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 |
|
#include "../utils/emms.h" |
44 |
|
|
45 |
|
#define INITIAL_SKIP_THRESH (10) |
46 |
|
#define FINAL_SKIP_THRESH (50) |
47 |
|
#define MAX_SAD00_FOR_SKIP (20) |
48 |
|
#define MAX_CHROMA_SAD_FOR_SKIP (22) |
49 |
|
#define SKIP_THRESH_B (25) |
50 |
|
|
51 |
|
#define CHECK_CANDIDATE(X,Y,D) { \ |
52 |
|
(*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } |
53 |
|
|
54 |
|
#define iDiamondSize 2 |
55 |
|
|
56 |
static int32_t lambda_vec16[32] = /* rounded values for lambda param for weight of motion bits as in modified H.26L */ |
static __inline int |
57 |
{ 0, (int) (1.00235 + 0.5), (int) (1.15582 + 0.5), (int) (1.31976 + 0.5), |
d_mv_bits(int x, int y, const uint32_t iFcode) |
58 |
(int) (1.49591 + 0.5), (int) (1.68601 + 0.5), |
{ |
59 |
(int) (1.89187 + 0.5), (int) (2.11542 + 0.5), (int) (2.35878 + 0.5), |
int xb, yb; |
|
(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 */ |
|
60 |
|
|
61 |
|
if (x == 0) xb = 1; |
62 |
|
else { |
63 |
|
if (x < 0) x = -x; |
64 |
|
x += (1 << (iFcode - 1)) - 1; |
65 |
|
x >>= (iFcode - 1); |
66 |
|
if (x > 32) x = 32; |
67 |
|
xb = mvtab[x] + iFcode; |
68 |
|
} |
69 |
|
|
70 |
|
if (y == 0) yb = 1; |
71 |
|
else { |
72 |
|
if (y < 0) y = -y; |
73 |
|
y += (1 << (iFcode - 1)) - 1; |
74 |
|
y >>= (iFcode - 1); |
75 |
|
if (y > 32) y = 32; |
76 |
|
yb = mvtab[y] + iFcode; |
77 |
|
} |
78 |
|
return xb + yb; |
79 |
|
} |
80 |
|
|
|
// mv.length table |
|
|
static const uint32_t mvtab[33] = { |
|
|
1, 2, 3, 4, 6, 7, 7, 7, |
|
|
9, 9, 9, 10, 10, 10, 10, 10, |
|
|
10, 10, 10, 10, 10, 10, 10, 10, |
|
|
10, 11, 11, 11, 11, 11, 11, 12, 12 |
|
|
}; |
|
81 |
|
|
82 |
|
/* CHECK_CANDIATE FUNCTIONS START */ |
83 |
|
|
84 |
static __inline uint32_t |
static void |
85 |
mv_bits(int32_t component, |
CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
const uint32_t iFcode) |
|
86 |
{ |
{ |
87 |
if (component == 0) |
int32_t * const sad = data->temp; |
88 |
return 1; |
int t; |
89 |
|
const uint8_t * Reference; |
90 |
|
|
91 |
if (component < 0) |
if (( x > data->max_dx) || ( x < data->min_dx) |
92 |
component = -component; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
93 |
|
|
94 |
if (iFcode == 1) { |
switch ( ((x&1)<<1) + (y&1) ) { |
95 |
if (component > 32) |
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
96 |
component = 32; |
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
97 |
|
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
98 |
return mvtab[component] + 1; |
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
99 |
} |
} |
100 |
|
|
101 |
component += (1 << (iFcode - 1)) - 1; |
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, sad+1); |
|
component >>= (iFcode - 1); |
|
102 |
|
|
103 |
if (component > 32) |
t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
104 |
component = 32; |
data->temp[0] += lambda_vec16[data->iQuant] * t; |
105 |
|
data->temp[1] += lambda_vec8[data->iQuant] * t; |
106 |
|
|
107 |
return mvtab[component] + 1 + iFcode - 1; |
if (data->temp[0] < data->iMinSAD[0]) { |
108 |
} |
data->iMinSAD[0] = data->temp[0]; |
109 |
|
data->currentMV[0].x = x; data->currentMV[0].y = y; |
110 |
|
*dir = Direction; } |
111 |
|
|
112 |
|
if (data->temp[1] < data->iMinSAD[1]) { |
113 |
|
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
114 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
115 |
|
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
116 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
117 |
|
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
118 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
119 |
|
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
120 |
|
|
|
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)); |
|
121 |
} |
} |
122 |
|
|
123 |
static __inline uint32_t |
static void |
124 |
calc_delta_8(const int32_t dx, |
CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
const int32_t dy, |
|
|
const uint32_t iFcode, |
|
|
const uint32_t iQuant) |
|
125 |
{ |
{ |
126 |
return NEIGH_TEND_8X8 * lambda_vec8[iQuant] * (mv_bits(dx, iFcode) + |
int32_t sad; |
127 |
mv_bits(dy, iFcode)); |
const uint8_t * Reference; |
|
} |
|
128 |
|
|
129 |
bool |
if (( x > data->max_dx) || ( x < data->min_dx) |
130 |
MotionEstimation(MBParam * const pParam, |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|
FRAMEINFO * const current, |
|
|
FRAMEINFO * const reference, |
|
|
const IMAGE * const pRefH, |
|
|
const IMAGE * const pRefV, |
|
|
const IMAGE * const pRefHV, |
|
|
const uint32_t iLimit) |
|
|
{ |
|
|
const uint32_t iWcount = pParam->mb_width; |
|
|
const uint32_t iHcount = pParam->mb_height; |
|
|
MACROBLOCK *const pMBs = current->mbs; |
|
|
MACROBLOCK *const prevMBs = reference->mbs; |
|
|
const IMAGE *const pCurrent = ¤t->image; |
|
|
const IMAGE *const pRef = &reference->image; |
|
131 |
|
|
132 |
static const VECTOR zeroMV = { 0, 0 }; |
switch ( ((x&1)<<1) + (y&1) ) |
133 |
VECTOR predMV; |
{ |
134 |
|
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
135 |
|
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
136 |
|
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
137 |
|
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
138 |
|
} |
139 |
|
|
140 |
int32_t x, y; |
sad = lambda_vec16[data->iQuant] * |
141 |
int32_t iIntra = 0; |
d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
142 |
VECTOR pmv; |
sad += sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
143 |
|
|
144 |
if (sadInit) |
if (sad < *(data->iMinSAD)) { |
145 |
(*sadInit) (); |
*(data->iMinSAD) = sad; |
146 |
|
data->currentMV[0].x = x; data->currentMV[0].y = y; |
147 |
|
*dir = Direction; } |
148 |
|
} |
149 |
|
|
150 |
for (y = 0; y < iHcount; y++) { |
static void |
151 |
for (x = 0; x < iWcount; x ++) { |
CheckCandidate16no4vI(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
152 |
|
{ |
153 |
|
int32_t sad; |
154 |
|
|
155 |
MACROBLOCK *const pMB = &pMBs[x + y * iWcount]; |
if (( x > data->max_dx) || ( x < data->min_dx) |
156 |
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
157 |
|
|
158 |
predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
sad = lambda_vec16[data->iQuant] * |
159 |
|
d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
160 |
|
|
161 |
pMB->sad16 = |
sad += sad16(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), |
162 |
SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
data->iEdgedWidth, 256*4096); |
|
x, y, predMV.x, predMV.y, predMV.x, predMV.y, |
|
|
current->motion_flags, current->quant, |
|
|
current->fcode, pParam, pMBs, prevMBs, &pMB->mv16, |
|
|
&pMB->pmvs[0]); |
|
163 |
|
|
164 |
if (0 < (pMB->sad16 - MV16_INTER_BIAS)) { |
if (sad < *(data->iMinSAD)) { |
165 |
int32_t deviation; |
*(data->iMinSAD) = sad; |
166 |
|
data->currentMV[0].x = x; data->currentMV[0].y = y; |
167 |
|
*dir = Direction; } |
168 |
|
} |
169 |
|
|
|
deviation = |
|
|
dev16(pCurrent->y + x * 16 + y * 16 * pParam->edged_width, |
|
|
pParam->edged_width); |
|
170 |
|
|
171 |
if (deviation < (pMB->sad16 - MV16_INTER_BIAS)) { |
static void |
172 |
pMB->mode = MODE_INTRA; |
CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) |
173 |
pMB->mv16 = pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = |
{ |
174 |
pMB->mvs[3] = zeroMV; |
int32_t sad; |
175 |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = |
const int xb = data->currentMV[1].x; |
176 |
pMB->sad8[3] = 0; |
const int yb = data->currentMV[1].y; |
177 |
|
const uint8_t *ReferenceF, *ReferenceB; |
178 |
|
|
179 |
iIntra++; |
if (( xf > data->max_dx) || ( xf < data->min_dx) |
180 |
if (iIntra >= iLimit) |
|| ( yf > data->max_dy) || (yf < data->min_dy)) return; |
|
return 1; |
|
181 |
|
|
182 |
continue; |
switch ( ((xf&1)<<1) + (yf&1) ) { |
183 |
|
case 0 : ReferenceF = data->Ref + xf/2 + (yf/2)*(data->iEdgedWidth); break; |
184 |
|
case 1 : ReferenceF = data->RefV + xf/2 + ((yf-1)/2)*(data->iEdgedWidth); break; |
185 |
|
case 2 : ReferenceF = data->RefH + (xf-1)/2 + (yf/2)*(data->iEdgedWidth); break; |
186 |
|
default : ReferenceF = data->RefHV + (xf-1)/2 + ((yf-1)/2)*(data->iEdgedWidth); break; |
187 |
} |
} |
|
} |
|
|
|
|
|
pmv = pMB->pmvs[0]; |
|
|
if (current->global_flags & XVID_INTER4V) |
|
|
if ((!(current->global_flags & XVID_LUMIMASKING) || |
|
|
pMB->dquant == NO_CHANGE)) { |
|
|
int32_t sad8 = IMV16X16 * current->quant; |
|
|
|
|
|
if (sad8 < pMB->sad16) { |
|
|
sad8 += pMB->sad8[0] = |
|
|
SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, |
|
|
pCurrent, 2 * x, 2 * y, |
|
|
pMB->mv16.x, pMB->mv16.y, 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]); |
|
|
} |
|
|
|
|
|
/* decide: MODE_INTER or MODE_INTER4V |
|
|
mpeg4: if (sad8 < pMB->sad16 - nb/2+1) use_inter4v |
|
|
*/ |
|
188 |
|
|
189 |
if (sad8 < pMB->sad16) { |
switch ( ((xb&1)<<1) + (yb&1) ) { |
190 |
pMB->mode = MODE_INTER4V; |
case 0 : ReferenceB = data->bRef + xb/2 + (yb/2)*(data->iEdgedWidth); break; |
191 |
pMB->sad8[0] *= 4; |
case 1 : ReferenceB = data->bRefV + xb/2 + ((yb-1)/2)*(data->iEdgedWidth); break; |
192 |
pMB->sad8[1] *= 4; |
case 2 : ReferenceB = data->bRefH + (xb-1)/2 + (yb/2)*(data->iEdgedWidth); break; |
193 |
pMB->sad8[2] *= 4; |
default : ReferenceB = data->bRefHV + (xb-1)/2 + ((yb-1)/2)*(data->iEdgedWidth); break; |
|
pMB->sad8[3] *= 4; |
|
|
continue; |
|
194 |
} |
} |
195 |
|
|
196 |
} |
sad = lambda_vec16[data->iQuant] * |
197 |
|
( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) + |
198 |
|
d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode) ); |
199 |
|
|
200 |
pMB->mode = MODE_INTER; |
sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
|
pMB->pmvs[0] = pmv; /* pMB->pmvs[1] = pMB->pmvs[2] = pMB->pmvs[3] are not needed for INTER */ |
|
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mv16; |
|
|
pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = |
|
|
pMB->sad16; |
|
|
} |
|
|
} |
|
201 |
|
|
202 |
return 0; |
if (sad < *(data->iMinSAD)) { |
203 |
|
*(data->iMinSAD) = sad; |
204 |
|
data->currentMV->x = xf; data->currentMV->y = yf; |
205 |
|
*dir = Direction; } |
206 |
} |
} |
207 |
|
|
208 |
|
static void |
209 |
#define CHECK_MV16_ZERO {\ |
CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
if ( (0 <= max_dx) && (0 >= min_dx) \ |
|
|
&& (0 <= max_dy) && (0 >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0, 0 , iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); \ |
|
|
iSAD += calc_delta_16(-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; } } \ |
|
|
} |
|
|
|
|
|
/* too slow and not fully functional at the moment */ |
|
|
/* |
|
|
int32_t ZeroSearch16( |
|
|
const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const IMAGE * const pCur, |
|
|
const int x, const int y, |
|
|
const uint32_t MotionFlags, |
|
|
const uint32_t iQuant, |
|
|
const uint32_t iFcode, |
|
|
MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const MACROBLOCK * const prevMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV) |
|
210 |
{ |
{ |
211 |
const int32_t iEdgedWidth = pParam->edged_width; |
int32_t sad; |
212 |
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
int k; |
213 |
int32_t iSAD; |
const uint8_t *ReferenceF; |
214 |
VECTOR pred; |
const uint8_t *ReferenceB; |
215 |
|
VECTOR mvs, b_mvs; |
216 |
|
|
217 |
|
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
218 |
|
|
219 |
pred = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1); |
220 |
|
|
221 |
iSAD = sad16( cur, |
for (k = 0; k < 4; k++) { |
222 |
get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0,0, iEdgedWidth), |
mvs.x = data->directmvF[k].x + x; |
223 |
iEdgedWidth, MV_MAX_ERROR); |
b_mvs.x = ((x == 0) ? |
224 |
if (iSAD <= iQuant * 96) |
data->directmvB[k].x |
225 |
iSAD -= MV16_00_BIAS; |
: mvs.x - data->referencemv[k].x); |
226 |
|
|
227 |
currMV->x = 0; |
mvs.y = data->directmvF[k].y + y; |
