionsP(VECTOR * const pmv, int x, int y, const int iWcount, |
621 |
int32_t |
const int iHcount, const MACROBLOCK * const prevMB) |
|
Halfpel16_Refine(const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, |
|
|
const int y, |
|
|
VECTOR * const currMV, |
|
|
int32_t iMinSAD, |
|
|
const int center_x, |
|
|
const int center_y, |
|
|
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) |
|
622 |
{ |
{ |
|
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
|
623 |
|
|
624 |
int32_t iSAD; |
//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself |
|
VECTOR backupMV = *currMV; |
|
625 |
|
|
626 |
CHECK_MV16_CANDIDATE(backupMV.x - 1, backupMV.y - 1); |
if ( (y != 0) && (x != (iWcount-1)) ) { // [5] top-right neighbour |
627 |
CHECK_MV16_CANDIDATE(backupMV.x, backupMV.y - 1); |
pmv[5].x = EVEN(pmv[3].x); |
628 |
CHECK_MV16_CANDIDATE(backupMV.x + 1, backupMV.y - 1); |
pmv[5].y = EVEN(pmv[3].y); } |
629 |
CHECK_MV16_CANDIDATE(backupMV.x - 1, backupMV.y); |
else pmv[5].x = pmv[5].y = 0; |
|
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); |
|
630 |
|
|
631 |
return iMinSAD; |
if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour |
632 |
} |
else pmv[3].x = pmv[3].y = 0; |
|
|
|
|
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
|
|
|
|
|
|
|
|
|
|
|
int32_t |
|
|
PMVfastSearch16(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, /* start is searched first, so it should contain the most */ |
|
|
const int start_y, /* likely motion vector for this block */ |
|
|
const int center_x, /* center is from where length of MVs is measured */ |
|
|
const int center_y, |
|
|
const uint32_t MotionFlags, |
|
|
const uint32_t iQuant, |
|
|
const uint32_t iFcode, |
|
|
const MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const MACROBLOCK * const prevMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV) |
|
|
{ |
|
|
const uint32_t iWcount = pParam->mb_width; |
|
|
const int32_t iWidth = pParam->width; |
|
|
const int32_t iHeight = pParam->height; |
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
|
|
|
|
const uint8_t *cur = pCur->y + x * 16 + y * 16 * iEdgedWidth; |
|
|
|
|
|
int32_t iDiamondSize; |
|
|
|
|
|
int32_t min_dx; |
|
|
int32_t max_dx; |
|
|
int32_t min_dy; |
|
|
int32_t max_dy; |
|
|
|
|
|
int32_t iFound; |
|
|
|
|
|
VECTOR newMV; |
|
|
VECTOR backupMV; /* just for PMVFAST */ |
|
|
|
|
|
VECTOR pmv[4]; |
|
|
int32_t psad[4]; |
|
|
|
|
|
MainSearch16FuncPtr MainSearchPtr; |
|
|
|
|
|
const MACROBLOCK *const prevMB = prevMBs + x + y * iWcount; |
|
|
|
|
|
int32_t threshA, threshB; |
|
|
int32_t bPredEq; |
|
|
int32_t iMinSAD, iSAD; |
|
|
|
|
|
/* Get maximum range */ |
|
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 16, iWidth, iHeight, |
|
|
iFcode); |
|
|
|
|
|
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
|
|
|
|
|
if (!(MotionFlags & PMV_HALFPEL16)) { |
|
|
min_dx = EVEN(min_dx); |
|
|
max_dx = EVEN(max_dx); |
|
|
min_dy = EVEN(min_dy); |
|
|
max_dy = EVEN(max_dy); |
|
|
} |
|
|
|
|
|
/* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
|
|
//bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
|
|
bPredEq = get_pmvdata2(pMBs, iWcount, 0, x, y, 0, pmv, psad); |
|
|
|
|
|
if ((x == 0) && (y == 0)) { |
|
|
threshA = 512; |
|
|
threshB = 1024; |
|
|
} else { |
|
|
threshA = psad[0]; |
|
|
threshB = threshA + 256; |
|
|
if (threshA < 512) |
|
|
threshA = 512; |
|
|
if (threshA > 1024) |
|
|
threshA = 1024; |
|
|
if (threshB > 1792) |
|
|
threshB = 1792; |
|
|
} |
|
|
|
|
|
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. |
|
|
*/ |
|
|
|
|
|
currMV->x = start_x; |
|
|
currMV->y = start_y; |
|
|
|
|
|
if (!(MotionFlags & PMV_HALFPEL16)) { /* This should NOT be necessary! */ |
|
|
currMV->x = EVEN(currMV->x); |
|
|
currMV->y = EVEN(currMV->y); |
|
|
} |
|
|
|
|
|
if (currMV->x > max_dx) { |
|
|
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; |
|
|
} |
|
|
|
|
|
iMinSAD = |
|
|
sad16(cur, |
|
|
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, |
|
|
iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); |
|
|
iMinSAD += |
|
|
calc_delta_16(currMV->x - center_x, currMV->y - center_y, |
|
|
(uint8_t) iFcode, iQuant); |
|
|
|
|
|
if ((iMinSAD < 256) || |
|
|
((MVequal(*currMV, prevMB->mvs[0])) && |
|
|
((int32_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; |
|
|
} |
|
|
} |
|
|
|
|
|
if (MotionFlags & PMV_QUICKSTOP16) |
|
|
goto PMVfast16_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfast16_Terminate_with_Refine; |
|
|
} |
|
|
|
|
|
|
|
|
/* Step 2 (lazy eval): Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
|
|
vector of the median. |
|
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
|
|
*/ |
|
|
|
|
|
if ((bPredEq) && (MVequal(pmv[0], prevMB->mvs[0]))) |
|
|
iFound = 2; |
|
|
|
|
|
/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
|
|
Otherwise select large Diamond Search. |
|
|
*/ |
|
|
|
|
|
if ((!MVzero(pmv[0])) || (threshB < 1536) || (bPredEq)) |
|
|
iDiamondSize = 1; // halfpel! |
|
|
else |
|
|
iDiamondSize = 2; // halfpel! |
|
|
|
|
|
if (!(MotionFlags & PMV_HALFPELDIAMOND16)) |
|
|
iDiamondSize *= 2; |
|
|
|
|
|
/* |
|
|
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. |
|
|
*/ |
|
|
|
|
|
// (0,0) is always possible |
|
|
|
|
|
if (!MVzero(pmv[0])) |
|
|
CHECK_MV16_ZERO; |
|
|
|
|
|
// previous frame MV is always possible |
|
|
|
|
|
if (!MVzero(prevMB->mvs[0])) |
|
|
if (!MVequal(prevMB->mvs[0], pmv[0])) |
|
|
CHECK_MV16_CANDIDATE(prevMB->mvs[0].x, prevMB->mvs[0].y); |
|
|
|
|
|
// left neighbour, if allowed |
|
|
|
|
|
if (!MVzero(pmv[1])) |
|
|
if (!MVequal(pmv[1], prevMB->mvs[0])) |
|
|
if (!MVequal(pmv[1], pmv[0])) { |
|
|
if (!(MotionFlags & PMV_HALFPEL16)) { |
|
|
pmv[1].x = EVEN(pmv[1].x); |
|
|
pmv[1].y = EVEN(pmv[1].y); |
|
|
} |
|
|
|
|
|
CHECK_MV16_CANDIDATE(pmv[1].x, pmv[1].y); |
|
|
} |
|
|
// top neighbour, if allowed |
|
|
if (!MVzero(pmv[2])) |
|
|
if (!MVequal(pmv[2], prevMB->mvs[0])) |
|
|
if (!MVequal(pmv[2], pmv[0])) |
|
|
if (!MVequal(pmv[2], pmv[1])) { |
|
|
if (!(MotionFlags & PMV_HALFPEL16)) { |
|
|
pmv[2].x = EVEN(pmv[2].x); |
|
|
pmv[2].y = EVEN(pmv[2].y); |
|
|
} |
|
|
CHECK_MV16_CANDIDATE(pmv[2].x, pmv[2].y); |
|
|
|
|
|
// top right neighbour, if allowed |
|
|
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); |
|
|
} |
|
|
} |
|
|
|
|
|
if ((MVzero(*currMV)) && |
|
|
(!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96) */ ) |
|
|
iMinSAD -= MV16_00_BIAS; |
|
|
|
|
|
|
|
|
/* Step 6: If MinSAD <= thresa goto Step 10. |
|
|
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
|
|
*/ |
|
|
|
|
|
if ((iMinSAD <= threshA) || |
|
|
(MVequal(*currMV, prevMB->mvs[0]) && |
|
|
((int32_t) iMinSAD < prevMB->sad16))) { |
|
|
if (MotionFlags & PMV_QUICKSTOP16) |
|
|
goto PMVfast16_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfast16_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. |
|
|
*/ |
|
|
|
|
|
if (MotionFlags & PMV_USESQUARES16) |
|
|
MainSearchPtr = Square16_MainSearch; |
|
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
|
|
MainSearchPtr = AdvDiamond16_MainSearch; |
|
|
else |
|
|
MainSearchPtr = Diamond16_MainSearch; |
|
|
|
|
|
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
|
|
|
|
|
|
|
|
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
|
|
iSAD = |
|
|
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, |
|
|
currMV->x, currMV->y, iMinSAD, &newMV, center_x, center_y, |
|
|
min_dx, max_dx, |
|
|
min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, |
|
|
iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) { |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
|
|
|
if (MotionFlags & PMV_EXTSEARCH16) { |
|
|
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
|
|
|
|
|
if (!(MVequal(pmv[0], backupMV))) { |
|
|
iSAD = |
|
|
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, |
|
|
center_x, center_y, iMinSAD, &newMV, center_x, center_y, |
|
|
min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
|
|
iDiamondSize, iFcode, iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) { |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
|
|
|
|
if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { |
|
|
iSAD = |
|
|
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, 0, 0, |
|
|
iMinSAD, &newMV, center_x, center_y, |
|
|
min_dx, max_dx, min_dy, max_dy, |
|
|
iEdgedWidth, iDiamondSize, iFcode, |
|
|
iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) { |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
|
} |
|
|
|
|
|
/* |
|
|
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, center_x, center_y, min_dx, max_dx, min_dy, max_dy, |
|
|
iFcode, iQuant, iEdgedWidth); |
|
|
|
|
|
PMVfast16_Terminate_without_Refine: |
|
|
currPMV->x = currMV->x - center_x; |
|
|
currPMV->y = currMV->y - center_y; |
|
|
return iMinSAD; |
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
int32_t |
|
|
Diamond8_MainSearch(const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, |
|
|
const int y, |
|
|
int32_t start_x, |
|
|
int32_t start_y, |
|
|
int32_t iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const int center_x, |
|
|
const int center_y, |
|
|
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 iDirectionBackup; |
|
|
int32_t iSAD; |
|
|
VECTOR backupMV; |
|
|
|
|
|
backupMV.x = start_x; |
|
|
backupMV.y = start_y; |
|
|
|
|
|
/* 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! |
|
|
iDirectionBackup = iDirection; |
|
|
|
|
|
if (iDirectionBackup != 2) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y, 1); |
|
|
if (iDirectionBackup != 1) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y, 2); |
|
|
if (iDirectionBackup != 4) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, |
|
|
backupMV.y - iDiamondSize, 3); |
|
|
if (iDirectionBackup != 3) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, |
|
|
backupMV.y + iDiamondSize, 4); |
|
|
} |
|
|
} else { |
|
|
currMV->x = start_x; |
|
|
currMV->y = start_y; |
|
|
} |
|
|
return iMinSAD; |
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
int32_t |
|
|
Square8_MainSearch(const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, |
|
|
const int y, |
|
|
int32_t start_x, |
|
|
int32_t start_y, |
|
|
int32_t iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const int center_x, |
|
|
const int center_y, |
|
|
const int32_t min_dx, |
|
|
const int32_t max_dx, |
|
|
const int32_t min_dy, |
|
|
const int32_t max_dy, |
|
|
const int32_t iEdgedWidth, |
|
|
const int32_t iDiamondSize, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
int iFound) |
|
|
{ |
|
|
/* Do a square search around given starting point, return SAD of best */ |
|
|
|
|
|
int32_t iDirection = 0; |
|
|
int32_t iSAD; |
|
|
VECTOR backupMV; |
|
|
|
|
|
backupMV.x = start_x; |
|
|
backupMV.y = start_y; |
|
|
|
|
|
/* 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_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); |
|
|
|
|
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x - iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 5); |
|
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x - iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 6); |
|
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x + iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 7); |
|
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x + iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 8); |
|
|
|
|
|
|
|
|
if (iDirection) { |
|
|
while (!iFound) { |
|
|
iFound = 1; |
|
|
backupMV = *currMV; |
|
|
|
|
|
switch (iDirection) { |
|
|
case 1: |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y, 1); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 5); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 7); |
|
|
break; |
|
|
case 2: |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, |
|
|
2); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 6); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 8); |
|
|
break; |
|
|
|
|
|
case 3: |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, |
|
|
4); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 7); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 8); |
|
|
break; |
|
|
|
|
|
case 4: |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, |
|
|
3); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 5); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 6); |
|
|
break; |
|
|
|
|
|
case 5: |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, |
|
|
1); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, |
|
|
3); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 5); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 6); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 7); |
|
|
break; |
|
|
|
|
|
case 6: |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, |
|
|
2); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, |
|
|
3); |
|
|
|
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 5); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 6); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 8); |
|
|
|
|
|
break; |
|
|
|
|
|
case 7: |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y, 1); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, |
|
|
4); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 5); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 7); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 8); |
|
|
break; |
|
|
|
|
|
case 8: |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, |
|
|
2); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, |
|
|
4); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 6); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 7); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 8); |
|
|
break; |
|
|
default: |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, |
|
|
1); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, |
|
|
2); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, |
|
|
3); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, |
|
|
4); |
|
|
|
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 5); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 6); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 7); |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 8); |
|
|
break; |
|
|
} |
|
|
} |
|
|
} else { |
|
|
currMV->x = start_x; |
|
|
currMV->y = start_y; |
|
|
} |
|
|
return iMinSAD; |
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
int32_t |
|
|
Halfpel8_Refine_c(const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, |
|
|
const int y, |
|
|
VECTOR * const currMV, |
|
|
int32_t iMinSAD, |
|
|
const int center_x, |
|
|
const int center_y, |
|
|
const int32_t min_dx, |
|
|
const int32_t max_dx, |
|
|
const int32_t min_dy, |
|
|
const int32_t max_dy, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
const int32_t iEdgedWidth) |
|
|
{ |
|
|
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
|
|
|
|
|
int32_t iSAD; |
|
|
VECTOR backupMV = *currMV; |
|
|
|
|
|
CHECK_MV8_CANDIDATE(backupMV.x - 1, backupMV.y - 1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x, backupMV.y - 1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x + 1, backupMV.y - 1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x - 1, backupMV.y); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x + 1, backupMV.y); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x - 1, backupMV.y + 1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x, backupMV.y + 1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x + 1, backupMV.y + 1); |
|
|
|
|
|
return iMinSAD; |
|
|
} |
|
|
|
|
|
|
|
|
#define PMV_HALFPEL8 (PMV_HALFPELDIAMOND8|PMV_HALFPELREFINE8) |
|
|
|
|
|
int32_t |
|
|
PMVfastSearch8(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 int center_x, |
|
|
const int center_y, |
|
|
const uint32_t MotionFlags, |
|
|
const uint32_t iQuant, |
|
|
const uint32_t iFcode, |
|
|
const MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const MACROBLOCK * const prevMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV) |
|
|
{ |
|
|
const uint32_t iWcount = pParam->mb_width; |
|
|
const int32_t iWidth = pParam->width; |
|
|
const int32_t iHeight = pParam->height; |
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
|
|
|
|
const uint8_t *cur = pCur->y + x * 8 + y * 8 * iEdgedWidth; |
|
|
|
|
|
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; |
|
|
|
|
|
// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
|
|
const MACROBLOCK *const prevMB = prevMBs + (x >> 1) + (y >> 1) * iWcount; |
|
|
|
|
|
int32_t threshA, threshB; |
|
|
int32_t iFound, bPredEq; |
|
|
int32_t iMinSAD, iSAD; |
|
|
|
|
|
int32_t iSubBlock = (y & 1) + (y & 1) + (x & 1); |
|
|
|
|
|
MainSearch8FuncPtr MainSearchPtr; |
|
|
|
|
|
/* Init variables */ |
|
|
startMV.x = start_x; |
|
|
startMV.y = start_y; |
|
|
|
|
|
/* Get maximum range */ |
|
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 8, iWidth, iHeight, |
|
|
iFcode); |
|
|
|
|
|
if (!(MotionFlags & PMV_HALFPELDIAMOND8)) { |
|
|
min_dx = EVEN(min_dx); |
|
|
max_dx = EVEN(max_dx); |
|
|
min_dy = EVEN(min_dy); |
|
|
max_dy = EVEN(max_dy); |
|
|
} |
|
|
|
|
|
/* because we might use IF (dx>max_dx) THEN dx=max_dx; */ |
|
|
//bPredEq = get_pmvdata(pMBs, (x >> 1), (y >> 1), iWcount, iSubBlock, pmv, psad); |
|
|
bPredEq = get_pmvdata2(pMBs, iWcount, 0, (x >> 1), (y >> 1), iSubBlock, pmv, psad); |
|
|
|
|
|
if ((x == 0) && (y == 0)) { |
|
|
threshA = 512 / 4; |
|
|
threshB = 1024 / 4; |
|
|
|
|
|
} 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; |
|
|
} |
|
|
|
|
|
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 |
|
|
|
|
|
if (MotionFlags & PMV_USESQUARES8) |
|
|
MainSearchPtr = Square8_MainSearch; |
|
|
else |
|
|
|
|
|
if (MotionFlags & PMV_ADVANCEDDIAMOND8) |
|
|
MainSearchPtr = AdvDiamond8_MainSearch; |
|
|
else |
|
|
MainSearchPtr = Diamond8_MainSearch; |
|
|
|
|
|
|
|
|
*currMV = startMV; |
|
|
|
|
|
iMinSAD = |
|
|
sad8(cur, |
|
|
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, |
|
|
iEdgedWidth), iEdgedWidth); |
|
|
iMinSAD += |
|
|
calc_delta_8(currMV->x - center_x, currMV->y - center_y, |
|
|
(uint8_t) iFcode, iQuant); |
|
|
|
|
|
if ((iMinSAD < 256 / 4) || ((MVequal(*currMV, prevMB->mvs[iSubBlock])) |
|
|
&& ((int32_t) iMinSAD < |
|
|
prevMB->sad8[iSubBlock]))) { |
|
|
if (MotionFlags & PMV_QUICKSTOP16) |
|
|
goto PMVfast8_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfast8_Terminate_with_Refine; |
|
|
} |
|
|
|
|
|
/* Step 2 (lazy eval): Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
|
|
vector of the median. |
|
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
|
|
*/ |
|
|
|
|
|
if ((bPredEq) && (MVequal(pmv[0], prevMB->mvs[iSubBlock]))) |
|
|
iFound = 2; |
|
|
|
|
|
/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
|
|
Otherwise select large Diamond Search. |
|
|
*/ |
|
|
|
|
|
if ((!MVzero(pmv[0])) || (threshB < 1536 / 4) || (bPredEq)) |
|
|
iDiamondSize = 1; // 1 halfpel! |
|
|
else |
|
|
iDiamondSize = 2; // 2 halfpel = 1 full pixel! |
|
|
|
|
|
if (!(MotionFlags & PMV_HALFPELDIAMOND8)) |
|
|
iDiamondSize *= 2; |
|
|
|
|
|
|
|
|
/* |
|
|
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
|
|
Also calculate (0,0) but do not subtract offset. |
|
|
Let MinSAD be the smallest SAD up to this point. |
|
|
If MV is (0,0) subtract offset. |
|
|
*/ |
|
|
|
|
|
// the median prediction might be even better than mv16 |
|
633 |
|
|
634 |
if (!MVequal(pmv[0], startMV)) |
if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour |
635 |
CHECK_MV8_CANDIDATE(center_x, center_y); |
else pmv[4].x = pmv[4].y = 0; |
636 |
|
|
637 |
// (0,0) if needed |
// [1] median prediction |
638 |
if (!MVzero(pmv[0])) |
pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); |
|
if (!MVzero(startMV)) |
|
|
CHECK_MV8_ZERO; |
|
|
|
|
|
// previous frame MV if needed |
|
|
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); |
|
|
|
|
|
if ((iMinSAD <= threshA) || |
|
|
(MVequal(*currMV, prevMB->mvs[iSubBlock]) && |
|
|
((int32_t) iMinSAD < prevMB->sad8[iSubBlock]))) { |
|
|
if (MotionFlags & PMV_QUICKSTOP16) |
|
|
goto PMVfast8_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfast8_Terminate_with_Refine; |
|
|
} |
|
|
|
|
|
// 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])) { |
|
|
if (!(MotionFlags & PMV_HALFPEL8)) { |
|
|
pmv[1].x = EVEN(pmv[1].x); |
|
|
pmv[1].y = EVEN(pmv[1].y); |
|
|
} |
|
|
CHECK_MV8_CANDIDATE(pmv[1].x, pmv[1].y); |
|
|
} |
|
|
// top neighbour, if allowed and needed |
|
|
if (!MVzero(pmv[2])) |
|
|
if (!MVequal(pmv[2], startMV)) |
|
|
if (!MVequal(pmv[2], prevMB->mvs[iSubBlock])) |
|
|
if (!MVequal(pmv[2], pmv[0])) |
|
|
if (!MVequal(pmv[2], pmv[1])) { |
|
|
if (!(MotionFlags & PMV_HALFPEL8)) { |
|
|
pmv[2].x = EVEN(pmv[2].x); |
|
|
pmv[2].y = EVEN(pmv[2].y); |
|
|
} |
|
|
CHECK_MV8_CANDIDATE(pmv[2].x, pmv[2].y); |
|
|
|
|
|
// top right neighbour, if allowed and needed |
|
|
if (!MVzero(pmv[3])) |
|
|
if (!MVequal(pmv[3], startMV)) |
|
|
if (!MVequal(pmv[3], prevMB->mvs[iSubBlock])) |
|
|
if (!MVequal(pmv[3], pmv[0])) |
|
|
if (!MVequal(pmv[3], pmv[1])) |
|
|
if (!MVequal(pmv[3], pmv[2])) { |
|
|
if (! |
|
|
(MotionFlags & |
|
|
PMV_HALFPEL8)) { |
|
|
pmv[3].x = EVEN(pmv[3].x); |
|
|
pmv[3].y = EVEN(pmv[3].y); |
|
|
} |
|
|
CHECK_MV8_CANDIDATE(pmv[3].x, |
|
|
pmv[3].y); |
|
|
} |
|
|
} |
|
|
|
|
|
if ((MVzero(*currMV)) && |
|
|
(!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96) */ ) |
|
|
iMinSAD -= MV8_00_BIAS; |
|
639 |
|
|
640 |
|
pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask |
641 |
|
|
642 |
/* Step 6: If MinSAD <= thresa goto Step 10. |
pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
643 |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
pmv[2].y = EVEN(prevMB->mvs[0].y); |
|
*/ |
|
644 |
|
|
645 |
if ((iMinSAD <= threshA) || |
if ((x != iWcount-1) && (y != iHcount-1)) { |
646 |
(MVequal(*currMV, prevMB->mvs[iSubBlock]) && |
pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame |
647 |
((int32_t) iMinSAD < prevMB->sad8[iSubBlock]))) { |
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); } |
648 |
if (MotionFlags & PMV_QUICKSTOP16) |
else pmv[6].x = pmv[6].y = 0; |
|
goto PMVfast8_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfast8_Terminate_with_Refine; |
|
649 |
} |
} |
650 |
|
|
651 |
/************ (Diamond Search) **************/ |
static void |
652 |
/* |
SearchP(const uint8_t * const pRef, |
|
Step 7: Perform Diamond search, with either the small or large diamond. |
|
|
If Found=2 only examine one Diamond pattern, and afterwards goto step 10 |
|
|
Step 8: If small diamond, iterate small diamond search pattern until motion vector lies in the center of the diamond. |
|
|
If center then goto step 10. |
|
|
Step 9: If large diamond, iterate large diamond search pattern until motion vector lies in the center. |
|
|
Refine by using small diamond and goto step 10. |
|
|
*/ |
|
|
|
|
|
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
|
|
|
|
|
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
|
|
iSAD = |
|
|
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV->x, |
|
|
currMV->y, iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, |
|
|
min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, |
|
|
iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) { |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
|
|
|
if (MotionFlags & PMV_EXTSEARCH8) { |
|
|
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
|
|
|
|
|
if (!(MVequal(pmv[0], backupMV))) { |
|
|
iSAD = |
|
|
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, |
|
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, center_x, center_y, |
|
|
