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 |
54 |
#define MV16_THRESHOLD 192 |
#define MV16_THRESHOLD 192 |
55 |
#define MV8_THRESHOLD 56 |
#define MV8_THRESHOLD 56 |
56 |
|
|
57 |
|
#define NEIGH_MOVE_THRESH 0 |
58 |
|
// how much a block's MV must differ from his neighbour |
59 |
|
// to be search for INTER4V. The more, the faster... |
60 |
|
|
61 |
/* sad16(0,0) bias; mpeg4 spec suggests nb/2+1 */ |
/* sad16(0,0) bias; mpeg4 spec suggests nb/2+1 */ |
62 |
/* nb = vop pixels * 2^(bpp-8) */ |
/* nb = vop pixels * 2^(bpp-8) */ |
63 |
#define MV16_00_BIAS (128+1) |
#define MV16_00_BIAS (128+1) |
64 |
|
#define MV8_00_BIAS (0) |
65 |
|
|
66 |
/* INTER bias for INTER/INTRA decision; mpeg4 spec suggests 2*nb */ |
/* INTER bias for INTER/INTRA decision; mpeg4 spec suggests 2*nb */ |
67 |
#define INTER_BIAS 512 |
#define MV16_INTER_BIAS 512 |
68 |
|
|
69 |
/* Parameters which control inter/inter4v decision */ |
/* Parameters which control inter/inter4v decision */ |
70 |
#define IMV16X16 5 |
#define IMV16X16 5 |
73 |
#define NEIGH_TEND_16X16 2 |
#define NEIGH_TEND_16X16 2 |
74 |
#define NEIGH_TEND_8X8 2 |
#define NEIGH_TEND_8X8 2 |
75 |
|
|
|
|
|
76 |
// fast ((A)/2)*2 |
// fast ((A)/2)*2 |
77 |
#define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) |
#define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) |
78 |
|
|
79 |
|
#define MVzero(A) ( ((A).x)==(0) && ((A).y)==(0) ) |
80 |
|
#define MVequal(A,B) ( ((A).x)==((B).x) && ((A).y)==((B).y) ) |
81 |
|
|
82 |
|
int32_t PMVfastSearch16( |
83 |
|
const uint8_t * const pRef, |
84 |
|
const uint8_t * const pRefH, |
85 |
|
const uint8_t * const pRefV, |
86 |
|
const uint8_t * const pRefHV, |
87 |
|
const IMAGE * const pCur, |
88 |
|
const int x, const int y, |
89 |
|
const uint32_t MotionFlags, |
90 |
|
const uint32_t iQuant, |
91 |
|
const uint32_t iFcode, |
92 |
|
const MBParam * const pParam, |
93 |
|
const MACROBLOCK * const pMBs, |
94 |
|
const MACROBLOCK * const prevMBs, |
95 |
|
VECTOR * const currMV, |
96 |
|
VECTOR * const currPMV); |
97 |
|
|
98 |
#define MIN(X, Y) ((X)<(Y)?(X):(Y)) |
int32_t EPZSSearch16( |
99 |
#define MAX(X, Y) ((X)>(Y)?(X):(Y)) |
const uint8_t * const pRef, |
100 |
#define ABS(X) (((X)>0)?(X):-(X)) |
const uint8_t * const pRefH, |
101 |
#define SIGN(X) (((X)>0)?1:-1) |
const uint8_t * const pRefV, |
102 |
|
const uint8_t * const pRefHV, |
103 |
|
const IMAGE * const pCur, |
104 |
|
const int x, const int y, |
105 |
|
const uint32_t MotionFlags, |
106 |
|
const uint32_t iQuant, |
107 |
|
const uint32_t iFcode, |
108 |
|
const MBParam * const pParam, |
109 |
|
const MACROBLOCK * const pMBs, |
110 |
|
const MACROBLOCK * const prevMBs, |
111 |
|
VECTOR * const currMV, |
112 |
|
VECTOR * const currPMV); |
113 |
|
|
114 |
|
|
115 |
int32_t PMVfastSearch8( |
int32_t PMVfastSearch8( |
119 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
120 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
121 |
const int x, const int y, |
const int x, const int y, |
122 |
const int start_x, int start_y, |
const int start_x, const int start_y, |
123 |
const uint32_t iQuality, |
const uint32_t MotionFlags, |
124 |
MBParam * const pParam, |
const uint32_t iQuant, |
125 |
MACROBLOCK * const pMBs, |
const uint32_t iFcode, |
126 |
|
const MBParam * const pParam, |
127 |
|
const MACROBLOCK * const pMBs, |
128 |
|
const MACROBLOCK * const prevMBs, |
129 |
VECTOR * const currMV, |
VECTOR * const currMV, |
130 |
VECTOR * const currPMV); |
VECTOR * const currPMV); |
131 |
|
|
132 |
int32_t PMVfastSearch16( |
int32_t EPZSSearch8( |
133 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
134 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
135 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
136 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
137 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
138 |
const int x, const int y, |
const int x, const int y, |
139 |
const uint32_t iQuality, |
const int start_x, const int start_y, |
140 |
MBParam * const pParam, |
const uint32_t MotionFlags, |
141 |
MACROBLOCK * const pMBs, |
const uint32_t iQuant, |
142 |
|
const uint32_t iFcode, |
143 |
|
const MBParam * const pParam, |
144 |
|
const MACROBLOCK * const pMBs, |
145 |
|
const MACROBLOCK * const prevMBs, |
146 |
VECTOR * const currMV, |
VECTOR * const currMV, |
147 |
VECTOR * const currPMV); |
VECTOR * const currPMV); |
148 |
|
|
149 |
|
|
150 |
|
typedef int32_t (MainSearch16Func)( |
151 |
|
const uint8_t * const pRef, |
152 |
|
const uint8_t * const pRefH, |
153 |
|
const uint8_t * const pRefV, |
154 |
|
const uint8_t * const pRefHV, |
155 |
|
const uint8_t * const cur, |
156 |
|
const int x, const int y, |
157 |
|
int32_t startx, int32_t starty, |
158 |
|
int32_t iMinSAD, |
159 |
|
VECTOR * const currMV, |
160 |
|
const VECTOR * const pmv, |
161 |
|
const int32_t min_dx, const int32_t max_dx, |
162 |
|
const int32_t min_dy, const int32_t max_dy, |
163 |
|
const int32_t iEdgedWidth, |
164 |
|
const int32_t iDiamondSize, |
165 |
|
const int32_t iFcode, |
166 |
|
const int32_t iQuant, |
167 |
|
int iFound); |
168 |
|
|
169 |
|
typedef MainSearch16Func* MainSearch16FuncPtr; |
170 |
|
|
|
/* diamond search stuff |
|
|
keep the the sequence in circular order (so optimization works) |
|
|
*/ |
|
171 |
|
|
172 |
typedef struct |
typedef int32_t (MainSearch8Func)( |
173 |
{ |
const uint8_t * const pRef, |
174 |
int32_t dx; |
const uint8_t * const pRefH, |
175 |
int32_t dy; |
const uint8_t * const pRefV, |
176 |
} |
const uint8_t * const pRefHV, |
177 |
DPOINT; |
const uint8_t * const cur, |
178 |
|
const int x, const int y, |
179 |
|
int32_t startx, int32_t starty, |
180 |
|
int32_t iMinSAD, |
181 |
|
VECTOR * const currMV, |
182 |
|
const VECTOR * const pmv, |
183 |
|
const int32_t min_dx, const int32_t max_dx, |
184 |
|
const int32_t min_dy, const int32_t max_dy, |
185 |
|
const int32_t iEdgedWidth, |
186 |
|
const int32_t iDiamondSize, |
187 |
|
const int32_t iFcode, |
188 |
|
const int32_t iQuant, |
189 |
|
int iFound); |
190 |
|
|
191 |
|
typedef MainSearch8Func* MainSearch8FuncPtr; |
192 |
|
|
193 |
static const DPOINT diamond_small[4] = |
static int32_t lambda_vec16[32] = /* rounded values for lambda param for weight of motion bits as in modified H.26L */ |
194 |
{ |
{ 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), |
195 |
{0, 1}, {1, 0}, {0, -1}, {-1, 0} |
(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), |
196 |
}; |
(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), |
197 |
|
(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), |
198 |
|
(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), |
199 |
|
(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), |
200 |
|
(int)(36.4949+0.5) }; |
201 |
|
|
202 |
|
static int32_t *lambda_vec8 = lambda_vec16; /* same table for INTER and INTER4V for now*/ |
203 |
|
|
|
static const DPOINT diamond_large[8] = |
|
|
{ |
|
|
{0, 2}, {1, 1}, {2, 0}, {1, -1}, {0, -2}, {-1, -1}, {-2, 0}, {-1, 1} |
|
|
}; |
|
204 |
|
|
205 |
|
|
206 |
// mv.length table |
// mv.length table |
238 |
} |
} |
239 |
|
|
240 |
|
|
241 |
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) |
|
|
|
|
242 |
{ |
{ |
243 |
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)); |
244 |
} |
} |
245 |
|
|
246 |
|
static __inline uint32_t calc_delta_8(const int32_t dx, const int32_t dy, const uint32_t iFcode, const uint32_t iQuant) |
247 |
|
|
|
|
|
|
|
|
|
/* 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) |
|
248 |
{ |
{ |
249 |
const int search_range = 32 << (fcode - 1); |
return NEIGH_TEND_8X8 * lambda_vec8[iQuant] * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); |
|
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)); |
|
250 |
} |
} |
251 |
|
|
252 |
|
|
|
/* 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; |
|
|
} |
|
|
} |
|
|
|
|
253 |
|
|
|
/* This is somehow a copy of get_ref, but with MV instead of X,Y */ |
|
254 |
|
|
|
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) |
|
|
{ |
|
|
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; |
|
|
} |
|
|
} |
|
255 |
|
|
256 |
#ifndef SEARCH16 |
#ifndef SEARCH16 |
257 |
#define SEARCH16 PMVfastSearch16 |
#define SEARCH16 PMVfastSearch16 |
258 |
|
//#define SEARCH16 FullSearch16 |
259 |
|
//#define SEARCH16 EPZSSearch16 |
260 |
#endif |
#endif |
261 |
|
|
262 |
#ifndef SEARCH8 |
#ifndef SEARCH8 |
263 |
#define SEARCH8 PMVfastSearch8 |
#define SEARCH8 PMVfastSearch8 |
264 |
|
//#define SEARCH8 EPZSSearch8 |
265 |
#endif |
#endif |
266 |
|
|
267 |
bool MotionEstimation( |
bool MotionEstimation( |
|
MACROBLOCK * const pMBs, |
|
268 |
MBParam * const pParam, |
MBParam * const pParam, |
269 |
const IMAGE * const pRef, |
FRAMEINFO * const current, |
270 |
|
FRAMEINFO * const reference, |
271 |
const IMAGE * const pRefH, |
const IMAGE * const pRefH, |
272 |
const IMAGE * const pRefV, |
const IMAGE * const pRefV, |
273 |
const IMAGE * const pRefHV, |
const IMAGE * const pRefHV, |