228 |
currMV->y = 0; |
b_mvs.y = ((y == 0) ? |
229 |
currPMV->x = -pred.x; |
data->directmvB[k].y |
230 |
currPMV->y = -pred.y; |
: mvs.y - data->referencemv[k].y); |
231 |
|
|
232 |
return iSAD; |
if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
233 |
|
|| ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) |
234 |
|
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
235 |
|
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
236 |
|
|
237 |
|
switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
238 |
|
case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
239 |
|
case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
240 |
|
case 2 : ReferenceF = data->RefH + (mvs.x-1)/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
241 |
|
default : ReferenceF = data->RefHV + (mvs.x-1)/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
242 |
} |
} |
|
*/ |
|
|
|
|
|
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); |
|
243 |
|
|
244 |
if (iDirection) { |
switch ( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { |
245 |
while (!iFound) { |
case 0 : ReferenceB = data->bRef + b_mvs.x/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
246 |
iFound = 1; |
case 1 : ReferenceB = data->bRefV + b_mvs.x/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
247 |
backupMV = *currMV; |
case 2 : ReferenceB = data->bRefH + (b_mvs.x-1)/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
248 |
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); |
|
249 |
} |
} |
250 |
} else { |
|
251 |
currMV->x = start_x; |
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
252 |
currMV->y = start_y; |
ReferenceF + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
253 |
} |
ReferenceB + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
254 |
return iMinSAD; |
data->iEdgedWidth); |
255 |
|
if (sad > *(data->iMinSAD)) return; |
256 |
} |
} |
257 |
|
|
258 |
int32_t |
if (sad < *(data->iMinSAD)) { |
259 |
Square16_MainSearch(const uint8_t * const pRef, |
*(data->iMinSAD) = sad; |
260 |
const uint8_t * const pRefH, |
data->currentMV->x = x; data->currentMV->y = y; |
261 |
const uint8_t * const pRefV, |
*dir = Direction; } |
262 |
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 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_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, backupMV.y, 1); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, backupMV.y, 2); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, 3); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, 4); |
|
|
|
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 5); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 6); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, |
|
|
backupMV.y - iDiamondSize, 7); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 8); |
|
263 |
|
|
264 |
|
static void |
265 |
|
CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
266 |
|
{ |
267 |
|
int32_t sad; |
268 |
|
const uint8_t *ReferenceF; |
269 |
|
const uint8_t *ReferenceB; |
270 |
|
VECTOR mvs, b_mvs; |
271 |
|
|
272 |
if (iDirection) { |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
|
while (!iFound) { |
|
|
iFound = 1; |
|
|
backupMV = *currMV; |
|
273 |
|
|
274 |
switch (iDirection) { |
sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1); |
|
case 1: |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y, 1); |
|
|
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; |
|
275 |
|
|
276 |
case 3: |
mvs.x = data->directmvF[0].x + x; |
277 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, |
b_mvs.x = ((x == 0) ? |
278 |
4); |
data->directmvB[0].x |
279 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
: mvs.x - data->referencemv[0].x); |
|
backupMV.y - iDiamondSize, 7); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 8); |
|
|
break; |
|
280 |
|
|
281 |
case 4: |
mvs.y = data->directmvF[0].y + y; |
282 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, |
b_mvs.y = ((y == 0) ? |
283 |
3); |
data->directmvB[0].y |
284 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
: mvs.y - data->referencemv[0].y); |
|
backupMV.y - iDiamondSize, 5); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 6); |
|
|
break; |
|
285 |
|
|
286 |
case 5: |
if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
287 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, |
|| ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) |
288 |
1); |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
289 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, |
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
|
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, 7); |
|
|
break; |
|
290 |
|
|
291 |
case 6: |
switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
292 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, |
case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
293 |
2); |
case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
294 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, |
case 2 : ReferenceF = data->RefH + (mvs.x-1)/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
295 |
3); |
default : ReferenceF = data->RefHV + (mvs.x-1)/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
296 |
|
} |
|
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); |
|
297 |
|
|
298 |
break; |
switch ( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { |
299 |
|
case 0 : ReferenceB = data->bRef + b_mvs.x/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
300 |
|
case 1 : ReferenceB = data->bRefV + b_mvs.x/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
301 |
|
case 2 : ReferenceB = data->bRefH + (b_mvs.x-1)/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
302 |
|
default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
303 |
|
} |
304 |
|
|
305 |
case 7: |
sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, |
|
|
backupMV.y, 1); |
|
|
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, 7); |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, |
|
|
backupMV.y + iDiamondSize, 8); |
|
|
break; |
|
306 |
|
|
307 |
case 8: |
if (sad < *(data->iMinSAD)) { |
308 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, |
*(data->iMinSAD) = sad; |
309 |
2); |
data->currentMV->x = x; data->currentMV->y = y; |
310 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, |
*dir = Direction; } |
|
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; |
|
311 |
} |
} |
312 |
|
|
313 |
|
static void |
314 |
|
CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
315 |
|
{ |
316 |
|
int32_t sad; |
317 |
|
const uint8_t * Reference; |
318 |
|
|
319 |
int32_t |
if (( x > data->max_dx) || ( x < data->min_dx) |
320 |
Full16_MainSearch(const uint8_t * const pRef, |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, |
|
|
const int y, |
|
|
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); |
|
321 |
|
|
322 |
return iMinSAD; |
switch ( ((x&1)<<1) + (y&1) ) |
323 |
|
{ |
324 |
|
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
325 |
|
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
326 |
|
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
327 |
|
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
328 |
} |
} |
329 |
|
|
330 |
int32_t |
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
331 |
AdvDiamond16_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, |
|
|
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) |
|
|
{ |
|
332 |
|
|
333 |
int32_t iSAD; |
if (sad < *(data->iMinSAD)) { |
334 |
|
*(data->iMinSAD) = sad; |
335 |
|
data->currentMV->x = x; data->currentMV->y = y; |
336 |
|
*dir = Direction; } |
337 |
|
} |
338 |
|
|
339 |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
/* CHECK_CANDIATE FUNCTIONS END */ |
340 |
|
|
341 |
if (iDirection) { |
/* MAINSEARCH FUNCTIONS START */ |
|
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; |
|
342 |
|
|
343 |
do { |
static void |
344 |
iDirection = 0; |
AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection) |
345 |
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_MV16_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); |
|
346 |
|
|
347 |
if (bDirection & 2) |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
|
CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, start_y, 2); |
|
348 |
|
|
349 |
if (bDirection & 4) |
int iDirection; |
|
CHECK_MV16_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); |
|
350 |
|
|
351 |
if (bDirection & 8) |
do { |
352 |
CHECK_MV16_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); |
iDirection = 0; |
353 |
|
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
354 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
355 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
356 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
357 |
|
|
358 |
/* now we're doing diagonal checks near our candidate */ |
/* now we're doing diagonal checks near our candidate */ |
359 |
|
|
360 |
if (iDirection) //checking if anything found |
if (iDirection) { //checking if anything found |
|
{ |
|
361 |
bDirection = iDirection; |
bDirection = iDirection; |
362 |
iDirection = 0; |
iDirection = 0; |
363 |
start_x = currMV->x; |
x = data->currentMV->x; y = data->currentMV->y; |
364 |
start_y = currMV->y; |
if (bDirection & 3) { //our candidate is left or right |
365 |
if (bDirection & 3) //our candidate is left or right |
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
366 |
{ |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
367 |
CHECK_MV16_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); |
} else { // what remains here is up or down |
368 |
CHECK_MV16_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
369 |
} 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); |
|
|
} |
|
370 |
|
|
371 |
if (iDirection) { |
if (iDirection) { |
372 |
bDirection += iDirection; |
bDirection += iDirection; |
373 |
start_x = currMV->x; |
x = data->currentMV->x; y = data->currentMV->y; } |
374 |
start_y = currMV->y; |
} else { //about to quit, eh? not so fast.... |
|
} |
|
|
} else //about to quit, eh? not so fast.... |
|
|
{ |
|
375 |
switch (bDirection) { |
switch (bDirection) { |
376 |
case 2: |
case 2: |
377 |
CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
378 |
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); |
|
379 |
break; |
break; |
380 |
case 1: |
case 1: |
381 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
382 |
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); |
|
383 |
break; |
break; |
384 |
case 2 + 4: |
case 2 + 4: |
385 |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
386 |
start_y - iDiamondSize, 1 + 4); |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
387 |
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); |
|
388 |
break; |
break; |
389 |
case 4: |
case 4: |
390 |
CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
391 |
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); |
|
392 |
break; |
break; |
393 |
case 8: |
case 8: |
394 |
CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
395 |
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); |
|
396 |
break; |
break; |
397 |
case 1 + 4: |
case 1 + 4: |
398 |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
399 |
start_y + iDiamondSize, 1 + 8); |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
400 |
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); |
|
401 |
break; |
break; |
402 |
case 2 + 8: |
case 2 + 8: |
403 |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
404 |
start_y - iDiamondSize, 1 + 4); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
405 |
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); |
|
406 |
break; |
break; |
407 |
case 1 + 8: |
case 1 + 8: |
408 |
CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
409 |