min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
|
|
iDiamondSize, iFcode, iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) { |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
|
|
|
|
if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { |
|
|
iSAD = |
|
|
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, 0, 0, |
|
|
iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, min_dy, |
|
|
max_dy, iEdgedWidth, iDiamondSize, iFcode, |
|
|
iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) { |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
|
} |
|
|
|
|
|
/* Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
|
|
By performing an optional local half-pixel search, we can refine this result even further. |
|
|
*/ |
|
|
|
|
|
PMVfast8_Terminate_with_Refine: |
|
|
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
|
|
iMinSAD = |
|
|
Halfpel8_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, |
|
|
iMinSAD, center_x, center_y, min_dx, max_dx, min_dy, max_dy, |
|
|
iFcode, iQuant, iEdgedWidth); |
|
|
|
|
|
|
|
|
PMVfast8_Terminate_without_Refine: |
|
|
currPMV->x = currMV->x - center_x; |
|
|
currPMV->y = currMV->y - center_y; |
|
|
|
|
|
return iMinSAD; |
|
|
} |
|
|
|
|
|
int32_t |
|
|
EPZSSearch16(const uint8_t * const pRef, |
|
653 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
654 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
655 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
656 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
657 |
const int x, |
const int x, |
658 |
const int y, |
const int y, |
|
const int start_x, |
|
|
const int start_y, |
|
|
const int center_x, |
|
|
const int center_y, |
|
659 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
660 |
const uint32_t iQuant, |
const uint32_t iQuant, |
661 |
const uint32_t iFcode, |
const uint32_t iFcode, |
662 |
const MBParam * const pParam, |
const MBParam * const pParam, |
663 |
const MACROBLOCK * const pMBs, |
const MACROBLOCK * const pMBs, |
664 |
const MACROBLOCK * const prevMBs, |
const MACROBLOCK * const prevMBs, |
665 |
VECTOR * const currMV, |
int inter4v, |
666 |
VECTOR * const currPMV) |
MACROBLOCK * const pMB) |
667 |
{ |
{ |
|
const uint32_t iWcount = pParam->mb_width; |
|
|
const uint32_t iHcount = pParam->mb_height; |
|
668 |
|
|
|
const int32_t iWidth = pParam->width; |
|
|
const int32_t iHeight = pParam->height; |
|
669 |
const int32_t iEdgedWidth = pParam->edged_width; |
const int32_t iEdgedWidth = pParam->edged_width; |
670 |
|
|
671 |
const uint8_t *cur = pCur->y + x * 16 + y * 16 * iEdgedWidth; |
int i, iDirection = 255, mask, threshA; |
672 |
|
int32_t temp[5]; |
673 |
int32_t min_dx; |
VECTOR currentMV[5], pmv[7]; |
674 |
int32_t max_dx; |
int32_t psad[4], iMinSAD[5]; |
675 |
int32_t min_dy; |
MainSearchFunc * MainSearchPtr; |
676 |
int32_t max_dy; |
SearchData Data; |
677 |
|
|
678 |
VECTOR newMV; |
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, psad); //has to be changed to get_pmv(2)() |
679 |
VECTOR backupMV; |
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 16, |
680 |
|
pParam->width, pParam->height, iFcode); |
681 |
VECTOR pmv[4]; |
|
682 |
int32_t psad[8]; |
Data.predMV = pmv[0]; |
683 |
|
Data.Cur = pCur->y + (x + y * iEdgedWidth) * 16; |
684 |
static MACROBLOCK *oldMBs = NULL; |
Data.iEdgedWidth = iEdgedWidth; |
685 |
|
Data.currentMV = currentMV; |
686 |
// const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
Data.iMinSAD = iMinSAD; |
687 |
const MACROBLOCK *const prevMB = prevMBs + x + y * iWcount; |
Data.Ref = pRef + (x + iEdgedWidth*y)*16; |
688 |
MACROBLOCK *oldMB = NULL; |
Data.RefH = pRefH + (x + iEdgedWidth*y) * 16; |
689 |
|
Data.RefV = pRefV + (x + iEdgedWidth*y) * 16; |
690 |
int32_t thresh2; |
Data.RefHV = pRefHV + (x + iEdgedWidth*y) * 16; |
691 |
int32_t bPredEq; |
Data.temp = temp; |
|
int32_t iMinSAD, iSAD = 9999; |
|
|
|
|
|
MainSearch16FuncPtr MainSearchPtr; |
|
|
|
|
|
if (oldMBs == NULL) { |
|
|
oldMBs = (MACROBLOCK *) calloc(iWcount * iHcount, sizeof(MACROBLOCK)); |
|
|
// fprintf(stderr,"allocated %d bytes for oldMBs\n",iWcount*iHcount*sizeof(MACROBLOCK)); |
|
|
} |
|
|
oldMB = oldMBs + x + y * iWcount; |
|
|
|
|
|
/* Get maximum range */ |
|
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 16, iWidth, iHeight, |
|
|
iFcode); |
|
692 |
|
|
693 |
if (!(MotionFlags & PMV_HALFPEL16)) { |
Data.iQuant = iQuant; |
694 |
min_dx = EVEN(min_dx); |
Data.iFcode = iFcode; |
|
max_dx = EVEN(max_dx); |
|
|
min_dy = EVEN(min_dy); |
|
|
max_dy = EVEN(max_dy); |
|
|
} |
|
|
/* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
|
|
//bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
|
|
bPredEq = get_pmvdata2(pMBs, iWcount, 0, x, y, 0, pmv, psad); |
|
|
|
|
|
/* 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->x = start_x; |
|
|
currMV->y = start_y; |
|
695 |
|
|
696 |
if (!(MotionFlags & PMV_HALFPEL16)) { |
if (!(MotionFlags & PMV_HALFPEL16)) { |
697 |
currMV->x = EVEN(currMV->x); |
Data.min_dx = EVEN(Data.min_dx); |
698 |
currMV->y = EVEN(currMV->y); |
Data.max_dx = EVEN(Data.max_dx); |
699 |
} |
Data.min_dy = EVEN(Data.min_dy); |
700 |
|
Data.max_dy = EVEN(Data.max_dy); } |
|
if (currMV->x > max_dx) |
|
|
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 ******************/ |
|
|
|
|
|
iMinSAD = |
|
|
sad16(cur, |
|
|
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, |
|
|
iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); |
|
|
iMinSAD += |
|
|
calc_delta_16(currMV->x - center_x, currMV->y - center_y, |
|
|
(uint8_t) iFcode, iQuant); |
|
|
|
|
|
// thresh1 is fixed to 256 |
|
|
if ((iMinSAD < 256) || |
|
|
((MVequal(*currMV, prevMB->mvs[0])) && |
|
|
((int32_t) iMinSAD < prevMB->sad16))) { |
|
|
if (MotionFlags & PMV_QUICKSTOP16) |
|
|
goto EPZS16_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto EPZS16_Terminate_with_Refine; |
|
|
} |
|
|
|
|
|
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
|
|
|
|
|
// previous frame MV |
|
|
CHECK_MV16_CANDIDATE(prevMB->mvs[0].x, prevMB->mvs[0].y); |
|
|
|
|
|
// set threshhold based on Min of Prediction and SAD of collocated block |
|
|
// CHECK_MV16 always uses iSAD for the SAD of last vector to check, so now iSAD is what we want |
|
|
|
|
|
if ((x == 0) && (y == 0)) { |
|
|
thresh2 = 512; |
|
|
} else { |
|
|
/* T_k = 1.2 * MIN(SAD_top,SAD_left,SAD_topleft,SAD_coll) +128; [Tourapis, 2002] */ |
|
|
|
|
|
thresh2 = MIN(psad[0], iSAD) * 6 / 5 + 128; |
|
|
} |
|
701 |
|
|
702 |
// MV=(0,0) is often a good choice |
for(i = 0; i < 5; i++) currentMV[i].x = currentMV[i].y = 0; |
703 |
|
|
704 |
CHECK_MV16_ZERO; |
i = d_mv_bits(pmv[0].x, pmv[0].y, iFcode); |
705 |
|
|
706 |
|
iMinSAD[0] = pMB->sad16 + lambda_vec16[iQuant] * i; |
707 |
|
iMinSAD[1] = pMB->sad8[0] + lambda_vec8[iQuant] * i; |
708 |
|
iMinSAD[2] = pMB->sad8[1]; |
709 |
|
iMinSAD[3] = pMB->sad8[2]; |
710 |
|
iMinSAD[4] = pMB->sad8[3]; |
711 |
|
|
712 |
// left neighbour, if allowed |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
|
if (x != 0) { |
|
|
if (!(MotionFlags & PMV_HALFPEL16)) { |
|
|
pmv[1].x = EVEN(pmv[1].x); |
|
|
pmv[1].y = EVEN(pmv[1].y); |
|
|
} |
|
|
CHECK_MV16_CANDIDATE(pmv[1].x, pmv[1].y); |
|
|
} |
|
|
// top neighbour, if allowed |
|
|
if (y != 0) { |
|
|
if (!(MotionFlags & PMV_HALFPEL16)) { |
|
|
pmv[2].x = EVEN(pmv[2].x); |
|
|
pmv[2].y = EVEN(pmv[2].y); |
|
|
} |
|
|
CHECK_MV16_CANDIDATE(pmv[2].x, pmv[2].y); |
|
713 |
|
|
714 |
// top right neighbour, if allowed |
if ((x == 0) && (y == 0)) threshA = 512; |
715 |
if ((uint32_t) x != (iWcount - 1)) { |
else { |
716 |
if (!(MotionFlags & PMV_HALFPEL16)) { |
threshA = psad[0] + 20; |
717 |
pmv[3].x = EVEN(pmv[3].x); |
if (threshA < 512) threshA = 512; |
718 |
pmv[3].y = EVEN(pmv[3].y); |
if (threshA > 1024) threshA = 1024; } |
719 |
} |
|
720 |
CHECK_MV16_CANDIDATE(pmv[3].x, pmv[3].y); |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
721 |
|
prevMBs + x + y * pParam->mb_width); |
722 |
|
|
723 |
|
if (inter4v) CheckCandidate = CheckCandidate16; |
724 |
|
else CheckCandidate = CheckCandidate16no4v; |
725 |
|
|
726 |
|
/* main loop. checking all predictions */ |
727 |
|
|
728 |
|
for (i = 1; i < 7; i++) { |
729 |
|
if (!(mask = make_mask(pmv, i)) ) continue; |
730 |
|
CheckCandidate16(pmv[i].x, pmv[i].y, mask, &iDirection, &Data); |
731 |
|
if (iMinSAD[0] < threshA) break; |
732 |
|
} |
733 |
|
|
734 |
|
if ((iMinSAD[0] <= threshA) || |
735 |
|
(MVequal(currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
736 |
|
(iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { |
737 |
|
inter4v = 0; |
738 |
|
if (MotionFlags & PMV_QUICKSTOP16) goto PMVfast16_Terminate_without_Refine; |
739 |
|
if (MotionFlags & PMV_EARLYSTOP16) { |
740 |
|
CheckCandidate = CheckCandidate16no4v; // I sure hope it's faster |
741 |
|
goto PMVfast16_Terminate_with_Refine; |
742 |
} |
} |
743 |
} |
} |
744 |
|
|
|
/* Terminate if MinSAD <= T_2 |
|
|
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
|
|
*/ |
|
|
|
|
|
if ((iMinSAD <= thresh2) |
|
|
|| (MVequal(*currMV, prevMB->mvs[0]) && |
|
|
((int32_t) iMinSAD <= prevMB->sad16))) { |
|
|
if (MotionFlags & PMV_QUICKSTOP16) |
|
|
goto EPZS16_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto EPZS16_Terminate_with_Refine; |
|
|
} |
|
|
|
|
|
/***** predictor SET C: acceleration MV (new!), neighbours in prev. frame(new!) ****/ |
|
|
|
|
|
backupMV = prevMB->mvs[0]; // collocated MV |
|
|
backupMV.x += (prevMB->mvs[0].x - oldMB->mvs[0].x); // acceleration X |
|
|
backupMV.y += (prevMB->mvs[0].y - oldMB->mvs[0].y); // acceleration Y |
|
|
|
|
|
CHECK_MV16_CANDIDATE(backupMV.x, backupMV.y); |
|
|
|
|
|
// left neighbour |
|
|
if (x != 0) |
|
|
CHECK_MV16_CANDIDATE((prevMB - 1)->mvs[0].x, (prevMB - 1)->mvs[0].y); |
|
|
|
|
|
// top neighbour |
|
|
if (y != 0) |
|
|
CHECK_MV16_CANDIDATE((prevMB - iWcount)->mvs[0].x, |
|
|
(prevMB - iWcount)->mvs[0].y); |
|
|
|
|
|
// right neighbour, if allowed (this value is not written yet, so take it from pMB->mvs |
|
|
|
|
|
if ((uint32_t) x != iWcount - 1) |
|
|
CHECK_MV16_CANDIDATE((prevMB + 1)->mvs[0].x, (prevMB + 1)->mvs[0].y); |
|
|
|
|
|
// bottom neighbour, dito |
|
|
if ((uint32_t) y != iHcount - 1) |
|
|
CHECK_MV16_CANDIDATE((prevMB + iWcount)->mvs[0].x, |
|
|
(prevMB + iWcount)->mvs[0].y); |
|
|
|
|
|
/* Terminate if MinSAD <= T_3 (here T_3 = T_2) */ |
|
|
if (iMinSAD <= thresh2) { |
|
|
if (MotionFlags & PMV_QUICKSTOP16) |
|
|
goto EPZS16_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto EPZS16_Terminate_with_Refine; |
|
|
} |
|
|
|
|
|
/************ (if Diamond Search) **************/ |
|
|
|
|
|
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
|
|
|
|
745 |
if (MotionFlags & PMV_USESQUARES16) |
if (MotionFlags & PMV_USESQUARES16) |
746 |
MainSearchPtr = Square16_MainSearch; |
MainSearchPtr = SquareSearch; |
747 |
else |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
748 |
if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
MainSearchPtr = AdvDiamondSearch; |
749 |
MainSearchPtr = AdvDiamond16_MainSearch; |
else MainSearchPtr = DiamondSearch; |
|
else |
|
|
MainSearchPtr = Diamond16_MainSearch; |
|
|
|
|
|
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
|
|
|
|
|
iSAD = |
|
|
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV->x, |
|
|
currMV->y, iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, |
|
|
min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); |
|
|
|
|
|
if (iSAD < iMinSAD) { |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
|
|
|
|
|
|
if (MotionFlags & PMV_EXTSEARCH16) { |
|
|
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
|
|
|
|
|
if (!(MVequal(pmv[0], backupMV))) { |
|
|
iSAD = |
|
|
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, |
|
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, center_x, center_y, |
|
|
