2 |
* |
* |
3 |
* Modifications: |
* Modifications: |
4 |
* |
* |
5 |
|
* 01.05.2002 updated MotionEstimationBVOP |
6 |
|
* 25.04.2002 partial prevMB conversion |
7 |
|
* 22.04.2002 remove some compile warning by chenm001 <chenm001@163.com> |
8 |
|
* 14.04.2002 added MotionEstimationBVOP() |
9 |
|
* 02.04.2002 add EPZS(^2) as ME algorithm, use PMV_USESQUARES to choose between |
10 |
|
* EPZS and EPZS^2 |
11 |
* 08.02.2002 split up PMVfast into three routines: PMVFast, PMVFast_MainLoop |
* 08.02.2002 split up PMVfast into three routines: PMVFast, PMVFast_MainLoop |
12 |
* PMVFast_Refine to support multiple searches with different start points |
* PMVFast_Refine to support multiple searches with different start points |
13 |
* 07.01.2002 uv-block-based interpolation |
* 07.01.2002 uv-block-based interpolation |
37 |
|
|
38 |
#include <assert.h> |
#include <assert.h> |
39 |
#include <stdio.h> |
#include <stdio.h> |
40 |
|
#include <stdlib.h> |
41 |
|
|
42 |
#include "../encoder.h" |
#include "../encoder.h" |
43 |
#include "../utils/mbfunctions.h" |
#include "../utils/mbfunctions.h" |
44 |
#include "../prediction/mbprediction.h" |
#include "../prediction/mbprediction.h" |
45 |
#include "../global.h" |
#include "../global.h" |
46 |
#include "../utils/timer.h" |
#include "../utils/timer.h" |
47 |
|
#include "motion.h" |
48 |
#include "sad.h" |
#include "sad.h" |
49 |
|
|
50 |
// very large value |
// very large value |
57 |
/* sad16(0,0) bias; mpeg4 spec suggests nb/2+1 */ |
/* sad16(0,0) bias; mpeg4 spec suggests nb/2+1 */ |
58 |
/* nb = vop pixels * 2^(bpp-8) */ |
/* nb = vop pixels * 2^(bpp-8) */ |
59 |
#define MV16_00_BIAS (128+1) |
#define MV16_00_BIAS (128+1) |
60 |
|
#define MV8_00_BIAS (0) |
61 |
|
|
62 |
/* INTER bias for INTER/INTRA decision; mpeg4 spec suggests 2*nb */ |
/* INTER bias for INTER/INTRA decision; mpeg4 spec suggests 2*nb */ |
63 |
#define INTER_BIAS 512 |
#define INTER_BIAS 512 |
74 |
#define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) |
#define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) |
75 |
|
|
76 |
|
|
77 |
#define MIN(X, Y) ((X)<(Y)?(X):(Y)) |
int32_t PMVfastSearch16( |
78 |
#define MAX(X, Y) ((X)>(Y)?(X):(Y)) |
const uint8_t * const pRef, |
79 |
#define ABS(X) (((X)>0)?(X):-(X)) |
const uint8_t * const pRefH, |
80 |
#define SIGN(X) (((X)>0)?1:-1) |
const uint8_t * const pRefV, |
81 |
|
const uint8_t * const pRefHV, |
82 |
|
const IMAGE * const pCur, |
83 |
|
const int x, const int y, |
84 |
|
const uint32_t MotionFlags, |
85 |
|
const uint32_t iQuant, |
86 |
|
const uint32_t iFcode, |
87 |
|
const MBParam * const pParam, |
88 |
|
const MACROBLOCK * const pMBs, |
89 |
|
const MACROBLOCK * const prevMBs, |
90 |
|
VECTOR * const currMV, |
91 |
|
VECTOR * const currPMV); |
92 |
|
|
93 |
|
int32_t EPZSSearch16( |
94 |
|
const uint8_t * const pRef, |
95 |
|
const uint8_t * const pRefH, |
96 |
|
const uint8_t * const pRefV, |
97 |
|
const uint8_t * const pRefHV, |
98 |
|
const IMAGE * const pCur, |
99 |
|
const int x, const int y, |
100 |
|
const uint32_t MotionFlags, |
101 |
|
const uint32_t iQuant, |
102 |
|
const uint32_t iFcode, |
103 |
|
const MBParam * const pParam, |
104 |
|
const MACROBLOCK * const pMBs, |
105 |
|
const MACROBLOCK * const prevMBs, |
106 |
|
VECTOR * const currMV, |
107 |
|
VECTOR * const currPMV); |
108 |
|
|
109 |
|
|
110 |
int32_t PMVfastSearch8( |
int32_t PMVfastSearch8( |
114 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
115 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
116 |
const int x, const int y, |
const int x, const int y, |
117 |
const int start_x, int start_y, |
const int start_x, const int start_y, |
118 |
const uint32_t iQuality, |
const uint32_t MotionFlags, |
119 |
MBParam * const pParam, |
const uint32_t iQuant, |
120 |
MACROBLOCK * const pMBs, |
const uint32_t iFcode, |
121 |
|
const MBParam * const pParam, |
122 |
|
const MACROBLOCK * const pMBs, |
123 |
|
const MACROBLOCK * const prevMBs, |
124 |
VECTOR * const currMV, |
VECTOR * const currMV, |
125 |
VECTOR * const currPMV); |
VECTOR * const currPMV); |
126 |
|
|
127 |
int32_t PMVfastSearch16( |
int32_t EPZSSearch8( |
128 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
129 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
130 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
131 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
132 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
133 |
const int x, const int y, |
const int x, const int y, |
134 |
const uint32_t iQuality, |
const int start_x, const int start_y, |
135 |
MBParam * const pParam, |
const uint32_t MotionFlags, |
136 |
MACROBLOCK * const pMBs, |
const uint32_t iQuant, |
137 |
|
const uint32_t iFcode, |
138 |
|
const MBParam * const pParam, |
139 |
|
const MACROBLOCK * const pMBs, |
140 |
|
const MACROBLOCK * const prevMBs, |
141 |
VECTOR * const currMV, |
VECTOR * const currMV, |
142 |
VECTOR * const currPMV); |
VECTOR * const currPMV); |
143 |
|
|
144 |
|
|
145 |
|
typedef int32_t (MainSearch16Func)( |
146 |
|
const uint8_t * const pRef, |
147 |
|
const uint8_t * const pRefH, |
148 |
|
const uint8_t * const pRefV, |
149 |
|
const uint8_t * const pRefHV, |
150 |
|
const uint8_t * const cur, |
151 |
|
const int x, const int y, |
152 |
|
int32_t startx, int32_t starty, |
153 |
|
int32_t iMinSAD, |
154 |
|
VECTOR * const currMV, |
155 |
|
const VECTOR * const pmv, |
156 |
|
const int32_t min_dx, const int32_t max_dx, |
157 |
|
const int32_t min_dy, const int32_t max_dy, |
158 |
|
const int32_t iEdgedWidth, |
159 |
|
const int32_t iDiamondSize, |
160 |
|
const int32_t iFcode, |
161 |
|
const int32_t iQuant, |
162 |
|
int iFound); |
163 |
|
|
164 |
/* |
typedef MainSearch16Func* MainSearch16FuncPtr; |
|
* diamond search stuff |
|
|
* keep the the sequence in circular order (so optimization works) |
|
|
*/ |
|
165 |
|
|
|
typedef struct |
|
|
{ |
|
|
int32_t dx; |
|
|
int32_t dy; |
|
|
} |
|
|
DPOINT; |
|
166 |
|
|
167 |
|
typedef int32_t (MainSearch8Func)( |
168 |
|
const uint8_t * const pRef, |
169 |
|
const uint8_t * const pRefH, |
170 |
|
const uint8_t * const pRefV, |
171 |
|
const uint8_t * const pRefHV, |
172 |
|
const uint8_t * const cur, |
173 |
|
const int x, const int y, |
174 |
|
int32_t startx, int32_t starty, |
175 |
|
int32_t iMinSAD, |
176 |
|
VECTOR * const currMV, |
177 |
|
const VECTOR * const pmv, |
178 |
|
const int32_t min_dx, const int32_t max_dx, |
179 |
|
const int32_t min_dy, const int32_t max_dy, |
180 |
|
const int32_t iEdgedWidth, |
181 |
|
const int32_t iDiamondSize, |
182 |
|
const int32_t iFcode, |
183 |
|
const int32_t iQuant, |
184 |
|
int iFound); |
185 |
|
|
186 |
static const DPOINT diamond_small[4] = |
typedef MainSearch8Func* MainSearch8FuncPtr; |
|
{ |
|
|
{0, 1}, {1, 0}, {0, -1}, {-1, 0} |
|
|
}; |
|
187 |
|
|
188 |
|
static int32_t lambda_vec16[32] = /* rounded values for lambda param for weight of motion bits as in modified H.26L */ |
189 |
|
{ 0 ,(int)(1.00235+0.5), (int)(1.15582+0.5), (int)(1.31976+0.5), (int)(1.49591+0.5), (int)(1.68601+0.5), |
190 |
|
(int)(1.89187+0.5), (int)(2.11542+0.5), (int)(2.35878+0.5), (int)(2.62429+0.5), (int)(2.91455+0.5), |
191 |
|
(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), |
192 |
|
(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), |
193 |
|
(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), |
194 |
|
(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), |
195 |
|
(int)(36.4949+0.5) }; |
196 |
|
|
197 |
|
static int32_t *lambda_vec8 = lambda_vec16; /* same table for INTER and INTER4V for now*/ |
198 |
|
|
|
static const DPOINT diamond_large[8] = |
|
|
{ |
|
|
{0, 2}, {1, 1}, {2, 0}, {1, -1}, {0, -2}, {-1, -1}, {-2, 0}, {-1, 1} |
|
|
}; |
|
199 |
|
|
200 |
|
|
201 |
// mv.length table |
// mv.length table |
233 |
} |
} |
234 |
|
|
235 |
|
|
236 |
static __inline uint32_t calc_delta_16(const int32_t dx, const int32_t dy, const uint32_t iFcode) |
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 * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); |
|
|
} |
|
|
|
|
|
static __inline uint32_t calc_delta_8(const int32_t dx, const int32_t dy, const uint32_t iFcode) |
|
|
|
|
237 |
{ |
{ |
238 |
return NEIGH_TEND_8X8 * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); |
return NEIGH_TEND_16X16 * lambda_vec16[iQuant] * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); |
239 |
} |
} |
240 |
|
|
241 |
|
static __inline uint32_t calc_delta_8(const int32_t dx, const int32_t dy, const uint32_t iFcode, const uint32_t iQuant) |
242 |
|
|
|
|
|
|
|
|
|
/* calculate the min/max range (in halfpixels) |
|
|
relative to the _MACROBLOCK_ position |
|
|
*/ |
|
|
|
|
|
static void __inline get_range( |
|
|
int32_t * const min_dx, int32_t * const max_dx, |
|
|
int32_t * const min_dy, int32_t * const max_dy, |
|
|
const uint32_t x, const uint32_t y, |
|
|
const uint32_t block_sz, // block dimension, 8 or 16 |
|
|
const uint32_t width, const uint32_t height, |
|
|
const uint32_t fcode) |
|
243 |
{ |
{ |
244 |
|
return NEIGH_TEND_8X8 * lambda_vec8[iQuant] * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); |
|
const int search_range = 32 << (fcode - 1); |
|
|
const int high = search_range - 1; |
|
|
const int low = -search_range; |
|
|
|
|
|
// convert full-pixel measurements to half pixel |
|
|
const int hp_width = 2 * width; |
|
|
const int hp_height = 2 * height; |
|
|
const int hp_edge = 2 * block_sz; |