start_y - iDiamondSize, 2 + 4); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
410 |
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); |
|
411 |
break; |
break; |
412 |
default: //1+2+4+8 == we didn't find anything at all |
default: //1+2+4+8 == we didn't find anything at all |
413 |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
414 |
start_y - iDiamondSize, 1 + 4); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
415 |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
416 |
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); |
|
417 |
break; |
break; |
418 |
} |
} |
419 |
if (!iDirection) |
if (!iDirection) break; //ok, the end. really |
|
break; //ok, the end. really |
|
|
else { |
|
420 |
bDirection = iDirection; |
bDirection = iDirection; |
421 |
start_x = currMV->x; |
x = data->currentMV->x; y = data->currentMV->y; |
|
start_y = currMV->y; |
|
|
} |
|
422 |
} |
} |
423 |
} |
} |
424 |
while (1); //forever |
while (1); //forever |
425 |
} |
} |
|
return iMinSAD; |
|
|
} |
|
|
|
|
|
#define CHECK_MV16_F_INTERPOL(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); } } \ |
|
|
} |
|
|
|
|
|
#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, |
|
426 |
|
|
427 |
const int x, |
static void |
428 |
const int y, |
SquareSearch(int x, int y, const SearchData * const data, int bDirection) |
429 |
|
{ |
430 |
const int f_start_x, |
int iDirection; |
|
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, |
|
431 |
|
|
432 |
const int x, |
do { |
433 |
const int y, |
iDirection = 0; |
434 |
|
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64); |
435 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128); |
436 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32); |
437 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128); |
438 |
|
if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64); |
439 |
|
if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128); |
440 |
|
if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128); |
441 |
|
if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128); |
442 |
|
|
443 |
const int TRB, |
bDirection = iDirection; |
444 |
const int TRD, |
x = data->currentMV->x; y = data->currentMV->y; |
445 |
|
} while (iDirection); |
446 |
|
} |
447 |
|
|
448 |
const int start_x, |
static void |
449 |
const int start_y, |
DiamondSearch(int x, int y, const SearchData * const data, int bDirection) |
450 |
|
{ |
451 |
|
|
452 |
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, |
|
453 |
|
|
454 |
const int32_t min_dx, |
int iDirection; |
|
const int32_t max_dx, |
|
|
const int32_t min_dy, |
|
|
const int32_t max_dy, |
|
455 |
|
|
456 |
const int32_t iEdgedWidth, |
do { |
457 |
const int32_t iDiamondSize, |
iDirection = 0; |
458 |
|
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
459 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
460 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
461 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
462 |
|
|
463 |
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 */ |
|
464 |
|
|
465 |
int32_t iSAD; |
if (iDirection) { //checking if anything found |
466 |
|
bDirection = iDirection; |
467 |
|
iDirection = 0; |
468 |
|
x = data->currentMV->x; y = data->currentMV->y; |
469 |
|
if (bDirection & 3) { //our candidate is left or right |
470 |
|
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
471 |
|
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
472 |
|
} else { // what remains here is up or down |
473 |
|
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
474 |
|
CHECK_CANDIDATE(x - iDiamondSize, y, 1); } |
475 |
|
|
476 |
VECTOR backupMV; |
bDirection += iDirection; |
477 |
|
x = data->currentMV->x; y = data->currentMV->y; |
478 |
|
} |
479 |
|
} |
480 |
|
while (iDirection); |
481 |
|
} |
482 |
|
|
483 |
currMV->x = start_x; |
/* MAINSEARCH FUNCTIONS END */ |
|
currMV->y = start_y; |
|
484 |
|
|
485 |
/* 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 */ |
486 |
|
|
487 |
do |
static void |
488 |
|
HalfpelRefine(const SearchData * const data) |
489 |
{ |
{ |
490 |
iFound = 1; |
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
491 |
|
|
492 |
backupMV = *currMV; |
VECTOR backupMV = *(data->currentMV); |
493 |
|
int iDirection; //not needed |
494 |
|
|
495 |
CHECK_MV16_DIRECT_FOUND(backupMV.x - iDiamondSize, backupMV.y); |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); |
496 |
CHECK_MV16_DIRECT_FOUND(backupMV.x + iDiamondSize, backupMV.y); |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); |
497 |
CHECK_MV16_DIRECT_FOUND(backupMV.x, backupMV.y - iDiamondSize); |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); |
498 |
CHECK_MV16_DIRECT_FOUND(backupMV.x, backupMV.y + iDiamondSize); |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0); |
499 |
|
|
500 |
} while (!iFound); |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); |
501 |
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); |
502 |
|
|
503 |
return iMinSAD; |
CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); |
504 |
|
CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); |
505 |
} |
} |
506 |
|
|
507 |
|
static __inline int |
508 |
|
SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
509 |
|
const int x, const int y, |
510 |
|
const uint32_t iEdgedWidth, const uint32_t iQuant) |
511 |
|
|
|
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) |
|
512 |
{ |
{ |
513 |
|
/* keep repeating checks for all b-frames before this P frame, |
514 |
|
to make sure that SKIP is possible (todo) |
515 |
|
how: if skip is not possible set sad00 to a very high value */ |
516 |
|
|
517 |
|
uint32_t sadC = sad8(current->u + x*8 + y*(iEdgedWidth/2)*8, |
518 |
|
reference->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); |
519 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
520 |
|
sadC += sad8(current->v + (x + y*(iEdgedWidth/2))*8, |
521 |
|
reference->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); |
522 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
523 |
|
|
524 |
int32_t iSAD; |
return 1; |
525 |
|
} |
526 |
|
|
527 |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
static __inline void |
528 |
|
SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
529 |
|
{ |
530 |
|
pMB->mode = MODE_NOT_CODED; |
531 |
|
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
532 |
|
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; |
533 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
534 |
|
} |
535 |
|
|
536 |
if (iDirection) { |
bool |
537 |
CHECK_MV8_CANDIDATE(start_x - iDiamondSize, start_y); |
MotionEstimation(MBParam * const pParam, |
538 |
CHECK_MV8_CANDIDATE(start_x + iDiamondSize, start_y); |
FRAMEINFO * const current, |
539 |
CHECK_MV8_CANDIDATE(start_x, start_y - iDiamondSize); |
FRAMEINFO * const reference, |
540 |
CHECK_MV8_CANDIDATE(start_x, start_y + iDiamondSize); |
const IMAGE * const pRefH, |
541 |
} else { |
const IMAGE * const pRefV, |
542 |
int bDirection = 1 + 2 + 4 + 8; |
const IMAGE * const pRefHV, |
543 |
|
const uint32_t iLimit) |
544 |
|
{ |
545 |
|
MACROBLOCK *const pMBs = current->mbs; |
546 |
|
const IMAGE *const pCurrent = ¤t->image; |
547 |
|
const IMAGE *const pRef = &reference->image; |
548 |
|
|
549 |
do { |
const VECTOR zeroMV = { 0, 0 }; |
550 |
iDirection = 0; |
|
551 |
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) |
uint32_t x, y; |
552 |
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); |
uint32_t iIntra = 0; |
553 |
|
int32_t InterBias, quant = current->quant; |
554 |
|
|
555 |
|
// some pre-initialized thingies for SearchP |
556 |
|
int32_t temp[5]; |
557 |
|
VECTOR currentMV[5]; |
558 |
|
int32_t iMinSAD[5]; |
559 |
|
SearchData Data; |
560 |
|
Data.iEdgedWidth = pParam->edged_width; |
561 |
|
Data.currentMV = currentMV; |
562 |
|
Data.iMinSAD = iMinSAD; |
563 |
|
Data.temp = temp; |
564 |
|
Data.iFcode = current->fcode; |
565 |
|
|
566 |
|
if (sadInit) (*sadInit) (); |
567 |
|
|
568 |
|
for (y = 0; y < pParam->mb_height; y++) { |
569 |
|
for (x = 0; x < pParam->mb_width; x++) { |
570 |
|
|
571 |
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
572 |
|
int32_t sad00 = pMB->sad16 |
573 |
|
= sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, |
574 |
|
pRef->y + (x + y * pParam->edged_width) * 16, |
575 |
|
pParam->edged_width, pMB->sad8 ); |
576 |
|
|
577 |
|
if (!(current->global_flags & XVID_LUMIMASKING)) { |
578 |
|
pMB->dquant = NO_CHANGE; |
579 |
|
pMB->quant = current->quant; } |
580 |
|
else |
581 |
|
if (pMB->dquant != NO_CHANGE) { |
582 |
|
quant += DQtab[pMB->dquant]; |
583 |
|
if (quant > 31) quant = 31; |
584 |
|
else if (quant < 1) quant = 1; |
585 |
|
pMB->quant = quant; |
586 |
|
} |
587 |
|
|
588 |
if (bDirection & 2) |
//initial skip decision |
589 |
CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, start_y, 2); |
|
590 |
|
if ((pMB->dquant == NO_CHANGE) && (sad00 <= MAX_SAD00_FOR_SKIP * pMB->quant) |
591 |
|
&& (SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) ) { |
592 |
|
if (pMB->sad16 < pMB->quant * INITIAL_SKIP_THRESH) { |
593 |
|
SkipMacroblockP(pMB, sad00); |
594 |
|
continue; |
595 |
|
} |
596 |
|
} else sad00 = 256*4096; // skip not allowed - for final skip decision |
597 |
|
|
598 |
if (bDirection & 4) |
SearchP(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
599 |
CHECK_MV8_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); |
y, current->motion_flags, pMB->quant, |
600 |
|
&Data, pParam, pMBs, reference->mbs, |
601 |
|
current->global_flags & XVID_INTER4V, pMB); |
602 |
|
|
603 |
|
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
604 |
|
if (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
605 |
|
if ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) |
606 |
|
{ SkipMacroblockP(pMB, sad00); continue; } |
607 |
|
|
608 |
if (bDirection & 8) |
/* finally, intra decision */ |
|
CHECK_MV8_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); |
|
609 |
|
|
610 |
/* now we're doing diagonal checks near our candidate */ |
InterBias = MV16_INTER_BIAS; |
611 |
|
if (pMB->quant > 8) InterBias += 50 * (pMB->quant - 8); // to make high quants work |
612 |
|
if (y != 0) |
613 |
|
if ((pMB - pParam->mb_width)->mode == MODE_INTER ) InterBias -= 50; |
614 |
|
if (x != 0) |
615 |
|
if ((pMB - 1)->mode == MODE_INTER ) InterBias -= 50; |
616 |
|
|
617 |
if (iDirection) //checking if anything found |
if (InterBias < pMB->sad16) { |
618 |
{ |
const int32_t deviation = |
619 |
bDirection = iDirection; |
dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
620 |
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); |
|
|
} |
|
621 |
|
|
622 |
if (iDirection) { |
if (deviation < (pMB->sad16 - InterBias)) { |
623 |
bDirection += iDirection; |
if (++iIntra >= iLimit) return 1; |
624 |
start_x = currMV->x; |
pMB->mode = MODE_INTRA; |
625 |
start_y = currMV->y; |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = |
626 |
} |
pMB->mvs[3] = zeroMV; |
627 |
} else //about to quit, eh? not so fast.... |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = |
628 |
{ |
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; |
|
629 |
} |
} |
630 |
} |
} |
631 |
} |
} |
|
while (1); //forever |
|
632 |
} |
} |
633 |
return iMinSAD; |
return 0; |
634 |
} |
} |
635 |
|
|
636 |
|
|
637 |
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); |
|
638 |
|
|
639 |
return iMinSAD; |
static __inline int |
640 |
|
make_mask(const VECTOR * const pmv, const int i) |
641 |
|
{ |
642 |
|
int mask = 255, j; |
643 |
|
for (j = 0; j < i; j++) { |
644 |
|
if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already |
645 |
|
if (pmv[i].x == pmv[j].x) { |
646 |
|
if (pmv[i].y == pmv[j].y + iDiamondSize) { mask &= ~4; continue; } |
647 |
|
if (pmv[i].y == pmv[j].y - iDiamondSize) { mask &= ~8; continue; } |
648 |
|