min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
|
|
2, iFcode, iQuant, 0); |
|
|
} |
|
|
|
|
|
if (iSAD < iMinSAD) { |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
|
|
|
if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { |
|
|
iSAD = |
|
|
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, 0, 0, |
|
|
iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, min_dy, |
|
|
max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); |
|
|
|
|
|
if (iSAD < iMinSAD) { |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
|
} |
|
|
|
|
|
/*************** Choose best MV found **************/ |
|
|
|
|
|
EPZS16_Terminate_with_Refine: |
|
|
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
|
|
iMinSAD = |
|
|
Halfpel16_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, |
|
|
iMinSAD, center_x, center_y, min_dx, max_dx, min_dy, max_dy, |
|
|
iFcode, iQuant, iEdgedWidth); |
|
|
|
|
|
EPZS16_Terminate_without_Refine: |
|
|
|
|
|
*oldMB = *prevMB; |
|
|
|
|
|
currPMV->x = currMV->x - center_x; |
|
|
currPMV->y = currMV->y - center_y; |
|
|
return iMinSAD; |
|
|
} |
|
|
|
|
|
|
|
|
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 int center_x, |
|
|
const int center_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) |
|
|
{ |
|
|
/* Please not that EPZS might not be a good choice for 8x8-block motion search ! */ |
|
|
|
|
|
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 = 1; |
|
|
|
|
|
int32_t min_dx; |
|
|
int32_t max_dx; |
|
|
int32_t min_dy; |
|
|
int32_t max_dy; |
|
|
|
|
|
VECTOR newMV; |
|
|
VECTOR backupMV; |
|
|
|
|
|
VECTOR pmv[4]; |
|
|
int32_t psad[8]; |
|
|
|
|
|
const int32_t iSubBlock = ((y & 1) << 1) + (x & 1); |
|
|
|
|
|
// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
|
|
const MACROBLOCK *const prevMB = prevMBs + (x >> 1) + (y >> 1) * iWcount; |
|
|
|
|
|
int32_t bPredEq; |
|
|
int32_t iMinSAD, iSAD = 9999; |
|
|
|
|
|
MainSearch8FuncPtr MainSearchPtr; |
|
|
|
|
|
/* Get maximum range */ |
|
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 8, iWidth, iHeight, |
|
|
iFcode); |
|
|
|
|
|
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
|
|
|
|
|
if (!(MotionFlags & PMV_HALFPEL8)) { |
|
|
min_dx = EVEN(min_dx); |
|
|
max_dx = EVEN(max_dx); |
|
|
min_dy = EVEN(min_dy); |
|
|
max_dy = EVEN(max_dy); |
|
|
} |
|
|
/* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
|
|
//bPredEq = get_pmvdata(pMBs, x >> 1, y >> 1, iWcount, iSubBlock, pmv[0].x, pmv[0].y, psad); |
|
|
bPredEq = get_pmvdata2(pMBs, iWcount, 0, x >> 1, y >> 1, iSubBlock, pmv, psad); |
|
|
|
|
|
|
|
|
/* 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 |
|
|
|
|
|
|
|
|
if (!(MotionFlags & PMV_HALFPEL8)) { |
|
|
currMV->x = EVEN(currMV->x); |
|
|
currMV->y = EVEN(currMV->y); |
|
|
} |
|
|
|
|
|
if (currMV->x > max_dx) |
|
|
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 ******************/ |
|
|
|
|
|
|
|
|
iMinSAD = |
|
|
sad8(cur, |
|
|
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, |
|
|
iEdgedWidth), iEdgedWidth); |
|
|
iMinSAD += |
|
|
calc_delta_8(currMV->x - center_x, currMV->y - center_y, |
|
|
(uint8_t) iFcode, iQuant); |
|
|
|
|
|
|
|
|
// thresh1 is fixed to 256 |
|
|
if (iMinSAD < 256 / 4) { |
|
|
if (MotionFlags & PMV_QUICKSTOP8) |
|
|
goto EPZS8_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP8) |
|
|
goto EPZS8_Terminate_with_Refine; |
|
|
} |
|
|
|
|
|
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
|
|
|
|
750 |
|
|
751 |
// MV=(0,0) is often a good choice |
(*MainSearchPtr)(currentMV->x, currentMV->y, &Data, iDirection); |
|
CHECK_MV8_ZERO; |
|
752 |
|
|
753 |
// previous frame MV |
/* extended search, diamond starting in 0,0 and in prediction. |
754 |
CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x, prevMB->mvs[iSubBlock].y); |
note that this search is/might be done in halfpel positions, |
755 |
|
which makes it more different than the diamond above */ |
756 |
|
|
757 |
// left neighbour, if allowed |
if (MotionFlags & PMV_EXTSEARCH16) { |
758 |
if (psad[1] != MV_MAX_ERROR) { |
int32_t bSAD; |
759 |
if (!(MotionFlags & PMV_HALFPEL8)) { |
VECTOR startMV = Data.predMV, backupMV = currentMV[0]; |
760 |
pmv[1].x = EVEN(pmv[1].x); |
if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? |
761 |
pmv[1].y = EVEN(pmv[1].y); |
startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); |
762 |
} |
if (!(MVequal(startMV, backupMV))) { |
763 |
CHECK_MV8_CANDIDATE(pmv[1].x, pmv[1].y); |
bSAD = iMinSAD[0]; iMinSAD[0] = MV_MAX_ERROR; |
764 |
|
|
765 |
|
CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, &Data); |
766 |
|
(*MainSearchPtr)(startMV.x, startMV.y, &Data, 255); |
767 |
|
if (bSAD < iMinSAD[0]) { |
768 |
|
currentMV[0] = backupMV; |
769 |
|
iMinSAD[0] = bSAD; } |
770 |
|
} |
771 |
|
|
772 |
|
backupMV = currentMV[0]; |
773 |
|
if (MotionFlags & PMV_HALFPELREFINE16) startMV.x = startMV.y = 1; |
774 |
|
else startMV.x = startMV.y = 0; |
775 |
|
if (!(MVequal(startMV, backupMV))) { |
776 |
|
bSAD = iMinSAD[0]; iMinSAD[0] = MV_MAX_ERROR; |
777 |
|
|
778 |
|
CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, &Data); |
779 |
|
(*MainSearchPtr)(startMV.x, startMV.y, &Data, 255); |
780 |
|
if (bSAD < iMinSAD[0]) { |
781 |
|
currentMV[0] = backupMV; |
782 |
|
iMinSAD[0] = bSAD; } |
783 |
} |
} |
|
// 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); |
|
784 |
} |
} |
|
CHECK_MV8_CANDIDATE(pmv[2].x, pmv[2].y); |
|
785 |
|
|
786 |
// top right neighbour, if allowed |
PMVfast16_Terminate_with_Refine: |
787 |
if (psad[3] != MV_MAX_ERROR) { |
|
788 |
if (!(MotionFlags & PMV_HALFPEL8)) { |
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(&Data); |
|
pmv[3].x = EVEN(pmv[3].x); |
|
|
pmv[3].y = EVEN(pmv[3].y); |
|
|
} |
|
|
CHECK_MV8_CANDIDATE(pmv[3].x, pmv[3].y); |
|
|
} |
|
|
} |
|
789 |
|
|
790 |
/* // this bias is zero anyway, at the moment! |
PMVfast16_Terminate_without_Refine: |
791 |
|
|
792 |
if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) ) // && (iMinSAD <= iQuant * 96) |
if (inter4v) |
793 |
iMinSAD -= MV8_00_BIAS; |
for(i = 0; i < 4; i++) |
794 |
|
Search8(&Data, 2*x+(i&1), 2*y+(i>>1), MotionFlags, pParam, pMB, pMBs, i); |
795 |
|
|
796 |
|
if (!(inter4v) || |
797 |
|
(iMinSAD[0] < iMinSAD[1] + iMinSAD[2] + iMinSAD[3] + iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
798 |
|
// INTER MODE |
799 |
|
pMB->mode = MODE_INTER; |
800 |
|
pMB->mv16 = pMB->mvs[0] = pMB->mvs[1] |
801 |
|
= pMB->mvs[2] = pMB->mvs[3] = currentMV[0]; |
802 |
|
|
803 |
*/ |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
804 |
|
pMB->sad8[2] = pMB->sad8[3] = iMinSAD[0]; |
805 |
|
|
806 |
/* Terminate if MinSAD <= T_2 |
pMB->pmvs[0].x = currentMV[0].x - Data.predMV.x; |
807 |
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
pMB->pmvs[0].y = currentMV[0].y - Data.predMV.y; |
808 |
*/ |
} else { |
809 |
|
// INTER4V MODE; all other things are already set in Search8 |
810 |
|
pMB->mode = MODE_INTER4V; |
811 |
|
pMB->sad16 = iMinSAD[1] + iMinSAD[2] + iMinSAD[3] + iMinSAD[4] + IMV16X16 * iQuant; |
812 |
|
} |
813 |
|
|
|
if (iMinSAD < 512 / 4) { /* T_2 == 512/4 hardcoded */ |
|
|
if (MotionFlags & PMV_QUICKSTOP8) |
|
|
goto EPZS8_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP8) |
|
|
goto EPZS8_Terminate_with_Refine; |
|
814 |
} |
} |
815 |
|
|
816 |
/************ (Diamond Search) **************/ |
static void |
817 |
|
Search8(const SearchData * const OldData, |
818 |
|
const int x, const int y, |
819 |
|
const uint32_t MotionFlags, |
820 |
|
const MBParam * const pParam, |
821 |
|
MACROBLOCK * const pMB, |
822 |
|
const MACROBLOCK * const pMBs, |
823 |
|
const int block) |
824 |
|
{ |
825 |
|
SearchData Data; |
826 |
|
|
827 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
Data.predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
828 |
|
Data.iMinSAD = OldData->iMinSAD + 1 + block; |
829 |
|
Data.currentMV = OldData->currentMV+1+block; |
830 |
|
Data.iFcode = OldData->iFcode; |
831 |
|
Data.iQuant = OldData->iQuant; |
832 |
|
|
833 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8)) |
if (block != 0) |
834 |
iDiamondSize *= 2; |
*(Data.iMinSAD) += lambda_vec8[Data.iQuant] * |
835 |
|
d_mv_bits( Data.currentMV->x - Data.predMV.x, |
836 |
|
Data.currentMV->y - Data.predMV.y, |
837 |
|
Data.iFcode); |
838 |
|
|
|
/* default: use best prediction as starting point for one call of EPZS_MainSearch */ |
|
839 |
|
|
840 |
// there is no EPZS^2 for inter4v at the moment |
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { |
841 |
|
|
842 |
if (MotionFlags & PMV_USESQUARES8) |
Data.Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
843 |
MainSearchPtr = Square8_MainSearch; |
Data.RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
844 |
else |
Data.RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
845 |
|
Data.RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
846 |
|
|
847 |
if (MotionFlags & PMV_ADVANCEDDIAMOND8) |
Data.iEdgedWidth = pParam->edged_width; |
|
MainSearchPtr = AdvDiamond8_MainSearch; |
|
|
else |
|
|
MainSearchPtr = Diamond8_MainSearch; |
|
848 |
|
|
849 |
iSAD = |
Data.Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV->x, |
|
|
currMV->y, iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, |
|
|
min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, |
|
|
iQuant, 0); |
|
850 |
|
|
851 |
|
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 8, |
852 |
|
pParam->width, pParam->height, OldData->iFcode); |
853 |
|
|
854 |
if (iSAD < iMinSAD) { |
CheckCandidate = CheckCandidate8; |
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
855 |
|
|
856 |
if (MotionFlags & PMV_EXTSEARCH8) { |
if (MotionFlags & PMV_EXTSEARCH8) { |
|
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
|
857 |
|
|
858 |
if (!(MVequal(pmv[0], backupMV))) { |
MainSearchFunc *MainSearchPtr; |
859 |
iSAD = |
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
860 |
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, |
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
861 |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, center_x, center_y, |
else MainSearchPtr = DiamondSearch; |
862 |
min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
|
863 |
iDiamondSize, iFcode, iQuant, 0); |
(*MainSearchPtr)(Data.currentMV->x, Data.currentMV->y, &Data, 255); } |
864 |
|
|
865 |
if (iSAD < iMinSAD) { |
if (MotionFlags & PMV_HALFPELREFINE8) HalfpelRefine(&Data); |
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
866 |
} |
} |
867 |
|
|
868 |
|
pMB->pmvs[block].x = Data.currentMV->x - Data.predMV.x; |
869 |
|
pMB->pmvs[block].y = Data.currentMV->y - Data.predMV.y; |
870 |
|
pMB->mvs[block] = *(Data.currentMV); |
871 |
|
pMB->sad8[block] = 4 * (*(Data.iMinSAD)); |
872 |
} |
} |
873 |
|
|
874 |
if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { |
/* B-frames code starts here */ |
|
iSAD = |
|
|
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, 0, 0, |
|
|
iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, min_dy, |
|
|
max_dy, iEdgedWidth, iDiamondSize, iFcode, |
|
|
iQuant, 0); |
|
875 |
|
|
876 |
if (iSAD < iMinSAD) { |
static __inline VECTOR |
877 |
*currMV = newMV; |
ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) |
878 |
iMinSAD = iSAD; |
{ |
879 |
} |
/* the stupidiest function ever */ |
880 |
} |
if (mode == MODE_FORWARD) return pMB->mvs[0]; |
881 |
|
else return pMB->b_mvs[0]; |
882 |
} |
} |
883 |
|
|
884 |
/*************** Choose best MV found **************/ |
static void __inline |
885 |
|
PreparePredictionsBF(VECTOR * const pmv, const int x, const int y, |
886 |
|
const uint32_t iWcount, |
887 |
|
const MACROBLOCK * const pMB, |
888 |
|
const uint32_t mode_curr) |
889 |
|
{ |
890 |
|
|
891 |
EPZS8_Terminate_with_Refine: |
// [0] is prediction |
892 |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); |
|
iMinSAD = |
|
|