|
|
const int hp_x = 2 * (x) * block_sz; // we need _right end_ of block, not x-coordinate |
|
|
const int hp_y = 2 * (y) * block_sz; // same for _bottom end_ |
|
|
|
|
|
*max_dx = MIN(high, hp_width - hp_x); |
|
|
*max_dy = MIN(high, hp_height - hp_y); |
|
|
*min_dx = MAX(low, -(hp_edge + hp_x)); |
|
|
*min_dy = MAX(low, -(hp_edge + hp_y)); |
|
|
|
|
|
} |
|
|
|
|
|
|
|
|
/* |
|
|
* getref: calculate reference image pointer |
|
|
* the decision to use interpolation h/v/hv or the normal image is |
|
|
* based on dx & dy. |
|
|
*/ |
|
|
|
|
|
static __inline const uint8_t * get_ref( |
|
|
const uint8_t * const refn, |
|
|
const uint8_t * const refh, |
|
|
const uint8_t * const refv, |
|
|
const uint8_t * const refhv, |
|
|
const uint32_t x, const uint32_t y, |
|
|
const uint32_t block, // block dimension, 8 or 16 |
|
|
const int32_t dx, const int32_t dy, |
|
|
const uint32_t stride) |
|
|
{ |
|
|
|
|
|
switch ( ((dx&1)<<1) + (dy&1) ) // ((dx%2)?2:0)+((dy%2)?1:0) |
|
|
{ |
|
|
case 0 : return refn + (x*block+dx/2) + (y*block+dy/2)*stride; |
|
|
case 1 : return refv + (x*block+dx/2) + (y*block+(dy-1)/2)*stride; |
|
|
case 2 : return refh + (x*block+(dx-1)/2) + (y*block+dy/2)*stride; |
|
|
default : |
|
|
case 3 : return refhv + (x*block+(dx-1)/2) + (y*block+(dy-1)/2)*stride; |
|
|
} |
|
|
|
|
245 |
} |
} |
246 |
|
|
247 |
|
|
|
/* This is somehow a copy of get_ref, but with MV instead of X,Y */ |
|
|
|
|
|
static __inline const uint8_t * get_ref_mv( |
|
|
const uint8_t * const refn, |
|
|
const uint8_t * const refh, |
|
|
const uint8_t * const refv, |
|
|
const uint8_t * const refhv, |
|
|
const uint32_t x, const uint32_t y, |
|
|
const uint32_t block, // block dimension, 8 or 16 |
|
|
const VECTOR* mv, // measured in half-pel! |
|
|
const uint32_t stride) |
|
|
{ |
|
248 |
|
|
|
switch ( (((mv->x)&1)<<1) + ((mv->y)&1) ) |
|
|
{ |
|
|
case 0 : return refn + (x*block+(mv->x)/2) + (y*block+(mv->y)/2)*stride; |
|
|
case 1 : return refv + (x*block+(mv->x)/2) + (y*block+((mv->y)-1)/2)*stride; |
|
|
case 2 : return refh + (x*block+((mv->x)-1)/2) + (y*block+(mv->y)/2)*stride; |
|
|
default : |
|
|
case 3 : return refhv + (x*block+((mv->x)-1)/2) + (y*block+((mv->y)-1)/2)*stride; |
|
|
} |
|
249 |
|
|
|
} |
|
250 |
|
|
251 |
#ifndef SEARCH16 |
#ifndef SEARCH16 |
252 |
#define SEARCH16 PMVfastSearch16 |
#define SEARCH16 PMVfastSearch16 |
253 |
|
//#define SEARCH16 FullSearch16 |
254 |
|
//#define SEARCH16 EPZSSearch16 |
255 |
#endif |
#endif |
256 |
|
|
257 |
#ifndef SEARCH8 |
#ifndef SEARCH8 |
258 |
#define SEARCH8 PMVfastSearch8 |
#define SEARCH8 PMVfastSearch8 |
259 |
|
//#define SEARCH8 EPZSSearch8 |
260 |
#endif |
#endif |
261 |
|
|
262 |
bool MotionEstimation( |
bool MotionEstimation( |
|
MACROBLOCK * const pMBs, |
|
263 |
MBParam * const pParam, |
MBParam * const pParam, |
264 |
const IMAGE * const pRef, |
FRAMEINFO * const current, |
265 |
|
FRAMEINFO * const reference, |
266 |
const IMAGE * const pRefH, |
const IMAGE * const pRefH, |
267 |
const IMAGE * const pRefV, |
const IMAGE * const pRefV, |
268 |
const IMAGE * const pRefHV, |
const IMAGE * const pRefHV, |
|
IMAGE * const pCurrent, |
|
269 |
const uint32_t iLimit) |
const uint32_t iLimit) |
270 |
|
|
271 |
{ |
{ |
272 |
const uint32_t iWcount = pParam->mb_width; |
const uint32_t iWcount = pParam->mb_width; |
273 |
const uint32_t iHcount = pParam->mb_height; |
const uint32_t iHcount = pParam->mb_height; |
274 |
|
MACROBLOCK * pMBs = current->mbs; |
275 |
|
IMAGE * pCurrent = ¤t->image; |
276 |
|
|
277 |
|
MACROBLOCK * prevMBs = reference->mbs; // previous frame |
278 |
|
IMAGE * pRef = &reference->image; |
279 |
|
|
280 |
|
|
281 |
uint32_t i, j, iIntra = 0; |
uint32_t i, j, iIntra = 0; |
282 |
|
|
287 |
int32_t sad16; |
int32_t sad16; |
288 |
int32_t deviation; |
int32_t deviation; |
289 |
|
|
290 |
|
if (sadInit) |
291 |
|
(*sadInit)(); |
292 |
|
|
293 |
// note: i==horizontal, j==vertical |
// note: i==horizontal, j==vertical |
294 |
for (i = 0; i < iHcount; i++) |
for (i = 0; i < iHcount; i++) |
295 |
for (j = 0; j < iWcount; j++) |
for (j = 0; j < iWcount; j++) |
296 |
{ |
{ |
297 |
MACROBLOCK *pMB = &pMBs[j + i * iWcount]; |
MACROBLOCK *pMB = &pMBs[j + i * iWcount]; |
298 |
|
MACROBLOCK *prevMB = &prevMBs[j + i * iWcount]; |
299 |
|
|
300 |
sad16 = SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
sad16 = SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
301 |
j, i, pParam->motion_flags, |
j, i, current->motion_flags, current->quant, current->fcode, |
302 |
pParam, pMBs, &mv16, &pmv16); |
pParam, pMBs, prevMBs, &mv16, &pmv16); |
303 |
pMB->sad16=sad16; |
pMB->sad16=sad16; |
304 |
|
|
305 |
|
|
315 |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
316 |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; |
317 |
|
|
318 |
|
pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = 0; |
319 |
|
|
320 |
iIntra++; |
iIntra++; |
321 |
if(iIntra >= iLimit) |
if(iIntra >= iLimit) |
322 |
return 1; |
return 1; |
324 |
continue; |
continue; |
325 |
} |
} |
326 |
|
|
327 |
if (pParam->global_flags & XVID_INTER4V) |
if (current->global_flags & XVID_INTER4V) |
328 |
{ |
{ |
329 |
pMB->sad8[0] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
pMB->sad8[0] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
330 |
2 * j, 2 * i, mv16.x, mv16.y, pParam->motion_flags, |
2 * j, 2 * i, mv16.x, mv16.y, |
331 |
pParam, pMBs, &pMB->mvs[0], &pMB->pmvs[0]); |
current->motion_flags, current->quant, current->fcode, |
332 |
|
pParam, pMBs, prevMBs, &pMB->mvs[0], &pMB->pmvs[0]); |
333 |
|
|
334 |
pMB->sad8[1] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
pMB->sad8[1] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
335 |
2 * j + 1, 2 * i, mv16.x, mv16.y, pParam->motion_flags, |
2 * j + 1, 2 * i, mv16.x, mv16.y, |
336 |
pParam, pMBs, &pMB->mvs[1], &pMB->pmvs[1]); |
current->motion_flags, current->quant, current->fcode, |
337 |
|
pParam, pMBs, prevMBs, &pMB->mvs[1], &pMB->pmvs[1]); |
338 |
|
|
339 |
pMB->sad8[2] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
pMB->sad8[2] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
340 |
2 * j, 2 * i + 1, mv16.x, mv16.y, pParam->motion_flags, |
2 * j, 2 * i + 1, mv16.x, mv16.y, |
341 |
pParam, pMBs, &pMB->mvs[2], &pMB->pmvs[2]); |
current->motion_flags, current->quant, current->fcode, |
342 |
|
pParam, pMBs, prevMBs, &pMB->mvs[2], &pMB->pmvs[2]); |
343 |
|
|
344 |
pMB->sad8[3] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
pMB->sad8[3] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
345 |
2 * j + 1, 2 * i + 1, mv16.x, mv16.y, pParam->motion_flags, |
2 * j + 1, 2 * i + 1, mv16.x, mv16.y, |
346 |
pParam, pMBs, &pMB->mvs[3], &pMB->pmvs[3]); |
current->motion_flags, current->quant, current->fcode, |
347 |
|
pParam, pMBs, prevMBs, &pMB->mvs[3], &pMB->pmvs[3]); |
348 |
|
|
349 |
sad8 = pMB->sad8[0] + pMB->sad8[1] + pMB->sad8[2] + pMB->sad8[3]; |
sad8 = pMB->sad8[0] + pMB->sad8[1] + pMB->sad8[2] + pMB->sad8[3]; |
350 |
} |
} |
354 |
mpeg4: if (sad8 < sad16 - nb/2+1) use_inter4v |
mpeg4: if (sad8 < sad16 - nb/2+1) use_inter4v |
355 |
*/ |
*/ |
356 |
|
|
357 |
if (pMB->dquant == NO_CHANGE) { |
if (!(current->global_flags & XVID_LUMIMASKING) || pMB->dquant == NO_CHANGE) |
358 |
if (((pParam->global_flags & XVID_INTER4V)==0) || |
{ |
359 |
(sad16 < (sad8 + (int32_t)(IMV16X16 * pParam->quant)))) { |
if (((current->global_flags & XVID_INTER4V)==0) || |
360 |
|
(sad16 < (sad8 + (int32_t)(IMV16X16 * current->quant)))) |
361 |
|
{ |
362 |
|
|
363 |
sad8 = sad16; |
sad8 = sad16; |
364 |
pMB->mode = MODE_INTER; |
pMB->mode = MODE_INTER; |
365 |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; |
366 |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; |
367 |
|
pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad16; |
368 |
pMB->pmvs[0].x = pmv16.x; |
pMB->pmvs[0].x = pmv16.x; |
369 |
pMB->pmvs[0].y = pmv16.y; |
pMB->pmvs[0].y = pmv16.y; |
370 |
} |
} |
371 |
else |
else |
372 |
|
{ |
373 |
pMB->mode = MODE_INTER4V; |
pMB->mode = MODE_INTER4V; |
374 |
|
pMB->sad8[0] *= 4; |
375 |
|
pMB->sad8[1] *= 4; |
376 |
|
pMB->sad8[2] *= 4; |
377 |
|
pMB->sad8[3] *= 4; |
378 |
|
} |
379 |
} |
} |
380 |
else |
else |
381 |
{ |
{ |
383 |
pMB->mode = MODE_INTER; |
pMB->mode = MODE_INTER; |
384 |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; |
385 |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; |
386 |
|
pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad16; |
387 |
|
|
388 |
pMB->pmvs[0].x = pmv16.x; |
pMB->pmvs[0].x = pmv16.x; |
389 |
pMB->pmvs[0].y = pmv16.y; |
pMB->pmvs[0].y = pmv16.y; |
390 |
} |
} |
403 |
&& (0 <= max_dy) && (0 >= min_dy) ) \ |
&& (0 <= max_dy) && (0 >= min_dy) ) \ |
404 |
{ \ |
{ \ |
405 |
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0, 0 , iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); \ |
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0, 0 , iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); \ |
406 |
iSAD += calc_delta_16(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
iSAD += calc_delta_16(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode, iQuant);\ |
|
if (iSAD <= iQuant * 96) \ |
|
|
iSAD -= MV16_00_BIAS; \ |
|
407 |
if (iSAD < iMinSAD) \ |
if (iSAD < iMinSAD) \ |
408 |
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } } \ |
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } } \ |
409 |
} |
} |
410 |
|
|
411 |