|
IMAGE * const pCurrent, |
|
274 |
const uint32_t iLimit) |
const uint32_t iLimit) |
275 |
|
|
276 |
{ |
{ |
277 |
const uint32_t iWcount = pParam->mb_width; |
const uint32_t iWcount = pParam->mb_width; |
278 |
const uint32_t iHcount = pParam->mb_height; |
const uint32_t iHcount = pParam->mb_height; |
279 |
|
MACROBLOCK * const pMBs = current->mbs; |
280 |
|
MACROBLOCK * const prevMBs = reference->mbs; // previous frame |
281 |
|
|
282 |
uint32_t i, j, iIntra = 0; |
const IMAGE * const pCurrent = ¤t->image; |
283 |
|
const IMAGE * const pRef = &reference->image; |
284 |
|
|
285 |
VECTOR mv16; |
const VECTOR zeroMV = {0,0}; |
|
VECTOR pmv16; |
|
286 |
|
|
287 |
int32_t sad8 = 0; |
int32_t x, y; |
288 |
int32_t sad16; |
int32_t iIntra = 0; |
289 |
int32_t deviation; |
VECTOR pmv; |
290 |
|
|
291 |
// note: i==horizontal, j==vertical |
if (sadInit) |
292 |
for (i = 0; i < iHcount; i++) |
(*sadInit)(); |
|
for (j = 0; j < iWcount; j++) |
|
|
{ |
|
|
MACROBLOCK *pMB = &pMBs[j + i * iWcount]; |
|
293 |
|
|
294 |
sad16 = SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
for (y = 0; y < iHcount; y++) |
295 |
j, i, pParam->motion_flags, |
for (x = 0; x < iWcount; x++) |
296 |
pParam, pMBs, &mv16, &pmv16); |
{ |
297 |
pMB->sad16=sad16; |
MACROBLOCK* const pMB = &pMBs[x + y * iWcount]; |
298 |
|
|
299 |
|
pMB->sad16 = SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
300 |
|
x, y, current->motion_flags, current->quant, current->fcode, |
301 |
|
pParam, pMBs, prevMBs, &pMB->mv16, &pMB->pmvs[0]); |
302 |
|
|
|
/* decide: MODE_INTER or MODE_INTRA |
|
|
if (dev_intra < sad_inter - 2 * nb) use_intra |
|
|
*/ |
|
303 |
|
|
304 |
deviation = dev16(pCurrent->y + j*16 + i*16*pParam->edged_width, pParam->edged_width); |
if (0 < (pMB->sad16 - MV16_INTER_BIAS)) |
305 |
|
{ |
306 |
|
int32_t deviation; |
307 |
|
deviation = dev16(pCurrent->y + x*16 + y*16*pParam->edged_width, |
308 |
|
pParam->edged_width); |
309 |
|
|
310 |
if (deviation < (sad16 - INTER_BIAS)) |
if (deviation < (pMB->sad16 - MV16_INTER_BIAS)) |
311 |
{ |
{ |
312 |
pMB->mode = MODE_INTRA; |
pMB->mode = MODE_INTRA; |
313 |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
pMB->mv16 = pMB->mvs[0] = pMB->mvs[1] |
314 |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; |
= pMB->mvs[2] = pMB->mvs[3] = zeroMV; |
315 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] |
316 |
|
= pMB->sad8[2] = pMB->sad8[3] = 0; |
317 |
|
|
318 |
iIntra++; |
iIntra++; |
319 |
if(iIntra >= iLimit) |
if(iIntra >= iLimit) |
321 |
|
|
322 |
continue; |
continue; |
323 |
} |
} |
324 |
|
} |
325 |
|
pMB->mode = MODE_INTER; |
326 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mv16; |
327 |
|
pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = pMB->sad16; |
328 |
|
} |
329 |
|
|
330 |
|
// we try to do as few INTER4V-searches as possible. So we split ME in two parts, normal |
331 |
|
// SEARCH16 and only for special blocks SEARCH8. May this should be modified for quality |
332 |
|
// levels. |
333 |
|
|
|
if (pParam->global_flags & XVID_INTER4V) |
|
|
{ |
|
|
pMB->sad8[0] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
|
|
2 * j, 2 * i, mv16.x, mv16.y, pParam->motion_flags, |
|
|
pParam, pMBs, &pMB->mvs[0], &pMB->pmvs[0]); |
|
334 |
|
|
|
pMB->sad8[1] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
|
|
2 * j + 1, 2 * i, mv16.x, mv16.y, pParam->motion_flags, |
|
|
pParam, pMBs, &pMB->mvs[1], &pMB->pmvs[1]); |
|
335 |
|
|
336 |
pMB->sad8[2] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
if (current->global_flags & XVID_INTER4V) |
337 |
2 * j, 2 * i + 1, mv16.x, mv16.y, pParam->motion_flags, |
for (y = 0; y < iHcount; y++) |
338 |
pParam, pMBs, &pMB->mvs[2], &pMB->pmvs[2]); |
for (x = 0; x < iWcount; x++) |
339 |
|
{ |
340 |
|
MACROBLOCK* const pMB = &pMBs[x + y * iWcount]; |
341 |
|
|
342 |
|
if (pMB->mode == MODE_INTRA) |
343 |
|
continue; |
344 |
|
|
|
pMB->sad8[3] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
|
|
2 * j + 1, 2 * i + 1, mv16.x, mv16.y, pParam->motion_flags, |
|
|
pParam, pMBs, &pMB->mvs[3], &pMB->pmvs[3]); |
|
345 |
|
|
346 |
sad8 = pMB->sad8[0] + pMB->sad8[1] + pMB->sad8[2] + pMB->sad8[3]; |
if ( (!(current->global_flags & XVID_LUMIMASKING) || pMB->dquant == NO_CHANGE) ) |
347 |
} |
{ |
348 |
|
int32_t neigh=0; |
349 |
|
|
350 |
|
if (x>0) |
351 |
|
{ neigh += abs((pMB->mv16.x)-((pMB-1)->mv16.x)); |
352 |
|
neigh += abs((pMB->mv16.y)-((pMB-1)->mv16.y)); |
353 |
|
} |
354 |
|
if (y>0) |
355 |
|
{ neigh += abs((pMB->mv16.x)-((pMB-iWcount)->mv16.x)); |
356 |
|
neigh += abs((pMB->mv16.y)-((pMB-iWcount)->mv16.y)); |
357 |
|
} |
358 |
|
if (x<(iWcount-1)) |
359 |
|
{ neigh += abs((pMB->mv16.x)-((pMB+1)->mv16.x)); |
360 |
|
neigh += abs((pMB->mv16.y)-((pMB+1)->mv16.y)); |
361 |
|
} |
362 |
|
if (y<(iHcount-1)) |
363 |
|
{ neigh += abs((pMB->mv16.x)-((pMB+iHcount)->mv16.x)); |
364 |
|
neigh += abs((pMB->mv16.y)-((pMB+iHcount)->mv16.y)); |
365 |
|
} |
366 |
|
|
367 |
|
if (neigh > NEIGH_MOVE_THRESH) |
368 |
|
{ |
369 |
|
int32_t sad8 = IMV16X16 * current->quant; |
370 |
|
|
371 |
|
if (sad8 < pMB->sad16) |
372 |
|
sad8 += pMB->sad8[0] |
373 |
|
= SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
374 |
|
2*x, 2*y, pMB->mv16.x, pMB->mv16.y, |
375 |
|
current->motion_flags, current->quant, current->fcode, |
376 |
|
pParam, pMBs, prevMBs, &pMB->mvs[0], &pMB->pmvs[0]); |
377 |
|
|
378 |
|
if (sad8 < pMB->sad16) |
379 |
|
sad8 += pMB->sad8[1] |
380 |
|
= SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
381 |
|
2*x+1, 2*y, pMB->mv16.x, pMB->mv16.y, |
382 |
|
current->motion_flags, current->quant, current->fcode, |
383 |
|
pParam, pMBs, prevMBs, &pMB->mvs[1], &pMB->pmvs[1]); |
384 |
|
|
385 |
|
if (sad8 < pMB->sad16) |
386 |
|
sad8 += pMB->sad8[2] |
387 |
|
= SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
388 |
|
2*x, 2*y+1, pMB->mv16.x, pMB->mv16.y, |
389 |
|
current->motion_flags, current->quant, current->fcode, |
390 |
|
pParam, pMBs, prevMBs, &pMB->mvs[2], &pMB->pmvs[2]); |
391 |
|
|
392 |
|
if (sad8 < pMB->sad16) |
393 |
|
sad8 += pMB->sad8[3] |
394 |
|
= SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
395 |
|
2*x+1, 2*y+1, pMB->mv16.x, pMB->mv16.y, |
396 |
|
current->motion_flags, current->quant, current->fcode, |
397 |
|
pParam, pMBs, prevMBs, &pMB->mvs[3], &pMB->pmvs[3]); |
398 |
|
|
399 |
/* decide: MODE_INTER or MODE_INTER4V |
/* decide: MODE_INTER or MODE_INTER4V |
400 |
mpeg4: if (sad8 < sad16 - nb/2+1) use_inter4v |
mpeg4: if (sad8 < pMB->sad16 - nb/2+1) use_inter4v |
401 |
*/ |
*/ |
402 |
|
|
403 |
if (pMB->dquant == NO_CHANGE) { |
if (sad8 < pMB->sad16) |
|
if (((pParam->global_flags & XVID_INTER4V)==0) || |
|
|
(sad16 < (sad8 + (int32_t)(IMV16X16 * pParam->quant)))) { |
|
|
|
|
|
sad8 = sad16; |
|
|
pMB->mode = MODE_INTER; |
|
|
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; |
|
|
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; |
|
|
pMB->pmvs[0].x = pmv16.x; |
|
|
pMB->pmvs[0].y = pmv16.y; |
|
|
} |
|
|
else |
|
|
pMB->mode = MODE_INTER4V; |
|
|
} |
|
|
else |
|
404 |
{ |
{ |
405 |
sad8 = sad16; |
pMB->mode = MODE_INTER4V; |
406 |
pMB->mode = MODE_INTER; |
pMB->sad8[0] *= 4; |
407 |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; |
pMB->sad8[1] *= 4; |
408 |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; |
pMB->sad8[2] *= 4; |
409 |
pMB->pmvs[0].x = pmv16.x; |
pMB->sad8[3] *= 4; |
410 |
pMB->pmvs[0].y = pmv16.y; |
continue; |
411 |
} |
} |
412 |
|
|
413 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mv16; |
414 |
} |
} |
415 |
|
|
|
return 0; |
|
416 |
} |
} |
417 |
|
|
418 |
#define MVzero(A) ( ((A).x)==(0) && ((A).y)==(0) ) |
// get_pmv has to be called again, because inter4v changes predictors |
419 |
|
|
420 |
#define MVequal(A,B) ( ((A).x)==((B).x) && ((A).y)==((B).y) ) |
pmv = get_pmv(pMBs, x, y, pParam->mb_width, 0); |
421 |
|
pMB->pmvs[0].x = pMB->mv16.x - pmv.x; /* the other pmvs are only needed in INTER4V-mode */ |
422 |
|
pMB->pmvs[0].y = pMB->mv16.y - pmv.y; |
423 |
|
|
424 |
|
} |
425 |
|
|
426 |
|
return 0; |
427 |
|
} |
428 |
|
|
429 |
#define CHECK_MV16_ZERO {\ |
#define CHECK_MV16_ZERO {\ |
430 |
if ( (0 <= max_dx) && (0 >= min_dx) \ |
if ( (0 <= max_dx) && (0 >= min_dx) \ |
431 |
&& (0 <= max_dy) && (0 >= min_dy) ) \ |
&& (0 <= max_dy) && (0 >= min_dy) ) \ |
432 |
{ \ |
{ \ |
433 |
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); \ |
434 |
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; \ |
|
435 |
if (iSAD < iMinSAD) \ |
if (iSAD < iMinSAD) \ |
436 |
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } } \ |
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } } \ |
437 |
} |
} |
438 |
|
|
439 |
|
#define NOCHECK_MV16_CANDIDATE(X,Y) { \ |
440 |
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
441 |
|
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode, iQuant);\ |
442 |
|
if (iSAD < iMinSAD) \ |
443 |
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } \ |
444 |
|
} |
445 |
|
|
446 |