Halfpel8_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, |
|
|
iMinSAD, center_x, center_y, min_dx, max_dx, min_dy, max_dy, |
|
|
iFcode, iQuant, iEdgedWidth); |
|
893 |
|
|
894 |
EPZS8_Terminate_without_Refine: |
pmv[1].x = pmv[1].y = 0; // [1] is zero |
895 |
|
|
896 |
currPMV->x = currMV->x - center_x; |
pmv[2] = ChoosePred(pMB, mode_curr); |
897 |
currPMV->y = currMV->y - center_y; |
pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); |
898 |
return iMinSAD; |
|
899 |
} |
pmv[3].x = pmv[3].y = 0; |
900 |
|
if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour |
901 |
|
pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); |
902 |
|
pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); } |
903 |
|
|
904 |
|
if (y != 0) { |
905 |
|
pmv[4] = ChoosePred(pMB-iWcount, mode_curr); |
906 |
|
pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y); |
907 |
|
} else pmv[4].x = pmv[4].y = 0; |
908 |
|
|
909 |
|
if (x != 0) { |
910 |
|
pmv[5] = ChoosePred(pMB-1, mode_curr); |
911 |
|
pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
912 |
|
} else pmv[5].x = pmv[5].y = 0; |
913 |
|
|
914 |
|
if ((x != 0)&&(y != 0)) { |
915 |
|
pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); |
916 |
|
pmv[6].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
917 |
|
} else pmv[6].x = pmv[6].y = 0; |
918 |
|
|
919 |
|
// more? |
920 |
|
} |
921 |
|
|
922 |
|
|
923 |
int32_t |
/* search backward or forward, for b-frames */ |
924 |
PMVfastIntSearch16(const uint8_t * const pRef, |
static void |
925 |
|
SearchBF( const uint8_t * const pRef, |
926 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
927 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
928 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
929 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
930 |
const int x, |
const int x, const int y, |
|
const int y, |
|
|
const int start_x, /* start should be most likely vector */ |
|
|
const int start_y, |
|
|
const int center_x, /* center is from where length of MVs is measured */ |
|
|
const int center_y, |
|
931 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
932 |
const uint32_t iQuant, |
const uint32_t iQuant, |
933 |
const uint32_t iFcode, |
const uint32_t iFcode, |
934 |
const MBParam * const pParam, |
const MBParam * const pParam, |
935 |
const MACROBLOCK * const pMBs, |
MACROBLOCK * const pMB, |
936 |
const MACROBLOCK * const prevMBs, |
const VECTOR * const predMV, |
937 |
VECTOR * const currMV, |
int32_t * const best_sad, |
938 |
VECTOR * const currPMV) |
const int32_t mode_current) |
939 |
{ |
{ |
|
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 * 16 + y * 16 * iEdgedWidth; |
|
|
const VECTOR zeroMV = { 0, 0 }; |
|
|
|
|
|
int32_t iDiamondSize; |
|
|
|
|
|
int32_t min_dx; |
|
|
int32_t max_dx; |
|
|
int32_t min_dy; |
|
|
int32_t max_dy; |
|
|
|
|
|
int32_t iFound; |
|
940 |
|
|
941 |
VECTOR newMV; |
const int32_t iEdgedWidth = pParam->edged_width; |
|
VECTOR backupMV; |
|
|
|
|
|
VECTOR pmv[4]; |
|
|
int32_t psad[4]; |
|
942 |
|
|
943 |
MainSearch16FuncPtr MainSearchPtr; |
int i, iDirection, mask; |
944 |
|
VECTOR currentMV, pmv[7]; |
945 |
|
MainSearchFunc *MainSearchPtr; |
946 |
|
int32_t iMinSAD = MV_MAX_ERROR; |
947 |
|
SearchData Data; |
948 |
|
|
949 |
|
Data.iMinSAD = &iMinSAD; |
950 |
|
Data.Cur = pCur->y + (x + y * iEdgedWidth) * 16; |
951 |
|
Data.iEdgedWidth = iEdgedWidth; |
952 |
|
Data.currentMV = ¤tMV; |
953 |
|
Data.iMinSAD = &iMinSAD; |
954 |
|
Data.Ref = pRef + (x + y * iEdgedWidth) * 16; |
955 |
|
Data.RefH = pRefH + (x + y * iEdgedWidth) * 16; |
956 |
|
Data.RefV = pRefV + (x + y * iEdgedWidth) * 16; |
957 |
|
Data.RefHV = pRefHV + (x + y * iEdgedWidth) * 16; |
958 |
|
|
959 |
|
Data.iQuant = iQuant; |
960 |
|
Data.iFcode = iFcode; |
961 |
|
Data.predMV = *predMV; |
962 |
|
|
963 |
const MACROBLOCK *const prevMB = prevMBs + x + y * iWcount; |
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 16, |
964 |
MACROBLOCK *const pMB = pMBs + x + y * iWcount; |
pParam->width, pParam->height, iFcode); |
965 |
|
|
966 |
int32_t threshA, threshB; |
if (!(MotionFlags & PMV_HALFPEL16)) { |
967 |
int32_t bPredEq; |
Data.min_dx = EVEN(Data.min_dx); |
968 |
int32_t iMinSAD, iSAD; |
Data.max_dx = EVEN(Data.max_dx); |
969 |
|
Data.min_dy = EVEN(Data.min_dy); |
970 |
|
Data.max_dy = EVEN(Data.max_dy); } // no-halpel and b-frames. do we need it? |
971 |
|
|
972 |
|
|
973 |
/* Get maximum range */ |
pmv[0] = Data.predMV; |
974 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 16, iWidth, iHeight, |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, |
975 |
iFcode); |
pMB, mode_current); |
976 |
|
|
977 |
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
currentMV.x = currentMV.y = 0; |
978 |
|
|
979 |
if ((x == 0) && (y == 0)) { |
CheckCandidate = CheckCandidate16no4v; |
|
threshA = 512; |
|
|
threshB = 1024; |
|
980 |
|
|
981 |
bPredEq = 0; |
// main loop. checking all predictions |
982 |
psad[0] = psad[1] = psad[2] = psad[3] = 0; |
for (i = 0; i < 8; i++) { |
983 |
*currMV = pmv[0] = pmv[1] = pmv[2] = pmv[3] = zeroMV; |
if (!(mask = make_mask(pmv, i)) ) continue; |
984 |
|
CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, &Data); |
985 |
|
} |
986 |
|
|
987 |
} else { |
if (MotionFlags & PMV_USESQUARES16) |
988 |
|
MainSearchPtr = SquareSearch; |
989 |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
990 |
|
MainSearchPtr = AdvDiamondSearch; |
991 |
|
else MainSearchPtr = DiamondSearch; |
992 |
|
|
993 |
bPredEq = get_ipmvdata(pMBs, iWcount, 0, x, y, 0, pmv, psad); |
(*MainSearchPtr)(currentMV.x, currentMV.y, &Data, 255); |
994 |
|
|
995 |
threshA = psad[0]; |
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(&Data); |
|
threshB = threshA + 256; |
|
|
if (threshA < 512) |
|
|
threshA = 512; |
|
|
if (threshA > 1024) |
|
|
threshA = 1024; |
|
|
if (threshB > 1792) |
|
|
threshB = 1792; |
|
|
|
|
|
*currMV = pmv[0]; /* current best := prediction */ |
|
|
} |
|
|
|
|
|
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. |
|
|
*/ |
|
996 |
|
|
997 |
if (currMV->x > max_dx) { |
// three bits are needed to code backward mode. four for forward |
998 |
currMV->x = EVEN(max_dx); |
// we treat the bits just like they were vector's |
999 |
} |
if (mode_current == MODE_FORWARD) iMinSAD += 4 * lambda_vec16[iQuant]; |
1000 |
if (currMV->x < min_dx) { |
else iMinSAD += 3 * lambda_vec16[iQuant]; |
|
currMV->x = EVEN(min_dx); |
|
|
} |
|
|
if (currMV->y > max_dy) { |
|
|
currMV->y = EVEN(max_dy); |
|
|
} |
|
|
if (currMV->y < min_dy) { |
|
|
currMV->y = EVEN(min_dy); |
|
|
} |
|
1001 |
|
|
|
iMinSAD = |
|
|
sad16(cur, |
|
|
get_iref_mv(pRef, x, y, 16, currMV, |
|
|
iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); |
|
|
iMinSAD += |
|
|
calc_delta_16(currMV->x - center_x, currMV->y - center_y, |
|
|
(uint8_t) iFcode, iQuant); |
|
1002 |
|
|
1003 |
if ((iMinSAD < 256) || |
if (iMinSAD < *best_sad) { |
1004 |
((MVequal(*currMV, prevMB->i_mvs[0])) && |
*best_sad = iMinSAD; |
1005 |
((int32_t) iMinSAD < prevMB->i_sad16))) { |
pMB->mode = mode_current; |
1006 |
if (iMinSAD < 2 * iQuant) // high chances for SKIP-mode |
pMB->pmvs[0].x = currentMV.x - predMV->x; |
1007 |
{ |
pMB->pmvs[0].y = currentMV.y - predMV->y; |
1008 |
if (!MVzero(*currMV)) { |
if (mode_current == MODE_FORWARD) pMB->mvs[0] = currentMV; |
1009 |
iMinSAD += MV16_00_BIAS; |
else pMB->b_mvs[0] = currentMV; |
|
CHECK_MV16_ZERO; // (0,0) saves space for letterboxed pictures |
|
|
iMinSAD -= MV16_00_BIAS; |
|
|
} |
|
1010 |
} |
} |
1011 |
|
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfastInt16_Terminate_with_Refine; |
|
1012 |
} |
} |
1013 |
|
|
1014 |
|
static int32_t |
1015 |
|
SearchDirect(const uint8_t * const f_Ref, |
1016 |
|
const uint8_t * const f_RefH, |
1017 |
|
const uint8_t * const f_RefV, |
1018 |
|
const uint8_t * const f_RefHV, |
1019 |
|
const uint8_t * const b_Ref, |
1020 |
|
const uint8_t * const b_RefH, |
1021 |
|
const uint8_t * const b_RefV, |
1022 |
|
const uint8_t * const b_RefHV, |
1023 |
|
const IMAGE * const pCur, |
1024 |
|
const int x, const int y, |
1025 |
|
const uint32_t MotionFlags, |
1026 |
|
const uint32_t iQuant, |
1027 |
|
const int32_t TRB, const int32_t TRD, |
1028 |
|
const MBParam * const pParam, |
1029 |
|
MACROBLOCK * const pMB, |
1030 |
|
const MACROBLOCK * const b_mb, |
1031 |
|
int32_t * const best_sad) |
1032 |
|
|
1033 |
|
{ |
1034 |
|
const uint32_t iEdgedWidth = pParam->edged_width; |
1035 |
|
int32_t iMinSAD = 0, skip_sad; |
1036 |
|
int k; |
1037 |
|
VECTOR currentMV; |
1038 |
|
MainSearchFunc *MainSearchPtr; |
1039 |
|
SearchData Data; |
1040 |
|
|
1041 |
|
Data.iMinSAD = &iMinSAD; |
1042 |
|
Data.Cur = pCur->y + x * 16 + y * 16 * iEdgedWidth; |
1043 |
|
Data.iEdgedWidth = iEdgedWidth; |
1044 |
|
Data.currentMV = ¤tMV; |
1045 |
|
Data.iQuant = iQuant; |
1046 |
|
Data.referencemv = b_mb->mvs; |
1047 |
|
|
1048 |
|
Data.Ref= f_Ref + (x + iEdgedWidth*y) * 16; |
1049 |
|
Data.RefH = f_RefH + (x + iEdgedWidth*y) * 16; |
1050 |
|
Data.RefV = f_RefV + (x + iEdgedWidth*y) * 16; |
1051 |
|
Data.RefHV = f_RefHV + (x + iEdgedWidth*y) * 16; |
1052 |
|
Data.bRef = b_Ref + (x + iEdgedWidth*y) * 16; |
1053 |
|
Data.bRefH = b_RefH + (x + iEdgedWidth*y) * 16; |
1054 |
|
Data.bRefV = b_RefV + (x + iEdgedWidth*y) * 16; |
1055 |
|
Data.bRefHV = b_RefHV + (x + iEdgedWidth*y) * 16; |
1056 |
|
/* |
1057 |
|
//What we do here is a complicated version of CheckCandidateDirect(0,0); |
1058 |
|
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 16, pParam->width, pParam->height, 19); |
1059 |
|
|
|
/* Step 2 (lazy eval): Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
|
|
vector of the median. |
|
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
|
|
*/ |
|
|
|
|
|
if ((bPredEq) && (MVequal(pmv[0], prevMB->i_mvs[0]))) |
|
|
iFound = 2; |
|
|
|
|
|
/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
|
|
Otherwise select large Diamond Search. |
|
1060 |
*/ |
*/ |
1061 |
|
Data.max_dx = 2 * pParam->width - 2 * (x) * 16; |
1062 |
|
Data.max_dy = 2 * pParam->height - 2 * (y) * 16; |
1063 |
|
Data.min_dx = -(2 * 16 + 2 * (x) * 16); |
1064 |
|
Data.min_dy = -(2 * 16 + 2 * (y) * 16); |
1065 |
|
|
1066 |
if ((!MVzero(pmv[0])) || (threshB < 1536) || (bPredEq)) |
for (k = 0; k < 4; k++) { |
1067 |
iDiamondSize = 2; // halfpel units! |
pMB->mvs[k].x = Data.directmvF[k].x = ((TRB * Data.referencemv[k].x) / TRD); |
1068 |
else |
pMB->b_mvs[k].x = Data.directmvB[k].x = ((TRB - TRD) * Data.referencemv[k].x) / TRD; |
1069 |
iDiamondSize = 4; // halfpel units! |
pMB->mvs[k].y = Data.directmvF[k].y = ((TRB * Data.referencemv[k].y) / TRD); |
1070 |
|
pMB->b_mvs[k].y = Data.directmvB[k].y = ((TRB - TRD) * Data.referencemv[k].y) / TRD; |
1071 |
|
|
1072 |
|
if (( pMB->mvs[k].x > Data.max_dx ) || ( pMB->mvs[k].x < Data.min_dx ) |
1073 |
|
|| ( pMB->mvs[k].y > Data.max_dy ) || ( pMB->mvs[k].y < Data.min_dy ) |
1074 |
|
|| ( pMB->b_mvs[k].x > Data.max_dx ) || ( pMB->b_mvs[k].x < Data.min_dx ) |
1075 |
|
|| ( pMB->b_mvs[k].y > Data.max_dy ) || ( pMB->b_mvs[k].y < Data.min_dy )) { |
1076 |
/* |
/* |
1077 |
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
fprintf(debug, "\nERROR - out of range : vector %d,%d and %d,%d\n", pMB->mvs[k].x, pMB->mvs[k].y,pMB->b_mvs[k].x,pMB->b_mvs[k].y ); |
1078 |
Also calculate (0,0) but do not subtract offset. |
fprintf(debug, " range is x: %d..%d y: %d..%d \n", Data.min_dx,Data.max_dx,Data.min_dy,Data.max_dy); |
1079 |
Let MinSAD be the smallest SAD up to this point. |
fprintf(debug,"macroblock %d, %d \n", x, y); |
1080 |
If MV is (0,0) subtract offset. |
fprintf(debug, "direct MV is %d,%d \n", directmv[k].x, directmv[k].y); |
1081 |
*/ |
*/ |
1082 |
|
*best_sad = 256*4096; // in that case, we won't use direct mode |
1083 |
|
pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV" |
1084 |
|