} else |
649 |
|
if (pmv[i].y == pmv[j].y) { |
650 |
|
if (pmv[i].x == pmv[j].x + iDiamondSize) { mask &= ~1; continue; } |
651 |
|
if (pmv[i].x == pmv[j].x - iDiamondSize) { mask &= ~2; continue; } |
652 |
|
} |
653 |
|
} |
654 |
|
return mask; |
655 |
} |
} |
656 |
|
|
657 |
Halfpel8_RefineFuncPtr Halfpel8_Refine; |
static __inline void |
658 |
|
PreparePredictionsP(VECTOR * const pmv, int x, int y, const int iWcount, |
659 |
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) |
|
660 |
{ |
{ |
|
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
|
661 |
|
|
662 |
int32_t iSAD; |
//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself |
|
VECTOR backupMV = *currMV; |
|
663 |
|
|
664 |
CHECK_MV16_CANDIDATE(backupMV.x - 1, backupMV.y - 1); |
if ( (y != 0) && (x != (iWcount-1)) ) { // [5] top-right neighbour |
665 |
CHECK_MV16_CANDIDATE(backupMV.x, backupMV.y - 1); |
pmv[5].x = EVEN(pmv[3].x); |
666 |
CHECK_MV16_CANDIDATE(backupMV.x + 1, backupMV.y - 1); |
pmv[5].y = EVEN(pmv[3].y); |
667 |
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); |
|
668 |
|
|
669 |
return iMinSAD; |
if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour |
670 |
} |
else pmv[3].x = pmv[3].y = 0; |
671 |
|
|
672 |
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour |
673 |
|
else pmv[4].x = pmv[4].y = 0; |
674 |
|
|
675 |
|
// [1] median prediction |
676 |
|
pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); |
677 |
|
|
678 |
|
pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask |
679 |
|
|
680 |
int32_t |
pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
681 |
PMVfastSearch16(const uint8_t * const pRef, |
pmv[2].y = EVEN(prevMB->mvs[0].y); |
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const IMAGE * const pCur, |
|
|
const int x, |
|
|
const int y, |
|
|
const int start_x, /* 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 |
|
|
|
|
|
if (!MVequal(pmv[0], startMV)) |
|
|
CHECK_MV8_CANDIDATE(center_x, center_y); |
|
|
|
|
|
// (0,0) if needed |
|
|
if (!MVzero(pmv[0])) |
|
|
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; |
|
|
|
|
|
|
|
|
/* 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[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; |
|
|
} |
|
|
|
|
|
/************ (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. |
|
|
*/ |
|
|
|
|
|
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, |
|
|
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 uint32_t iHcount = pParam->mb_height; |
|
|
|
|
|
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 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]; |
|
|
|
|
|
static MACROBLOCK *oldMBs = NULL; |
|
|
|
|
|
// const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
|
|
const MACROBLOCK *const prevMB = prevMBs + x + y * iWcount; |
|
|
MACROBLOCK *oldMB = NULL; |
|
|
|
|
|
int32_t thresh2; |
|
|
int32_t bPredEq; |
|
|
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); |
|
|
|
|
|
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); |
|
|
|
|
|
/* 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; |
|
|
|
|
|
if (!(MotionFlags & PMV_HALFPEL16)) { |
|
|
currMV->x = EVEN(currMV->x); |
|
|
currMV->y = EVEN(currMV->y); |
|
|
} |
|
682 |
|
|
683 |
if (currMV->x > max_dx) |
if ((x != iWcount-1) && (y != iHcount-1)) { |
684 |
currMV->x = max_dx; |
pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame |
685 |
if (currMV->x < min_dx) |
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
686 |
currMV->x = min_dx; |
} else pmv[6].x = pmv[6].y = 0; |
|
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; |
|
687 |
} |
} |
688 |
|
|
689 |
// MV=(0,0) is often a good choice |
static void |
690 |
|
SearchP(const uint8_t * const pRef, |
|
CHECK_MV16_ZERO; |
|
|
|
|
|
|
|
|
// left neighbour, if allowed |
|
|
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); |
|
|
|
|
|
// top right neighbour, if allowed |
|
|
if ((uint32_t) x != (iWcount - 1)) { |
|
|
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); |
|
|
} |
|
|
} |
|
|
|
|
|
/* 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 */ |
|
|
|
|
|
if (MotionFlags & PMV_USESQUARES16) |
|
|
MainSearchPtr = Square16_MainSearch; |
|
|
else |
|
|
if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
|
|
MainSearchPtr = AdvDiamond16_MainSearch; |
|
|
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, |
|
691 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
692 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
693 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
694 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
695 |
const int x, |
const int x, |
696 |
const int y, |
const int y, |
|
const int start_x, |
|
|
const int start_y, |
|
|
const int center_x, |
|
|
const int center_y, |
|
697 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
698 |
const uint32_t iQuant, |
const uint32_t iQuant, |
699 |
const uint32_t iFcode, |
SearchData * const Data, |
700 |
const MBParam * const pParam, |
const MBParam * const pParam, |
701 |
const MACROBLOCK * const pMBs, |
const MACROBLOCK * const pMBs, |
702 |
const MACROBLOCK * const prevMBs, |
const MACROBLOCK * const prevMBs, |
703 |
VECTOR * const currMV, |
int inter4v, |
704 |
VECTOR * const currPMV) |
MACROBLOCK * const pMB) |
705 |
{ |
{ |
|
/* 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); |
|
706 |
|
|
707 |
|
int i, iDirection = 255, mask, threshA; |
708 |
|
VECTOR pmv[7]; |
709 |
|
|
710 |
/* Step 4: Calculate SAD around the Median prediction. |
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); //has to be changed to get_pmv(2)() |
711 |
MinSAD=SAD |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
712 |
If Motion Vector equal to Previous frame motion vector |
pParam->width, pParam->height, Data->iFcode); |
|
and MinSAD<PrevFrmSAD goto Step 10. |
|
|
If SAD<=256 goto Step 10. |
|
|
*/ |
|
713 |
|
|
714 |
// Prepare for main loop |
Data->predMV = pmv[0]; |
715 |
|
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16; |
716 |
|
Data->Ref = pRef + (x + Data->iEdgedWidth*y)*16; |
717 |
|
Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16; |
718 |
|
Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16; |
719 |
|
Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16; |
720 |
|
|
721 |
|
Data->iQuant = iQuant; |
722 |
|
|
723 |
if (!(MotionFlags & PMV_HALFPEL8)) { |
if (!(MotionFlags & PMV_HALFPEL16)) { |
724 |
currMV->x = EVEN(currMV->x); |
Data->min_dx = EVEN(Data->min_dx); |
725 |
currMV->y = EVEN(currMV->y); |
Data->max_dx = EVEN(Data->max_dx); |
726 |
} |
Data->min_dy = EVEN(Data->min_dy); |
727 |
|
Data->max_dy = EVEN(Data->max_dy); } |
728 |
|
|
729 |
|
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
730 |
|
|
731 |
|
if (inter4v) CheckCandidate = CheckCandidate16; |
732 |
|
else CheckCandidate = CheckCandidate16no4v; |
733 |
|
|
734 |
|
for(i = 0; i < 5; i++) Data->currentMV[i].x = Data->currentMV[i].y = 0; |
735 |
|
|
736 |
|
i = d_mv_bits(Data->predMV.x, Data->predMV.y, Data->iFcode); |
737 |
|
Data->iMinSAD[0] = pMB->sad16 + lambda_vec16[iQuant] * i; |
738 |
|
Data->iMinSAD[1] = pMB->sad8[0] + lambda_vec8[iQuant] * i; |
739 |
|
Data->iMinSAD[2] = pMB->sad8[1]; |
740 |
|
Data->iMinSAD[3] = pMB->sad8[2]; |
741 |
|
Data->iMinSAD[4] = pMB->sad8[3]; |
742 |
|
|
743 |
if (currMV->x > max_dx) |
if ((x == 0) && (y == 0)) threshA = 512; |
744 |
currMV->x = max_dx; |
else { |
745 |
if (currMV->x < min_dx) |
threshA = Data->temp[0]; // that's when we keep this SAD atm |
746 |
currMV->x = min_dx; |
if (threshA < 512) threshA = 512; |
747 |
if (currMV->y > max_dy) |
if (threshA > 1024) threshA = 1024; } |
|
currMV->y = max_dy; |
|
|
if (currMV->y < min_dy) |
|
|
currMV->y = min_dy; |
|
748 |
|
|
749 |
/***************** This is predictor SET A: only median prediction ******************/ |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
750 |
|
prevMBs + x + y * pParam->mb_width); |
751 |
|
|
752 |
|
if (inter4v) CheckCandidate = CheckCandidate16; |
753 |
|
else CheckCandidate = CheckCandidate16no4v; |
754 |
|
|
|
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); |
|
755 |
|
|
756 |
|
/* main loop. checking all predictions */ |
757 |
|
|
758 |
// thresh1 is fixed to 256 |
for (i = 1; i < 7; i++) { |
759 |
if (iMinSAD < 256 / 4) { |
if (!(mask = make_mask(pmv, i)) ) continue; |
760 |
if (MotionFlags & PMV_QUICKSTOP8) |
(*CheckCandidate)(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
761 |
goto EPZS8_Terminate_without_Refine; |
if (Data->iMinSAD[0] <= threshA) break; |
|
if (MotionFlags & PMV_EARLYSTOP8) |
|
|
goto EPZS8_Terminate_with_Refine; |
|
762 |
} |
} |
763 |
|
|
764 |
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
if ((Data->iMinSAD[0] <= threshA) || |
765 |
|
(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
766 |
|
(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { |
767 |
// MV=(0,0) is often a good choice |
inter4v = 0; |
768 |
CHECK_MV8_ZERO; |
} else { |
769 |
|
|
770 |
// previous frame MV |
MainSearchFunc * MainSearchPtr; |
771 |
CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x, prevMB->mvs[iSubBlock].y); |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
772 |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
773 |
|
else MainSearchPtr = DiamondSearch; |
774 |
|
|
775 |
|
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
776 |
|
|
777 |
|
/* extended search, diamond starting in 0,0 and in prediction. |
778 |
|
note that this search is/might be done in halfpel positions, |
779 |
|
which makes it more different than the diamond above */ |
780 |
|
|
781 |
// left neighbour, if allowed |
if (MotionFlags & PMV_EXTSEARCH16) { |
782 |
if (psad[1] != MV_MAX_ERROR) { |
int32_t bSAD; |
783 |
if (!(MotionFlags & PMV_HALFPEL8)) { |
VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; |
784 |
pmv[1].x = EVEN(pmv[1].x); |
if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? |
785 |
pmv[1].y = EVEN(pmv[1].y); |
startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); |
786 |
|
if (!(MVequal(startMV, backupMV))) { |
787 |
|
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
788 |
|
|
789 |
|
CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, Data); |
790 |
|
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
791 |
|
if (bSAD < Data->iMinSAD[0]) { |
792 |
|
Data->currentMV[0] = backupMV; |
793 |
|
Data->iMinSAD[0] = bSAD; } |
794 |
|
} |
795 |
|
|
796 |
|
backupMV = Data->currentMV[0]; |
797 |
|
if (MotionFlags & PMV_HALFPELREFINE16) startMV.x = startMV.y = 1; |
798 |
|
else startMV.x = startMV.y = 0; |
799 |
|
if (!(MVequal(startMV, backupMV))) { |
800 |
|
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
801 |
|
|
802 |
|
CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, Data); |
803 |
|
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
804 |
|
if (bSAD < Data->iMinSAD[0]) { |
805 |
|
Data->currentMV[0] = backupMV; |
806 |
|
Data->iMinSAD[0] = bSAD; } |
807 |
} |
} |
|
CHECK_MV8_CANDIDATE(pmv[1].x, pmv[1].y); |
|
808 |
} |
} |
|
// 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); |
|
809 |
} |
} |
|
CHECK_MV8_CANDIDATE(pmv[2].x, pmv[2].y); |
|
810 |
|
|
811 |
// top right neighbour, if allowed |
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data); |
812 |
if (psad[3] != MV_MAX_ERROR) { |
|
813 |
if (!(MotionFlags & PMV_HALFPEL8)) { |
if (inter4v) { |
814 |
pmv[3].x = EVEN(pmv[3].x); |
SearchData Data8; |
815 |
pmv[3].y = EVEN(pmv[3].y); |
Data8.iFcode = Data->iFcode; |
816 |
} |
Data8.iQuant = Data->iQuant; |
817 |
CHECK_MV8_CANDIDATE(pmv[3].x, pmv[3].y); |
Data8.iEdgedWidth = Data->iEdgedWidth; |
818 |
|
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
819 |
|
Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); |
820 |
|
Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); |
821 |
|
Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); |
822 |
|
} |
823 |
|
|
824 |
|
if (!(inter4v) || |
825 |
|
(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
826 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
827 |
|
// INTER MODE |
828 |
|
pMB->mode = MODE_INTER; |
829 |
|
pMB->mvs[0] = pMB->mvs[1] |
830 |
|
= pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
831 |
|
|
832 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
833 |
|
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
834 |
|
|
835 |
|
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
836 |
|
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
837 |
|
} else { |
838 |
|
// INTER4V MODE; all other things are already set in Search8 |
839 |
|
pMB->mode = MODE_INTER4V; |
840 |
|
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + |
841 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant; |
842 |
} |
} |
843 |
|
|
844 |
} |
} |
845 |
|
|
846 |
/* // this bias is zero anyway, at the moment! |
static void |
847 |
|
Search8(const SearchData * const OldData, |
848 |
|
const int x, const int y, |
849 |
|
const uint32_t MotionFlags, |
850 |
|
const MBParam * const pParam, |
851 |
|
MACROBLOCK * const pMB, |
852 |
|
const MACROBLOCK * const pMBs, |
853 |
|
const int block, |
854 |
|
SearchData * const Data) |
855 |
|
{ |
856 |
|
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
857 |
|
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
858 |
|
Data->currentMV = OldData->currentMV + 1 + block; |
859 |
|
|
860 |
if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) ) // && (iMinSAD <= iQuant * 96) |
if (block != 0) |
861 |
iMinSAD -= MV8_00_BIAS; |
*(Data->iMinSAD) += lambda_vec8[Data->iQuant] * |
862 |
|
d_mv_bits( Data->currentMV->x - Data->predMV.x, |
863 |
|
Data->currentMV->y - Data->predMV.y, |
864 |
|
Data->iFcode); |
865 |
|
|
866 |
*/ |
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { |
867 |
|
|
868 |
/* Terminate if MinSAD <= T_2 |
Data->Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
869 |
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
Data->RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
870 |
*/ |
Data->RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
871 |
|
Data->RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
872 |
|
|
873 |
if (iMinSAD < 512 / 4) { /* T_2 == 512/4 hardcoded */ |
Data->Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
if (MotionFlags & PMV_QUICKSTOP8) |
|
|
goto EPZS8_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP8) |
|
|
goto EPZS8_Terminate_with_Refine; |
|
|
} |
|
874 |
|
|
875 |
/************ (Diamond Search) **************/ |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
876 |
|
pParam->width, pParam->height, OldData->iFcode); |
877 |
|
|
878 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
CheckCandidate = CheckCandidate8; |
879 |
|
|
880 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8)) |
if (MotionFlags & PMV_EXTSEARCH8) { |
|
iDiamondSize *= 2; |
|
881 |
|
|
882 |
/* default: use best prediction as starting point for one call of EPZS_MainSearch */ |
MainSearchFunc *MainSearchPtr; |
883 |
|
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
884 |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
885 |
|
else MainSearchPtr = DiamondSearch; |
886 |
|
|
887 |
// there is no EPZS^2 for inter4v at the moment |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); } |
888 |
|
|
889 |
if (MotionFlags & PMV_USESQUARES8) |
if (MotionFlags & PMV_HALFPELREFINE8) HalfpelRefine(Data); |
890 |
MainSearchPtr = Square8_MainSearch; |
} |
|
else |
|
891 |
|
|
892 |
if (MotionFlags & PMV_ADVANCEDDIAMOND8) |
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
893 |
MainSearchPtr = AdvDiamond8_MainSearch; |
pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; |
894 |
else |
pMB->mvs[block] = *(Data->currentMV); |
895 |
MainSearchPtr = Diamond8_MainSearch; |
pMB->sad8[block] = 4 * (*Data->iMinSAD); |
896 |
|
|
897 |
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, 0); |
|
898 |
|
|
899 |
|
/* B-frames code starts here */ |
900 |
|
|
901 |
if (iSAD < iMinSAD) { |
static __inline VECTOR |
902 |
*currMV = newMV; |
ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) |
903 |
iMinSAD = iSAD; |
{ |
904 |
|
/* the stupidiest function ever */ |
905 |
|
if (mode == MODE_FORWARD) return pMB->mvs[0]; |
906 |
|
else return pMB->b_mvs[0]; |
907 |
} |
} |
908 |
|
|
909 |
if (MotionFlags & PMV_EXTSEARCH8) { |
static void __inline |
910 |
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
PreparePredictionsBF(VECTOR * const pmv, const int x, const int y, |
911 |
|
const uint32_t iWcount, |
912 |
|
const MACROBLOCK * const pMB, |
913 |
|
const uint32_t mode_curr) |
914 |
|
{ |
915 |
|
|
916 |
if (!(MVequal(pmv[0], backupMV))) { |
// [0] is prediction |
917 |
iSAD = |
pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); |
|
(*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, 0); |
|
918 |
|
|
919 |
if (iSAD < iMinSAD) { |
pmv[1].x = pmv[1].y = 0; // [1] is zero |
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
920 |
|
|
921 |
if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { |
pmv[2] = ChoosePred(pMB, mode_curr); |
922 |
iSAD = |
pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); |
|
(*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); |
|
923 |
|
|
924 |
if (iSAD < iMinSAD) { |
if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour |
925 |
*currMV = newMV; |
pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); |
926 |
iMinSAD = iSAD; |
pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); |
927 |
} |
} else pmv[3].x = pmv[3].y = 0; |
|
} |
|
|
} |
|
928 |
|
|
929 |
/*************** Choose best MV found **************/ |
if (y != 0) { |
930 |
|
pmv[4] = ChoosePred(pMB-iWcount, mode_curr); |
931 |
|
pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y); |
932 |
|
} else pmv[4].x = pmv[4].y = 0; |
933 |
|
|
934 |
EPZS8_Terminate_with_Refine: |
if (x != 0) { |
935 |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
pmv[5] = ChoosePred(pMB-1, mode_curr); |
936 |
iMinSAD = |
pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
937 |
Halfpel8_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, |
} else pmv[5].x = pmv[5].y = 0; |
|
iMinSAD, center_x, center_y, min_dx, max_dx, min_dy, max_dy, |
|
|
iFcode, iQuant, iEdgedWidth); |
|
938 |
|
|
939 |
EPZS8_Terminate_without_Refine: |
if ((x != 0)&&(y != 0)) { |
940 |
|
pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); |
941 |
|
pmv[6].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
942 |
|
} else pmv[6].x = pmv[6].y = 0; |
943 |
|
|
944 |
currPMV->x = currMV->x - center_x; |
// more? |
|
currPMV->y = currMV->y - center_y; |
|
|
return iMinSAD; |
|
945 |
} |
} |
946 |
|
|
947 |
|
|
948 |
|
/* search backward or forward, for b-frames */ |
949 |
int32_t |
static void |
950 |
PMVfastIntSearch16(const uint8_t * const pRef, |
SearchBF( const uint8_t * const pRef, |
951 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
952 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
953 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
954 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
955 |
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, |
|
956 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
|
const uint32_t iQuant, |
|
957 |
const uint32_t iFcode, |
const uint32_t iFcode, |
958 |
const MBParam * const pParam, |
const MBParam * const pParam, |
959 |
const MACROBLOCK * const pMBs, |
MACROBLOCK * const pMB, |
960 |
const MACROBLOCK * const prevMBs, |
const VECTOR * const predMV, |
961 |
VECTOR * const currMV, |
int32_t * const best_sad, |
962 |
VECTOR * const currPMV) |
const int32_t mode_current, |
963 |
|
SearchData * const Data) |
964 |
{ |
{ |
|
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 }; |
|
965 |
|
|
966 |
int32_t iDiamondSize; |
const int32_t iEdgedWidth = pParam->edged_width; |
967 |
|
|
968 |
int32_t min_dx; |
int i, iDirection, mask; |
969 |
int32_t max_dx; |
VECTOR pmv[7]; |
970 |
int32_t min_dy; |
MainSearchFunc *MainSearchPtr; |
971 |
int32_t max_dy; |
*Data->iMinSAD = MV_MAX_ERROR; |
972 |
|
Data->iFcode = iFcode; |
973 |
|
|
974 |
int32_t iFound; |
Data->Ref = pRef + (x + y * iEdgedWidth) * 16; |
975 |
|
Data->RefH = pRefH + (x + y * iEdgedWidth) * 16; |
976 |
|
Data->RefV = pRefV + (x + y * iEdgedWidth) * 16; |
977 |
|
Data->RefHV = pRefHV + (x + y * iEdgedWidth) * 16; |
978 |
|
|
979 |
VECTOR newMV; |
Data->predMV = *predMV; |
|
VECTOR backupMV; |
|
980 |
|
|
981 |
VECTOR pmv[4]; |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
982 |
int32_t psad[4]; |
pParam->width, pParam->height, iFcode); |
983 |
|
|
984 |
MainSearch16FuncPtr MainSearchPtr; |
pmv[0] = Data->predMV; |
985 |
|
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
986 |
|
|
987 |
const MACROBLOCK *const prevMB = prevMBs + x + y * iWcount; |
Data->currentMV->x = Data->currentMV->y = 0; |
|
MACROBLOCK *const pMB = pMBs + x + y * iWcount; |
|
988 |
|
|
989 |
int32_t threshA, threshB; |
CheckCandidate = CheckCandidate16no4v; |
|
int32_t bPredEq; |
|
|
int32_t iMinSAD, iSAD; |
|
990 |
|
|
991 |
|
// main loop. checking all predictions |
992 |
|
for (i = 0; i < 8; i++) { |
993 |
|
if (!(mask = make_mask(pmv, i)) ) continue; |
994 |
|
CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
995 |
|
} |
996 |
|
|
997 |
/* Get maximum range */ |
if (MotionFlags & PMV_USESQUARES16) |
998 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 16, iWidth, iHeight, |
MainSearchPtr = SquareSearch; |
999 |
iFcode); |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
1000 |
|
MainSearchPtr = AdvDiamondSearch; |
1001 |
|
else MainSearchPtr = DiamondSearch; |
1002 |
|
|
1003 |
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1004 |
|
|
1005 |
if ((x == 0) && (y == 0)) { |
HalfpelRefine(Data); |
|
threshA = 512; |
|
|
threshB = 1024; |
|
1006 |
|
|
1007 |
bPredEq = 0; |
// three bits are needed to code backward mode. four for forward |
1008 |
psad[0] = psad[1] = psad[2] = psad[3] = 0; |
// we treat the bits just like they were vector's |
1009 |
*currMV = pmv[0] = pmv[1] = pmv[2] = pmv[3] = zeroMV; |
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * lambda_vec16[Data->iQuant]; |
1010 |
|
else *Data->iMinSAD += 3 * lambda_vec16[Data->iQuant]; |
1011 |
|
|
|
} else { |
|
|
threshA = psad[0]; |
|
|
threshB = threshA + 256; |
|
|
if (threshA < 512) |
|
|
threshA = 512; |
|
|
if (threshA > 1024) |
|
|
threshA = 1024; |
|
|
if (threshB > 1792) |
|
|
threshB = 1792; |
|
1012 |
|
|
1013 |
bPredEq = get_ipmvdata(pMBs, iWcount, 0, x, y, 0, pmv, psad); |
if (*Data->iMinSAD < *best_sad) { |
1014 |
*currMV = pmv[0]; /* current best := prediction */ |
*best_sad = *Data->iMinSAD; |
1015 |
|
pMB->mode = mode_current; |
1016 |
|
pMB->pmvs[0].x = Data->currentMV->x - predMV->x; |
1017 |
|
pMB->pmvs[0].y = Data->currentMV->y - predMV->y; |
1018 |
|
if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; |
1019 |
|
else pMB->b_mvs[0] = *Data->currentMV; |
1020 |
} |
} |
1021 |
|
|
1022 |
iFound = 0; |
} |
1023 |
|
|
1024 |
/* Step 4: Calculate SAD around the Median prediction. |
static int32_t |
1025 |
MinSAD=SAD |
SearchDirect(const IMAGE * const f_Ref, |
1026 |
If Motion Vector equal to Previous frame motion vector |
const uint8_t * const f_RefH, |
1027 |
and MinSAD<PrevFrmSAD goto Step 10. |
const uint8_t * const f_RefV, |
1028 |
If SAD<=256 goto Step 10. |
const uint8_t * const f_RefHV, |
1029 |
*/ |
const IMAGE * const b_Ref, |
1030 |
|
const uint8_t * const b_RefH, |
1031 |
|
const uint8_t * const b_RefV, |
1032 |
|
const uint8_t * const b_RefHV, |
1033 |
|
const IMAGE * const pCur, |
1034 |
|
const int x, const int y, |
1035 |
|
const uint32_t MotionFlags, |
1036 |
|