#define CHECK_MV16_CANDIDATE(X,Y) { \ |
#define CHECK_MV16_CANDIDATE(X,Y) { \ |
447 |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
448 |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
449 |
{ \ |
{ \ |
450 |
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); \ |
451 |
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);\ |
452 |
if (iSAD < iMinSAD) \ |
if (iSAD < iMinSAD) \ |
453 |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ |
454 |
} |
} |
458 |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
459 |
{ \ |
{ \ |
460 |
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); \ |
461 |
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);\ |
462 |
if (iSAD < iMinSAD) \ |
if (iSAD < iMinSAD) \ |
463 |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ |
464 |
} |
} |
468 |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
469 |
{ \ |
{ \ |
470 |
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); \ |
471 |
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);\ |
472 |
if (iSAD < iMinSAD) \ |
if (iSAD < iMinSAD) \ |
473 |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
474 |
} |
} |
476 |
|
|
477 |
#define CHECK_MV8_ZERO {\ |
#define CHECK_MV8_ZERO {\ |
478 |
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); \ |
479 |
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);\ |
480 |
if (iSAD < iMinSAD) \ |
if (iSAD < iMinSAD) \ |
481 |
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } \ |
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } \ |
482 |
} |
} |
483 |
|
|
484 |
|
#define NOCHECK_MV8_CANDIDATE(X,Y) \ |
485 |
|
{ \ |
486 |
|
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
487 |
|
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode, iQuant);\ |
488 |
|
if (iSAD < iMinSAD) \ |
489 |
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } \ |
490 |
|
} |
491 |
|
|
492 |
#define CHECK_MV8_CANDIDATE(X,Y) { \ |
#define CHECK_MV8_CANDIDATE(X,Y) { \ |
493 |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
494 |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
495 |
{ \ |
{ \ |
496 |
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); \ |
497 |
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);\ |
498 |
if (iSAD < iMinSAD) \ |
if (iSAD < iMinSAD) \ |
499 |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ |
500 |
} |
} |
504 |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
505 |
{ \ |
{ \ |
506 |
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); \ |
507 |
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);\ |
508 |
if (iSAD < iMinSAD) \ |
if (iSAD < iMinSAD) \ |
509 |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ |
510 |
} |
} |
514 |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
515 |
{ \ |
{ \ |
516 |
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); \ |
517 |
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);\ |
518 |
if (iSAD < iMinSAD) \ |
if (iSAD < iMinSAD) \ |
519 |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
520 |
} |
} |
529 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
530 |
const int x, const int y, |
const int x, const int y, |
531 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
532 |
|
const uint32_t iQuant, |
533 |
|
const uint32_t iFcode, |
534 |
MBParam * const pParam, |
MBParam * const pParam, |
535 |
MACROBLOCK * const pMBs, |
const MACROBLOCK * const pMBs, |
536 |
|
const MACROBLOCK * const prevMBs, |
537 |
VECTOR * const currMV, |
VECTOR * const currMV, |
538 |
VECTOR * const currPMV) |
VECTOR * const currPMV) |
539 |
{ |
{ |
540 |
const int32_t iEdgedWidth = pParam->edged_width; |
const int32_t iEdgedWidth = pParam->edged_width; |
|
const int32_t iQuant = pParam->quant; |
|
541 |
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
542 |
int32_t iSAD; |
int32_t iSAD; |
543 |
int32_t pred_x,pred_y; |
int32_t pred_x,pred_y; |
560 |
} |
} |
561 |
*/ |
*/ |
562 |
|
|
563 |
int32_t PMVfastSearch16_MainSearch( |
int32_t Diamond16_MainSearch( |
564 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
565 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
566 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
617 |
return iMinSAD; |
return iMinSAD; |
618 |
} |
} |
619 |
|
|
620 |
int32_t PMVfastSearch16_Refine( |
int32_t Square16_MainSearch( |
621 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
622 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
623 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
624 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
625 |
const uint8_t * const cur, |
const uint8_t * const cur, |
626 |
const int x, const int y, |
const int x, const int y, |
627 |
VECTOR * const currMV, |
int32_t startx, int32_t starty, |
628 |
int32_t iMinSAD, |
int32_t iMinSAD, |
629 |
|
VECTOR * const currMV, |
630 |
const VECTOR * const pmv, |
const VECTOR * const pmv, |
631 |
const int32_t min_dx, const int32_t max_dx, |
const int32_t min_dx, const int32_t max_dx, |
632 |
const int32_t min_dy, const int32_t max_dy, |
const int32_t min_dy, const int32_t max_dy, |
633 |
|
const int32_t iEdgedWidth, |
634 |
|
const int32_t iDiamondSize, |
635 |
const int32_t iFcode, |
const int32_t iFcode, |
636 |
const int32_t iQuant, |
const int32_t iQuant, |
637 |
const int32_t iEdgedWidth) |
int iFound) |
638 |
{ |
{ |
639 |
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
/* Do a square search around given starting point, return SAD of best */ |
640 |
|
|
641 |
|
int32_t iDirection=0; |
642 |
int32_t iSAD; |
int32_t iSAD; |
643 |
VECTOR backupMV = *currMV; |
VECTOR backupMV; |
644 |
|
backupMV.x = startx; |
645 |
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y-1); |
backupMV.y = starty; |
|
CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y-1); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y-1); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y+1); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y+1); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y+1); |
|
|
|
|
|
return iMinSAD; |
|
|
} |
|
|
|
|
|
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
|
646 |
|
|
647 |
int32_t PMVfastSearch16( |
/* It's one search with full square pattern, and new parts for all following diamonds */ |
|
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, |
|
|
MBParam * const pParam, |
|
|
MACROBLOCK * const pMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV) |
|
|
{ |
|
|
const uint32_t iWcount = pParam->mb_width; |
|
|
const int32_t iFcode = pParam->fixed_code; |
|
|
const int32_t iQuant = pParam->quant; |
|
|
const int32_t iWidth = pParam->width; |
|
|
const int32_t iHeight = pParam->height; |
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
648 |
|
|
649 |
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
/* new direction are extra, so 1-4 is normal diamond |
650 |
|
537 |
651 |
|
1*2 |
652 |
|
648 |
653 |
|
*/ |
654 |
|
|
655 |
int32_t iDiamondSize; |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
656 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
657 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
658 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
659 |
|
|
660 |
int32_t min_dx; |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
661 |
int32_t max_dx; |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
662 |
int32_t min_dy; |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
663 |
int32_t max_dy; |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
664 |
|
|
|
int32_t iFound; |
|
665 |
|
|
666 |
VECTOR newMV; |
if (iDirection) |
667 |
VECTOR backupMV; /* just for PMVFAST */ |
while (!iFound) |
668 |
|
{ |
669 |
|
iFound = 1; |
670 |
|
backupMV=*currMV; |
671 |
|
|
672 |
VECTOR pmv[4]; |
switch (iDirection) |
673 |
int32_t psad[4]; |
{ |
674 |
|
case 1: |
675 |
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
676 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
677 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
678 |
|
break; |
679 |
|
case 2: |
680 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
681 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
682 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
683 |
|
break; |
684 |
|
|
685 |
MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
case 3: |
686 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
687 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
688 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
689 |
|
break; |
690 |
|
|
691 |
static int32_t threshA,threshB; |
case 4: |
692 |
int32_t bPredEq; |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
693 |
int32_t iMinSAD,iSAD; |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
694 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
695 |
|
break; |
696 |
|
|
697 |
/* Get maximum range */ |
case 5: |
698 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
699 |
x, y, 16, iWidth, iHeight, iFcode); |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