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; /* because backwards and interpol might rely on this */ |
1085 |
|
return 0; } |
1086 |
|
|
1087 |
|
|
1088 |
|
if (b_mb->mode != MODE_INTER4V) { |
1089 |
|
iMinSAD = sad16bi(Data.Cur, |
1090 |
|
get_ref_mv(f_Ref, f_RefH, f_RefV, f_RefHV, |
1091 |
|
x, y, 16, &pMB->mvs[0], iEdgedWidth), |
1092 |
|
get_ref_mv(b_Ref, b_RefH, b_RefV, b_RefHV, |
1093 |
|
x, y, 16, &pMB->b_mvs[0], iEdgedWidth), iEdgedWidth); |
1094 |
|
|
1095 |
|
Data.directmvF[1] = Data.directmvF[2] = Data.directmvF[3] = Data.directmvF[0]; |
1096 |
|
Data.directmvB[1] = Data.directmvB[2] = Data.directmvB[3] = Data.directmvB[0]; |
1097 |
|
break; |
1098 |
|
} |
1099 |
|
iMinSAD += sad8bi(Data.Cur + (k&1)*8 + (k>>1)* 8 * iEdgedWidth, |
1100 |
|
get_ref_mv(f_Ref, f_RefH, f_RefV, f_RefHV, |
1101 |
|
(2*x+(k&1)), (2*y+(k>>1)), 8, &pMB->mvs[k], iEdgedWidth), |
1102 |
|
get_ref_mv(b_Ref, b_RefH, b_RefV, b_RefHV, |
1103 |
|
(2*x+(k&1)), (2*y+(k>>1)), 8, &pMB->b_mvs[k], iEdgedWidth), |
1104 |
|
iEdgedWidth); |
1105 |
|
} |
1106 |
|
|
1107 |
|
// skip decision |
1108 |
|
if (iMinSAD < (int32_t)iQuant * SKIP_THRESH_B) { |
1109 |
|
pMB->mode = MODE_DIRECT_NONE_MV; |
1110 |
|
return iMinSAD; } |
1111 |
|
|
1112 |
|
skip_sad = iMinSAD; |
1113 |
|
iMinSAD += 2 * lambda_vec16[iQuant]; // 2 bits needed to code vector 0,0 |
1114 |
|
currentMV.x = currentMV.y = 0; |
1115 |
|
if (b_mb->mode == MODE_INTER4V) |
1116 |
|
CheckCandidate = CheckCandidateDirect; |
1117 |
|
else CheckCandidate = CheckCandidateDirectno4v; |
1118 |
|
|
1119 |
// (0,0) is often a good choice |
// DIRECT MODE DELTA VECTOR SEARCH. |
1120 |
|
// This has to be made more effective, but at the moment I'm happy it's running at all |
1121 |
|
|
1122 |
if (!MVzero(pmv[0])) |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1123 |
CHECK_MV16_ZERO; |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1124 |
|
else MainSearchPtr = DiamondSearch; |
1125 |
|
|
1126 |
// previous frame MV is always possible |
(*MainSearchPtr)(0, 0, &Data, 255); |
1127 |
|
|
1128 |
if (!MVzero(prevMB->i_mvs[0])) |
HalfpelRefine(&Data); |
|
if (!MVequal(prevMB->i_mvs[0], pmv[0])) |
|
|
CHECK_MV16_CANDIDATE(prevMB->i_mvs[0].x, prevMB->i_mvs[0].y); |
|
|
|
|
|
// left neighbour, if allowed |
|
|
|
|
|
if (!MVzero(pmv[1])) |
|
|
if (!MVequal(pmv[1], prevMB->i_mvs[0])) |
|
|
if (!MVequal(pmv[1], pmv[0])) |
|
|
CHECK_MV16_CANDIDATE(pmv[1].x, pmv[1].y); |
|
|
|
|
|
// top neighbour, if allowed |
|
|
if (!MVzero(pmv[2])) |
|
|
if (!MVequal(pmv[2], prevMB->i_mvs[0])) |
|
|
if (!MVequal(pmv[2], pmv[0])) |
|
|
if (!MVequal(pmv[2], pmv[1])) |
|
|
CHECK_MV16_CANDIDATE(pmv[2].x, pmv[2].y); |
|
|
|
|
|
// top right neighbour, if allowed |
|
|
if (!MVzero(pmv[3])) |
|
|
if (!MVequal(pmv[3], prevMB->i_mvs[0])) |
|
|
if (!MVequal(pmv[3], pmv[0])) |
|
|
if (!MVequal(pmv[3], pmv[1])) |
|
|
if (!MVequal(pmv[3], pmv[2])) |
|
|
CHECK_MV16_CANDIDATE(pmv[3].x, |
|
|
pmv[3].y); |
|
|
|
|
|
if ((MVzero(*currMV)) && |
|
|
(!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96) */ ) |
|
|
iMinSAD -= MV16_00_BIAS; |
|
1129 |
|
|
1130 |
|
iMinSAD += 1 * lambda_vec16[iQuant]; // one bit is needed to code direct mode. we treat this bit just like it was vector's |
1131 |
|
*best_sad = iMinSAD; |
1132 |
|
|
1133 |
/* Step 6: If MinSAD <= thresa goto Step 10. |
if (b_mb->mode == MODE_INTER4V) |
1134 |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
pMB->mode = MODE_DIRECT; |
1135 |
*/ |
else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
1136 |
|
|
1137 |
if ((iMinSAD <= threshA) || |
pMB->pmvs[3] = currentMV; |
|
(MVequal(*currMV, prevMB->i_mvs[0]) && |
|
|
((int32_t) iMinSAD < prevMB->i_sad16))) { |
|
1138 |
|
|
1139 |
if (MotionFlags & PMV_EARLYSTOP16) |
for (k = 0; k < 4; k++) { |
1140 |
goto PMVfastInt16_Terminate_with_Refine; |
pMB->mvs[k].x = Data.directmvF[k].x + currentMV.x; |
1141 |
|
pMB->b_mvs[k].x = ((currentMV.x == 0) |
1142 |
|
? Data.directmvB[k].x |
1143 |
|
: pMB->mvs[k].x - Data.referencemv[k].x); |
1144 |
|
pMB->mvs[k].y = (Data.directmvF[k].y + currentMV.y); |
1145 |
|
pMB->b_mvs[k].y = ((currentMV.y == 0) |
1146 |
|
? Data.directmvB[k].y |
1147 |
|
: pMB->mvs[k].y - Data.referencemv[k].y); |
1148 |
|
if (b_mb->mode != MODE_INTER4V) { |
1149 |
|
pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; |
1150 |
|
pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; |
1151 |
|
break; |
1152 |
} |
} |
1153 |
|
} |
1154 |
|
return 0;//skip_sad; |
1155 |
|
} |
1156 |
|
|
1157 |
|
static __inline void |
1158 |
|
SearchInterpolate(const uint8_t * const f_Ref, |
1159 |
|
const uint8_t * const f_RefH, |
1160 |
|
const uint8_t * const f_RefV, |
1161 |
|
const uint8_t * const f_RefHV, |
1162 |
|
const uint8_t * const b_Ref, |
1163 |
|
const uint8_t * const b_RefH, |
1164 |
|
const uint8_t * const b_RefV, |
1165 |
|
const uint8_t * const b_RefHV, |
1166 |
|
const IMAGE * const pCur, |
1167 |
|
const int x, const int y, |
1168 |
|
const uint32_t fcode, |
1169 |
|
const uint32_t bcode, |
1170 |
|
const uint32_t MotionFlags, |
1171 |
|
const uint32_t iQuant, |
1172 |
|
const MBParam * const pParam, |
1173 |
|
const VECTOR * const f_predMV, |
1174 |
|
const VECTOR * const b_predMV, |
1175 |
|
MACROBLOCK * const pMB, |
1176 |
|
int32_t * const best_sad) |
1177 |
|
|
1178 |
|
{ |
1179 |
|
/* Interpolated MC motion vector search, this is tedious and more complicated because there are |
1180 |
|
two values for everything, always one for backward and one for forward ME. Still, we don't gain |
1181 |
|
much from this search, maybe it should simply be skipped and simply current i_sad16 value used |
1182 |
|
as "optimal". */ |
1183 |
|
|
1184 |
/************ (Diamond Search) **************/ |
const int32_t iEdgedWidth = pParam->edged_width; |
|
/* |
|
|
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. |
|
|
*/ |
|
|
|
|
|
if (MotionFlags & PMV_USESQUARES16) |
|
|
MainSearchPtr = Square16_MainSearch; |
|
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
|
|
MainSearchPtr = AdvDiamond16_MainSearch; |
|
|
else |
|
|
MainSearchPtr = Diamond16_MainSearch; |
|
|
|
|
|
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
|
1185 |
|
|
1186 |
|
int iDirection, i, j; |
1187 |
|
int32_t iMinSAD = 256*4096; |
1188 |
|
VECTOR currentMV[3]; |
1189 |
|
SearchData fData, bData; |
1190 |
|
|
|
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
|
|
iSAD = |
|
|
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV->x, |
|
|
currMV->y, iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, |
|
|
min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, |
|
|
iQuant, iFound); |
|
1191 |
|
|
1192 |
if (iSAD < iMinSAD) { |
fData.iMinSAD = bData.iMinSAD = &iMinSAD; |
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
1193 |
|
|
1194 |
if (MotionFlags & PMV_EXTSEARCH16) { |
fData.Cur = bData.Cur = pCur->y + (x + y * iEdgedWidth) * 16; |
1195 |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
fData.iEdgedWidth = bData.iEdgedWidth = iEdgedWidth; |
1196 |
|
fData.currentMV = currentMV; bData.currentMV = currentMV + 1; |
1197 |
|
fData.iQuant = bData.iQuant = iQuant; |
1198 |
|
fData.iFcode = bData.bFcode = fcode; fData.bFcode = bData.iFcode = bcode; |
1199 |
|
|
|
if (!(MVequal(pmv[0], backupMV))) { |
|
|
iSAD = |
|
|
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, |
|
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, center_x, center_y, |
|
|
min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
|
|
iDiamondSize, iFcode, iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) { |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
1200 |
|
|
1201 |
if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { |
bData.bRef = fData.Ref = f_Ref + (x + y * iEdgedWidth) * 16; |
1202 |
iSAD = |
bData.bRefH = fData.RefH = f_RefH + (x + y * iEdgedWidth) * 16; |
1203 |
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, 0, 0, |
bData.bRefV = fData.RefV = f_RefV + (x + y * iEdgedWidth) * 16; |
1204 |
iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, min_dy, |
bData.bRefHV = fData.RefHV = f_RefHV + (x + y * iEdgedWidth) * 16; |
1205 |
max_dy, iEdgedWidth, iDiamondSize, iFcode, |
bData.Ref = fData.bRef = b_Ref + (x + y * iEdgedWidth) * 16; |
1206 |
iQuant, iFound); |
bData.RefH = fData.bRefH = b_RefH + (x + y * iEdgedWidth) * 16; |
1207 |
|
bData.RefV = fData.bRefV = b_RefV + (x + y * iEdgedWidth) * 16; |
1208 |
if (iSAD < iMinSAD) { |
bData.RefHV = fData.bRefHV = b_RefHV + (x + y * iEdgedWidth) * 16; |
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
|
} |
|
1209 |
|
|
1210 |
/* |
bData.bpredMV = fData.predMV = *f_predMV; |
1211 |
Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
fData.bpredMV = bData.predMV = *b_predMV; |
|
*/ |
|
1212 |
|
|
|
PMVfastInt16_Terminate_with_Refine: |
|
1213 |
|
|
1214 |
pMB->i_mvs[0] = pMB->i_mvs[1] = pMB->i_mvs[2] = pMB->i_mvs[3] = pMB->i_mv16 = *currMV; |
currentMV[0] = pMB->mvs[0]; |
1215 |
pMB->i_sad8[0] = pMB->i_sad8[1] = pMB->i_sad8[2] = pMB->i_sad8[3] = pMB->i_sad16 = iMinSAD; |
currentMV[1] = pMB->b_mvs[0]; |
1216 |
|
get_range(&fData.min_dx, &fData.max_dx, &fData.min_dy, &fData.max_dy, x, y, 16, pParam->width, pParam->height, fcode); |
1217 |
|
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode); |
1218 |
|
|
1219 |
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
CheckCandidateInt(currentMV[0].x, currentMV[0].y, 255, &iDirection, &fData); |
|
iMinSAD = |
|
|
Halfpel16_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, |
|
|
iMinSAD, center_x, center_y, min_dx, max_dx, min_dy, max_dy, |
|
|
iFcode, iQuant, iEdgedWidth); |
|
1220 |
|
|
1221 |
pmv[0] = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); // get _REAL_ prediction (halfpel possible) |
//diamond. I wish we could use normal mainsearch functions (square, advdiamond) |
1222 |
|
|
1223 |
PMVfastInt16_Terminate_without_Refine: |
do { |
1224 |
currPMV->x = currMV->x - center_x; |
iDirection = 255; |
1225 |
currPMV->y = currMV->y - center_y; |
// forward MV moves |
1226 |
return iMinSAD; |
i = currentMV[0].x; j = currentMV[0].y; |
1227 |
|
|
1228 |
|
CheckCandidateInt(i + 2, j, 0, &iDirection, &fData); |
1229 |
|
CheckCandidateInt(i, j + 2, 0, &iDirection, &fData); |
1230 |
|
CheckCandidateInt(i - 2, j, 0, &iDirection, &fData); |
1231 |
|
CheckCandidateInt(i, j - 2, 0, &iDirection, &fData); |
1232 |
|
|
1233 |
|
// backward MV moves |
1234 |
|
i = currentMV[1].x; j = currentMV[1].y; |
1235 |
|
currentMV[2] = currentMV[0]; |
1236 |
|
|
1237 |
|
CheckCandidateInt(i + 2, j, 0, &iDirection, &bData); |
1238 |
|
CheckCandidateInt(i, j + 2, 0, &iDirection, &bData); |
1239 |
|
CheckCandidateInt(i - 2, j, 0, &iDirection, &bData); |
1240 |
|
CheckCandidateInt(i, j - 2, 0, &iDirection, &bData); |
1241 |
|
|
1242 |
|
} while (!(iDirection)); |
1243 |
|
|
1244 |
|
/* halfpel refinement. luckly we can use normal halfpel function for it */ |
1245 |
|
|
1246 |
|
if (MotionFlags & PMV_HALFPELREFINE16) { |
1247 |
|
CheckCandidate = CheckCandidateInt; |
1248 |
|
HalfpelRefine(&fData); |
1249 |
|
currentMV[2] = currentMV[0]; |
1250 |
|
HalfpelRefine(&bData); |
1251 |
|
} |
1252 |
|
|
1253 |
|
// two bits are needed to code interpolate mode. we treat the bits just like they were vector's |
1254 |
|
iMinSAD += 2 * lambda_vec16[iQuant]; |
1255 |
|
if (iMinSAD < *best_sad) { |
1256 |
|
*best_sad = iMinSAD; |
1257 |
|
pMB->mvs[0] = currentMV[0]; |
1258 |
|
pMB->b_mvs[0] = currentMV[1]; |
1259 |
|
pMB->mode = MODE_INTERPOLATE; |
1260 |
|
|
1261 |
|
pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; |
1262 |
|
pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; |
1263 |
|
pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; |
1264 |
|
pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; |
1265 |
|
} |
1266 |
} |
} |
|
|
|
|
|
|
|
|
|
|