|
#define NOCHECK_MV16_CANDIDATE(X,Y) { \ |
412 |
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
413 |
|
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode, iQuant);\ |
414 |
|
if (iSAD < iMinSAD) \ |
415 |
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } \ |
416 |
|
} |
417 |
|
|
418 |
#define CHECK_MV16_CANDIDATE(X,Y) { \ |
#define CHECK_MV16_CANDIDATE(X,Y) { \ |
419 |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
420 |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
421 |
{ \ |
{ \ |
422 |
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
423 |
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode, iQuant);\ |
424 |
if (iSAD < iMinSAD) \ |
if (iSAD < iMinSAD) \ |
425 |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ |
426 |
} |
} |
430 |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
431 |
{ \ |
{ \ |
432 |
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
433 |
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode, iQuant);\ |
434 |
if (iSAD < iMinSAD) \ |
if (iSAD < iMinSAD) \ |
435 |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ |
436 |
} |
} |
440 |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
441 |
{ \ |
{ \ |
442 |
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
443 |
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode, iQuant);\ |
444 |
if (iSAD < iMinSAD) \ |
if (iSAD < iMinSAD) \ |
445 |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
446 |
} |
} |
448 |
|
|
449 |
#define CHECK_MV8_ZERO {\ |
#define CHECK_MV8_ZERO {\ |
450 |
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, 0, 0 , iEdgedWidth), iEdgedWidth); \ |
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, 0, 0 , iEdgedWidth), iEdgedWidth); \ |
451 |
iSAD += calc_delta_8(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
iSAD += calc_delta_8(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode, iQuant);\ |
452 |
if (iSAD < iMinSAD) \ |
if (iSAD < iMinSAD) \ |
453 |
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } \ |
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } \ |
454 |
} |
} |
455 |
|
|
456 |
|
#define NOCHECK_MV8_CANDIDATE(X,Y) \ |
457 |
|
{ \ |
458 |
|
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
459 |
|
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode, iQuant);\ |
460 |
|
if (iSAD < iMinSAD) \ |
461 |
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } \ |
462 |
|
} |
463 |
|
|
464 |
#define CHECK_MV8_CANDIDATE(X,Y) { \ |
#define CHECK_MV8_CANDIDATE(X,Y) { \ |
465 |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
466 |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
467 |
{ \ |
{ \ |
468 |
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
469 |
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode, iQuant);\ |
470 |
if (iSAD < iMinSAD) \ |
if (iSAD < iMinSAD) \ |
471 |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ |
472 |
} |
} |
476 |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
477 |
{ \ |
{ \ |
478 |
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
479 |
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode, iQuant);\ |
480 |
if (iSAD < iMinSAD) \ |
if (iSAD < iMinSAD) \ |
481 |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ |
482 |
} |
} |
486 |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
487 |
{ \ |
{ \ |
488 |
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
489 |
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode, iQuant);\ |
490 |
if (iSAD < iMinSAD) \ |
if (iSAD < iMinSAD) \ |
491 |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
492 |
} |
} |
501 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
502 |
const int x, const int y, |
const int x, const int y, |
503 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
504 |
|
const uint32_t iQuant, |
505 |
|
const uint32_t iFcode, |
506 |
MBParam * const pParam, |
MBParam * const pParam, |
507 |
MACROBLOCK * const pMBs, |
const MACROBLOCK * const pMBs, |
508 |
|
const MACROBLOCK * const prevMBs, |
509 |
VECTOR * const currMV, |
VECTOR * const currMV, |
510 |
VECTOR * const currPMV) |
VECTOR * const currPMV) |
511 |
{ |
{ |
512 |
const int32_t iEdgedWidth = pParam->edged_width; |
const int32_t iEdgedWidth = pParam->edged_width; |
|
const int32_t iQuant = pParam->quant; |
|
513 |
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
514 |
int32_t iSAD; |
int32_t iSAD; |
515 |
int32_t pred_x,pred_y; |
int32_t pred_x,pred_y; |
532 |
} |
} |
533 |
*/ |
*/ |
534 |
|
|
535 |
int32_t PMVfastSearch16_MainSearch( |
int32_t Diamond16_MainSearch( |
536 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
537 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
538 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
589 |
return iMinSAD; |
return iMinSAD; |
590 |
} |
} |
591 |
|
|
592 |
int32_t PMVfastSearch16_Refine( |
int32_t Square16_MainSearch( |
593 |
|
const uint8_t * const pRef, |
594 |
|
const uint8_t * const pRefH, |
595 |
|
const uint8_t * const pRefV, |
596 |
|
const uint8_t * const pRefHV, |
597 |
|
const uint8_t * const cur, |
598 |
|
const int x, const int y, |
599 |
|
int32_t startx, int32_t starty, |
600 |
|
int32_t iMinSAD, |
601 |
|
VECTOR * const currMV, |
602 |
|
const VECTOR * const pmv, |
603 |
|
const int32_t min_dx, const int32_t max_dx, |
604 |
|
const int32_t min_dy, const int32_t max_dy, |
605 |
|
const int32_t iEdgedWidth, |
606 |
|
const int32_t iDiamondSize, |
607 |
|
const int32_t iFcode, |
608 |
|
const int32_t iQuant, |
609 |
|
int iFound) |
610 |
|
{ |
611 |
|
/* Do a square search around given starting point, return SAD of best */ |
612 |
|
|
613 |
|
int32_t iDirection=0; |
614 |
|
int32_t iSAD; |
615 |
|
VECTOR backupMV; |
616 |
|
backupMV.x = startx; |
617 |
|
backupMV.y = starty; |
618 |
|
|
619 |
|
/* It's one search with full square pattern, and new parts for all following diamonds */ |
620 |
|
|
621 |
|
/* new direction are extra, so 1-4 is normal diamond |
622 |
|
537 |
623 |
|
1*2 |
624 |
|
648 |
625 |
|
*/ |
626 |
|
|
627 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
628 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
629 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
630 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
631 |
|
|
632 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
633 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
634 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
635 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
636 |
|
|
637 |
|
|
638 |
|
if (iDirection) |
639 |
|
while (!iFound) |
640 |
|
{ |
641 |
|
iFound = 1; |
642 |
|
backupMV=*currMV; |
643 |
|
|
644 |
|
switch (iDirection) |
645 |
|
{ |
646 |
|
case 1: |
647 |
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
648 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
649 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
650 |
|
break; |
651 |
|
case 2: |
652 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
653 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
654 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
655 |
|
break; |
656 |
|
|
657 |
|
case 3: |
658 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
659 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
660 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
661 |
|
break; |
662 |
|
|
663 |
|
case 4: |
664 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
665 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
666 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
667 |
|
break; |
668 |
|
|
669 |
|
case 5: |
670 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
671 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
672 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
673 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
674 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
675 |
|
break; |
676 |
|
|
677 |
|
case 6: |
678 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
679 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
680 |
|
|
681 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
682 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
683 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
684 |
|
|
685 |
|
break; |
686 |
|
|
687 |
|
case 7: |
688 |
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
689 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
690 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
691 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
692 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
693 |
|
break; |
694 |
|
|
695 |
|
case 8: |
696 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
697 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
698 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
699 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
700 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
701 |
|
break; |
702 |
|
default: |
703 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
704 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
705 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
706 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
707 |
|
|
708 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
709 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
710 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
711 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
712 |
|
break; |
713 |