700 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
701 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
702 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
703 |
|
break; |
704 |
|
|
705 |
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
case 6: |
706 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
707 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
708 |
|
|
709 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
710 |
{ min_dx = EVEN(min_dx); |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
711 |
max_dx = EVEN(max_dx); |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
|
min_dy = EVEN(min_dy); |
|
|
max_dy = EVEN(max_dy); |
|
|
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
|
712 |
|
|
713 |
|
break; |
714 |
|
|
715 |
bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
case 7: |
716 |
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
717 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
718 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
719 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
720 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
721 |
|
break; |
722 |
|
|
723 |
if ((x==0) && (y==0) ) |
case 8: |
724 |
{ |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
725 |
threshA = 512; |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
726 |
threshB = 1024; |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
727 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
728 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
729 |
|
break; |
730 |
|
default: |
731 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
732 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
733 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
734 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
735 |
|
|
736 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
737 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
738 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
739 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
740 |
|
break; |
741 |
|
} |
742 |
} |
} |
743 |
else |
else |
744 |
{ |
{ |
745 |
threshA = psad[0]; |
currMV->x = startx; |
746 |
threshB = threshA+256; |
currMV->y = starty; |
747 |
if (threshA< 512) threshA = 512; |
} |
748 |
if (threshA>1024) threshA = 1024; |
return iMinSAD; |
|
if (threshB>1792) threshB = 1792; |
|
749 |
} |
} |
750 |
|
|
|
iFound=0; |
|
|
|
|
|
/* 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[0]) ) ) |
|
|
iFound=2; |
|
|
|
|
|
/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
|
|
Otherwise select large Diamond Search. |
|
|
*/ |
|
751 |
|
|
752 |
if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536) || (bPredEq) ) |
int32_t Full16_MainSearch( |
753 |
iDiamondSize=1; // halfpel! |
const uint8_t * const pRef, |
754 |
else |
const uint8_t * const pRefH, |
755 |
iDiamondSize=2; // halfpel! |
const uint8_t * const pRefV, |
756 |
|
const uint8_t * const pRefHV, |
757 |
|
const uint8_t * const cur, |
758 |
|
const int x, const int y, |
759 |
|
int32_t startx, int32_t starty, |
760 |
|
int32_t iMinSAD, |
761 |
|
VECTOR * const currMV, |
762 |
|
const VECTOR * const pmv, |
763 |
|
const int32_t min_dx, const int32_t max_dx, |
764 |
|
const int32_t min_dy, const int32_t max_dy, |
765 |
|
const int32_t iEdgedWidth, |
766 |
|
const int32_t iDiamondSize, |
767 |
|
const int32_t iFcode, |
768 |
|
const int32_t iQuant, |
769 |
|
int iFound) |
770 |
|
{ |
771 |
|
int32_t iSAD; |
772 |
|
int32_t dx,dy; |
773 |
|
VECTOR backupMV; |
774 |
|
backupMV.x = startx; |
775 |
|
backupMV.y = starty; |
776 |
|
|
777 |
if (!(MotionFlags & PMV_HALFPELDIAMOND16) ) |
for (dx = min_dx; dx<=max_dx; dx+=iDiamondSize) |
778 |
iDiamondSize*=2; |
for (dy = min_dy; dy<= max_dy; dy+=iDiamondSize) |
779 |
|
NOCHECK_MV16_CANDIDATE(dx,dy); |
780 |
|
|
781 |
|
return iMinSAD; |
782 |
|
} |
783 |
|
|
784 |
|
int32_t Full8_MainSearch( |
785 |
|
const uint8_t * const pRef, |
786 |
|
const uint8_t * const pRefH, |
787 |
|
const uint8_t * const pRefV, |
788 |
|
const uint8_t * const pRefHV, |
789 |
|
const uint8_t * const cur, |
790 |
|
const int x, const int y, |
791 |
|
int32_t startx, int32_t starty, |
792 |
|
int32_t iMinSAD, |
793 |
|
VECTOR * const currMV, |
794 |
|
const VECTOR * const pmv, |
795 |
|
const int32_t min_dx, const int32_t max_dx, |
796 |
|
const int32_t min_dy, const int32_t max_dy, |
797 |
|
const int32_t iEdgedWidth, |
798 |
|
const int32_t iDiamondSize, |
799 |
|
const int32_t iFcode, |
800 |
|
const int32_t iQuant, |
801 |
|
int iFound) |
802 |
|
{ |
803 |
|
int32_t iSAD; |
804 |
|
int32_t dx,dy; |
805 |
|
VECTOR backupMV; |
806 |
|
backupMV.x = startx; |
807 |
|
backupMV.y = starty; |
808 |
|
|
809 |
|
for (dx = min_dx; dx<=max_dx; dx+=iDiamondSize) |
810 |
|
for (dy = min_dy; dy<= max_dy; dy+=iDiamondSize) |
811 |
|
NOCHECK_MV8_CANDIDATE(dx,dy); |
812 |
|
|
813 |
|
return iMinSAD; |
814 |
|
} |
815 |
|
|
816 |
|
|
817 |
|
|
818 |
|
int32_t Halfpel16_Refine( |
819 |
|
const uint8_t * const pRef, |
820 |
|
const uint8_t * const pRefH, |
821 |
|
const uint8_t * const pRefV, |
822 |
|
const uint8_t * const pRefHV, |
823 |
|
const uint8_t * const cur, |
824 |
|
const int x, const int y, |
825 |
|
VECTOR * const currMV, |
826 |
|
int32_t iMinSAD, |
827 |
|
const VECTOR * const pmv, |
828 |
|
const int32_t min_dx, const int32_t max_dx, |
829 |
|
const int32_t min_dy, const int32_t max_dy, |
830 |
|
const int32_t iFcode, |
831 |
|
const int32_t iQuant, |
832 |
|
const int32_t iEdgedWidth) |
833 |
|
{ |
834 |
|
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
835 |
|
|
836 |
|
int32_t iSAD; |
837 |
|
VECTOR backupMV = *currMV; |
838 |
|
|
839 |
|
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y-1); |
840 |
|
CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y-1); |
841 |
|
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y-1); |
842 |
|
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y); |
843 |
|
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y); |
844 |
|
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y+1); |
845 |
|
CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y+1); |
846 |
|
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y+1); |
847 |
|
|
848 |
|
return iMinSAD; |
849 |
|
} |
850 |
|
|
851 |
|
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
852 |
|
|
853 |
|
|
854 |
|
int32_t PMVfastSearch16( |
855 |
|
const uint8_t * const pRef, |
856 |
|
const uint8_t * const pRefH, |
857 |
|
const uint8_t * const pRefV, |
858 |
|
const uint8_t * const pRefHV, |
859 |
|
const IMAGE * const pCur, |
860 |
|
const int x, const int y, |
861 |
|
const uint32_t MotionFlags, |
862 |
|
const uint32_t iQuant, |
863 |
|
const uint32_t iFcode, |
864 |
|
const MBParam * const pParam, |
865 |
|
const MACROBLOCK * const pMBs, |
866 |
|
const MACROBLOCK * const prevMBs, |
867 |
|
VECTOR * const currMV, |
868 |
|
VECTOR * const currPMV) |
869 |
|
{ |
870 |
|
const uint32_t iWcount = pParam->mb_width; |
871 |
|
const int32_t iWidth = pParam->width; |
872 |
|
const int32_t iHeight = pParam->height; |
873 |
|
const int32_t iEdgedWidth = pParam->edged_width; |
874 |
|
|
875 |
|
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
876 |
|
|
877 |
|
int32_t iDiamondSize; |
878 |
|
|
879 |
|
int32_t min_dx; |
880 |
|
int32_t max_dx; |
881 |
|
int32_t min_dy; |
882 |
|
int32_t max_dy; |
883 |
|
|
884 |
|
int32_t iFound; |
885 |
|
|
886 |
|
VECTOR newMV; |
887 |
|
VECTOR backupMV; /* just for PMVFAST */ |
888 |
|
|
889 |
|
VECTOR pmv[4]; |
890 |
|
int32_t psad[4]; |
891 |
|
|
892 |
|
// const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
893 |
|
const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; |
894 |
|
|
895 |
|
static int32_t threshA,threshB; |
896 |
|
int32_t bPredEq; |
897 |
|
int32_t iMinSAD,iSAD; |
898 |
|
|
899 |
|
/* Get maximum range */ |
900 |
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
901 |
|
x, y, 16, iWidth, iHeight, iFcode); |
902 |
|
|
903 |
|
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
904 |
|
|
905 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
906 |
|
{ min_dx = EVEN(min_dx); |
907 |
|
max_dx = EVEN(max_dx); |
908 |
|
min_dy = EVEN(min_dy); |
909 |
|
max_dy = EVEN(max_dy); |
910 |
|
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
911 |
|
|
912 |
|
|
913 |
|
bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
914 |
|
|
915 |
|
if ((x==0) && (y==0) ) |
916 |
|
{ |
917 |
|
threshA = 512; |
918 |
|
threshB = 1024; |
919 |
|
|
920 |
|
} |
921 |
|
else |
922 |
|
{ |
923 |
|
threshA = psad[0]; |
924 |
|
threshB = threshA+256; |
925 |
|
if (threshA< 512) threshA = 512; |
926 |
|
if (threshA>1024) threshA = 1024; |
927 |
|
if (threshB>1792) threshB = 1792; |
928 |
|
} |
929 |
|
|
930 |
|
iFound=0; |
931 |
|
|
932 |
/* Step 4: Calculate SAD around the Median prediction. |
/* Step 4: Calculate SAD around the Median prediction. |
933 |
MinSAD=SAD |
MinSAD=SAD |
936 |