const int32_t TRB, const int32_t TRD, |
1037 |
|
const MBParam * const pParam, |
1038 |
|
MACROBLOCK * const pMB, |
1039 |
|
const MACROBLOCK * const b_mb, |
1040 |
|
int32_t * const best_sad, |
1041 |
|
SearchData * const Data) |
1042 |
|
|
1043 |
|
{ |
1044 |
|
int32_t skip_sad; |
1045 |
|
int k; |
1046 |
|
|
1047 |
|
MainSearchFunc *MainSearchPtr; |
1048 |
|
|
1049 |
|
*Data->iMinSAD = 256*4096; |
1050 |
|
Data->referencemv = b_mb->mvs; |
1051 |
|
|
1052 |
|
Data->Ref = f_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
1053 |
|
Data->RefH = f_RefH + (x + Data->iEdgedWidth*y) * 16; |
1054 |
|
Data->RefV = f_RefV + (x + Data->iEdgedWidth*y) * 16; |
1055 |
|
Data->RefHV = f_RefHV + (x + Data->iEdgedWidth*y) * 16; |
1056 |
|
Data->bRef = b_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
1057 |
|
Data->bRefH = b_RefH + (x + Data->iEdgedWidth*y) * 16; |
1058 |
|
Data->bRefV = b_RefV + (x + Data->iEdgedWidth*y) * 16; |
1059 |
|
Data->bRefHV = b_RefHV + (x + Data->iEdgedWidth*y) * 16; |
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 (currMV->x > max_dx) { |
for (k = 0; k < 4; k++) { |
1067 |
currMV->x = EVEN(max_dx); |
pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); |
1068 |
} |
pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD; |
1069 |
if (currMV->x < min_dx) { |
pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); |
1070 |
currMV->x = EVEN(min_dx); |
pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; |
1071 |
|
|
1072 |
|
if ( ( pMB->b_mvs[k].x > Data->max_dx ) || ( pMB->b_mvs[k].x < Data->min_dx ) |
1073 |
|
|| ( pMB->b_mvs[k].y > Data->max_dy ) || ( pMB->b_mvs[k].y < Data->min_dy )) { |
1074 |
|
|
1075 |
|
*best_sad = 256*4096; // in that case, we won't use direct mode |
1076 |
|
pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV" |
1077 |
|
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
1078 |
|
return 0; |
1079 |
} |
} |
1080 |
if (currMV->y > max_dy) { |
if (b_mb->mode != MODE_INTER4V) { |
1081 |
currMV->y = EVEN(max_dy); |
pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; |
1082 |
|
pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0]; |
1083 |
|
Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0]; |
1084 |
|
Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0]; |
1085 |
|
break; |
1086 |
} |
} |
|
if (currMV->y < min_dy) { |
|
|
currMV->y = EVEN(min_dy); |
|
1087 |
} |
} |
1088 |
|
|
1089 |
iMinSAD = |
if (b_mb->mode == MODE_INTER4V) |
1090 |
sad16(cur, |
CheckCandidate = CheckCandidateDirect; |
1091 |
get_iref_mv(pRef, x, y, 16, currMV, |
else CheckCandidate = CheckCandidateDirectno4v; |
|
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->i_mvs[0])) && |
|
|
((int32_t) iMinSAD < prevMB->i_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; |
|
|
} |
|
|
} |
|
1092 |
|
|
1093 |
if (MotionFlags & PMV_EARLYSTOP16) |
(*CheckCandidate)(0, 0, 255, &k, Data); |
|
goto PMVfastInt16_Terminate_with_Refine; |
|
|
} |
|
1094 |
|
|
1095 |
|
// skip decision |
1096 |
|
if (*Data->iMinSAD - 2 * lambda_vec16[Data->iQuant] < (int32_t)Data->iQuant * SKIP_THRESH_B) { |
1097 |
|
//checking chroma. everything copied from MC |
1098 |
|
//this is not full chroma compensation, only it's fullpel approximation. should work though |
1099 |
|
int sum, dx, dy, b_dx, b_dy; |
1100 |
|
|
1101 |
/* Step 2 (lazy eval): Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
1102 |
vector of the median. |
dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
|
|
*/ |
|
1103 |
|
|
1104 |
if ((bPredEq) && (MVequal(pmv[0], prevMB->i_mvs[0]))) |
sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
1105 |
iFound = 2; |
dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
1106 |
|
|
1107 |
/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
sum = pMB->b_mvs[0].x + pMB->b_mvs[1].x + pMB->b_mvs[2].x + pMB->b_mvs[3].x; |
1108 |
Otherwise select large Diamond Search. |
b_dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
*/ |
|
1109 |
|
|
1110 |
if ((!MVzero(pmv[0])) || (threshB < 1536) || (bPredEq)) |
sum = pMB->b_mvs[0].y + pMB->b_mvs[1].y + pMB->b_mvs[2].y + pMB->b_mvs[3].y; |
1111 |
iDiamondSize = 2; // halfpel units! |
b_dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
else |
|
|
iDiamondSize = 4; // halfpel units! |
|
1112 |
|
|
1113 |
/* |
sum = sad8bi(pCur->u + 8*x + 8*y*(Data->iEdgedWidth/2), |
1114 |
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
f_Ref->u + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
1115 |
Also calculate (0,0) but do not subtract offset. |
b_Ref->u + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
1116 |
Let MinSAD be the smallest SAD up to this point. |
Data->iEdgedWidth/2); |
1117 |
If MV is (0,0) subtract offset. |
sum += sad8bi(pCur->v + 8*x + 8*y*(Data->iEdgedWidth/2), |
1118 |
*/ |
f_Ref->v + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
1119 |
|
b_Ref->v + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
1120 |
// (0,0) is often a good choice |
Data->iEdgedWidth/2); |
|
|
|
|
if (!MVzero(pmv[0])) |
|
|
CHECK_MV16_ZERO; |
|
|
|
|
|
// previous frame MV is always possible |
|
|
|
|
|
if (!MVzero(prevMB->i_mvs[0])) |
|
|
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; |
|
|
|
|
|
|
|
|
/* 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->i_mvs[0]) && |
|
|
((int32_t) iMinSAD < prevMB->i_sad16))) { |
|
|
|
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfastInt16_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. |
|
|
*/ |
|
1121 |
|
|
1122 |
if (MotionFlags & PMV_USESQUARES16) |
if ((uint32_t) sum < MAX_CHROMA_SAD_FOR_SKIP * Data->iQuant) { |
1123 |
MainSearchPtr = Square16_MainSearch; |
pMB->mode = MODE_DIRECT_NONE_MV; |
1124 |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
return *Data->iMinSAD; |
1125 |
MainSearchPtr = AdvDiamond16_MainSearch; |
} |
1126 |
else |
} |
|
MainSearchPtr = Diamond16_MainSearch; |
|
1127 |
|
|
1128 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
skip_sad = *Data->iMinSAD; |
1129 |
|
|
1130 |
|
// DIRECT MODE DELTA VECTOR SEARCH. |
1131 |
|
// This has to be made more effective, but at the moment I'm happy it's running at all |
1132 |
|
|
1133 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1134 |
iSAD = |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1135 |
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV->x, |
else MainSearchPtr = DiamondSearch; |
|
currMV->y, iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, |
|
|
min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, |
|
|
iQuant, iFound); |
|
1136 |
|
|
1137 |
if (iSAD < iMinSAD) { |
(*MainSearchPtr)(0, 0, Data, 255); |
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
1138 |
|
|
1139 |
if (MotionFlags & PMV_EXTSEARCH16) { |
HalfpelRefine(Data); |
|
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
|
1140 |
|
|
1141 |
if (!(MVequal(pmv[0], backupMV))) { |
*Data->iMinSAD += 1 * lambda_vec16[Data->iQuant]; // one bit is needed to code direct mode. we treat this bit just like it was vector's |
1142 |
iSAD = |
*best_sad = *Data->iMinSAD; |
|
(*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); |
|
1143 |
|
|
1144 |
if (iSAD < iMinSAD) { |
if (b_mb->mode == MODE_INTER4V) |
1145 |
*currMV = newMV; |
pMB->mode = MODE_DIRECT; |
1146 |
iMinSAD = iSAD; |
else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
|
} |
|
|
} |
|
1147 |
|
|
1148 |
if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { |
pMB->pmvs[3] = *Data->currentMV; |
|
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); |
|
1149 |
|
|
1150 |
if (iSAD < iMinSAD) { |
for (k = 0; k < 4; k++) { |
1151 |
*currMV = newMV; |
pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x; |
1152 |
iMinSAD = iSAD; |
pMB->b_mvs[k].x = ((Data->currentMV->x == 0) |
1153 |
|
? Data->directmvB[k].x |
1154 |
|
: pMB->mvs[k].x - Data->referencemv[k].x); |
1155 |
|
pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y); |
1156 |
|
pMB->b_mvs[k].y = ((Data->currentMV->y == 0) |
1157 |
|
? Data->directmvB[k].y |
1158 |
|
: pMB->mvs[k].y - Data->referencemv[k].y); |
1159 |
|
if (b_mb->mode != MODE_INTER4V) { |
1160 |
|
pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; |
1161 |
|
pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; |
1162 |
|
break; |
1163 |
} |
} |
1164 |
} |
} |
1165 |
|
return skip_sad; |
1166 |
} |
} |
1167 |
|
|
|
/* |
|
|
Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
|
|
*/ |
|
1168 |
|
|
1169 |
PMVfastInt16_Terminate_with_Refine: |
static __inline void |
1170 |
|
SearchInterpolate(const uint8_t * const f_Ref, |
1171 |
|
const uint8_t * const f_RefH, |
1172 |
|
const uint8_t * const f_RefV, |
1173 |
|
const uint8_t * const f_RefHV, |
1174 |
|
const uint8_t * const b_Ref, |
1175 |
|
const uint8_t * const b_RefH, |
1176 |
|
const uint8_t * const b_RefV, |
1177 |
|
const uint8_t * const b_RefHV, |
1178 |
|
const IMAGE * const pCur, |
1179 |
|
const int x, const int y, |
1180 |
|
const uint32_t fcode, |
1181 |
|
const uint32_t bcode, |
1182 |
|
const uint32_t MotionFlags, |
1183 |
|
const MBParam * const pParam, |
1184 |
|
const VECTOR * const f_predMV, |
1185 |
|
const VECTOR * const b_predMV, |
1186 |
|
MACROBLOCK * const pMB, |
1187 |
|
int32_t * const best_sad, |
1188 |
|
SearchData * const fData) |
1189 |
|
|
1190 |
pMB->i_mvs[0] = pMB->i_mvs[1] = pMB->i_mvs[2] = pMB->i_mvs[3] = pMB->i_mv16 = *currMV; |
{ |
|
pMB->i_sad8[0] = pMB->i_sad8[1] = pMB->i_sad8[2] = pMB->i_sad8[3] = pMB->i_sad16 = iMinSAD; |
|
1191 |
|
|
1192 |
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
const int32_t iEdgedWidth = pParam->edged_width; |
|
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); |
|
1193 |
|
|
1194 |
pmv[0] = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); // get _REAL_ prediction (halfpel possible) |
int iDirection, i, j; |
1195 |
|
SearchData bData; |
1196 |
|
|
1197 |
PMVfastInt16_Terminate_without_Refine: |
bData.iMinSAD = fData->iMinSAD; |
1198 |
currPMV->x = currMV->x - center_x; |
*bData.iMinSAD = 4096*256; |
1199 |
currPMV->y = currMV->y - center_y; |
bData.Cur = fData->Cur; |
1200 |
return iMinSAD; |
fData->iEdgedWidth = bData.iEdgedWidth = iEdgedWidth; |
1201 |
} |
bData.currentMV = fData->currentMV + 1; |
1202 |
|
bData.iQuant = fData->iQuant; |
1203 |
|
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
1204 |
|
|
1205 |
|
bData.bRef = fData->Ref = f_Ref + (x + y * iEdgedWidth) * 16; |
1206 |
|
bData.bRefH = fData->RefH = f_RefH + (x + y * iEdgedWidth) * 16; |
1207 |
|
bData.bRefV = fData->RefV = f_RefV + (x + y * iEdgedWidth) * 16; |
1208 |
|
bData.bRefHV = fData->RefHV = f_RefHV + (x + y * iEdgedWidth) * 16; |
1209 |
|
bData.Ref = fData->bRef = b_Ref + (x + y * iEdgedWidth) * 16; |
1210 |
|
bData.RefH = fData->bRefH = b_RefH + (x + y * iEdgedWidth) * 16; |
1211 |
|
bData.RefV = fData->bRefV = b_RefV + (x + y * iEdgedWidth) * 16; |
1212 |
|
bData.RefHV = fData->bRefHV = b_RefHV + (x + y * iEdgedWidth) * 16; |
1213 |
|
|
1214 |
|
bData.bpredMV = fData->predMV = *f_predMV; |
1215 |
|
fData->bpredMV = bData.predMV = *b_predMV; |
1216 |
|
|
1217 |
|
fData->currentMV[0] = pMB->mvs[0]; |
1218 |
|
fData->currentMV[1] = pMB->b_mvs[0]; |
1219 |
|
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode); |
1220 |
|
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode); |
1221 |
|
|
1222 |
|
if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; |
1223 |
|
if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dy; |
1224 |
|
if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dx; |
1225 |
|
if (fData->currentMV[0].y > fData->min_dy) fData->currentMV[0].y = fData->min_dy; |
1226 |
|
|
1227 |
|
if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx; |
1228 |
|
if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dy; |
1229 |
|
if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dx; |
1230 |
|