|
} |
714 |
|
} |
715 |
|
else |
716 |
|
{ |
717 |
|
currMV->x = startx; |
718 |
|
currMV->y = starty; |
719 |
|
} |
720 |
|
return iMinSAD; |
721 |
|
} |
722 |
|
|
723 |
|
|
724 |
|
int32_t Full16_MainSearch( |
725 |
|
const uint8_t * const pRef, |
726 |
|
const uint8_t * const pRefH, |
727 |
|
const uint8_t * const pRefV, |
728 |
|
const uint8_t * const pRefHV, |
729 |
|
const uint8_t * const cur, |
730 |
|
const int x, const int y, |
731 |
|
int32_t startx, int32_t starty, |
732 |
|
int32_t iMinSAD, |
733 |
|
VECTOR * const currMV, |
734 |
|
const VECTOR * const pmv, |
735 |
|
const int32_t min_dx, const int32_t max_dx, |
736 |
|
const int32_t min_dy, const int32_t max_dy, |
737 |
|
const int32_t iEdgedWidth, |
738 |
|
const int32_t iDiamondSize, |
739 |
|
const int32_t iFcode, |
740 |
|
const int32_t iQuant, |
741 |
|
int iFound) |
742 |
|
{ |
743 |
|
int32_t iSAD; |
744 |
|
int32_t dx,dy; |
745 |
|
VECTOR backupMV; |
746 |
|
backupMV.x = startx; |
747 |
|
backupMV.y = starty; |
748 |
|
|
749 |
|
for (dx = min_dx; dx<=max_dx; dx+=iDiamondSize) |
750 |
|
for (dy = min_dy; dy<= max_dy; dy+=iDiamondSize) |
751 |
|
NOCHECK_MV16_CANDIDATE(dx,dy); |
752 |
|
|
753 |
|
return iMinSAD; |
754 |
|
} |
755 |
|
|
756 |
|
int32_t Full8_MainSearch( |
757 |
|
const uint8_t * const pRef, |
758 |
|
const uint8_t * const pRefH, |
759 |
|
const uint8_t * const pRefV, |
760 |
|
const uint8_t * const pRefHV, |
761 |
|
const uint8_t * const cur, |
762 |
|
const int x, const int y, |
763 |
|
int32_t startx, int32_t starty, |
764 |
|
int32_t iMinSAD, |
765 |
|
VECTOR * const currMV, |
766 |
|
const VECTOR * const pmv, |
767 |
|
const int32_t min_dx, const int32_t max_dx, |
768 |
|
const int32_t min_dy, const int32_t max_dy, |
769 |
|
const int32_t iEdgedWidth, |
770 |
|
const int32_t iDiamondSize, |
771 |
|
const int32_t iFcode, |
772 |
|
const int32_t iQuant, |
773 |
|
int iFound) |
774 |
|
{ |
775 |
|
int32_t iSAD; |
776 |
|
int32_t dx,dy; |
777 |
|
VECTOR backupMV; |
778 |
|
backupMV.x = startx; |
779 |
|
backupMV.y = starty; |
780 |
|
|
781 |
|
for (dx = min_dx; dx<=max_dx; dx+=iDiamondSize) |
782 |
|
for (dy = min_dy; dy<= max_dy; dy+=iDiamondSize) |
783 |
|
NOCHECK_MV8_CANDIDATE(dx,dy); |
784 |
|
|
785 |
|
return iMinSAD; |
786 |
|
} |
787 |
|
|
788 |
|
|
789 |
|
|
790 |
|
int32_t Halfpel16_Refine( |
791 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
792 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
793 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
822 |
|
|
823 |
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
824 |
|
|
825 |
|
|
826 |
int32_t PMVfastSearch16( |
int32_t PMVfastSearch16( |
827 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
828 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
831 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
832 |
const int x, const int y, |
const int x, const int y, |
833 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
834 |
MBParam * const pParam, |
const uint32_t iQuant, |
835 |
MACROBLOCK * const pMBs, |
const uint32_t iFcode, |
836 |
|
const MBParam * const pParam, |
837 |
|
const MACROBLOCK * const pMBs, |
838 |
|
const MACROBLOCK * const prevMBs, |
839 |
VECTOR * const currMV, |
VECTOR * const currMV, |
840 |
VECTOR * const currPMV) |
VECTOR * const currPMV) |
841 |
{ |
{ |
842 |
const uint32_t iWcount = pParam->mb_width; |
const uint32_t iWcount = pParam->mb_width; |
|
const int32_t iFcode = pParam->fixed_code; |
|
|
const int32_t iQuant = pParam->quant; |
|
843 |
const int32_t iWidth = pParam->width; |
const int32_t iWidth = pParam->width; |
844 |
const int32_t iHeight = pParam->height; |
const int32_t iHeight = pParam->height; |
845 |
const int32_t iEdgedWidth = pParam->edged_width; |
const int32_t iEdgedWidth = pParam->edged_width; |
861 |
VECTOR pmv[4]; |
VECTOR pmv[4]; |
862 |
int32_t psad[4]; |
int32_t psad[4]; |
863 |
|
|
864 |
MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
865 |
|
const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; |
866 |
|
|
867 |
static int32_t threshA,threshB; |
static int32_t threshA,threshB; |
868 |
int32_t bPredEq; |
int32_t bPredEq; |
901 |
|
|
902 |
iFound=0; |
iFound=0; |
903 |
|
|
904 |
/* Step 2: Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
/* Step 4: Calculate SAD around the Median prediction. |
905 |
vector of the median. |
MinSAD=SAD |
906 |
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
If Motion Vector equal to Previous frame motion vector |
907 |
*/ |
and MinSAD<PrevFrmSAD goto Step 10. |
908 |
|
If SAD<=256 goto Step 10. |
|
if ((bPredEq) && (MVequal(pmv[0],pMB->mvs[0]) ) ) |
|
|
iFound=2; |
|
|
|
|
|
/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
|
|
Otherwise select large Diamond Search. |
|
909 |
*/ |
*/ |
910 |
|
|
|
if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536) || (bPredEq) ) |
|
|
iDiamondSize=1; // halfpel! |
|
|
else |
|
|
iDiamondSize=2; // halfpel! |
|
|
|
|
|
if (!(MotionFlags & PMV_HALFPELDIAMOND16) ) |
|
|
iDiamondSize*=2; |
|
|
|
|
|
/* Step 4: Calculate SAD around the Median prediction. |
|
|
MinSAD=SAD |
|
|
If Motion Vector equal to Previous frame motion vector |
|
|
and MinSAD<PrevFrmSAD goto Step 10. |
|
|
If SAD<=256 goto Step 10. |
|
|
*/ |
|
|
|
|
|
|
|
|
// Prepare for main loop |
|
|
|
|
911 |
*currMV=pmv[0]; /* current best := prediction */ |
*currMV=pmv[0]; /* current best := prediction */ |
912 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
if (!(MotionFlags & PMV_HALFPEL16 )) |
913 |
{ /* This should NOT be necessary! */ |
{ /* This should NOT be necessary! */ |
935 |
iMinSAD = sad16( cur, |
iMinSAD = sad16( cur, |
936 |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
937 |
iEdgedWidth, MV_MAX_ERROR); |
iEdgedWidth, MV_MAX_ERROR); |
938 |
iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; |
iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode, iQuant); |
939 |
|
|
940 |
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,pMB->mvs[0])) && (iMinSAD < pMB->sad16) ) ) |
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,prevMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
941 |
|
{ |
942 |
|
if (iMinSAD < 2*iQuant) // high chances for SKIP-mode |
943 |
{ |
{ |
944 |
|
if (!MVzero(*currMV)) |
945 |
|
{ |
946 |
|
iMinSAD += MV16_00_BIAS; |
947 |
|
CHECK_MV16_ZERO; // (0,0) saves space for letterboxed pictures |
948 |
|
iMinSAD -= MV16_00_BIAS; |
949 |
|
} |
950 |
|
} |
951 |
|
|
952 |
if (MotionFlags & PMV_QUICKSTOP16) |
if (MotionFlags & PMV_QUICKSTOP16) |
953 |
goto step10b; |
goto PMVfast16_Terminate_without_Refine; |
954 |
if (MotionFlags & PMV_EARLYSTOP16) |
if (MotionFlags & PMV_EARLYSTOP16) |
955 |
goto step10; |
goto PMVfast16_Terminate_with_Refine; |
956 |
} |
} |
957 |
|
|
958 |
|
|
959 |
|
/* Step 2 (lazy eval): Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
960 |
|
vector of the median. |
961 |
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
962 |
|
*/ |
963 |
|
|
964 |
|
if ((bPredEq) && (MVequal(pmv[0],prevMB->mvs[0]) ) ) |
965 |
|
iFound=2; |
966 |
|
|
967 |
|
/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
968 |
|
Otherwise select large Diamond Search. |
969 |
|
*/ |
970 |
|
|
971 |
|
if ( (!MVzero(pmv[0])) || (threshB<1536) || (bPredEq) ) |
972 |
|
iDiamondSize=1; // halfpel! |
973 |
|
else |
974 |
|
iDiamondSize=2; // halfpel! |
975 |
|
|
976 |
|
if (!(MotionFlags & PMV_HALFPELDIAMOND16) ) |
977 |
|
iDiamondSize*=2; |
978 |
|
|
979 |
/* |
/* |
980 |
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
981 |
Also calculate (0,0) but do not subtract offset. |
Also calculate (0,0) but do not subtract offset. |
982 |
Let MinSAD be the smallest SAD up to this point. |
Let MinSAD be the smallest SAD up to this point. |
983 |
If MV is (0,0) subtract offset. ******** WHAT'S THIS 'OFFSET' ??? *********** |
If MV is (0,0) subtract offset. |
984 |
*/ |
*/ |
985 |
|
|
986 |
// (0,0) is always possible |
// (0,0) is always possible |
987 |
|
|
988 |
|
if (!MVzero(pmv[0])) |
989 |
CHECK_MV16_ZERO; |
CHECK_MV16_ZERO; |
990 |
|
|
991 |
// previous frame MV is always possible |
// previous frame MV is always possible |
992 |
CHECK_MV16_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); |
|
993 |
|
if (!MVzero(prevMB->mvs[0])) |
994 |
|
if (!MVequal(prevMB->mvs[0],pmv[0])) |
995 |
|
CHECK_MV16_CANDIDATE(prevMB->mvs[0].x,prevMB->mvs[0].y); |
996 |
|
|
997 |
// left neighbour, if allowed |
// left neighbour, if allowed |
998 |
if (x != 0) |
|
999 |
|
if (!MVzero(pmv[1])) |
1000 |
|
if (!MVequal(pmv[1],prevMB->mvs[0])) |
1001 |
|
if (!MVequal(pmv[1],pmv[0])) |
1002 |
{ |
{ |
1003 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
if (!(MotionFlags & PMV_HALFPEL16 )) |
1004 |
{ pmv[1].x = EVEN(pmv[1].x); |
{ pmv[1].x = EVEN(pmv[1].x); |
1005 |
pmv[1].y = EVEN(pmv[1].y); |
pmv[1].y = EVEN(pmv[1].y); |
1006 |
} |
} |
1007 |
|
|
1008 |
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
1009 |
} |
} |
1010 |
|
|
1011 |
// top neighbour, if allowed |
// top neighbour, if allowed |
1012 |
if (y != 0) |
if (!MVzero(pmv[2])) |
1013 |
|
if (!MVequal(pmv[2],prevMB->mvs[0])) |
1014 |
|
if (!MVequal(pmv[2],pmv[0])) |
1015 |
|
if (!MVequal(pmv[2],pmv[1])) |
1016 |
{ |
{ |
1017 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
if (!(MotionFlags & PMV_HALFPEL16 )) |
1018 |
{ pmv[2].x = EVEN(pmv[2].x); |
{ pmv[2].x = EVEN(pmv[2].x); |
1021 |
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