If SAD<=256 goto Step 10. |
If SAD<=256 goto Step 10. |
937 |
*/ |
*/ |
938 |
|
|
|
|
|
|
// Prepare for main loop |
|
|
|
|
939 |
*currMV=pmv[0]; /* current best := prediction */ |
*currMV=pmv[0]; /* current best := prediction */ |
940 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
if (!(MotionFlags & PMV_HALFPEL16 )) |
941 |
{ /* This should NOT be necessary! */ |
{ /* This should NOT be necessary! */ |
963 |
iMinSAD = sad16( cur, |
iMinSAD = sad16( cur, |
964 |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
965 |
iEdgedWidth, MV_MAX_ERROR); |
iEdgedWidth, MV_MAX_ERROR); |
966 |
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); |
967 |
|
|
968 |
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,pMB->mvs[0])) && (iMinSAD < pMB->sad16) ) ) |
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,prevMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
969 |
|
{ |
970 |
|
if (iMinSAD < 2*iQuant) // high chances for SKIP-mode |
971 |
|
{ |
972 |
|
if (!MVzero(*currMV)) |
973 |
{ |
{ |
974 |
|
iMinSAD += MV16_00_BIAS; |
975 |
|
CHECK_MV16_ZERO; // (0,0) saves space for letterboxed pictures |
976 |
|
iMinSAD -= MV16_00_BIAS; |
977 |
|
} |
978 |
|
} |
979 |
|
|
980 |
if (MotionFlags & PMV_QUICKSTOP16) |
if (MotionFlags & PMV_QUICKSTOP16) |
981 |
goto step10b; |
goto PMVfast16_Terminate_without_Refine; |
982 |
if (MotionFlags & PMV_EARLYSTOP16) |
if (MotionFlags & PMV_EARLYSTOP16) |
983 |
goto step10; |
goto PMVfast16_Terminate_with_Refine; |
984 |
} |
} |
985 |
|
|
986 |
|
|
987 |
|
/* Step 2 (lazy eval): Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
988 |
|
vector of the median. |
989 |
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
990 |
|
*/ |
991 |
|
|
992 |
|
if ((bPredEq) && (MVequal(pmv[0],prevMB->mvs[0]) ) ) |
993 |
|
iFound=2; |
994 |
|
|
995 |
|
/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
996 |
|
Otherwise select large Diamond Search. |
997 |
|
*/ |
998 |
|
|
999 |
|
if ( (!MVzero(pmv[0])) || (threshB<1536) || (bPredEq) ) |
1000 |
|
iDiamondSize=1; // halfpel! |
1001 |
|
else |
1002 |
|
iDiamondSize=2; // halfpel! |
1003 |
|
|
1004 |
|
if (!(MotionFlags & PMV_HALFPELDIAMOND16) ) |
1005 |
|
iDiamondSize*=2; |
1006 |
|
|
1007 |
/* |
/* |
1008 |
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. |
1009 |
Also calculate (0,0) but do not subtract offset. |
Also calculate (0,0) but do not subtract offset. |
1010 |
Let MinSAD be the smallest SAD up to this point. |
Let MinSAD be the smallest SAD up to this point. |
1011 |
If MV is (0,0) subtract offset. ******** WHAT'S THIS 'OFFSET' ??? *********** |
If MV is (0,0) subtract offset. |
1012 |
*/ |
*/ |
1013 |
|
|
1014 |
// (0,0) is always possible |
// (0,0) is always possible |
1015 |
|
|
1016 |
|
if (!MVzero(pmv[0])) |
1017 |
CHECK_MV16_ZERO; |
CHECK_MV16_ZERO; |
1018 |
|
|
1019 |
// previous frame MV is always possible |
// previous frame MV is always possible |
1020 |
CHECK_MV16_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); |
|
1021 |
|
if (!MVzero(prevMB->mvs[0])) |
1022 |
|
if (!MVequal(prevMB->mvs[0],pmv[0])) |
1023 |
|
CHECK_MV16_CANDIDATE(prevMB->mvs[0].x,prevMB->mvs[0].y); |
1024 |
|
|
1025 |
// left neighbour, if allowed |
// left neighbour, if allowed |
1026 |
if (x != 0) |
|
1027 |
|
if (!MVzero(pmv[1])) |
1028 |
|
if (!MVequal(pmv[1],prevMB->mvs[0])) |
1029 |
|
if (!MVequal(pmv[1],pmv[0])) |
1030 |
{ |
{ |
1031 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
if (!(MotionFlags & PMV_HALFPEL16 )) |
1032 |
{ pmv[1].x = EVEN(pmv[1].x); |
{ pmv[1].x = EVEN(pmv[1].x); |
1033 |
pmv[1].y = EVEN(pmv[1].y); |
pmv[1].y = EVEN(pmv[1].y); |
1034 |
} |
} |
1035 |
|
|
1036 |
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
1037 |
} |
} |
1038 |
|
|
1039 |
// top neighbour, if allowed |
// top neighbour, if allowed |
1040 |
if (y != 0) |
if (!MVzero(pmv[2])) |
1041 |
|
if (!MVequal(pmv[2],prevMB->mvs[0])) |
1042 |
|
if (!MVequal(pmv[2],pmv[0])) |
1043 |
|
if (!MVequal(pmv[2],pmv[1])) |
1044 |
{ |
{ |
1045 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
if (!(MotionFlags & PMV_HALFPEL16 )) |
1046 |
{ pmv[2].x = EVEN(pmv[2].x); |
{ pmv[2].x = EVEN(pmv[2].x); |
1049 |
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
1050 |
|
|
1051 |
// top right neighbour, if allowed |
// top right neighbour, if allowed |
1052 |
if (x != (iWcount-1)) |
if (!MVzero(pmv[3])) |
1053 |
|
if (!MVequal(pmv[3],prevMB->mvs[0])) |
1054 |
|
if (!MVequal(pmv[3],pmv[0])) |
1055 |
|
if (!MVequal(pmv[3],pmv[1])) |
1056 |
|
if (!MVequal(pmv[3],pmv[2])) |
1057 |
{ |
{ |
1058 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
if (!(MotionFlags & PMV_HALFPEL16 )) |
1059 |
{ pmv[3].x = EVEN(pmv[3].x); |
{ pmv[3].x = EVEN(pmv[3].x); |
1063 |
} |
} |
1064 |
} |
} |
1065 |
|
|
1066 |
|
if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96)*/ ) |
1067 |
|
iMinSAD -= MV16_00_BIAS; |
1068 |
|
|
1069 |
|
|
1070 |
/* Step 6: If MinSAD <= thresa goto Step 10. |
/* Step 6: If MinSAD <= thresa goto Step 10. |
1071 |
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. |
1072 |
*/ |
*/ |
1073 |
|
|
1074 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,pMB->mvs[0]) && (iMinSAD < pMB->sad16) ) ) |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,prevMB->mvs[0]) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
1075 |
{ |
{ |
1076 |
if (MotionFlags & PMV_QUICKSTOP16) |
if (MotionFlags & PMV_QUICKSTOP16) |
1077 |
goto step10b; |
goto PMVfast16_Terminate_without_Refine; |
1078 |
if (MotionFlags & PMV_EARLYSTOP16) |
if (MotionFlags & PMV_EARLYSTOP16) |
1079 |
goto step10; |
goto PMVfast16_Terminate_with_Refine; |
1080 |
} |
} |
1081 |
|
|
1082 |
|
|
1093 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
1094 |
|
|
1095 |
/* 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 */ |
1096 |
iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1097 |
x, y, |
x, y, |
1098 |
currMV->x, currMV->y, iMinSAD, &newMV, |
currMV->x, currMV->y, iMinSAD, &newMV, |
1099 |
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); |
1109 |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
1110 |
|
|
1111 |
if (!(MVequal(pmv[0],backupMV)) ) |
if (!(MVequal(pmv[0],backupMV)) ) |
1112 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1113 |
x, y, |
x, y, |
1114 |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
1115 |
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); |
1122 |
} |
} |
1123 |
|
|
1124 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
1125 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1126 |
x, y, |
x, y, |
1127 |
0, 0, iMinSAD, &newMV, |
0, 0, iMinSAD, &newMV, |
1128 |
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); |
1139 |
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. |
1140 |
*/ |
*/ |
1141 |
|
|
1142 |
step10: |
PMVfast16_Terminate_with_Refine: |
1143 |
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
1144 |
iMinSAD = PMVfastSearch16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
iMinSAD = Halfpel16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
1145 |
x, y, |
x, y, |
1146 |
currMV, iMinSAD, |
currMV, iMinSAD, |
1147 |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
1148 |
|
|
1149 |
step10b: |
PMVfast16_Terminate_without_Refine: |
1150 |
currPMV->x = currMV->x - pmv[0].x; |
currPMV->x = currMV->x - pmv[0].x; |
1151 |
currPMV->y = currMV->y - pmv[0].y; |
currPMV->y = currMV->y - pmv[0].y; |
1152 |
return iMinSAD; |
return iMinSAD; |
1157 |
|
|
1158 |
|
|
1159 |
|
|
1160 |
int32_t PMVfastSearch8_MainSearch( |
int32_t Diamond8_MainSearch( |
1161 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
1162 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1163 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
1214 |
return iMinSAD; |
return iMinSAD; |
1215 |
} |
} |
1216 |
|
|
1217 |
int32_t PMVfastSearch8_Refine( |
int32_t Halfpel8_Refine( |
1218 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
1219 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1220 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
1257 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
1258 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
1259 |
const int x, const int y, |
const int x, const int y, |
1260 |
const int start_x, int start_y, |
const int start_x, const int start_y, |
1261 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
1262 |
MBParam * const pParam, |
const uint32_t iQuant, |
1263 |
MACROBLOCK * const pMBs, |
const uint32_t iFcode, |
1264 |
|
const MBParam * const pParam, |
1265 |
|
const MACROBLOCK * const pMBs, |
1266 |
|
const MACROBLOCK * const prevMBs, |
1267 |
VECTOR * const currMV, |
VECTOR * const currMV, |
1268 |
VECTOR * const currPMV) |
VECTOR * const currPMV) |
1269 |
{ |
{ |
1270 |
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; |
|
1271 |
const int32_t iWidth = pParam->width; |
const int32_t iWidth = pParam->width; |
1272 |
const int32_t iHeight = pParam->height; |
const int32_t iHeight = pParam->height; |