if (fData->currentMV[1].y > bData.min_dy) fData->currentMV[1].y = bData.min_dy; |
1231 |
|
|
1232 |
|
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
1233 |
|
|
1234 |
|
//diamond. I wish we could use normal mainsearch functions (square, advdiamond) |
1235 |
|
|
1236 |
|
do { |
1237 |
|
iDirection = 255; |
1238 |
|
// forward MV moves |
1239 |
|
i = fData->currentMV[0].x; j = fData->currentMV[0].y; |
1240 |
|
|
1241 |
|
CheckCandidateInt(i + 1, j, 0, &iDirection, fData); |
1242 |
|
CheckCandidateInt(i, j + 1, 0, &iDirection, fData); |
1243 |
|
CheckCandidateInt(i - 1, j, 0, &iDirection, fData); |
1244 |
|
CheckCandidateInt(i, j - 1, 0, &iDirection, fData); |
1245 |
|
|
1246 |
|
// backward MV moves |
1247 |
|
i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
1248 |
|
fData->currentMV[2] = fData->currentMV[0]; |
1249 |
|
|
1250 |
|
CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); |
1251 |
|
CheckCandidateInt(i, j + 1, 0, &iDirection, &bData); |
1252 |
|
CheckCandidateInt(i - 1, j, 0, &iDirection, &bData); |
1253 |
|
CheckCandidateInt(i, j - 1, 0, &iDirection, &bData); |
1254 |
|
|
1255 |
|
} while (!(iDirection)); |
1256 |
|
|
1257 |
|
// two bits are needed to code interpolate mode. we treat the bits just like they were vector's |
1258 |
|
*fData->iMinSAD += 2 * lambda_vec16[fData->iQuant]; |
1259 |
|
if (*fData->iMinSAD < *best_sad) { |
1260 |
|
*best_sad = *fData->iMinSAD; |
1261 |
|
pMB->mvs[0] = fData->currentMV[0]; |
1262 |
|
pMB->b_mvs[0] = fData->currentMV[1]; |
1263 |
|
pMB->mode = MODE_INTERPOLATE; |
1264 |
|
|
1265 |
|
pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; |
1266 |
|
pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; |
1267 |
|
pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; |
1268 |
|
pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; |
1269 |
|
} |
1270 |
|
} |
1271 |
|
|
|
/* *********************************************************** |
|
|
bvop motion estimation |
|
|
***************************************************************/ |
|
1272 |
|
|
1273 |
void |
void |
1274 |
MotionEstimationBVOP(MBParam * const pParam, |
MotionEstimationBVOP(MBParam * const pParam, |
1288 |
const IMAGE * const b_refV, |
const IMAGE * const b_refV, |
1289 |
const IMAGE * const b_refHV) |
const IMAGE * const b_refHV) |
1290 |
{ |
{ |
1291 |
const int mb_width = pParam->mb_width; |
uint32_t i, j; |
1292 |
const int mb_height = pParam->mb_height; |
int32_t best_sad, skip_sad; |
1293 |
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; |
|
|
|
|
1294 |
static const VECTOR zeroMV={0,0}; |
static const VECTOR zeroMV={0,0}; |
1295 |
|
|
|
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; |
|
|
|
|
1296 |
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; |
|
1297 |
|
|
1298 |
int min_dx, max_dx, min_dy, max_dy; |
const int32_t TRB = time_pp - time_bp; |
1299 |
int f_min_dx, f_max_dx, f_min_dy, f_max_dy; |
const int32_t TRD = time_pp; |
1300 |
int b_min_dx, b_max_dx, b_min_dy, b_max_dy; |
|
1301 |
|
// some pre-inintialized data for the rest of the search |
|
int f_count=0; |
|
|
int b_count=0; |
|
|
int i_count=0; |
|
|
int d_count=0; |
|
1302 |
|
|
1303 |
const int64_t TRB = (int32_t)time_pp - (int32_t)time_bp; |
SearchData Data; |
1304 |
const int64_t TRD = (int32_t)time_pp; |
int32_t iMinSAD; |
1305 |
|
VECTOR currentMV[3]; |
1306 |
|
Data.iEdgedWidth = pParam->edged_width; |
1307 |
|
Data.currentMV = currentMV; |
1308 |
|
Data.iMinSAD = &iMinSAD; |
1309 |
|
Data.iQuant = frame->quant; |
1310 |
|
|
|
// fprintf(stderr,"TRB = %lld TRD = %lld time_bp =%d time_pp =%d\n\n",TRB,TRD,time_bp,time_pp); |
|
1311 |
// note: i==horizontal, j==vertical |
// note: i==horizontal, j==vertical |
|
for (j = 0; j < mb_height; j++) { |
|
1312 |
|
|
1313 |
f_predMV = zeroMV; /* prediction is reset at left boundary */ |
for (j = 0; j < pParam->mb_height; j++) { |
|
b_predMV = zeroMV; |
|
1314 |
|
|
1315 |
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]; |
|
1316 |
|
|
1317 |
mb->deltamv=zeroMV; |
for (i = 0; i < pParam->mb_width; i++) { |
1318 |
|
MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; |
1319 |
|
const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; |
1320 |
|
|
1321 |
/* 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 */ |
1322 |
|
if (b_mb->mode == MODE_NOT_CODED) { |
1323 |
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; |
|
1324 |
continue; |
continue; |
1325 |
} |
} |
1326 |
|
|
1327 |
if (b_mb->mode == MODE_INTER4V) |
Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
1328 |
{ |
/* direct search comes first, because it (1) checks for SKIP-mode |
1329 |
d_sad16 = 0; |
and (2) sets very good predictions for forward and backward search */ |
1330 |
/* same method of scaling as in decoder.c, so we copy from there */ |
|
1331 |
for (k = 0; k < 4; k++) { |
skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
1332 |
|
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
1333 |
mb->directmv[k] = b_mb->mvs[k]; |
&frame->image, |
1334 |
|
i, j, |
1335 |
mb->mvs[k].x = (int32_t) ((TRB * mb->directmv[k].x) / TRD + mb->deltamv.x); |
frame->motion_flags, |
1336 |
mb->b_mvs[k].x = (int32_t) ((mb->deltamv.x == 0) |
TRB, TRD, |
1337 |
? ((TRB - TRD) * mb->directmv[k].x) / TRD |
pParam, |
1338 |
: mb->mvs[k].x - mb->directmv[k].x); |
pMB, b_mb, |
1339 |
|
&best_sad, |
1340 |
mb->mvs[k].y = (int32_t) ((TRB * mb->directmv[k].y) / TRD + mb->deltamv.y); |
&Data); |
|
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); |
|
|
} |
|
|
} |
|
|
else |
|
|
{ |
|
|
mb->directmv[3] = mb->directmv[2] = mb->directmv[1] = |
|
|
mb->directmv[0] = b_mb->mvs[0]; |
|
1341 |
|
|
1342 |
mb->mvs[0].x = (int32_t) ((TRB * mb->directmv[0].x) / TRD + mb->deltamv.x); |
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
|
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); |
|
|
|
|
|
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); |
|
1343 |
|
|
1344 |
} |
// best_sad = 256*4096; //uncomment to disable Directsearch. |
1345 |
d_sad16 += calc_delta_16(mb->deltamv.x, mb->deltamv.y, 1, frame->quant); |
// To disable any other mode, just comment the function call |
1346 |
|
|
1347 |
// forward search |
// forward search |
1348 |
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, |
1349 |
&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 */ |
|
1350 |
frame->motion_flags, |
frame->motion_flags, |
1351 |
frame->quant, frame->fcode, pParam, |
frame->fcode, pParam, |
1352 |
f_mbs, f_mbs, |
pMB, &f_predMV, &best_sad, |
1353 |
&mb->mvs[0], &pmv_dontcare); |
MODE_FORWARD, &Data); |
|
|
|
1354 |
|
|
1355 |
// backward search |
// backward search |
1356 |
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, |
1357 |
&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 */ |
|
1358 |
frame->motion_flags, |
frame->motion_flags, |
1359 |
frame->quant, frame->bcode, pParam, |
frame->bcode, pParam, |
1360 |
b_mbs, b_mbs, |
pMB, &b_predMV, &best_sad, |
1361 |
&mb->b_mvs[0], &pmv_dontcare); |
MODE_BACKWARD, &Data); |
1362 |
|
|
1363 |
i_sad16 = |
// interpolate search comes last, because it uses data from forward and backward as prediction |
1364 |
sad16bi(frame->image.y + i * 16 + j * 16 * edged_width, |
|
1365 |
get_ref_mv(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
SearchInterpolate(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); |
|
|
|
|
|
/* Interpolated MC motion vector search, this is tedious and more complicated because there are |
|
|
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". */ |
|
|
|
|
|
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, |
|
|
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, |
|
1366 |
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
1367 |
|
&frame->image, |
1368 |
i, j, |
i, j, |
1369 |
TRB,TRD, |
frame->fcode, frame->bcode, |
1370 |
mb->deltamv.x, mb->deltamv.y, |
frame->motion_flags, |
1371 |
d_sad16, |
pParam, |
1372 |
&mb->deltamv, |
&f_predMV, &b_predMV, |
1373 |
mb->directmv, // this has to be pre-initialized with b_mb->mvs[] |
pMB, &best_sad, |
1374 |
min_dx, max_dx, min_dy, max_dy, |
&Data); |
1375 |
edged_width, 1, frame->quant, 0); // equiv to halfpel refine |
|
1376 |
|
switch (pMB->mode) { |
1377 |
|
case MODE_FORWARD: |
1378 |
// i_sad16 = 65535; /* remove the comment to disable any of the MODEs */ |
f_count++; |
1379 |
// f_sad16 = 65535; |
f_predMV = pMB->mvs[0]; |
1380 |
// b_sad16 = 65535; |
break; |
1381 |
// d_sad16 = 65535; |
case MODE_BACKWARD: |
1382 |
|
b_count++; |
1383 |
if (f_sad16 < b_sad16) { |
b_predMV = pMB->b_mvs[0]; |
1384 |
best_sad = f_sad16; |
break; |
1385 |
mb->mode = MODE_FORWARD; |
case MODE_INTERPOLATE: |
1386 |
} else { |
i_count++; |
1387 |
best_sad = b_sad16; |
f_predMV = pMB->mvs[0]; |
1388 |
mb->mode = MODE_BACKWARD; |
b_predMV = pMB->b_mvs[0]; |
1389 |
|
break; |
1390 |
|
case MODE_DIRECT: |
1391 |
|
case MODE_DIRECT_NO4V: |
1392 |
|
d_count++; |
1393 |
|
break; |
1394 |
|
default: |
1395 |
|
break; |
1396 |
|
} |
1397 |
|
} |
1398 |
} |
} |
1399 |
|
|
1400 |
if (i_sad16 < best_sad) { |
// fprintf(debug,"B-Stat: F: %04d B: %04d I: %04d D: %04d, N: %04d\n", |
1401 |
best_sad = i_sad16; |
// f_count,b_count,i_count,d_count,n_count); |
1402 |
mb->mode = MODE_INTERPOLATE; |
|
1403 |
} |
} |
1404 |
|
|
1405 |
if (d_sad16 < best_sad) { |
/* Hinted ME starts here */ |
1406 |
|
|
1407 |
if (b_mb->mode == MODE_INTER4V) |
static __inline void |
1408 |
|
Search8hinted( const SearchData * const OldData, |
1409 |
|
const int x, const int y, |
1410 |
|
const uint32_t MotionFlags, |
1411 |
|
const MBParam * const pParam, |
1412 |
|
MACROBLOCK * const pMB, |
1413 |
|
const MACROBLOCK * const pMBs, |
1414 |
|
const int block) |
1415 |
{ |
{ |
1416 |
|
SearchData Data; |
1417 |
|
MainSearchFunc *MainSearchPtr; |
1418 |
|
|
1419 |
/* 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); |
1420 |
/* for the bitstream, the value mb->deltamv is read directly */ |
Data.iMinSAD = OldData->iMinSAD + 1 + block; |
1421 |
|
Data.currentMV = OldData->currentMV+1+block; |
1422 |
|
Data.iFcode = OldData->iFcode; |
1423 |
|
Data.iQuant = OldData->iQuant; |
1424 |
|
|
1425 |
for (k = 0; k < 4; k++) { |
Data.Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1426 |
|
Data.RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1427 |
|
Data.RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1428 |
|
Data.RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1429 |
|
Data.iEdgedWidth = pParam->edged_width; |
1430 |
|
Data.Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1431 |
|
|
1432 |
|
CheckCandidate = CheckCandidate8; |
1433 |
|
|
1434 |
|
if (block != 0) |
1435 |
|
*(Data.iMinSAD) += lambda_vec8[Data.iQuant] * |
1436 |
|
d_mv_bits( Data.currentMV->x - Data.predMV.x, |
1437 |
|
Data.currentMV->y - Data.predMV.y, |
1438 |
|
Data.iFcode); |
1439 |
|
|
1440 |
|
|
1441 |
|
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 8, |
1442 |
|
pParam->width, pParam->height, OldData->iFcode); |
1443 |
|
|
1444 |
|
if (pMB->mode == MODE_INTER4V) { |
1445 |
|
int dummy; |
1446 |
|
CheckCandidate8(pMB->mvs[block].x, pMB->mvs[block].y, 0, &dummy, &Data); } |
1447 |
|
|
1448 |
|
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
1449 |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
1450 |
|
else MainSearchPtr = DiamondSearch; |
1451 |
|
|
1452 |
|
(*MainSearchPtr)(Data.currentMV->x, Data.currentMV->y, &Data, 255); |
1453 |
|
|
1454 |
|
if (MotionFlags & PMV_HALFPELREFINE8) HalfpelRefine(&Data); |
1455 |
|
|
1456 |
|
pMB->pmvs[block].x = Data.currentMV->x - Data.predMV.x; |
1457 |
|
pMB->pmvs[block].y = Data.currentMV->y - Data.predMV.y; |
1458 |
|
pMB->mvs[block] = *(Data.currentMV); |
1459 |
|
pMB->sad8[block] = 4 * (*(Data.iMinSAD)); |
1460 |
|
} |
1461 |
|
|
1462 |
|
|
1463 |
|
static void |
1464 |
|
SearchPhinted ( const uint8_t * const pRef, |