1022 |
|
|
1023 |
// top right neighbour, if allowed |
// top right neighbour, if allowed |
1024 |
if (x != (iWcount-1)) |
if (!MVzero(pmv[3])) |
1025 |
|
if (!MVequal(pmv[3],prevMB->mvs[0])) |
1026 |
|
if (!MVequal(pmv[3],pmv[0])) |
1027 |
|
if (!MVequal(pmv[3],pmv[1])) |
1028 |
|
if (!MVequal(pmv[3],pmv[2])) |
1029 |
{ |
{ |
1030 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
if (!(MotionFlags & PMV_HALFPEL16 )) |
1031 |
{ pmv[3].x = EVEN(pmv[3].x); |
{ pmv[3].x = EVEN(pmv[3].x); |
1035 |
} |
} |
1036 |
} |
} |
1037 |
|
|
1038 |
|
if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96)*/ ) |
1039 |
|
iMinSAD -= MV16_00_BIAS; |
1040 |
|
|
1041 |
|
|
1042 |
/* Step 6: If MinSAD <= thresa goto Step 10. |
/* Step 6: If MinSAD <= thresa goto Step 10. |
1043 |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
1044 |
*/ |
*/ |
1045 |
|
|
1046 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,pMB->mvs[0]) && (iMinSAD < pMB->sad16) ) ) |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,prevMB->mvs[0]) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
1047 |
{ |
{ |
1048 |
if (MotionFlags & PMV_QUICKSTOP16) |
if (MotionFlags & PMV_QUICKSTOP16) |
1049 |
goto step10b; |
goto PMVfast16_Terminate_without_Refine; |
1050 |
if (MotionFlags & PMV_EARLYSTOP16) |
if (MotionFlags & PMV_EARLYSTOP16) |
1051 |
goto step10; |
goto PMVfast16_Terminate_with_Refine; |
1052 |
} |
} |
1053 |
|
|
1054 |
|
|
1065 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
1066 |
|
|
1067 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
1068 |
iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1069 |
x, y, |
x, y, |
1070 |
currMV->x, currMV->y, iMinSAD, &newMV, |
currMV->x, currMV->y, iMinSAD, &newMV, |
1071 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1081 |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
1082 |
|
|
1083 |
if (!(MVequal(pmv[0],backupMV)) ) |
if (!(MVequal(pmv[0],backupMV)) ) |
1084 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1085 |
x, y, |
x, y, |
1086 |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
1087 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1094 |
} |
} |
1095 |
|
|
1096 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
1097 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1098 |
x, y, |
x, y, |
1099 |
0, 0, iMinSAD, &newMV, |
0, 0, iMinSAD, &newMV, |
1100 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1111 |
Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
1112 |
*/ |
*/ |
1113 |
|
|
1114 |
step10: |
PMVfast16_Terminate_with_Refine: |
1115 |
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
1116 |
iMinSAD = PMVfastSearch16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
iMinSAD = Halfpel16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
1117 |
x, y, |
x, y, |
1118 |
currMV, iMinSAD, |
currMV, iMinSAD, |
1119 |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
1120 |
|
|
1121 |
step10b: |
PMVfast16_Terminate_without_Refine: |
1122 |
currPMV->x = currMV->x - pmv[0].x; |
currPMV->x = currMV->x - pmv[0].x; |
1123 |
currPMV->y = currMV->y - pmv[0].y; |
currPMV->y = currMV->y - pmv[0].y; |
1124 |
return iMinSAD; |
return iMinSAD; |
1129 |
|
|
1130 |
|
|
1131 |
|
|
1132 |
int32_t PMVfastSearch8_MainSearch( |
int32_t Diamond8_MainSearch( |
1133 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
1134 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1135 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
1186 |
return iMinSAD; |
return iMinSAD; |
1187 |
} |
} |
1188 |
|
|
1189 |
int32_t PMVfastSearch8_Refine( |
int32_t Halfpel8_Refine( |
1190 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
1191 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1192 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
1229 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
1230 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
1231 |
const int x, const int y, |
const int x, const int y, |
1232 |
const int start_x, int start_y, |
const int start_x, const int start_y, |
1233 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
1234 |
MBParam * const pParam, |
const uint32_t iQuant, |
1235 |
MACROBLOCK * const pMBs, |
const uint32_t iFcode, |
1236 |
|
const MBParam * const pParam, |
1237 |
|
const MACROBLOCK * const pMBs, |
1238 |
|
const MACROBLOCK * const prevMBs, |
1239 |
VECTOR * const currMV, |
VECTOR * const currMV, |
1240 |
VECTOR * const currPMV) |
VECTOR * const currPMV) |
1241 |
{ |
{ |
1242 |
const uint32_t iWcount = pParam->mb_width; |
const uint32_t iWcount = pParam->mb_width; |
|
|
|
|
const int32_t iFcode = pParam->fixed_code; |
|
|
const int32_t iQuant = pParam->quant; |
|
1243 |
const int32_t iWidth = pParam->width; |
const int32_t iWidth = pParam->width; |
1244 |
const int32_t iHeight = pParam->height; |
const int32_t iHeight = pParam->height; |
1245 |
const int32_t iEdgedWidth = pParam->edged_width; |
const int32_t iEdgedWidth = pParam->edged_width; |
1257 |
int32_t psad[4]; |
int32_t psad[4]; |
1258 |
VECTOR newMV; |
VECTOR newMV; |
1259 |
VECTOR backupMV; |
VECTOR backupMV; |
1260 |
|
VECTOR startMV; |
1261 |
|
|
1262 |
MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
1263 |
|
const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; |
1264 |
|
|
1265 |
static int32_t threshA,threshB; |
static int32_t threshA,threshB; |
1266 |
int32_t iFound,bPredEq; |
int32_t iFound,bPredEq; |
1267 |
int32_t iMinSAD,iSAD; |
int32_t iMinSAD,iSAD; |
1268 |
|
|
1269 |
int32_t iSubBlock = ((y&1)<<1) + (x&1); |
int32_t iSubBlock = (y&1)+(y&1) + (x&1); |
1270 |
|
|
1271 |
|
/* Init variables */ |
1272 |
|
startMV.x = start_x; |
1273 |
|
startMV.y = start_y; |
1274 |
|
|
1275 |
/* Get maximum range */ |
/* Get maximum range */ |
1276 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
1277 |
x, y, 8, iWidth, iHeight, iFcode); |
x, y, 8, iWidth, iHeight, iFcode); |
1278 |
|
|
|
/* we work with abs. MVs, not relative to prediction, so range is relative to 0,0 */ |
|
|
|
|
1279 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8 )) |
if (!(MotionFlags & PMV_HALFPELDIAMOND8 )) |
1280 |
{ min_dx = EVEN(min_dx); |
{ min_dx = EVEN(min_dx); |
1281 |
max_dx = EVEN(max_dx); |
max_dx = EVEN(max_dx); |
1303 |
|
|
1304 |
iFound=0; |
iFound=0; |
1305 |
|
|
|
/* Step 2: Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
|
|
vector of the median. |
|
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
|
|
*/ |
|
|
|
|
|
if ((bPredEq) && (MVequal(pmv[0],pMB->mvs[iSubBlock]) ) ) |
|
|
iFound=2; |
|
|
|
|
|
/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
|
|
Otherwise select large Diamond Search. |
|
|
*/ |
|
|
|
|
|
if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536/4) || (bPredEq) ) |
|
|
iDiamondSize=1; // 1 halfpel! |
|
|
else |
|
|
iDiamondSize=2; // 2 halfpel = 1 full pixel! |
|
|
|
|
|
if (!(MotionFlags & PMV_HALFPELDIAMOND8) ) |
|
|
iDiamondSize*=2; |
|
|
|
|
1306 |
/* Step 4: Calculate SAD around the Median prediction. |
/* Step 4: Calculate SAD around the Median prediction. |
1307 |
MinSAD=SAD |
MinSAD=SAD |
1308 |
If Motion Vector equal to Previous frame motion vector |
If Motion Vector equal to Previous frame motion vector |
1313 |
|
|
1314 |
// Prepare for main loop |
// Prepare for main loop |
1315 |
|
|
1316 |
currMV->x=start_x; /* start with mv16 */ |
*currMV = startMV; |
|
currMV->y=start_y; |
|
1317 |
|
|
1318 |
iMinSAD = sad8( cur, |
iMinSAD = sad8( cur, |
1319 |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
1320 |
iEdgedWidth); |
iEdgedWidth); |
1321 |
iMinSAD += calc_delta_8(currMV->x - pmv[0].x, currMV->y - pmv[0].y, (uint8_t)iFcode) * iQuant; |
iMinSAD += calc_delta_8(currMV->x - pmv[0].x, currMV->y - pmv[0].y, (uint8_t)iFcode, iQuant); |
1322 |
|
|
1323 |
if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,pMB->mvs[iSubBlock])) && (iMinSAD < pMB->sad8[iSubBlock]) ) ) |
if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,prevMB->mvs[iSubBlock])) |
1324 |
|
&& ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
1325 |
{ |
{ |
1326 |
if (MotionFlags & PMV_QUICKSTOP8) |
if (MotionFlags & PMV_QUICKSTOP16) |
1327 |
goto step10_8b; |
goto PMVfast8_Terminate_without_Refine; |
1328 |
if (MotionFlags & PMV_EARLYSTOP8) |
if (MotionFlags & PMV_EARLYSTOP16) |
1329 |
goto step10_8; |
goto PMVfast8_Terminate_with_Refine; |
1330 |
} |
} |
1331 |
|
|
1332 |
|
/* Step 2 (lazy eval): Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
1333 |
|
vector of the median. |
1334 |
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
1335 |
|
*/ |
1336 |
|
|
1337 |
|
if ((bPredEq) && (MVequal(pmv[0],prevMB->mvs[iSubBlock]) ) ) |
1338 |
|
iFound=2; |
1339 |
|
|
1340 |
|
/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
1341 |
|
Otherwise select large Diamond Search. |
1342 |
|
*/ |
1343 |
|
|
1344 |
|
if ( (!MVzero(pmv[0])) || (threshB<1536/4) || (bPredEq) ) |
1345 |
|
iDiamondSize=1; // 1 halfpel! |
1346 |
|
else |
1347 |
|
iDiamondSize=2; // 2 halfpel = 1 full pixel! |
1348 |
|
|
1349 |
|
if (!(MotionFlags & PMV_HALFPELDIAMOND8) ) |
1350 |
|
iDiamondSize*=2; |
1351 |
|
|
1352 |
|
|
1353 |
/* |
/* |
1354 |
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
1355 |
Also calculate (0,0) but do not subtract offset. |
Also calculate (0,0) but do not subtract offset. |
1356 |
Let MinSAD be the smallest SAD up to this point. |
Let MinSAD be the smallest SAD up to this point. |