1273 |
const int32_t iEdgedWidth = pParam->edged_width; |
const int32_t iEdgedWidth = pParam->edged_width; |
1285 |
int32_t psad[4]; |
int32_t psad[4]; |
1286 |
VECTOR newMV; |
VECTOR newMV; |
1287 |
VECTOR backupMV; |
VECTOR backupMV; |
1288 |
|
VECTOR startMV; |
1289 |
|
|
1290 |
MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
1291 |
|
const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; |
1292 |
|
|
1293 |
static int32_t threshA,threshB; |
static int32_t threshA,threshB; |
1294 |
int32_t iFound,bPredEq; |
int32_t iFound,bPredEq; |
1295 |
int32_t iMinSAD,iSAD; |
int32_t iMinSAD,iSAD; |
1296 |
|
|
1297 |
int32_t iSubBlock = ((y&1)<<1) + (x&1); |
int32_t iSubBlock = (y&1)+(y&1) + (x&1); |
1298 |
|
|
1299 |
|
/* Init variables */ |
1300 |
|
startMV.x = start_x; |
1301 |
|
startMV.y = start_y; |
1302 |
|
|
1303 |
/* Get maximum range */ |
/* Get maximum range */ |
1304 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
1305 |
x, y, 8, iWidth, iHeight, iFcode); |
x, y, 8, iWidth, iHeight, iFcode); |
1306 |
|
|
|
/* we work with abs. MVs, not relative to prediction, so range is relative to 0,0 */ |
|
|
|
|
1307 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8 )) |
if (!(MotionFlags & PMV_HALFPELDIAMOND8 )) |
1308 |
{ min_dx = EVEN(min_dx); |
{ min_dx = EVEN(min_dx); |
1309 |
max_dx = EVEN(max_dx); |
max_dx = EVEN(max_dx); |
1331 |
|
|
1332 |
iFound=0; |
iFound=0; |
1333 |
|
|
|
/* 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; |
|
|
|
|
1334 |
/* Step 4: Calculate SAD around the Median prediction. |
/* Step 4: Calculate SAD around the Median prediction. |
1335 |
MinSAD=SAD |
MinSAD=SAD |
1336 |
If Motion Vector equal to Previous frame motion vector |
If Motion Vector equal to Previous frame motion vector |
1341 |
|
|
1342 |
// Prepare for main loop |
// Prepare for main loop |
1343 |
|
|
1344 |
currMV->x=start_x; /* start with mv16 */ |
*currMV = startMV; |
|
currMV->y=start_y; |
|
1345 |
|
|
1346 |
iMinSAD = sad8( cur, |
iMinSAD = sad8( cur, |
1347 |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
1348 |
iEdgedWidth); |
iEdgedWidth); |
1349 |
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); |
1350 |
|
|
1351 |
if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,pMB->mvs[iSubBlock])) && (iMinSAD < pMB->sad8[iSubBlock]) ) ) |
if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,prevMB->mvs[iSubBlock])) |
1352 |
|
&& ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
1353 |
{ |
{ |
1354 |
if (MotionFlags & PMV_QUICKSTOP8) |
if (MotionFlags & PMV_QUICKSTOP16) |
1355 |
goto step10_8b; |
goto PMVfast8_Terminate_without_Refine; |
1356 |
if (MotionFlags & PMV_EARLYSTOP8) |
if (MotionFlags & PMV_EARLYSTOP16) |
1357 |
goto step10_8; |
goto PMVfast8_Terminate_with_Refine; |
1358 |
} |
} |
1359 |
|
|
1360 |
|
/* Step 2 (lazy eval): Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
1361 |
|
vector of the median. |
1362 |
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
1363 |
|
*/ |
1364 |
|
|
1365 |
|
if ((bPredEq) && (MVequal(pmv[0],prevMB->mvs[iSubBlock]) ) ) |
1366 |
|
iFound=2; |
1367 |
|
|
1368 |
|
/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
1369 |
|
Otherwise select large Diamond Search. |
1370 |
|
*/ |
1371 |
|
|
1372 |
|
if ( (!MVzero(pmv[0])) || (threshB<1536/4) || (bPredEq) ) |
1373 |
|
iDiamondSize=1; // 1 halfpel! |
1374 |
|
else |
1375 |
|
iDiamondSize=2; // 2 halfpel = 1 full pixel! |
1376 |
|
|
1377 |
|
if (!(MotionFlags & PMV_HALFPELDIAMOND8) ) |
1378 |
|
iDiamondSize*=2; |
1379 |
|
|
1380 |
|
|
1381 |
/* |
/* |
1382 |
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. |
1383 |
Also calculate (0,0) but do not subtract offset. |
Also calculate (0,0) but do not subtract offset. |
1384 |
Let MinSAD be the smallest SAD up to this point. |
Let MinSAD be the smallest SAD up to this point. |
1385 |
If MV is (0,0) subtract offset. ******** WHAT'S THIS 'OFFSET' ??? *********** |
If MV is (0,0) subtract offset. |
1386 |
*/ |
*/ |
1387 |
|
|
1388 |
// the prediction might be even better than mv16 |
// the median prediction might be even better than mv16 |
1389 |
|
|
1390 |
|
if (!MVequal(pmv[0],startMV)) |
1391 |
CHECK_MV8_CANDIDATE(pmv[0].x,pmv[0].y); |
CHECK_MV8_CANDIDATE(pmv[0].x,pmv[0].y); |
1392 |
|
|
1393 |
// (0,0) is always possible |
// (0,0) if needed |
1394 |
|
if (!MVzero(pmv[0])) |
1395 |
|
if (!MVzero(startMV)) |
1396 |
CHECK_MV8_ZERO; |
CHECK_MV8_ZERO; |
1397 |
|
|
1398 |
// previous frame MV is always possible |
// previous frame MV if needed |
1399 |
CHECK_MV8_CANDIDATE(pMB->mvs[iSubBlock].x,pMB->mvs[iSubBlock].y); |
if (!MVzero(prevMB->mvs[iSubBlock])) |
1400 |
|
if (!MVequal(prevMB->mvs[iSubBlock],startMV)) |
1401 |
|
if (!MVequal(prevMB->mvs[iSubBlock],pmv[0])) |
1402 |
|
CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x,prevMB->mvs[iSubBlock].y); |
1403 |
|
|
1404 |
// left neighbour, if allowed |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,prevMB->mvs[iSubBlock]) && ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
1405 |
if (psad[1] != MV_MAX_ERROR) |
{ |
1406 |
|
if (MotionFlags & PMV_QUICKSTOP16) |
1407 |
|
goto PMVfast8_Terminate_without_Refine; |
1408 |
|
if (MotionFlags & PMV_EARLYSTOP16) |
1409 |
|
goto PMVfast8_Terminate_with_Refine; |
1410 |
|
} |
1411 |
|
|
1412 |
|
|
1413 |
|
// left neighbour, if allowed and needed |
1414 |
|
if (!MVzero(pmv[1])) |
1415 |
|
if (!MVequal(pmv[1],startMV)) |
1416 |
|
if (!MVequal(pmv[1],prevMB->mvs[iSubBlock])) |
1417 |
|
if (!MVequal(pmv[1],pmv[0])) |
1418 |
{ |
{ |
1419 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
if (!(MotionFlags & PMV_HALFPEL8 )) |
1420 |
{ pmv[1].x = EVEN(pmv[1].x); |
{ pmv[1].x = EVEN(pmv[1].x); |
1423 |
CHECK_MV8_CANDIDATE(pmv[1].x,pmv[1].y); |
CHECK_MV8_CANDIDATE(pmv[1].x,pmv[1].y); |
1424 |
} |
} |
1425 |
|
|
1426 |
// top neighbour, if allowed |
// top neighbour, if allowed and needed |
1427 |
if (psad[2] != MV_MAX_ERROR) |
if (!MVzero(pmv[2])) |
1428 |
|
if (!MVequal(pmv[2],startMV)) |
1429 |
|
if (!MVequal(pmv[2],prevMB->mvs[iSubBlock])) |
1430 |
|
if (!MVequal(pmv[2],pmv[0])) |
1431 |
|
if (!MVequal(pmv[2],pmv[1])) |
1432 |
{ |
{ |
1433 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
if (!(MotionFlags & PMV_HALFPEL8 )) |
1434 |
{ pmv[2].x = EVEN(pmv[2].x); |
{ pmv[2].x = EVEN(pmv[2].x); |
1436 |
} |
} |
1437 |
CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); |
CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); |
1438 |
|
|
1439 |
// top right neighbour, if allowed |
// top right neighbour, if allowed and needed |
1440 |
if (psad[3] != MV_MAX_ERROR) |
if (!MVzero(pmv[3])) |
1441 |
|
if (!MVequal(pmv[3],startMV)) |
1442 |
|
if (!MVequal(pmv[3],prevMB->mvs[iSubBlock])) |
1443 |
|
if (!MVequal(pmv[3],pmv[0])) |
1444 |
|
if (!MVequal(pmv[3],pmv[1])) |
1445 |
|
if (!MVequal(pmv[3],pmv[2])) |
1446 |
{ |
{ |
1447 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
if (!(MotionFlags & PMV_HALFPEL8 )) |
1448 |
{ pmv[3].x = EVEN(pmv[3].x); |
{ pmv[3].x = EVEN(pmv[3].x); |
1452 |
} |
} |
1453 |
} |
} |
1454 |
|
|
1455 |
|
if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96) */ ) |
1456 |
|
iMinSAD -= MV8_00_BIAS; |
1457 |
|
|
1458 |
|
|
1459 |
/* Step 6: If MinSAD <= thresa goto Step 10. |
/* Step 6: If MinSAD <= thresa goto Step 10. |
1460 |
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. |
1461 |
*/ |
*/ |
1462 |
|
|
1463 |
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]) ) ) |
1464 |
{ |
{ |
1465 |
if (MotionFlags & PMV_QUICKSTOP8) |
if (MotionFlags & PMV_QUICKSTOP16) |
1466 |
goto step10_8b; |
goto PMVfast8_Terminate_without_Refine; |
1467 |
if (MotionFlags & PMV_EARLYSTOP8) |
if (MotionFlags & PMV_EARLYSTOP16) |
1468 |
goto step10_8; |
goto PMVfast8_Terminate_with_Refine; |
1469 |
} |
} |
1470 |
|
|
1471 |
/************ (Diamond Search) **************/ |
/************ (Diamond Search) **************/ |
1481 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
1482 |
|
|
1483 |
/* 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 */ |
1484 |
iSAD = PMVfastSearch8_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
iSAD = Diamond8_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1485 |
x, y, |
x, y, |
1486 |
currMV->x, currMV->y, iMinSAD, &newMV, |
currMV->x, currMV->y, iMinSAD, &newMV, |
1487 |
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); |
1497 |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
1498 |
|
|
1499 |
if (!(MVequal(pmv[0],backupMV)) ) |
if (!(MVequal(pmv[0],backupMV)) ) |
1500 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1501 |
x, y, |
x, y, |
1502 |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
1503 |
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); |
1510 |
} |
} |
1511 |
|
|
1512 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
1513 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1514 |
x, y, |
x, y, |
1515 |
0, 0, iMinSAD, &newMV, |
0, 0, iMinSAD, &newMV, |
1516 |