/* *********************************************************** |
|
|
bvop motion estimation |
|
|
***************************************************************/ |
|
1267 |
|
|
1268 |
void |
void |
1269 |
MotionEstimationBVOP(MBParam * const pParam, |
MotionEstimationBVOP(MBParam * const pParam, |
1283 |
const IMAGE * const b_refV, |
const IMAGE * const b_refV, |
1284 |
const IMAGE * const b_refHV) |
const IMAGE * const b_refHV) |
1285 |
{ |
{ |
1286 |
const int mb_width = pParam->mb_width; |
uint32_t i, j; |
1287 |
const int mb_height = pParam->mb_height; |
int32_t best_sad, skip_sad; |
1288 |
const int edged_width = pParam->edged_width; |
int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; |
|
|
|
|
const int32_t iWidth = pParam->width; |
|
|
const int32_t iHeight = pParam->height; |
|
|
|
|
|
int i, j, k; |
|
|
|
|
1289 |
static const VECTOR zeroMV={0,0}; |
static const VECTOR zeroMV={0,0}; |
1290 |
|
|
|
int f_sad16; /* forward (as usual) search */ |
|
|
int b_sad16; /* backward (only in b-frames) search */ |
|
|
int i_sad16; /* interpolated (both direction, b-frames only) */ |
|
|
int d_sad16; /* direct mode (assume almost linear motion) */ |
|
|
|
|
|
int best_sad; |
|
|
|
|
1291 |
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
|
VECTOR f_interpolMV, b_interpolMV; |
|
|
VECTOR pmv_dontcare; |
|
1292 |
|
|
1293 |
int min_dx, max_dx, min_dy, max_dy; |
const int32_t TRB = time_pp - time_bp; |
1294 |
int f_min_dx, f_max_dx, f_min_dy, f_max_dy; |
const int32_t TRD = time_pp; |
|
int b_min_dx, b_max_dx, b_min_dy, b_max_dy; |
|
|
|
|
|
int f_count=0; |
|
|
int b_count=0; |
|
|
int i_count=0; |
|
|
int d_count=0; |
|
1295 |
|
|
|
const int64_t TRB = (int32_t)time_pp - (int32_t)time_bp; |
|
|
const int64_t TRD = (int32_t)time_pp; |
|
|
|
|
|
// fprintf(stderr,"TRB = %lld TRD = %lld time_bp =%d time_pp =%d\n\n",TRB,TRD,time_bp,time_pp); |
|
1296 |
// note: i==horizontal, j==vertical |
// note: i==horizontal, j==vertical |
|
for (j = 0; j < mb_height; j++) { |
|
1297 |
|
|
1298 |
f_predMV = zeroMV; /* prediction is reset at left boundary */ |
for (j = 0; j < pParam->mb_height; j++) { |
|
b_predMV = zeroMV; |
|
1299 |
|
|
1300 |
for (i = 0; i < mb_width; i++) { |
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
|
MACROBLOCK *mb = &frame->mbs[i + j * mb_width]; |
|
|
const MACROBLOCK *f_mb = &f_mbs[i + j * mb_width]; |
|
|
const MACROBLOCK *b_mb = &b_mbs[i + j * mb_width]; |
|
1301 |
|
|
1302 |
mb->deltamv=zeroMV; |
for (i = 0; i < pParam->mb_width; i++) { |
1303 |
|
MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; |
1304 |
|
const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; |
1305 |
|
|
1306 |
/* special case, if collocated block is SKIPed: encoding is forward (0,0), cpb=0 without further ado */ |
/* special case, if collocated block is SKIPed: encoding is forward (0,0), cpb=0 without further ado */ |
1307 |
|
if (b_mb->mode == MODE_NOT_CODED) { |
1308 |
if (b_mb->mode == MODE_INTER && b_mb->cbp == 0 && |
pMB->mode = MODE_NOT_CODED; |
|
b_mb->mvs[0].x == 0 && b_mb->mvs[0].y == 0) { |
|
|
mb->mode = MODE_NOT_CODED; |
|
|
mb->b_mvs[0] = mb->mvs[0] = zeroMV; |
|
1309 |
continue; |
continue; |
1310 |
} |
} |
1311 |
|
|
1312 |
if (b_mb->mode == MODE_INTER4V) |
/* direct search comes first, because it (1) checks for SKIP-mode |
1313 |
{ |
and (2) sets very good predictions for forward and backward search */ |
|
d_sad16 = 0; |
|
|
/* same method of scaling as in decoder.c, so we copy from there */ |
|
|
for (k = 0; k < 4; k++) { |
|
1314 |
|
|
1315 |
mb->directmv[k] = b_mb->mvs[k]; |
skip_sad = SearchDirect(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
1316 |
|
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
1317 |
|
&frame->image, |
1318 |
|
i, j, |
1319 |
|
frame->motion_flags, |
1320 |
|
frame->quant, |
1321 |
|
TRB, TRD, |
1322 |
|
pParam, |
1323 |
|
pMB, b_mb, |
1324 |
|
&best_sad); |
1325 |
|
|
1326 |
mb->mvs[k].x = (int32_t) ((TRB * mb->directmv[k].x) / TRD + mb->deltamv.x); |
if (!(frame->global_flags & XVID_HALFPEL)) best_sad = skip_sad = 256*4096; |
|
mb->b_mvs[k].x = (int32_t) ((mb->deltamv.x == 0) |
|
|
? ((TRB - TRD) * mb->directmv[k].x) / TRD |
|
|
: mb->mvs[k].x - mb->directmv[k].x); |
|
|
|
|
|
mb->mvs[k].y = (int32_t) ((TRB * mb->directmv[k].y) / TRD + mb->deltamv.y); |
|
|
mb->b_mvs[k].y = (int32_t) ((mb->deltamv.y == 0) |
|
|
? ((TRB - TRD) * mb->directmv[k].y) / TRD |
|
|
: mb->mvs[k].y - mb->directmv[k].y); |
|
|
|
|
|
d_sad16 += |
|
|
sad8bi(frame->image.y + (2*i+(k&1))*8 + (2*j+(k>>1))*8*edged_width, |
|
|
get_ref_mv(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
(2*i+(k&1)), (2*j+(k>>1)), 8, &mb->mvs[k], edged_width), |
|
|
get_ref_mv(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
(2*i+(k&1)), (2*j+(k>>1)), 8, &mb->b_mvs[k], edged_width), |
|
|
edged_width); |
|
|
} |
|
|
} |
|
1327 |
else |
else |
1328 |
{ |
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
|
mb->directmv[3] = mb->directmv[2] = mb->directmv[1] = |
|
|
mb->directmv[0] = b_mb->mvs[0]; |
|
1329 |
|
|
1330 |
mb->mvs[0].x = (int32_t) ((TRB * mb->directmv[0].x) / TRD + mb->deltamv.x); |
// best_sad = 256*4096; //uncomment to disable Directsearch. |
1331 |
mb->b_mvs[0].x = (int32_t) ((mb->deltamv.x == 0) |
// To disable any other mode, just comment the function call |
|
? ((TRB - TRD) * mb->directmv[0].x) / TRD |
|
|
: mb->mvs[0].x - mb->directmv[0].x); |
|
|
|
|
|
mb->mvs[0].y = (int32_t) ((TRB * mb->directmv[0].y) / TRD + mb->deltamv.y); |
|
|
mb->b_mvs[0].y = (int32_t) ((mb->directmv[0].y == 0) |
|
|
? ((TRB - TRD) * mb->directmv[0].y) / TRD |
|
|
: mb->mvs[0].y - mb->directmv[0].y); |
|
|
|
|
|
d_sad16 = sad16bi(frame->image.y + i * 16 + j * 16 * edged_width, |
|
|
get_ref_mv(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
i, j, 16, &mb->mvs[0], edged_width), |
|
|
get_ref_mv(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, 16, &mb->b_mvs[0], edged_width), |
|
|
edged_width); |
|
|
|
|
|
} |
|
|
d_sad16 += calc_delta_16(mb->deltamv.x, mb->deltamv.y, 1, frame->quant); |
|
1332 |
|
|
1333 |
// forward search |
// forward search |
1334 |
f_sad16 = SEARCH16(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
1335 |
&frame->image, i, j, |
&frame->image, i, j, |
|
mb->mvs[0].x, mb->mvs[0].y, /* start point f_directMV */ |
|
|
f_predMV.x, f_predMV.y, /* center is f-prediction */ |
|
1336 |
frame->motion_flags, |
frame->motion_flags, |
1337 |
frame->quant, frame->fcode, pParam, |
frame->quant, frame->fcode, pParam, |
1338 |
f_mbs, f_mbs, |
pMB, &f_predMV, &best_sad, |
1339 |
&mb->mvs[0], &pmv_dontcare); |
MODE_FORWARD); |
|
|
|
1340 |
|
|
1341 |
// backward search |
// backward search |
1342 |
b_sad16 = SEARCH16(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
SearchBF(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
1343 |
&frame->image, i, j, |
&frame->image, i, j, |
|
mb->b_mvs[0].x, mb->b_mvs[0].y, /* start point b_directMV */ |
|
|
b_predMV.x, b_predMV.y, /* center is b-prediction */ |
|
1344 |
frame->motion_flags, |
frame->motion_flags, |
1345 |
frame->quant, frame->bcode, pParam, |
frame->quant, frame->bcode, pParam, |
1346 |
b_mbs, b_mbs, |
pMB, &b_predMV, &best_sad, |
1347 |
&mb->b_mvs[0], &pmv_dontcare); |
MODE_BACKWARD); |
|
|
|
|
i_sad16 = |
|
|
sad16bi(frame->image.y + i * 16 + j * 16 * edged_width, |
|
|
get_ref_mv(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
i, j, 16, &mb->mvs[0], edged_width), |
|
|
get_ref_mv(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, 16, &mb->b_mvs[0], edged_width), |
|
|
edged_width); |
|
|
i_sad16 += calc_delta_16(mb->mvs[0].x-f_predMV.x, mb->mvs[0].y-f_predMV.y, |
|
|
frame->fcode, frame->quant); |
|
|
i_sad16 += calc_delta_16(mb->b_mvs[0].x-b_predMV.x, mb->b_mvs[0].y-b_predMV.y, |
|
|
frame->bcode, frame->quant); |
|
|
|
|
|
get_range(&f_min_dx, &f_max_dx, &f_min_dy, &f_max_dy, i, j, 16, iWidth, iHeight, |
|
|
frame->fcode); |
|
|
get_range(&b_min_dx, &b_max_dx, &b_min_dy, &b_max_dy, i, j, 16, iWidth, iHeight, |
|
|
frame->bcode); |
|
1348 |
|
|
1349 |
/* Interpolated MC motion vector search, this is tedious and more complicated because there are |
// interpolate search comes last, because it uses data from forward and backward as prediction |
|
two values for everything, always one for backward and one for forward ME. Still, we don't gain |
|
|
much from this search, maybe it should simply be skipped and simply current i_sad16 value used |
|
|
as "optimal". */ |
|
1350 |
|
|
1351 |
i_sad16 = Diamond16_InterpolMainSearch( |
SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
frame->image.y + i * 16 + j * 16 * edged_width, |
|
|
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, |
|
|
mb->mvs[0].x, mb->mvs[0].y, |
|
|
mb->b_mvs[0].x, mb->b_mvs[0].y, |
|
|
i_sad16, |
|
|
&f_interpolMV, &b_interpolMV, |
|
|
f_predMV.x, f_predMV.y, b_predMV.x, b_predMV.y, |
|
|
f_min_dx, f_max_dx, f_min_dy, f_max_dy, |
|
|
b_min_dx, b_max_dx, b_min_dy, b_max_dy, |
|
|
edged_width, 2, |
|
|
frame->fcode, frame->bcode,frame->quant,0); |
|
|
|
|
|
i_sad16 = Diamond16_InterpolMainSearch( |
|
|
f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
frame->image.y + i * 16 + j * 16 * edged_width, |
|
|
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, |
|
|
f_interpolMV.x, f_interpolMV.y, |
|
|
b_interpolMV.x, b_interpolMV.y, |
|
|
i_sad16, |
|
|
&f_interpolMV, &b_interpolMV, |
|
|
f_predMV.x, f_predMV.y, b_predMV.x, b_predMV.y, |
|
|
f_min_dx, f_max_dx, f_min_dy, f_max_dy, |
|
|
b_min_dx, b_max_dx, b_min_dy, b_max_dy, |
|
|
edged_width, 1, |
|
|
frame->fcode, frame->bcode,frame->quant,0); // equiv to halfpel refine |
|
|
|
|
|
|
|
|
/* DIRECT MODE DELTA VECTOR SEARCH. |
|
|
This has to be made more effective, but at the moment I'm happy it's running at all */ |
|
|
|
|
|
/* There are two range restrictions for direct mode: deltaMV is limited to [-32,31] in halfpel units, and |
|
|
absolute vector must not lie outside of image dimensions. Constraint one is dealt with by CHECK_MV16_DIRECT |
|
|
and for constraint two we need distance to boundary. This is done by get_range very large fcode (hack!) */ |
|
|
|
|
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, i, j, 16, iWidth, iHeight, 19); |
|
|
|
|
|
d_sad16 = Diamond16_DirectMainSearch( |
|
|
f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
frame->image.y + i*16 + j*16*edged_width, |
|
|
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, |
|
|
TRB,TRD, |
|
|
0,0, |
|
|
d_sad16, |
|
|
&mb->deltamv, |
|
|
mb->directmv, // this has to be pre-initialized with b_mb->mvs[] |
|
|
min_dx, max_dx, min_dy, max_dy, |
|
|
edged_width, 2, frame->quant, 0); |
|
|
|
|
|
d_sad16 = Diamond16_DirectMainSearch( |
|
|
f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
frame->image.y + i*16 + j*16*edged_width, |
|
1352 |
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
1353 |
|
&frame->image, |
1354 |
i, j, |
i, j, |
1355 |
TRB,TRD, |
frame->fcode, frame->bcode, |
1356 |
mb->deltamv.x, mb->deltamv.y, |
frame->motion_flags, |
1357 |
d_sad16, |
frame->quant, pParam, |
1358 |
&mb->deltamv, |
&f_predMV, &b_predMV, |
1359 |
mb->directmv, // this has to be pre-initialized with b_mb->mvs[] |
pMB, &best_sad); |
1360 |
min_dx, max_dx, min_dy, max_dy, |
|
1361 |
edged_width, 1, frame->quant, 0); // equiv to halfpel refine |
switch (pMB->mode) { |
1362 |
|
case MODE_FORWARD: |
1363 |
|
f_count++; |
1364 |
// i_sad16 = 65535; /* remove the comment to disable any of the MODEs */ |
f_predMV = pMB->mvs[0]; |
1365 |
// f_sad16 = 65535; |
break; |
1366 |
// b_sad16 = 65535; |
case MODE_BACKWARD: |
1367 |
// d_sad16 = 65535; |
b_count++; |
1368 |
|
b_predMV = pMB->b_mvs[0]; |
1369 |
if (f_sad16 < b_sad16) { |
break; |
1370 |
best_sad = f_sad16; |
case MODE_INTERPOLATE: |
1371 |
mb->mode = MODE_FORWARD; |
i_count++; |
1372 |
} else { |