1357 |
If MV is (0,0) subtract offset. ******** WHAT'S THIS 'OFFSET' ??? *********** |
If MV is (0,0) subtract offset. |
1358 |
*/ |
*/ |
1359 |
|
|
1360 |
// the prediction might be even better than mv16 |
// the median prediction might be even better than mv16 |
1361 |
|
|
1362 |
|
if (!MVequal(pmv[0],startMV)) |
1363 |
CHECK_MV8_CANDIDATE(pmv[0].x,pmv[0].y); |
CHECK_MV8_CANDIDATE(pmv[0].x,pmv[0].y); |
1364 |
|
|
1365 |
// (0,0) is always possible |
// (0,0) if needed |
1366 |
|
if (!MVzero(pmv[0])) |
1367 |
|
if (!MVzero(startMV)) |
1368 |
CHECK_MV8_ZERO; |
CHECK_MV8_ZERO; |
1369 |
|
|
1370 |
// previous frame MV is always possible |
// previous frame MV if needed |
1371 |
CHECK_MV8_CANDIDATE(pMB->mvs[iSubBlock].x,pMB->mvs[iSubBlock].y); |
if (!MVzero(prevMB->mvs[iSubBlock])) |
1372 |
|
if (!MVequal(prevMB->mvs[iSubBlock],startMV)) |
1373 |
|
if (!MVequal(prevMB->mvs[iSubBlock],pmv[0])) |
1374 |
|
CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x,prevMB->mvs[iSubBlock].y); |
1375 |
|
|
1376 |
// left neighbour, if allowed |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,prevMB->mvs[iSubBlock]) && ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
1377 |
if (psad[1] != MV_MAX_ERROR) |
{ |
1378 |
|
if (MotionFlags & PMV_QUICKSTOP16) |
1379 |
|
goto PMVfast8_Terminate_without_Refine; |
1380 |
|
if (MotionFlags & PMV_EARLYSTOP16) |
1381 |
|
goto PMVfast8_Terminate_with_Refine; |
1382 |
|
} |
1383 |
|
|
1384 |
|
|
1385 |
|
// left neighbour, if allowed and needed |
1386 |
|
if (!MVzero(pmv[1])) |
1387 |
|
if (!MVequal(pmv[1],startMV)) |
1388 |
|
if (!MVequal(pmv[1],prevMB->mvs[iSubBlock])) |
1389 |
|
if (!MVequal(pmv[1],pmv[0])) |
1390 |
{ |
{ |
1391 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
if (!(MotionFlags & PMV_HALFPEL8 )) |
1392 |
{ pmv[1].x = EVEN(pmv[1].x); |
{ pmv[1].x = EVEN(pmv[1].x); |
1395 |
CHECK_MV8_CANDIDATE(pmv[1].x,pmv[1].y); |
CHECK_MV8_CANDIDATE(pmv[1].x,pmv[1].y); |
1396 |
} |
} |
1397 |
|
|
1398 |
// top neighbour, if allowed |
// top neighbour, if allowed and needed |
1399 |
if (psad[2] != MV_MAX_ERROR) |
if (!MVzero(pmv[2])) |
1400 |
|
if (!MVequal(pmv[2],startMV)) |
1401 |
|
if (!MVequal(pmv[2],prevMB->mvs[iSubBlock])) |
1402 |
|
if (!MVequal(pmv[2],pmv[0])) |
1403 |
|
if (!MVequal(pmv[2],pmv[1])) |
1404 |
{ |
{ |
1405 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
if (!(MotionFlags & PMV_HALFPEL8 )) |
1406 |
{ pmv[2].x = EVEN(pmv[2].x); |
{ pmv[2].x = EVEN(pmv[2].x); |
1408 |
} |
} |
1409 |
CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); |
CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); |
1410 |
|
|
1411 |
// top right neighbour, if allowed |
// top right neighbour, if allowed and needed |
1412 |
if (psad[3] != MV_MAX_ERROR) |
if (!MVzero(pmv[3])) |
1413 |
|
if (!MVequal(pmv[3],startMV)) |
1414 |
|
if (!MVequal(pmv[3],prevMB->mvs[iSubBlock])) |
1415 |
|
if (!MVequal(pmv[3],pmv[0])) |
1416 |
|
if (!MVequal(pmv[3],pmv[1])) |
1417 |
|
if (!MVequal(pmv[3],pmv[2])) |
1418 |
{ |
{ |
1419 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
if (!(MotionFlags & PMV_HALFPEL8 )) |
1420 |
{ pmv[3].x = EVEN(pmv[3].x); |
{ pmv[3].x = EVEN(pmv[3].x); |
1424 |
} |
} |
1425 |
} |
} |
1426 |
|
|
1427 |
|
if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96) */ ) |
1428 |
|
iMinSAD -= MV8_00_BIAS; |
1429 |
|
|
1430 |
|
|
1431 |
/* Step 6: If MinSAD <= thresa goto Step 10. |
/* Step 6: If MinSAD <= thresa goto Step 10. |
1432 |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
1433 |
*/ |
*/ |
1434 |
|
|
1435 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,pMB->mvs[iSubBlock]) && (iMinSAD < pMB->sad8[iSubBlock]) ) ) |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,prevMB->mvs[iSubBlock]) && ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
1436 |
{ |
{ |
1437 |
if (MotionFlags & PMV_QUICKSTOP8) |
if (MotionFlags & PMV_QUICKSTOP16) |
1438 |
goto step10_8b; |
goto PMVfast8_Terminate_without_Refine; |
1439 |
if (MotionFlags & PMV_EARLYSTOP8) |
if (MotionFlags & PMV_EARLYSTOP16) |
1440 |
goto step10_8; |
goto PMVfast8_Terminate_with_Refine; |
1441 |
} |
} |
1442 |
|
|
1443 |
/************ (Diamond Search) **************/ |
/************ (Diamond Search) **************/ |
1453 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
1454 |
|
|
1455 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
1456 |
iSAD = PMVfastSearch8_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
iSAD = Diamond8_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1457 |
x, y, |
x, y, |
1458 |
currMV->x, currMV->y, iMinSAD, &newMV, |
currMV->x, currMV->y, iMinSAD, &newMV, |
1459 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1469 |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
1470 |
|
|
1471 |
if (!(MVequal(pmv[0],backupMV)) ) |
if (!(MVequal(pmv[0],backupMV)) ) |
1472 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1473 |
x, y, |
x, y, |
1474 |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
1475 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1482 |
} |
} |
1483 |
|
|
1484 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
1485 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1486 |
x, y, |
x, y, |
1487 |
0, 0, iMinSAD, &newMV, |
0, 0, iMinSAD, &newMV, |
1488 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1499 |
By performing an optional local half-pixel search, we can refine this result even further. |
By performing an optional local half-pixel search, we can refine this result even further. |
1500 |
*/ |
*/ |
1501 |
|
|
1502 |
step10_8: |
PMVfast8_Terminate_with_Refine: |
1503 |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
1504 |
iMinSAD = PMVfastSearch8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
iMinSAD = Halfpel8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
1505 |
x, y, |
x, y, |
1506 |
currMV, iMinSAD, |
currMV, iMinSAD, |
1507 |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
1508 |
|
|
|
step10_8b: |
|
1509 |
|
|
1510 |
|
PMVfast8_Terminate_without_Refine: |
1511 |
currPMV->x = currMV->x - pmv[0].x; |
currPMV->x = currMV->x - pmv[0].x; |
1512 |
currPMV->y = currMV->y - pmv[0].y; |
currPMV->y = currMV->y - pmv[0].y; |
1513 |
|
|
1514 |
return iMinSAD; |
return iMinSAD; |
1515 |
} |
} |
1516 |
|
|
1517 |
|
int32_t EPZSSearch16( |
1518 |
|
const uint8_t * const pRef, |
1519 |
|
const uint8_t * const pRefH, |
1520 |
|
const uint8_t * const pRefV, |
1521 |
|
const uint8_t * const pRefHV, |
1522 |
|
const IMAGE * const pCur, |
1523 |
|
const int x, const int y, |
1524 |
|
const uint32_t MotionFlags, |
1525 |
|
const uint32_t iQuant, |
1526 |
|
const uint32_t iFcode, |
1527 |
|
const MBParam * const pParam, |
1528 |
|
const MACROBLOCK * const pMBs, |
1529 |
|
const MACROBLOCK * const prevMBs, |
1530 |
|
VECTOR * const currMV, |
1531 |
|
VECTOR * const currPMV) |
1532 |
|
{ |
1533 |
|
const uint32_t iWcount = pParam->mb_width; |
1534 |
|
const uint32_t iHcount = pParam->mb_height; |
1535 |
|
|
1536 |
|
const int32_t iWidth = pParam->width; |
1537 |
|
const int32_t iHeight = pParam->height; |
1538 |
|
const int32_t iEdgedWidth = pParam->edged_width; |
1539 |
|
|
1540 |
|
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
1541 |
|
|
1542 |
|
int32_t min_dx; |
1543 |
|
int32_t max_dx; |
1544 |
|
int32_t min_dy; |
1545 |
|
int32_t max_dy; |
1546 |
|
|
1547 |
|
VECTOR newMV; |
1548 |
|
VECTOR backupMV; |
1549 |
|
|
1550 |
|
VECTOR pmv[4]; |
1551 |
|
int32_t psad[8]; |
1552 |
|
|
1553 |
|
static MACROBLOCK * oldMBs = NULL; |
1554 |
|
const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
1555 |
|
const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; |
1556 |
|
MACROBLOCK * oldMB = NULL; |
1557 |
|
|
1558 |
|
static int32_t thresh2; |
1559 |
|
int32_t bPredEq; |
1560 |
|
int32_t iMinSAD,iSAD=9999; |
1561 |
|
|
1562 |
|
MainSearch16FuncPtr EPZSMainSearchPtr; |
1563 |
|
|
1564 |
|
if (oldMBs == NULL) |
1565 |
|
{ oldMBs = (MACROBLOCK*) calloc(iWcount*iHcount,sizeof(MACROBLOCK)); |
1566 |
|
// fprintf(stderr,"allocated %d bytes for oldMBs\n",iWcount*iHcount*sizeof(MACROBLOCK)); |
1567 |
|
} |
1568 |
|
oldMB = oldMBs + x + y * iWcount; |
1569 |
|
|
1570 |
|
/* Get maximum range */ |
1571 |
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
1572 |
|
x, y, 16, iWidth, iHeight, iFcode); |
1573 |
|
|
1574 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1575 |
|
{ min_dx = EVEN(min_dx); |
1576 |
|
max_dx = EVEN(max_dx); |
1577 |
|
min_dy = EVEN(min_dy); |
1578 |
|
max_dy = EVEN(max_dy); |
1579 |
|
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
1580 |
|
|
1581 |
|
bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
1582 |
|
|
1583 |
|
/* Step 4: Calculate SAD around the Median prediction. |
1584 |
|
MinSAD=SAD |
1585 |
|
If Motion Vector equal to Previous frame motion vector |
1586 |
|
and MinSAD<PrevFrmSAD goto Step 10. |
1587 |
|
If SAD<=256 goto Step 10. |
1588 |
|
*/ |
1589 |
|
|
1590 |
|
// Prepare for main loop |
1591 |
|
|
1592 |
|
*currMV=pmv[0]; /* current best := median prediction */ |
1593 |
|
if (!(MotionFlags & PMV_HALFPEL16)) |
1594 |
|
{ |
1595 |
|
currMV->x = EVEN(currMV->x); |
1596 |
|
currMV->y = EVEN(currMV->y); |
1597 |
|
} |
1598 |
|
|
1599 |
|
if (currMV->x > max_dx) |
1600 |
|
currMV->x=max_dx; |
1601 |
|
if (currMV->x < min_dx) |
1602 |
|
currMV->x=min_dx; |
1603 |
|
if (currMV->y > max_dy) |
1604 |
|
currMV->y=max_dy; |
1605 |
|
if (currMV->y < min_dy) |