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); |
1527 |
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. |
1528 |
*/ |
*/ |
1529 |
|
|
1530 |
step10_8: |
PMVfast8_Terminate_with_Refine: |
1531 |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
1532 |
iMinSAD = PMVfastSearch8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
iMinSAD = Halfpel8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
1533 |
x, y, |
x, y, |
1534 |
currMV, iMinSAD, |
currMV, iMinSAD, |
1535 |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
1536 |
|
|
|
step10_8b: |
|
1537 |
|
|
1538 |
|
PMVfast8_Terminate_without_Refine: |
1539 |
currPMV->x = currMV->x - pmv[0].x; |
currPMV->x = currMV->x - pmv[0].x; |
1540 |
currPMV->y = currMV->y - pmv[0].y; |
currPMV->y = currMV->y - pmv[0].y; |
1541 |
|
|
1542 |
return iMinSAD; |
return iMinSAD; |
1543 |
} |
} |
1544 |
|
|
1545 |
|
int32_t EPZSSearch16( |
1546 |
|
const uint8_t * const pRef, |
1547 |
|
const uint8_t * const pRefH, |
1548 |
|
const uint8_t * const pRefV, |
1549 |
|
const uint8_t * const pRefHV, |
1550 |
|
const IMAGE * const pCur, |
1551 |
|
const int x, const int y, |
1552 |
|
const uint32_t MotionFlags, |
1553 |
|
const uint32_t iQuant, |
1554 |
|
const uint32_t iFcode, |
1555 |
|
const MBParam * const pParam, |
1556 |
|
const MACROBLOCK * const pMBs, |
1557 |
|
const MACROBLOCK * const prevMBs, |
1558 |
|
VECTOR * const currMV, |
1559 |
|
VECTOR * const currPMV) |
1560 |
|
{ |
1561 |
|
const uint32_t iWcount = pParam->mb_width; |
1562 |
|
const uint32_t iHcount = pParam->mb_height; |
1563 |
|
|
1564 |
|
const int32_t iWidth = pParam->width; |
1565 |
|
const int32_t iHeight = pParam->height; |
1566 |
|
const int32_t iEdgedWidth = pParam->edged_width; |
1567 |
|
|
1568 |
|
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
1569 |
|
|
1570 |
|
int32_t min_dx; |
1571 |
|
int32_t max_dx; |
1572 |
|
int32_t min_dy; |
1573 |
|
int32_t max_dy; |
1574 |
|
|
1575 |
|
VECTOR newMV; |
1576 |
|
VECTOR backupMV; |
1577 |
|
|
1578 |
|
VECTOR pmv[4]; |
1579 |
|
int32_t psad[8]; |
1580 |
|
|
1581 |
|
static MACROBLOCK * oldMBs = NULL; |
1582 |
|
// const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
1583 |
|
const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; |
1584 |
|
MACROBLOCK * oldMB = NULL; |
1585 |
|
|
1586 |
|
static int32_t thresh2; |
1587 |
|
int32_t bPredEq; |
1588 |
|
int32_t iMinSAD,iSAD=9999; |
1589 |
|
|
1590 |
|
MainSearch16FuncPtr EPZSMainSearchPtr; |
1591 |
|
|
1592 |
|
if (oldMBs == NULL) |
1593 |
|
{ oldMBs = (MACROBLOCK*) calloc(iWcount*iHcount,sizeof(MACROBLOCK)); |
1594 |
|
// fprintf(stderr,"allocated %d bytes for oldMBs\n",iWcount*iHcount*sizeof(MACROBLOCK)); |
1595 |
|
} |
1596 |
|
oldMB = oldMBs + x + y * iWcount; |
1597 |
|
|
1598 |
|
/* Get maximum range */ |
1599 |
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
1600 |
|
x, y, 16, iWidth, iHeight, iFcode); |
1601 |
|
|
1602 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1603 |
|
{ min_dx = EVEN(min_dx); |
1604 |
|
max_dx = EVEN(max_dx); |
1605 |
|
min_dy = EVEN(min_dy); |
1606 |
|
max_dy = EVEN(max_dy); |
1607 |
|
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
1608 |
|
|
1609 |
|
bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
1610 |
|
|
1611 |
|
/* Step 4: Calculate SAD around the Median prediction. |
1612 |
|
MinSAD=SAD |
1613 |
|
If Motion Vector equal to Previous frame motion vector |
1614 |
|
and MinSAD<PrevFrmSAD goto Step 10. |
1615 |
|
If SAD<=256 goto Step 10. |
1616 |
|
*/ |
1617 |
|
|
1618 |
|
// Prepare for main loop |
1619 |
|
|
1620 |
|
*currMV=pmv[0]; /* current best := median prediction */ |
1621 |
|
if (!(MotionFlags & PMV_HALFPEL16)) |
1622 |
|
{ |
1623 |
|
currMV->x = EVEN(currMV->x); |
1624 |
|
currMV->y = EVEN(currMV->y); |
1625 |
|
} |
1626 |
|
|
1627 |
|
if (currMV->x > max_dx) |
1628 |
|
currMV->x=max_dx; |
1629 |
|
if (currMV->x < min_dx) |
1630 |
|
currMV->x=min_dx; |
1631 |
|
if (currMV->y > max_dy) |
1632 |
|
currMV->y=max_dy; |
1633 |
|
if (currMV->y < min_dy) |
1634 |
|
currMV->y=min_dy; |
1635 |
|
|
1636 |
|
/***************** This is predictor SET A: only median prediction ******************/ |
1637 |
|
|
1638 |
|
iMinSAD = sad16( cur, |
1639 |
|
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
1640 |
|
iEdgedWidth, MV_MAX_ERROR); |
1641 |
|
iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode, iQuant); |
1642 |
|
|
1643 |
|
// thresh1 is fixed to 256 |
1644 |
|
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV, prevMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
1645 |
|
{ |
1646 |
|
if (MotionFlags & PMV_QUICKSTOP16) |
1647 |
|
goto EPZS16_Terminate_without_Refine; |
1648 |
|
if (MotionFlags & PMV_EARLYSTOP16) |
1649 |
|
goto EPZS16_Terminate_with_Refine; |
1650 |
|
} |
1651 |
|
|
1652 |
|
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
1653 |
|
|
1654 |
|
// previous frame MV |
1655 |
|
CHECK_MV16_CANDIDATE(prevMB->mvs[0].x,prevMB->mvs[0].y); |
1656 |
|
|
1657 |
|
// set threshhold based on Min of Prediction and SAD of collocated block |
1658 |
|
// CHECK_MV16 always uses iSAD for the SAD of last vector to check, so now iSAD is what we want |
1659 |
|
|
1660 |
|
if ((x==0) && (y==0) ) |
1661 |
|
{ |
1662 |
|
thresh2 = 512; |
1663 |
|
} |
1664 |
|
else |
1665 |
|
{ |
1666 |
|
/* T_k = 1.2 * MIN(SAD_top,SAD_left,SAD_topleft,SAD_coll) +128; [Tourapis, 2002] */ |
1667 |
|
|
1668 |
|
thresh2 = MIN(psad[0],iSAD)*6/5 + 128; |
1669 |
|
} |
1670 |
|
|
1671 |
|
// MV=(0,0) is often a good choice |
1672 |
|
|
1673 |
|
CHECK_MV16_ZERO; |
1674 |
|
|
1675 |
|
|
1676 |
|
// left neighbour, if allowed |
1677 |
|
if (x != 0) |
1678 |
|
{ |
1679 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1680 |
|
{ pmv[1].x = EVEN(pmv[1].x); |
1681 |
|
pmv[1].y = EVEN(pmv[1].y); |
1682 |
|
} |
1683 |
|
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
1684 |
|
} |
1685 |
|
|
1686 |
|
// top neighbour, if allowed |
1687 |
|
if (y != 0) |
1688 |
|
{ |
1689 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1690 |
|
{ pmv[2].x = EVEN(pmv[2].x); |
1691 |
|
pmv[2].y = EVEN(pmv[2].y); |
1692 |
|
} |
1693 |
|
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
1694 |
|
|
1695 |
|
// top right neighbour, if allowed |
1696 |
|
if ((uint32_t)x != (iWcount-1)) |
1697 |
|
{ |
1698 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1699 |
|
{ pmv[3].x = EVEN(pmv[3].x); |
1700 |
|
pmv[3].y = EVEN(pmv[3].y); |
1701 |
|
} |
1702 |
|
CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); |
1703 |
|
} |
1704 |
|
} |
1705 |
|
|
1706 |
|
/* Terminate if MinSAD <= T_2 |
1707 |
|
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
1708 |
|
*/ |
1709 |
|
|
1710 |
|
if ( (iMinSAD <= thresh2) |
1711 |
|
|| ( MVequal(*currMV,prevMB->mvs[0]) && ((uint32_t)iMinSAD <= prevMB->sad16) ) ) |
1712 |
|
{ |
1713 |
|
if (MotionFlags & PMV_QUICKSTOP16) |
1714 |
|
goto EPZS16_Terminate_without_Refine; |
1715 |
|
if (MotionFlags & PMV_EARLYSTOP16) |
1716 |
|
goto EPZS16_Terminate_with_Refine; |
1717 |
|
} |
1718 |
|
|
1719 |
|
/***** predictor SET C: acceleration MV (new!), neighbours in prev. frame(new!) ****/ |
1720 |
|
|
1721 |
|
backupMV = prevMB->mvs[0]; // collocated MV |
1722 |
|
backupMV.x += (prevMB->mvs[0].x - oldMB->mvs[0].x ); // acceleration X |
1723 |
|
backupMV.y += (prevMB->mvs[0].y - oldMB->mvs[0].y ); // acceleration Y |
1724 |
|
|
1725 |
|
CHECK_MV16_CANDIDATE(backupMV.x,backupMV.y); |
1726 |
|
|
1727 |
|
// left neighbour |
1728 |
|
if (x != 0) |
1729 |
|
CHECK_MV16_CANDIDATE((prevMB-1)->mvs[0].x,(prevMB-1)->mvs[0].y); |
1730 |
|
|
1731 |
|
// top neighbour |
1732 |
|
if (y != 0) |
1733 |
|
CHECK_MV16_CANDIDATE((prevMB-iWcount)->mvs[0].x,(prevMB-iWcount)->mvs[0].y); |
1734 |
|
|
1735 |
|
// right neighbour, if allowed (this value is not written yet, so take it from pMB->mvs |
1736 |
|
|
1737 |
|
if ((uint32_t)x != iWcount-1) |
1738 |
|
CHECK_MV16_CANDIDATE((prevMB+1)->mvs[0].x,(prevMB+1)->mvs[0].y); |
1739 |
|
|
1740 |
|
// bottom neighbour, dito |
1741 |
|
if ((uint32_t)y != iHcount-1) |
1742 |
|
CHECK_MV16_CANDIDATE((prevMB+iWcount)->mvs[0].x,(prevMB+iWcount)->mvs[0].y); |
1743 |
|
|
1744 |
|
/* Terminate if MinSAD <= T_3 (here T_3 = T_2) */ |
1745 |
|
if (iMinSAD <= thresh2) |
1746 |
|
{ |
1747 |
|
if (MotionFlags & PMV_QUICKSTOP16) |
1748 |
|
goto EPZS16_Terminate_without_Refine; |
1749 |
|
if (MotionFlags & PMV_EARLYSTOP16) |
1750 |
|
goto EPZS16_Terminate_with_Refine; |
1751 |
|
} |
1752 |
|
|
1753 |
|
/************ (if Diamond Search) **************/ |
1754 |
|
|
1755 |
|
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
1756 |
|
|
1757 |
|
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
1758 |
|
|
1759 |
|
if (MotionFlags & PMV_USESQUARES16) |
1760 |
|
EPZSMainSearchPtr = Square16_MainSearch; |
1761 |
|
else |
1762 |
|
EPZSMainSearchPtr = Diamond16_MainSearch; |
1763 |
|
|
1764 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1765 |
|
x, y, |
1766 |
|