f_predMV = pMB->mvs[0]; |
1373 |
best_sad = b_sad16; |
b_predMV = pMB->b_mvs[0]; |
1374 |
mb->mode = MODE_BACKWARD; |
break; |
1375 |
|
case MODE_DIRECT: |
1376 |
|
case MODE_DIRECT_NO4V: |
1377 |
|
d_count++; |
1378 |
|
break; |
1379 |
|
default: |
1380 |
|
break; |
1381 |
|
} |
1382 |
} |
} |
1383 |
|
} |
1384 |
|
|
1385 |
|
// fprintf(debug,"B-Stat: F: %04d B: %04d I: %04d D: %04d, N: %04d\n", |
1386 |
|
// f_count,b_count,i_count,d_count,n_count); |
1387 |
|
|
|
if (i_sad16 < best_sad) { |
|
|
best_sad = i_sad16; |
|
|
mb->mode = MODE_INTERPOLATE; |
|
1388 |
} |
} |
1389 |
|
|
1390 |
if (d_sad16 < best_sad) { |
/* Hinted ME starts here */ |
1391 |
|
|
1392 |
if (b_mb->mode == MODE_INTER4V) |
static __inline void |
1393 |
|
Search8hinted( const SearchData * const OldData, |
1394 |
|
const int x, const int y, |
1395 |
|
const uint32_t MotionFlags, |
1396 |
|
const MBParam * const pParam, |
1397 |
|
MACROBLOCK * const pMB, |
1398 |
|
const MACROBLOCK * const pMBs, |
1399 |
|
const int block) |
1400 |
{ |
{ |
1401 |
|
SearchData Data; |
1402 |
|
MainSearchFunc *MainSearchPtr; |
1403 |
|
|
1404 |
/* how to calc vectors is defined in standard. mvs[] and b_mvs[] are only for motion compensation */ |
Data.predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
1405 |
/* for the bitstream, the value mb->deltamv is read directly */ |
Data.iMinSAD = OldData->iMinSAD + 1 + block; |
1406 |
|
Data.currentMV = OldData->currentMV+1+block; |
1407 |
|
Data.iFcode = OldData->iFcode; |
1408 |
|
Data.iQuant = OldData->iQuant; |
1409 |
|
|
1410 |
for (k = 0; k < 4; k++) { |
Data.Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1411 |
|
Data.RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1412 |
|
Data.RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1413 |
|
Data.RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1414 |
|
Data.iEdgedWidth = pParam->edged_width; |
1415 |
|
Data.Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1416 |
|
|
1417 |
mb->mvs[k].x = (int32_t) ((TRB * mb->directmv[k].x) / TRD + mb->deltamv.x); |
CheckCandidate = CheckCandidate8; |
1418 |
mb->b_mvs[k].x = (int32_t) ((mb->deltamv.x == 0) |
|
1419 |
? ((TRB - TRD) * mb->directmv[k].x) / TRD |
if (block != 0) |
1420 |
: mb->mvs[k].x - mb->directmv[k].x); |
*(Data.iMinSAD) += lambda_vec8[Data.iQuant] * |
1421 |
|
d_mv_bits( Data.currentMV->x - Data.predMV.x, |
1422 |
mb->mvs[k].y = (int32_t) ((TRB * mb->directmv[k].y) / TRD + mb->deltamv.y); |
Data.currentMV->y - Data.predMV.y, |
1423 |
mb->b_mvs[k].y = (int32_t) ((mb->deltamv.y == 0) |
Data.iFcode); |
1424 |
? ((TRB - TRD) * mb->directmv[k].y) / TRD |
|
1425 |
: mb->mvs[k].y - mb->directmv[k].y); |
|
1426 |
} |
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 8, |
1427 |
|
pParam->width, pParam->height, OldData->iFcode); |
1428 |
|
|
1429 |
|
if (pMB->mode == MODE_INTER4V) { |
1430 |
|
int dummy; |
1431 |
|
CheckCandidate8(pMB->mvs[block].x, pMB->mvs[block].y, 0, &dummy, &Data); } |
1432 |
|
|
1433 |
|
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
1434 |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
1435 |
|
else MainSearchPtr = DiamondSearch; |
1436 |
|
|
1437 |
|
(*MainSearchPtr)(Data.currentMV->x, Data.currentMV->y, &Data, 255); |
1438 |
|
|
1439 |
|
if (MotionFlags & PMV_HALFPELREFINE8) HalfpelRefine(&Data); |
1440 |
|
|
1441 |
|
pMB->pmvs[block].x = Data.currentMV->x - Data.predMV.x; |
1442 |
|
pMB->pmvs[block].y = Data.currentMV->y - Data.predMV.y; |
1443 |
|
pMB->mvs[block] = *(Data.currentMV); |
1444 |
|
pMB->sad8[block] = 4 * (*(Data.iMinSAD)); |
1445 |
} |
} |
|
else |
|
|
{ |
|
|
mb->mvs[0].x = (int32_t) ((TRB * mb->directmv[0].x) / TRD + mb->deltamv.x); |
|
1446 |
|
|
|
mb->b_mvs[0].x = (int32_t) ((mb->deltamv.x == 0) |
|
|
? ((TRB - TRD) * mb->directmv[0].x) / TRD |
|
|
: mb->mvs[0].x - mb->directmv[0].x); |
|
1447 |
|
|
1448 |
mb->mvs[0].y = (int32_t) ((TRB * mb->directmv[0].y) / TRD + mb->deltamv.y); |
static void |
1449 |
|
SearchPhinted ( const uint8_t * const pRef, |
1450 |
|
const uint8_t * const pRefH, |
1451 |
|
const uint8_t * const pRefV, |
1452 |
|
const uint8_t * const pRefHV, |
1453 |
|
const IMAGE * const pCur, |
1454 |
|
const int x, |
1455 |
|
const int y, |
1456 |
|
const uint32_t MotionFlags, |
1457 |
|
const uint32_t iQuant, |
1458 |
|
const uint32_t iFcode, |
1459 |
|
const MBParam * const pParam, |
1460 |
|
const MACROBLOCK * const pMBs, |
1461 |
|
int inter4v, |
1462 |
|
MACROBLOCK * const pMB) |
1463 |
|
{ |
1464 |
|
|
1465 |
|
const int32_t iEdgedWidth = pParam->edged_width; |
1466 |
|
|
1467 |
mb->b_mvs[0].y = (int32_t) ((mb->deltamv.y == 0) |
int i; |
1468 |
? ((TRB - TRD) * mb->directmv[0].y) / TRD |
VECTOR currentMV[5]; |
1469 |
: mb->mvs[0].y - mb->directmv[0].y); |
int32_t iMinSAD[5]; |
1470 |
|
int32_t temp[5]; |
1471 |
|
MainSearchFunc * MainSearchPtr; |
1472 |
|
SearchData Data; |
1473 |
|
|
1474 |
|
Data.predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
1475 |
|
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 16, |
1476 |
|
pParam->width, pParam->height, iFcode); |
1477 |
|
|
1478 |
|
Data.Cur = pCur->y + (x + y * iEdgedWidth) * 16; |
1479 |
|
Data.iEdgedWidth = iEdgedWidth; |
1480 |
|
Data.currentMV = currentMV; |
1481 |
|
Data.iMinSAD = iMinSAD; |
1482 |
|
Data.Ref = pRef + (x + iEdgedWidth*y)*16; |
1483 |
|
Data.RefH = pRefH + (x + iEdgedWidth*y) * 16; |
1484 |
|
Data.RefV = pRefV + (x + iEdgedWidth*y) * 16; |
1485 |
|
Data.RefHV = pRefHV + (x + iEdgedWidth*y) * 16; |
1486 |
|
Data.temp = temp; |
1487 |
|
Data.iQuant = iQuant; |
1488 |
|
Data.iFcode = iFcode; |
1489 |
|
|
1490 |
mb->mvs[3] = mb->mvs[2] = mb->mvs[1] = mb->mvs[0]; |
if (!(MotionFlags & PMV_HALFPEL16)) { |
1491 |
mb->b_mvs[3] = mb->b_mvs[2] = mb->b_mvs[1] = mb->b_mvs[0]; |
Data.min_dx = EVEN(Data.min_dx); |
1492 |
|
Data.max_dx = EVEN(Data.max_dx); |
1493 |
|
Data.min_dy = EVEN(Data.min_dy); |
1494 |
|
Data.max_dy = EVEN(Data.max_dy); |
1495 |
} |
} |
1496 |
|
|
1497 |
best_sad = d_sad16; |
for(i = 0; i < 5; i++) iMinSAD[i] = MV_MAX_ERROR; |
1498 |
mb->mode = MODE_DIRECT; |
|
1499 |
|
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
1500 |
|
|
1501 |
|
if (inter4v) |
1502 |
|
CheckCandidate = CheckCandidate16; |
1503 |
|
else CheckCandidate = CheckCandidate16no4v; |
1504 |
|
|
1505 |
|
|
1506 |
|
pMB->mvs[0].x = EVEN(pMB->mvs[0].x); |
1507 |
|
pMB->mvs[0].y = EVEN(pMB->mvs[0].y); |
1508 |
|
if (pMB->mvs[0].x > Data.max_dx) pMB->mvs[0].x = Data.max_dx; // this is in case iFcode changed |
1509 |
|
if (pMB->mvs[0].x < Data.min_dx) pMB->mvs[0].x = Data.min_dx; |
1510 |
|
if (pMB->mvs[0].y > Data.max_dy) pMB->mvs[0].y = Data.max_dy; |
1511 |
|
if (pMB->mvs[0].y < Data.min_dy) pMB->mvs[0].y = Data.min_dy; |
1512 |
|
|
1513 |
|
CheckCandidate16(pMB->mvs[0].x, pMB->mvs[0].y, 0, &i, &Data); |
1514 |
|
|
1515 |
|
if (pMB->mode == MODE_INTER4V) |
1516 |
|
for (i = 1; i < 4; i++) { // all four vectors will be used as four predictions for 16x16 search |
1517 |
|
pMB->mvs[i].x = EVEN(pMB->mvs[i].x); |
1518 |
|
pMB->mvs[i].y = EVEN(pMB->mvs[i].y); |
1519 |
|
if (!(make_mask(pMB->mvs, i))) |
1520 |
|
CheckCandidate16(pMB->mvs[i].x, pMB->mvs[i].y, 0, &i, &Data); |
1521 |
} |
} |
1522 |
|
|
1523 |
switch (mb->mode) |
if (MotionFlags & PMV_USESQUARES16) |
1524 |
{ |
MainSearchPtr = SquareSearch; |
1525 |
case MODE_FORWARD: |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
1526 |
f_count++; |
MainSearchPtr = AdvDiamondSearch; |
1527 |
f_predMV = mb->mvs[0]; |
else MainSearchPtr = DiamondSearch; |
|
break; |
|
|
case MODE_BACKWARD: |
|
|
b_count++; |
|
|
b_predMV = mb->b_mvs[0]; |
|
1528 |
|
|
1529 |
break; |
(*MainSearchPtr)(currentMV->x, currentMV->y, &Data, 255); |
1530 |
case MODE_INTERPOLATE: |
|
1531 |
i_count++; |
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(&Data); |
1532 |
mb->mvs[0] = f_interpolMV; |
|
1533 |
mb->b_mvs[0] = b_interpolMV; |
if (inter4v) |
1534 |
f_predMV = mb->mvs[0]; |
for(i = 0; i < 4; i++) |
1535 |
b_predMV = mb->b_mvs[0]; |
Search8hinted(&Data, 2*x+(i&1), 2*y+(i>>1), MotionFlags, pParam, pMB, pMBs, i); |
1536 |
break; |
|
1537 |
case MODE_DIRECT: |
if (!(inter4v) || |
1538 |
d_count++; |
(iMinSAD[0] < iMinSAD[1] + iMinSAD[2] + iMinSAD[3] + iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
1539 |
break; |
// INTER MODE |
1540 |
default: |
|
1541 |
break; |
pMB->mode = MODE_INTER; |
1542 |
|
pMB->mv16 = pMB->mvs[0] = pMB->mvs[1] |
1543 |
|
= pMB->mvs[2] = pMB->mvs[3] = currentMV[0]; |
1544 |
|
|
1545 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
1546 |
|
pMB->sad8[2] = pMB->sad8[3] = iMinSAD[0]; |
1547 |
|
|
1548 |
|
pMB->pmvs[0].x = currentMV[0].x - Data.predMV.x; |
1549 |
|
pMB->pmvs[0].y = currentMV[0].y - Data.predMV.y; |
1550 |
|
} else { |
1551 |
|
// INTER4V MODE; all other things are already set in Search8hinted |
1552 |
|
pMB->mode = MODE_INTER4V; |
1553 |
|
pMB->sad16 = iMinSAD[1] + iMinSAD[2] + iMinSAD[3] + iMinSAD[4] + IMV16X16 * iQuant; |
1554 |
} |
} |
1555 |
|
|
1556 |
} |
} |
1557 |
|
|
1558 |
|
void |
1559 |
|
MotionEstimationHinted( MBParam * const pParam, |
1560 |
|
FRAMEINFO * const current, |
1561 |
|
FRAMEINFO * const reference, |
1562 |
|
const IMAGE * const pRefH, |
1563 |
|
const IMAGE * const pRefV, |
1564 |
|
const IMAGE * const pRefHV) |
1565 |
|
{ |
1566 |
|
MACROBLOCK *const pMBs = current->mbs; |
1567 |
|
const IMAGE *const pCurrent = ¤t->image; |
1568 |
|
const IMAGE *const pRef = &reference->image; |
1569 |
|
|
1570 |
|
uint32_t x, y; |
1571 |
|
|
1572 |
|
if (sadInit) (*sadInit) (); |
1573 |
|
|
1574 |
|
for (y = 0; y < pParam->mb_height; y++) { |
1575 |
|
for (x = 0; x < pParam->mb_width; x++) { |
1576 |
|
int32_t sad00; |
1577 |
|
|
1578 |
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
1579 |
|
|
1580 |
|
//intra mode is copied from the first pass. At least for the time being |
1581 |
|
if ((pMB->mode == MODE_INTRA) || (pMB->mode == MODE_NOT_CODED) ) continue; |
1582 |
|
|
1583 |
|
if (!(current->global_flags & XVID_LUMIMASKING)) { |
1584 |
|
pMB->dquant = NO_CHANGE; |
1585 |
|
pMB->quant = current->quant; } |
1586 |
|
|
1587 |
|
if (pMB->dquant == NO_CHANGE) //no skip otherwise, anyway |
1588 |
|
sad00 = pMB->sad16 |
1589 |
|
= sad16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
1590 |
|
pRef->y + (x + y * pParam->edged_width) * 16, |
1591 |
|
pParam->edged_width, 256*4096 ); |
1592 |
|
else sad00 = 256*4096; |
1593 |
|
|
1594 |
|
|
1595 |
|
//initial skip decision |
1596 |
|
|
1597 |
|
if ( (pMB->dquant == NO_CHANGE) && (sad00 <= MAX_SAD00_FOR_SKIP * pMB->quant) |
1598 |
|
&& ( //(pMB->mode == MODE_NOT_CODED) || |
1599 |
|
(SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant) )) ) { |
1600 |
|
if (sad00 < pMB->quant * INITIAL_SKIP_THRESH) { |
1601 |
|
SkipMacroblockP(pMB, sad00); |
1602 |
|
continue; } //skipped |
1603 |
} |
} |
1604 |
|
else sad00 = 256*4096; |
1605 |
|
|
1606 |
|
if (pMB->mode == MODE_NOT_CODED) |
1607 |
|
SearchP( pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
1608 |
|
y, current->motion_flags, pMB->quant, |
1609 |
|
current->fcode, pParam, pMBs, reference->mbs, |
1610 |
|
current->global_flags & XVID_INTER4V, pMB); |
1611 |
|
|
1612 |
#ifdef _DEBUG_BFRAME_STAT |
else |
1613 |
fprintf(stderr,"B-Stat: F: %04d B: %04d I: %04d D: %04d\n", |
SearchPhinted(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
1614 |
f_count,b_count,i_count,d_count); |
y, current->motion_flags, pMB->quant, |
1615 |
#endif |
current->fcode, pParam, pMBs, |
1616 |
|
current->global_flags & XVID_INTER4V, pMB); |
1617 |
|
|
1618 |
|
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
1619 |
|
if (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
1620 |
|
if ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) |
1621 |
|
SkipMacroblockP(pMB, sad00); |
1622 |
|
|
1623 |
} |
} |
1624 |
|
} |
1625 |
|
} |
1626 |
|
|