1606 |
|
currMV->y=min_dy; |
1607 |
|
|
1608 |
|
/***************** This is predictor SET A: only median prediction ******************/ |
1609 |
|
|
1610 |
|
iMinSAD = sad16( cur, |
1611 |
|
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
1612 |
|
iEdgedWidth, MV_MAX_ERROR); |
1613 |
|
iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode, iQuant); |
1614 |
|
|
1615 |
|
// thresh1 is fixed to 256 |
1616 |
|
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV, prevMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
1617 |
|
{ |
1618 |
|
if (MotionFlags & PMV_QUICKSTOP16) |
1619 |
|
goto EPZS16_Terminate_without_Refine; |
1620 |
|
if (MotionFlags & PMV_EARLYSTOP16) |
1621 |
|
goto EPZS16_Terminate_with_Refine; |
1622 |
|
} |
1623 |
|
|
1624 |
|
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
1625 |
|
|
1626 |
|
// previous frame MV |
1627 |
|
CHECK_MV16_CANDIDATE(prevMB->mvs[0].x,prevMB->mvs[0].y); |
1628 |
|
|
1629 |
|
// set threshhold based on Min of Prediction and SAD of collocated block |
1630 |
|
// CHECK_MV16 always uses iSAD for the SAD of last vector to check, so now iSAD is what we want |
1631 |
|
|
1632 |
|
if ((x==0) && (y==0) ) |
1633 |
|
{ |
1634 |
|
thresh2 = 512; |
1635 |
|
} |
1636 |
|
else |
1637 |
|
{ |
1638 |
|
/* T_k = 1.2 * MIN(SAD_top,SAD_left,SAD_topleft,SAD_coll) +128; [Tourapis, 2002] */ |
1639 |
|
|
1640 |
|
thresh2 = MIN(psad[0],iSAD)*6/5 + 128; |
1641 |
|
} |
1642 |
|
|
1643 |
|
// MV=(0,0) is often a good choice |
1644 |
|
|
1645 |
|
CHECK_MV16_ZERO; |
1646 |
|
|
1647 |
|
|
1648 |
|
// left neighbour, if allowed |
1649 |
|
if (x != 0) |
1650 |
|
{ |
1651 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1652 |
|
{ pmv[1].x = EVEN(pmv[1].x); |
1653 |
|
pmv[1].y = EVEN(pmv[1].y); |
1654 |
|
} |
1655 |
|
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
1656 |
|
} |
1657 |
|
|
1658 |
|
// top neighbour, if allowed |
1659 |
|
if (y != 0) |
1660 |
|
{ |
1661 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1662 |
|
{ pmv[2].x = EVEN(pmv[2].x); |
1663 |
|
pmv[2].y = EVEN(pmv[2].y); |
1664 |
|
} |
1665 |
|
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
1666 |
|
|
1667 |
|
// top right neighbour, if allowed |
1668 |
|
if ((uint32_t)x != (iWcount-1)) |
1669 |
|
{ |
1670 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1671 |
|
{ pmv[3].x = EVEN(pmv[3].x); |
1672 |
|
pmv[3].y = EVEN(pmv[3].y); |
1673 |
|
} |
1674 |
|
CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); |
1675 |
|
} |
1676 |
|
} |
1677 |
|
|
1678 |
|
/* Terminate if MinSAD <= T_2 |
1679 |
|
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
1680 |
|
*/ |
1681 |
|
|
1682 |
|
if ( (iMinSAD <= thresh2) |
1683 |
|
|| ( MVequal(*currMV,prevMB->mvs[0]) && ((uint32_t)iMinSAD <= prevMB->sad16) ) ) |
1684 |
|
{ |
1685 |
|
if (MotionFlags & PMV_QUICKSTOP16) |
1686 |
|
goto EPZS16_Terminate_without_Refine; |
1687 |
|
if (MotionFlags & PMV_EARLYSTOP16) |
1688 |
|
goto EPZS16_Terminate_with_Refine; |
1689 |
|
} |
1690 |
|
|
1691 |
|
/***** predictor SET C: acceleration MV (new!), neighbours in prev. frame(new!) ****/ |
1692 |
|
|
1693 |
|
backupMV = prevMB->mvs[0]; // collocated MV |
1694 |
|
backupMV.x += (prevMB->mvs[0].x - oldMB->mvs[0].x ); // acceleration X |
1695 |
|
backupMV.y += (prevMB->mvs[0].y - oldMB->mvs[0].y ); // acceleration Y |
1696 |
|
|
1697 |
|
CHECK_MV16_CANDIDATE(backupMV.x,backupMV.y); |
1698 |
|
|
1699 |
|
// left neighbour |
1700 |
|
if (x != 0) |
1701 |
|
CHECK_MV16_CANDIDATE((prevMB-1)->mvs[0].x,(prevMB-1)->mvs[0].y); |
1702 |
|
|
1703 |
|
// top neighbour |
1704 |
|
if (y != 0) |
1705 |
|
CHECK_MV16_CANDIDATE((prevMB-iWcount)->mvs[0].x,(prevMB-iWcount)->mvs[0].y); |
1706 |
|
|
1707 |
|
// right neighbour, if allowed (this value is not written yet, so take it from pMB->mvs |
1708 |
|
|
1709 |
|
if ((uint32_t)x != iWcount-1) |
1710 |
|
CHECK_MV16_CANDIDATE((prevMB+1)->mvs[0].x,(prevMB+1)->mvs[0].y); |
1711 |
|
|
1712 |
|
// bottom neighbour, dito |
1713 |
|
if ((uint32_t)y != iHcount-1) |
1714 |
|
CHECK_MV16_CANDIDATE((prevMB+iWcount)->mvs[0].x,(prevMB+iWcount)->mvs[0].y); |
1715 |
|
|
1716 |
|
/* Terminate if MinSAD <= T_3 (here T_3 = T_2) */ |
1717 |
|
if (iMinSAD <= thresh2) |
1718 |
|
{ |
1719 |
|
if (MotionFlags & PMV_QUICKSTOP16) |
1720 |
|
goto EPZS16_Terminate_without_Refine; |
1721 |
|
if (MotionFlags & PMV_EARLYSTOP16) |
1722 |
|
goto EPZS16_Terminate_with_Refine; |
1723 |
|
} |
1724 |
|
|
1725 |
|
/************ (if Diamond Search) **************/ |
1726 |
|
|
1727 |
|
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
1728 |
|
|
1729 |
|
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
1730 |
|
|
1731 |
|
if (MotionFlags & PMV_USESQUARES16) |
1732 |
|
EPZSMainSearchPtr = Square16_MainSearch; |
1733 |
|
else |
1734 |
|
EPZSMainSearchPtr = Diamond16_MainSearch; |
1735 |
|
|
1736 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1737 |
|
x, y, |
1738 |
|
currMV->x, currMV->y, iMinSAD, &newMV, pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
1739 |
|
2, iFcode, iQuant, 0); |
1740 |
|
|
1741 |
|
if (iSAD < iMinSAD) |
1742 |
|
{ |
1743 |
|
*currMV = newMV; |
1744 |
|
iMinSAD = iSAD; |
1745 |
|
} |
1746 |
|
|
1747 |
|
|
1748 |
|
if (MotionFlags & PMV_EXTSEARCH16) |
1749 |
|
{ |
1750 |
|
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
1751 |
|
|
1752 |
|
if (!(MVequal(pmv[0],backupMV)) ) |
1753 |
|
{ |
1754 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1755 |
|
x, y, |
1756 |
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
1757 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); |
1758 |
|
} |
1759 |
|
|
1760 |
|
if (iSAD < iMinSAD) |
1761 |
|
{ |
1762 |
|
*currMV = newMV; |
1763 |
|
iMinSAD = iSAD; |
1764 |
|
} |
1765 |
|
|
1766 |
|
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
1767 |
|
{ |
1768 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1769 |
|
x, y, |
1770 |
|
0, 0, iMinSAD, &newMV, |
1771 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); |
1772 |
|
|
1773 |
|
if (iSAD < iMinSAD) |
1774 |
|
{ |
1775 |
|
*currMV = newMV; |
1776 |
|
iMinSAD = iSAD; |
1777 |
|
} |
1778 |
|
} |
1779 |
|
} |
1780 |
|
|
1781 |
|
/*************** Choose best MV found **************/ |
1782 |
|
|
1783 |
|
EPZS16_Terminate_with_Refine: |
1784 |
|
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
1785 |
|
iMinSAD = Halfpel16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
1786 |
|
x, y, |
1787 |
|
currMV, iMinSAD, |
1788 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
1789 |
|
|
1790 |
|
EPZS16_Terminate_without_Refine: |
1791 |
|
|
1792 |
|
*oldMB = *prevMB; |
1793 |
|
|
1794 |
|
currPMV->x = currMV->x - pmv[0].x; |
1795 |
|
currPMV->y = currMV->y - pmv[0].y; |
1796 |
|
return iMinSAD; |
1797 |
|
} |
1798 |
|
|
1799 |
|
|
1800 |
|
int32_t EPZSSearch8( |
1801 |
|
const uint8_t * const pRef, |
1802 |
|
const uint8_t * const pRefH, |
1803 |
|
const uint8_t * const pRefV, |
1804 |
|
const uint8_t * const pRefHV, |
1805 |
|
const IMAGE * const pCur, |
1806 |
|
const int x, const int y, |
1807 |
|
const int start_x, const int start_y, |
1808 |
|
const uint32_t MotionFlags, |
1809 |
|
const uint32_t iQuant, |
1810 |
|
const uint32_t iFcode, |
1811 |
|
const MBParam * const pParam, |
1812 |
|
const MACROBLOCK * const pMBs, |
1813 |
|
const MACROBLOCK * const prevMBs, |
1814 |
|
VECTOR * const currMV, |
1815 |
|
VECTOR * const currPMV) |
1816 |
|
{ |
1817 |
|
/* Please not that EPZS might not be a good choice for 8x8-block motion search ! */ |
1818 |
|
|
1819 |
|
const uint32_t iWcount = pParam->mb_width; |
1820 |
|
const int32_t iWidth = pParam->width; |
1821 |
|
const int32_t iHeight = pParam->height; |
1822 |
|
const int32_t iEdgedWidth = pParam->edged_width; |
1823 |
|
|
1824 |
|
const uint8_t * cur = pCur->y + x*8 + y*8*iEdgedWidth; |
1825 |
|
|
1826 |
|
int32_t iDiamondSize=1; |
1827 |
|
|
1828 |
|
int32_t min_dx; |
1829 |
|
int32_t max_dx; |
1830 |
|
int32_t min_dy; |
1831 |
|
int32_t max_dy; |
1832 |
|
|
1833 |
|
VECTOR newMV; |
1834 |
|
VECTOR backupMV; |
1835 |
|
|
1836 |
|
VECTOR pmv[4]; |
1837 |
|
int32_t psad[8]; |
1838 |
|
|
1839 |
|
const int32_t iSubBlock = ((y&1)<<1) + (x&1); |
1840 |
|
|
1841 |
|
const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
1842 |
|
const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; |
1843 |
|
|
1844 |
|
int32_t bPredEq; |
1845 |
|
int32_t iMinSAD,iSAD=9999; |
1846 |
|
|
1847 |
|
MainSearch8FuncPtr EPZSMainSearchPtr; |
1848 |
|
|
1849 |
|
/* Get maximum range */ |
1850 |
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
1851 |
|
x, y, 8, iWidth, iHeight, iFcode); |
1852 |
|
|
1853 |
|
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
1854 |
|
|
1855 |
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
1856 |
|
{ min_dx = EVEN(min_dx); |
1857 |
|
max_dx = EVEN(max_dx); |
1858 |
|
min_dy = EVEN(min_dy); |
1859 |
|
max_dy = EVEN(max_dy); |
1860 |
|
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
1861 |
|
|
1862 |
|
bPredEq = get_pmvdata(pMBs, x>>1, y>>1, iWcount, iSubBlock, pmv, psad); |
1863 |
|
|
1864 |
|
|
1865 |
|
/* Step 4: Calculate SAD around the Median prediction. |
1866 |
|
MinSAD=SAD |
1867 |
|
If Motion Vector equal to Previous frame motion vector |
1868 |
|
and MinSAD<PrevFrmSAD goto Step 10. |
1869 |
|
If SAD<=256 goto Step 10. |
1870 |
|
*/ |
1871 |
|
|
1872 |
|
// Prepare for main loop |
1873 |
|
|
1874 |
|
|
1875 |
|