1465 |
|
const uint8_t * const pRefH, |
1466 |
|
const uint8_t * const pRefV, |
1467 |
|
const uint8_t * const pRefHV, |
1468 |
|
const IMAGE * const pCur, |
1469 |
|
const int x, |
1470 |
|
const int y, |
1471 |
|
const uint32_t MotionFlags, |
1472 |
|
const uint32_t iQuant, |
1473 |
|
const MBParam * const pParam, |
1474 |
|
const MACROBLOCK * const pMBs, |
1475 |
|
int inter4v, |
1476 |
|
MACROBLOCK * const pMB, |
1477 |
|
SearchData * const Data) |
1478 |
|
{ |
1479 |
|
|
1480 |
|
const int32_t iEdgedWidth = pParam->edged_width; |
1481 |
|
|
1482 |
|
int i, t; |
1483 |
|
MainSearchFunc * MainSearchPtr; |
1484 |
|
|
1485 |
|
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
1486 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1487 |
|
pParam->width, pParam->height, Data->iFcode); |
1488 |
|
|
1489 |
|
Data->Cur = pCur->y + (x + y * iEdgedWidth) * 16; |
1490 |
|
Data->Ref = pRef + (x + iEdgedWidth*y)*16; |
1491 |
|
Data->RefH = pRefH + (x + iEdgedWidth*y) * 16; |
1492 |
|
Data->RefV = pRefV + (x + iEdgedWidth*y) * 16; |
1493 |
|
Data->RefHV = pRefHV + (x + iEdgedWidth*y) * 16; |
1494 |
|
Data->iQuant = iQuant; |
1495 |
|
|
1496 |
|
if (!(MotionFlags & PMV_HALFPEL16)) { |
1497 |
|
Data->min_dx = EVEN(Data->min_dx); |
1498 |
|
Data->max_dx = EVEN(Data->max_dx); |
1499 |
|
Data->min_dy = EVEN(Data->min_dy); |
1500 |
|
Data->max_dy = EVEN(Data->max_dy); |
1501 |
|
} |
1502 |
|
|
1503 |
|
for(i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
1504 |
|
|
1505 |
|
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
1506 |
|
|
1507 |
|
if (inter4v) |
1508 |
|
CheckCandidate = CheckCandidate16; |
1509 |
|
else CheckCandidate = CheckCandidate16no4v; |
1510 |
|
|
1511 |
|
|
1512 |
|
pMB->mvs[0].x = EVEN(pMB->mvs[0].x); |
1513 |
|
pMB->mvs[0].y = EVEN(pMB->mvs[0].y); |
1514 |
|
if (pMB->mvs[0].x > Data->max_dx) pMB->mvs[0].x = Data->max_dx; // this is in case iFcode changed |
1515 |
|
if (pMB->mvs[0].x < Data->min_dx) pMB->mvs[0].x = Data->min_dx; |
1516 |
|
if (pMB->mvs[0].y > Data->max_dy) pMB->mvs[0].y = Data->max_dy; |
1517 |
|
if (pMB->mvs[0].y < Data->min_dy) pMB->mvs[0].y = Data->min_dy; |
1518 |
|
|
1519 |
|
(*CheckCandidate)(pMB->mvs[0].x, pMB->mvs[0].y, 0, &t, Data); |
1520 |
|
|
1521 |
|
if (pMB->mode == MODE_INTER4V) |
1522 |
|
for (i = 1; i < 4; i++) { // all four vectors will be used as four predictions for 16x16 search |
1523 |
|
pMB->mvs[i].x = EVEN(pMB->mvs[i].x); |
1524 |
|
pMB->mvs[i].y = EVEN(pMB->mvs[i].y); |
1525 |
|
if (!(make_mask(pMB->mvs, i))) |
1526 |
|
(*CheckCandidate)(pMB->mvs[i].x, pMB->mvs[i].y, 0, &t, Data); |
1527 |
|
} |
1528 |
|
|
1529 |
|
if (MotionFlags & PMV_USESQUARES16) |
1530 |
|
MainSearchPtr = SquareSearch; |
1531 |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
1532 |
|
MainSearchPtr = AdvDiamondSearch; |
1533 |
|
else MainSearchPtr = DiamondSearch; |
1534 |
|
|
1535 |
|
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1536 |
|
|
1537 |
|
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data); |
1538 |
|
|
1539 |
mb->mvs[k].x = (int32_t) ((TRB * mb->directmv[k].x) / TRD + mb->deltamv.x); |
if (inter4v) |
1540 |
mb->b_mvs[k].x = (int32_t) ((mb->deltamv.x == 0) |
for(i = 0; i < 4; i++) |
1541 |
? ((TRB - TRD) * mb->directmv[k].x) / TRD |
Search8hinted(Data, 2*x+(i&1), 2*y+(i>>1), MotionFlags, pParam, pMB, pMBs, i); |
1542 |
: mb->mvs[k].x - mb->directmv[k].x); |
|
1543 |
|
if (!(inter4v) || |
1544 |
mb->mvs[k].y = (int32_t) ((TRB * mb->directmv[k].y) / TRD + mb->deltamv.y); |
(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] + |
1545 |
mb->b_mvs[k].y = (int32_t) ((mb->deltamv.y == 0) |
Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
1546 |
? ((TRB - TRD) * mb->directmv[k].y) / TRD |
// INTER MODE |
1547 |
: mb->mvs[k].y - mb->directmv[k].y); |
|
1548 |
|
pMB->mode = MODE_INTER; |
1549 |
|
pMB->mvs[0] = pMB->mvs[1] |
1550 |
|
= pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
1551 |
|
|
1552 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
1553 |
|
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
1554 |
|
|
1555 |
|
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
1556 |
|
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
1557 |
|
} else { |
1558 |
|
// INTER4V MODE; all other things are already set in Search8hinted |
1559 |
|
pMB->mode = MODE_INTER4V; |
1560 |
|
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] |
1561 |
|
+ Data->iMinSAD[4] + IMV16X16 * iQuant; |
1562 |
} |
} |
1563 |
|
|
1564 |
} |
} |
1565 |
|
|
1566 |
|
void |
1567 |
|
MotionEstimationHinted( MBParam * const pParam, |
1568 |
|
FRAMEINFO * const current, |
1569 |
|
FRAMEINFO * const reference, |
1570 |
|
const IMAGE * const pRefH, |
1571 |
|
const IMAGE * const pRefV, |
1572 |
|
const IMAGE * const pRefHV) |
1573 |
|
{ |
1574 |
|
MACROBLOCK *const pMBs = current->mbs; |
1575 |
|
const IMAGE *const pCurrent = ¤t->image; |
1576 |
|
const IMAGE *const pRef = &reference->image; |
1577 |
|
|
1578 |
|
uint32_t x, y; |
1579 |
|
int32_t temp[5], quant = current->quant; |
1580 |
|
int32_t iMinSAD[5]; |
1581 |
|
VECTOR currentMV[5]; |
1582 |
|
SearchData Data; |
1583 |
|
Data.iEdgedWidth = pParam->edged_width; |
1584 |
|
Data.currentMV = currentMV; |
1585 |
|
Data.iMinSAD = iMinSAD; |
1586 |
|
Data.temp = temp; |
1587 |
|
Data.iFcode = current->fcode; |
1588 |
|
|
1589 |
|
if (sadInit) (*sadInit) (); |
1590 |
|
|
1591 |
|
for (y = 0; y < pParam->mb_height; y++) { |
1592 |
|
for (x = 0; x < pParam->mb_width; x++) { |
1593 |
|
|
1594 |
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
1595 |
|
|
1596 |
|
//intra mode is copied from the first pass. At least for the time being |
1597 |
|
if ((pMB->mode == MODE_INTRA) || (pMB->mode == MODE_NOT_CODED) ) continue; |
1598 |
|
|
1599 |
|
|
1600 |
|
if (!(current->global_flags & XVID_LUMIMASKING)) { |
1601 |
|
pMB->dquant = NO_CHANGE; |
1602 |
|
pMB->quant = current->quant; } |
1603 |
else |
else |
1604 |
|
if (pMB->dquant != NO_CHANGE) { |
1605 |
|
quant += DQtab[pMB->dquant]; |
1606 |
|
if (quant > 31) quant = 31; |
1607 |
|
else if (quant < 1) quant = 1; |
1608 |
|
pMB->quant = quant; |
1609 |
|
} |
1610 |
|
|
1611 |
|
SearchPhinted(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
1612 |
|
y, current->motion_flags, pMB->quant, |
1613 |
|
pParam, pMBs, current->global_flags & XVID_INTER4V, pMB, |
1614 |
|
&Data); |
1615 |
|
|
1616 |
|
} |
1617 |
|
} |
1618 |
|
} |
1619 |
|
|
1620 |
|
static __inline int |
1621 |
|
MEanalyzeMB ( const uint8_t * const pRef, |
1622 |
|
const uint8_t * const pCur, |
1623 |
|
const int x, |
1624 |
|
const int y, |
1625 |
|
const MBParam * const pParam, |
1626 |
|
const MACROBLOCK * const pMBs, |
1627 |
|
MACROBLOCK * const pMB, |
1628 |
|
SearchData * const Data) |
1629 |
{ |
{ |
|
mb->mvs[0].x = (int32_t) ((TRB * mb->directmv[0].x) / TRD + mb->deltamv.x); |
|
1630 |
|
|
1631 |
mb->b_mvs[0].x = (int32_t) ((mb->deltamv.x == 0) |
int i, mask; |
1632 |
? ((TRB - TRD) * mb->directmv[0].x) / TRD |
VECTOR pmv[3]; |
|
: mb->mvs[0].x - mb->directmv[0].x); |
|
1633 |
|
|
1634 |
mb->mvs[0].y = (int32_t) ((TRB * mb->directmv[0].y) / TRD + mb->deltamv.y); |
*(Data->iMinSAD) = MV_MAX_ERROR; |
1635 |
|
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
1636 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1637 |
|
pParam->width, pParam->height, Data->iFcode); |
1638 |
|
|
1639 |
mb->b_mvs[0].y = (int32_t) ((mb->deltamv.y == 0) |
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
1640 |
? ((TRB - TRD) * mb->directmv[0].y) / TRD |
Data->Ref = pRef + (x + y * pParam->edged_width) * 16; |
|
: mb->mvs[0].y - mb->directmv[0].y); |
|
1641 |
|
|
1642 |
mb->mvs[3] = mb->mvs[2] = mb->mvs[1] = mb->mvs[0]; |
CheckCandidate = CheckCandidate16no4vI; |
1643 |
mb->b_mvs[3] = mb->b_mvs[2] = mb->b_mvs[1] = mb->b_mvs[0]; |
|
1644 |
} |
pmv[1].x = EVEN(pMB->mvs[0].x); |
1645 |
|
pmv[1].y = EVEN(pMB->mvs[0].y); |
1646 |
|
pmv[0].x = EVEN(Data->predMV.x); |
1647 |
|
pmv[0].y = EVEN(Data->predMV.y); |
1648 |
|
pmv[2].x = pmv[2].y = 0; |
1649 |
|
|
1650 |
best_sad = d_sad16; |
CheckCandidate16no4vI(pmv[0].x, pmv[0].y, 255, &i, Data); |
1651 |
mb->mode = MODE_DIRECT; |
if (!(mask = make_mask(pmv, 1))) |
1652 |
|
CheckCandidate16no4vI(pmv[1].x, pmv[1].y, mask, &i, Data); |
1653 |
|
if (!(mask = make_mask(pmv, 2))) |
1654 |
|
CheckCandidate16no4vI(0, 0, mask, &i, Data); |
1655 |
|
|
1656 |
|
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
1657 |
|
|
1658 |
|
pMB->mvs[0] = pMB->mvs[1] |
1659 |
|
= pMB->mvs[2] = pMB->mvs[3] = *Data->currentMV; // all, for future get_pmv() |
1660 |
|
|
1661 |
|
return *(Data->iMinSAD); |
1662 |
} |
} |
1663 |
|
|
1664 |
switch (mb->mode) |
#define INTRA_THRESH 1350 |
1665 |
|
#define INTER_THRESH 900 |
1666 |
|
|
1667 |
|
int |
1668 |
|
MEanalysis( const IMAGE * const pRef, |
1669 |
|
const IMAGE * const pCurrent, |
1670 |
|
MBParam * const pParam, |
1671 |
|
MACROBLOCK * const pMBs, |
1672 |
|
const uint32_t iFcode) |
1673 |
{ |
{ |
1674 |
case MODE_FORWARD: |
uint32_t x, y, intra = 0; |
1675 |
f_count++; |
int sSAD = 0; |
1676 |
f_predMV = mb->mvs[0]; |
|
1677 |
break; |
VECTOR currentMV; |
1678 |
case MODE_BACKWARD: |
int32_t iMinSAD; |
1679 |
b_count++; |
SearchData Data; |
1680 |
b_predMV = mb->b_mvs[0]; |
Data.iEdgedWidth = pParam->edged_width; |
1681 |
|
Data.currentMV = ¤tMV; |
1682 |
|
Data.iMinSAD = &iMinSAD; |
1683 |
|
Data.iFcode = iFcode; |
1684 |
|
Data.iQuant = 2; |
1685 |
|
|
1686 |
|
if (sadInit) (*sadInit) (); |
1687 |
|
|
1688 |
|
for (y = 0; y < pParam->mb_height-1; y++) { |
1689 |
|
for (x = 0; x < pParam->mb_width; x++) { |
1690 |
|
int sad, dev; |
1691 |
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
1692 |
|
|
1693 |
|
sad = MEanalyzeMB(pRef->y, pCurrent->y, x, y, |
1694 |
|
pParam, pMBs, pMB, &Data); |
1695 |
|
|
1696 |
|
if ( x != 0 && y != 0 && x != pParam->mb_width-1 ) { //no edge macroblocks, they just don't work |
1697 |
|
if (sad > INTRA_THRESH) { |
1698 |
|
dev = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
1699 |
|
pParam->edged_width); |
1700 |
|
if (dev + INTRA_THRESH < sad) intra++; |
1701 |
|
if (intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return 2; // I frame |
1702 |
|
} |
1703 |
|
sSAD += sad; |
1704 |
|
} |
1705 |
|
|
|
break; |
|
|
case MODE_INTERPOLATE: |
|
|
i_count++; |
|
|
mb->mvs[0] = f_interpolMV; |
|
|
mb->b_mvs[0] = b_interpolMV; |
|
|
f_predMV = mb->mvs[0]; |
|
|
b_predMV = mb->b_mvs[0]; |
|
|
break; |
|
|
case MODE_DIRECT: |
|
|
d_count++; |
|
|
break; |
|
|
default: |
|
|
break; |
|
1706 |
} |
} |
1707 |
|
} |
1708 |
|
sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); |
1709 |
|
if (sSAD > INTER_THRESH ) return 1; //P frame |
1710 |
|
emms(); |
1711 |
|
return 0; // B frame |
1712 |
|
|
1713 |
|
} |
1714 |
|
|
1715 |
|
int |
1716 |
|
FindFcode( const MBParam * const pParam, |
1717 |
|
const FRAMEINFO * const current) |
1718 |
|
{ |
1719 |
|
uint32_t x, y; |
1720 |
|
int max = 0, min = 0, i; |
1721 |
|
|
1722 |
|
for (y = 0; y < pParam->mb_height; y++) { |
1723 |
|
for (x = 0; x < pParam->mb_width; x++) { |
1724 |
|
|
1725 |
|
MACROBLOCK *pMB = ¤t->mbs[x + y * pParam->mb_width]; |
1726 |
|
for(i = 0; i < (pMB->mode == MODE_INTER4V ? 4:1); i++) { |
1727 |
|
if (pMB->mvs[i].x > max) max = pMB->mvs[i].x; |
1728 |
|
if (pMB->mvs[i].y > max) max = pMB->mvs[i].y; |
1729 |
|
|
1730 |
|
if (pMB->mvs[i].x < min) min = pMB->mvs[i].x; |
1731 |
|
if (pMB->mvs[i].y < min) min = pMB->mvs[i].y; |
1732 |
|
} |
1733 |
} |
} |
1734 |
} |
} |
1735 |
|
|
1736 |
#ifdef _DEBUG_BFRAME_STAT |
min = -min; |
1737 |
fprintf(stderr,"B-Stat: F: %04d B: %04d I: %04d D: %04d\n", |
max += 1; |
1738 |
f_count,b_count,i_count,d_count); |
if (min > max) max = min; |
|
#endif |
|
1739 |
|
|
1740 |
|
for (i = 1; (max > 32 << (i - 1)); i++); |
1741 |
|
return i; |
1742 |
} |
} |