if (!(MotionFlags & PMV_HALFPEL8)) |
1876 |
|
{ |
1877 |
|
currMV->x = EVEN(currMV->x); |
1878 |
|
currMV->y = EVEN(currMV->y); |
1879 |
|
} |
1880 |
|
|
1881 |
|
if (currMV->x > max_dx) |
1882 |
|
currMV->x=max_dx; |
1883 |
|
if (currMV->x < min_dx) |
1884 |
|
currMV->x=min_dx; |
1885 |
|
if (currMV->y > max_dy) |
1886 |
|
currMV->y=max_dy; |
1887 |
|
if (currMV->y < min_dy) |
1888 |
|
currMV->y=min_dy; |
1889 |
|
|
1890 |
|
/***************** This is predictor SET A: only median prediction ******************/ |
1891 |
|
|
1892 |
|
|
1893 |
|
iMinSAD = sad8( cur, |
1894 |
|
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
1895 |
|
iEdgedWidth); |
1896 |
|
iMinSAD += calc_delta_8(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode, iQuant); |
1897 |
|
|
1898 |
|
|
1899 |
|
// thresh1 is fixed to 256 |
1900 |
|
if (iMinSAD < 256/4 ) |
1901 |
|
{ |
1902 |
|
if (MotionFlags & PMV_QUICKSTOP8) |
1903 |
|
goto EPZS8_Terminate_without_Refine; |
1904 |
|
if (MotionFlags & PMV_EARLYSTOP8) |
1905 |
|
goto EPZS8_Terminate_with_Refine; |
1906 |
|
} |
1907 |
|
|
1908 |
|
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
1909 |
|
|
1910 |
|
|
1911 |
|
// MV=(0,0) is often a good choice |
1912 |
|
CHECK_MV8_ZERO; |
1913 |
|
|
1914 |
|
// previous frame MV |
1915 |
|
CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x,prevMB->mvs[iSubBlock].y); |
1916 |
|
|
1917 |
|
// left neighbour, if allowed |
1918 |
|
if (psad[1] != MV_MAX_ERROR) |
1919 |
|
{ |
1920 |
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
1921 |
|
{ pmv[1].x = EVEN(pmv[1].x); |
1922 |
|
pmv[1].y = EVEN(pmv[1].y); |
1923 |
|
} |
1924 |
|
CHECK_MV8_CANDIDATE(pmv[1].x,pmv[1].y); |
1925 |
|
} |
1926 |
|
|
1927 |
|
// top neighbour, if allowed |
1928 |
|
if (psad[2] != MV_MAX_ERROR) |
1929 |
|
{ |
1930 |
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
1931 |
|
{ pmv[2].x = EVEN(pmv[2].x); |
1932 |
|
pmv[2].y = EVEN(pmv[2].y); |
1933 |
|
} |
1934 |
|
CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); |
1935 |
|
|
1936 |
|
// top right neighbour, if allowed |
1937 |
|
if (psad[3] != MV_MAX_ERROR) |
1938 |
|
{ |
1939 |
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
1940 |
|
{ pmv[3].x = EVEN(pmv[3].x); |
1941 |
|
pmv[3].y = EVEN(pmv[3].y); |
1942 |
|
} |
1943 |
|
CHECK_MV8_CANDIDATE(pmv[3].x,pmv[3].y); |
1944 |
|
} |
1945 |
|
} |
1946 |
|
|
1947 |
|
/* // this bias is zero anyway, at the moment! |
1948 |
|
|
1949 |
|
if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) ) // && (iMinSAD <= iQuant * 96) |
1950 |
|
iMinSAD -= MV8_00_BIAS; |
1951 |
|
|
1952 |
|
*/ |
1953 |
|
|
1954 |
|
/* Terminate if MinSAD <= T_2 |
1955 |
|
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
1956 |
|
*/ |
1957 |
|
|
1958 |
|
if (iMinSAD < 512/4) /* T_2 == 512/4 hardcoded */ |
1959 |
|
{ |
1960 |
|
if (MotionFlags & PMV_QUICKSTOP8) |
1961 |
|
goto EPZS8_Terminate_without_Refine; |
1962 |
|
if (MotionFlags & PMV_EARLYSTOP8) |
1963 |
|
goto EPZS8_Terminate_with_Refine; |
1964 |
|
} |
1965 |
|
|
1966 |
|
/************ (Diamond Search) **************/ |
1967 |
|
|
1968 |
|
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
1969 |
|
|
1970 |
|
if (!(MotionFlags & PMV_HALFPELDIAMOND8)) |
1971 |
|
iDiamondSize *= 2; |
1972 |
|
|
1973 |
|
/* default: use best prediction as starting point for one call of EPZS_MainSearch */ |
1974 |
|
|
1975 |
|
/* // there is no EPZS^2 for inter4v at the moment |
1976 |
|
|
1977 |
|
if (MotionFlags & PMV_USESQUARES8) |
1978 |
|
EPZSMainSearchPtr = Square8_MainSearch; |
1979 |
|
else |
1980 |
|
*/ |
1981 |
|
|
1982 |
|
EPZSMainSearchPtr = Diamond8_MainSearch; |
1983 |
|
|
1984 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1985 |
|
x, y, |
1986 |
|
currMV->x, currMV->y, iMinSAD, &newMV, |
1987 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
1988 |
|
iDiamondSize, iFcode, iQuant, 0); |
1989 |
|
|
1990 |
|
|
1991 |
|
if (iSAD < iMinSAD) |
1992 |
|
{ |
1993 |
|
*currMV = newMV; |
1994 |
|
iMinSAD = iSAD; |
1995 |
|
} |
1996 |
|
|
1997 |
|
if (MotionFlags & PMV_EXTSEARCH8) |
1998 |
|
{ |
1999 |
|
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
2000 |
|
|
2001 |
|
if (!(MVequal(pmv[0],backupMV)) ) |
2002 |
|
{ |
2003 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
2004 |
|
x, y, |
2005 |
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
2006 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); |
2007 |
|
|
2008 |
|
if (iSAD < iMinSAD) |
2009 |
|
{ |
2010 |
|
*currMV = newMV; |
2011 |
|
iMinSAD = iSAD; |
2012 |
|
} |
2013 |
|
} |
2014 |
|
|
2015 |
|
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
2016 |
|
{ |
2017 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
2018 |
|
x, y, |
2019 |
|
0, 0, iMinSAD, &newMV, |
2020 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); |
2021 |
|
|
2022 |
|
if (iSAD < iMinSAD) |
2023 |
|
{ |
2024 |
|
*currMV = newMV; |
2025 |
|
iMinSAD = iSAD; |
2026 |
|
} |
2027 |
|
} |
2028 |
|
} |
2029 |
|
|
2030 |
|
/*************** Choose best MV found **************/ |
2031 |
|
|
2032 |
|
EPZS8_Terminate_with_Refine: |
2033 |
|
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
2034 |
|
iMinSAD = Halfpel8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
2035 |
|
x, y, |
2036 |
|
currMV, iMinSAD, |
2037 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
2038 |
|
|
2039 |
|
EPZS8_Terminate_without_Refine: |
2040 |
|
|
2041 |
|
currPMV->x = currMV->x - pmv[0].x; |
2042 |
|
currPMV->y = currMV->y - pmv[0].y; |
2043 |
|
return iMinSAD; |
2044 |
|
} |
2045 |
|
|
2046 |
|
|
2047 |
|
|
2048 |
|
|
2049 |
|
|
2050 |
|
/* *********************************************************** |
2051 |
|
bvop motion estimation |
2052 |
|
// TODO: need to incorporate prediction here (eg. sad += calc_delta_16) |
2053 |
|
***************************************************************/ |
2054 |
|
|
2055 |
|
|
2056 |
|
void MotionEstimationBVOP( |
2057 |
|
MBParam * const pParam, |
2058 |
|
FRAMEINFO * const frame, |
2059 |
|
|
2060 |
|
// forward (past) reference |
2061 |
|
const MACROBLOCK * const f_mbs, |
2062 |
|
const IMAGE * const f_ref, |
2063 |
|
const IMAGE * const f_refH, |
2064 |
|
const IMAGE * const f_refV, |
2065 |
|
const IMAGE * const f_refHV, |
2066 |
|
// backward (future) reference |
2067 |
|
const MACROBLOCK * const b_mbs, |
2068 |
|
const IMAGE * const b_ref, |
2069 |
|
const IMAGE * const b_refH, |
2070 |
|
const IMAGE * const b_refV, |
2071 |
|
const IMAGE * const b_refHV) |
2072 |
|
{ |
2073 |
|
const uint32_t mb_width = pParam->mb_width; |
2074 |
|
const uint32_t mb_height = pParam->mb_height; |
2075 |
|
const int32_t edged_width = pParam->edged_width; |
2076 |
|
|
2077 |
|
uint32_t i,j; |
2078 |
|
|
2079 |
|
int32_t f_sad16; |
2080 |
|
int32_t b_sad16; |
2081 |
|
int32_t i_sad16; |
2082 |
|
int32_t d_sad16; |
2083 |
|
int32_t best_sad; |
2084 |
|
|
2085 |
|
VECTOR pmv_dontcare; |
2086 |
|
|
2087 |
|
// note: i==horizontal, j==vertical |
2088 |
|
for (j = 0; j < mb_height; j++) |
2089 |
|
{ |
2090 |
|
for (i = 0; i < mb_width; i++) |
2091 |
|
{ |
2092 |
|
MACROBLOCK *mb = &frame->mbs[i + j*mb_width]; |
2093 |
|
const MACROBLOCK *f_mb = &f_mbs[i + j*mb_width]; |
2094 |
|
const MACROBLOCK *b_mb = &b_mbs[i + j*mb_width]; |
2095 |
|
|
2096 |
|
if (b_mb->mode == MODE_INTER |
2097 |
|
&& b_mb->cbp == 0 |
2098 |
|
&& b_mb->mvs[0].x == 0 |
2099 |
|
&& b_mb->mvs[0].y == 0) |
2100 |
|
{ |
2101 |
|
mb->mode = MODE_NOT_CODED; |
2102 |
|
mb->mvs[0].x = 0; |
2103 |
|
mb->mvs[0].y = 0; |
2104 |
|
mb->b_mvs[0].x = 0; |
2105 |
|
mb->b_mvs[0].y = 0; |
2106 |
|
continue; |
2107 |
|
} |
2108 |
|
|
2109 |
|
|
2110 |
|
// forward search |
2111 |
|
f_sad16 = SEARCH16(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
2112 |
|
&frame->image, |
2113 |
|
i, j, |
2114 |
|
frame->motion_flags, frame->quant, frame->fcode, |
2115 |
|
pParam, |
2116 |
|
f_mbs, f_mbs /* todo */, |
2117 |
|
&mb->mvs[0], &pmv_dontcare); // ignore pmv |
2118 |
|
|
2119 |
|
// backward search |
2120 |
|
b_sad16 = SEARCH16(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
2121 |
|
&frame->image, |
2122 |
|
i, j, |
2123 |
|
frame->motion_flags, frame->quant, frame->bcode, |
2124 |
|
pParam, |
2125 |
|
b_mbs, b_mbs, /* todo */ |
2126 |
|
&mb->b_mvs[0], &pmv_dontcare); // ignore pmv |
2127 |
|
|
2128 |
|
// interpolate search (simple, but effective) |
2129 |
|
i_sad16 = sad16bi_c( |
2130 |
|
frame->image.y + i*16 + j*16*edged_width, |
2131 |
|
get_ref(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
2132 |
|
i, j, 16, mb->mvs[0].x, mb->mvs[0].y, edged_width), |
2133 |
|
get_ref(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
2134 |
|
i, j, 16, mb->b_mvs[0].x, mb->b_mvs[0].x, edged_width), |
2135 |
|
edged_width); |
2136 |
|
|
2137 |
|
// TODO: direct search |
2138 |
|
// predictor + range of [-32,32] |
2139 |
|
d_sad16 = 65535; |
2140 |
|
|
2141 |
|
|
2142 |
|
if (f_sad16 < b_sad16) |
2143 |
|
{ |
2144 |
|
best_sad = f_sad16; |
2145 |
|
mb->mode = MODE_FORWARD; |
2146 |
|
} |
2147 |
|
else |
2148 |
|
{ |
2149 |
|
best_sad = b_sad16; |
2150 |
|
mb->mode = MODE_BACKWARD; |
2151 |
|
} |
2152 |
|
|
2153 |
|
if (i_sad16 < best_sad) |
2154 |
|
{ |
2155 |
|
best_sad = i_sad16; |
2156 |
|
mb->mode = MODE_INTERPOLATE; |
2157 |
|
} |
2158 |
|
|
2159 |
|
if (d_sad16 < best_sad) |
2160 |
|
{ |
2161 |
|
best_sad = d_sad16; |
2162 |
|
mb->mode = MODE_DIRECT; |
2163 |
|
} |
2164 |
|
|
2165 |
|
} |
2166 |
|
} |
2167 |
|
} |