currMV->x, currMV->y, iMinSAD, &newMV, pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
1767 |
|
2, iFcode, iQuant, 0); |
1768 |
|
|
1769 |
|
if (iSAD < iMinSAD) |
1770 |
|
{ |
1771 |
|
*currMV = newMV; |
1772 |
|
iMinSAD = iSAD; |
1773 |
|
} |
1774 |
|
|
1775 |
|
|
1776 |
|
if (MotionFlags & PMV_EXTSEARCH16) |
1777 |
|
{ |
1778 |
|
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
1779 |
|
|
1780 |
|
if (!(MVequal(pmv[0],backupMV)) ) |
1781 |
|
{ |
1782 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1783 |
|
x, y, |
1784 |
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
1785 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); |
1786 |
|
} |
1787 |
|
|
1788 |
|
if (iSAD < iMinSAD) |
1789 |
|
{ |
1790 |
|
*currMV = newMV; |
1791 |
|
iMinSAD = iSAD; |
1792 |
|
} |
1793 |
|
|
1794 |
|
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
1795 |
|
{ |
1796 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1797 |
|
x, y, |
1798 |
|
0, 0, iMinSAD, &newMV, |
1799 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); |
1800 |
|
|
1801 |
|
if (iSAD < iMinSAD) |
1802 |
|
{ |
1803 |
|
*currMV = newMV; |
1804 |
|
iMinSAD = iSAD; |
1805 |
|
} |
1806 |
|
} |
1807 |
|
} |
1808 |
|
|
1809 |
|
/*************** Choose best MV found **************/ |
1810 |
|
|
1811 |
|
EPZS16_Terminate_with_Refine: |
1812 |
|
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
1813 |
|
iMinSAD = Halfpel16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
1814 |
|
x, y, |
1815 |
|
currMV, iMinSAD, |
1816 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
1817 |
|
|
1818 |
|
EPZS16_Terminate_without_Refine: |
1819 |
|
|
1820 |
|
*oldMB = *prevMB; |
1821 |
|
|
1822 |
|
currPMV->x = currMV->x - pmv[0].x; |
1823 |
|
currPMV->y = currMV->y - pmv[0].y; |
1824 |
|
return iMinSAD; |
1825 |
|
} |
1826 |
|
|
1827 |
|
|
1828 |
|
int32_t EPZSSearch8( |
1829 |
|
const uint8_t * const pRef, |
1830 |
|
const uint8_t * const pRefH, |
1831 |
|
const uint8_t * const pRefV, |
1832 |
|
const uint8_t * const pRefHV, |
1833 |
|
const IMAGE * const pCur, |
1834 |
|
const int x, const int y, |
1835 |
|
const int start_x, const int start_y, |
1836 |
|
const uint32_t MotionFlags, |
1837 |
|
const uint32_t iQuant, |
1838 |
|
const uint32_t iFcode, |
1839 |
|
const MBParam * const pParam, |
1840 |
|
const MACROBLOCK * const pMBs, |
1841 |
|
const MACROBLOCK * const prevMBs, |
1842 |
|
VECTOR * const currMV, |
1843 |
|
VECTOR * const currPMV) |
1844 |
|
{ |
1845 |
|
/* Please not that EPZS might not be a good choice for 8x8-block motion search ! */ |
1846 |
|
|
1847 |
|
const uint32_t iWcount = pParam->mb_width; |
1848 |
|
const int32_t iWidth = pParam->width; |
1849 |
|
const int32_t iHeight = pParam->height; |
1850 |
|
const int32_t iEdgedWidth = pParam->edged_width; |
1851 |
|
|
1852 |
|
const uint8_t * cur = pCur->y + x*8 + y*8*iEdgedWidth; |
1853 |
|
|
1854 |
|
int32_t iDiamondSize=1; |
1855 |
|
|
1856 |
|
int32_t min_dx; |
1857 |
|
int32_t max_dx; |
1858 |
|
int32_t min_dy; |
1859 |
|
int32_t max_dy; |
1860 |
|
|
1861 |
|
VECTOR newMV; |
1862 |
|
VECTOR backupMV; |
1863 |
|
|
1864 |
|
VECTOR pmv[4]; |
1865 |
|
int32_t psad[8]; |
1866 |
|
|
1867 |
|
const int32_t iSubBlock = ((y&1)<<1) + (x&1); |
1868 |
|
|
1869 |
|
// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
1870 |
|
const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; |
1871 |
|
|
1872 |
|
int32_t bPredEq; |
1873 |
|
int32_t iMinSAD,iSAD=9999; |
1874 |
|
|
1875 |
|
MainSearch8FuncPtr EPZSMainSearchPtr; |
1876 |
|
|
1877 |
|
/* Get maximum range */ |
1878 |
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
1879 |
|
x, y, 8, iWidth, iHeight, iFcode); |
1880 |
|
|
1881 |
|
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
1882 |
|
|
1883 |
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
1884 |
|
{ min_dx = EVEN(min_dx); |
1885 |
|
max_dx = EVEN(max_dx); |
1886 |
|
min_dy = EVEN(min_dy); |
1887 |
|
max_dy = EVEN(max_dy); |
1888 |
|
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
1889 |
|
|
1890 |
|
bPredEq = get_pmvdata(pMBs, x>>1, y>>1, iWcount, iSubBlock, pmv, psad); |
1891 |
|
|
1892 |
|
|
1893 |
|
/* Step 4: Calculate SAD around the Median prediction. |
1894 |
|
MinSAD=SAD |
1895 |
|
If Motion Vector equal to Previous frame motion vector |
1896 |
|
and MinSAD<PrevFrmSAD goto Step 10. |
1897 |
|
If SAD<=256 goto Step 10. |
1898 |
|
*/ |
1899 |
|
|
1900 |
|
// Prepare for main loop |
1901 |
|
|
1902 |
|
|
1903 |
|
if (!(MotionFlags & PMV_HALFPEL8)) |
1904 |
|
{ |
1905 |
|
currMV->x = EVEN(currMV->x); |
1906 |
|
currMV->y = EVEN(currMV->y); |
1907 |
|
} |
1908 |
|
|
1909 |
|
if (currMV->x > max_dx) |
1910 |
|
currMV->x=max_dx; |
1911 |
|
if (currMV->x < min_dx) |
1912 |
|
currMV->x=min_dx; |
1913 |
|
if (currMV->y > max_dy) |
1914 |
|
currMV->y=max_dy; |
1915 |
|
if (currMV->y < min_dy) |
1916 |
|
currMV->y=min_dy; |
1917 |
|
|
1918 |
|
/***************** This is predictor SET A: only median prediction ******************/ |
1919 |
|
|
1920 |
|
|
1921 |
|
iMinSAD = sad8( cur, |
1922 |
|
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
1923 |
|
iEdgedWidth); |
1924 |
|
iMinSAD += calc_delta_8(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode, iQuant); |
1925 |
|
|
1926 |
|
|
1927 |
|
// thresh1 is fixed to 256 |
1928 |
|
if (iMinSAD < 256/4 ) |
1929 |
|
{ |
1930 |
|
if (MotionFlags & PMV_QUICKSTOP8) |
1931 |
|
goto EPZS8_Terminate_without_Refine; |
1932 |
|
if (MotionFlags & PMV_EARLYSTOP8) |
1933 |
|
goto EPZS8_Terminate_with_Refine; |
1934 |
|
} |
1935 |
|
|
1936 |
|
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
1937 |
|
|
1938 |
|
|
1939 |
|
// MV=(0,0) is often a good choice |
1940 |
|
CHECK_MV8_ZERO; |
1941 |
|
|
1942 |
|
// previous frame MV |
1943 |
|
CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x,prevMB->mvs[iSubBlock].y); |
1944 |
|
|
1945 |
|
// left neighbour, if allowed |
1946 |
|
if (psad[1] != MV_MAX_ERROR) |
1947 |
|
{ |
1948 |
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
1949 |
|
{ pmv[1].x = EVEN(pmv[1].x); |
1950 |
|
pmv[1].y = EVEN(pmv[1].y); |
1951 |
|
} |
1952 |
|
CHECK_MV8_CANDIDATE(pmv[1].x,pmv[1].y); |
1953 |
|
} |
1954 |
|
|
1955 |
|
// top neighbour, if allowed |
1956 |
|
if (psad[2] != MV_MAX_ERROR) |
1957 |
|
{ |
1958 |
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
1959 |
|
{ pmv[2].x = EVEN(pmv[2].x); |
1960 |
|
pmv[2].y = EVEN(pmv[2].y); |
1961 |
|
} |
1962 |
|
CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); |
1963 |
|
|
1964 |
|
// top right neighbour, if allowed |
1965 |
|
if (psad[3] != MV_MAX_ERROR) |
1966 |
|
{ |
1967 |
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
1968 |
|
{ pmv[3].x = EVEN(pmv[3].x); |
1969 |
|
pmv[3].y = EVEN(pmv[3].y); |
1970 |
|
} |
1971 |
|
CHECK_MV8_CANDIDATE(pmv[3].x,pmv[3].y); |
1972 |
|
} |
1973 |
|
} |
1974 |
|
|
1975 |
|
/* // this bias is zero anyway, at the moment! |
1976 |
|
|
1977 |
|
if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) ) // && (iMinSAD <= iQuant * 96) |
1978 |
|
iMinSAD -= MV8_00_BIAS; |
1979 |
|
|
1980 |
|
*/ |
1981 |
|
|
1982 |
|
/* Terminate if MinSAD <= T_2 |
1983 |
|
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
1984 |
|
*/ |
1985 |
|
|
1986 |
|
if (iMinSAD < 512/4) /* T_2 == 512/4 hardcoded */ |
1987 |
|
{ |
1988 |
|
if (MotionFlags & PMV_QUICKSTOP8) |
1989 |
|
goto EPZS8_Terminate_without_Refine; |
1990 |
|
if (MotionFlags & PMV_EARLYSTOP8) |
1991 |
|
goto EPZS8_Terminate_with_Refine; |
1992 |
|
} |
1993 |
|
|
1994 |
|
/************ (Diamond Search) **************/ |
1995 |
|
|
1996 |
|
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
1997 |
|
|
1998 |
|
if (!(MotionFlags & PMV_HALFPELDIAMOND8)) |
1999 |
|
iDiamondSize *= 2; |
2000 |
|
|
2001 |
|
/* default: use best prediction as starting point for one call of EPZS_MainSearch */ |
2002 |
|
|
2003 |
|
/* // there is no EPZS^2 for inter4v at the moment |
2004 |
|
|
2005 |
|
if (MotionFlags & PMV_USESQUARES8) |
2006 |
|
EPZSMainSearchPtr = Square8_MainSearch; |
2007 |
|
else |
2008 |
|
*/ |
2009 |
|
|
2010 |
|
EPZSMainSearchPtr = Diamond8_MainSearch; |
2011 |
|
|
2012 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
2013 |
|
x, y, |
2014 |
|
currMV->x, currMV->y, iMinSAD, &newMV, |
2015 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
2016 |
|
iDiamondSize, iFcode, iQuant, 0); |
2017 |
|
|
2018 |
|
|
2019 |
|
if (iSAD < iMinSAD) |
2020 |
|
{ |
2021 |
|
*currMV = newMV; |
2022 |
|
iMinSAD = iSAD; |
2023 |
|
} |
2024 |
|
|
2025 |
|
if (MotionFlags & PMV_EXTSEARCH8) |
2026 |
|
{ |
2027 |
|
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
2028 |
|
|
2029 |
|
if (!(MVequal(pmv[0],backupMV)) ) |
2030 |
|
{ |
2031 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
2032 |
|
x, y, |
2033 |
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
2034 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); |
2035 |
|
|
2036 |
|
if (iSAD < iMinSAD) |
2037 |
|
{ |
2038 |
|
*currMV = newMV; |
2039 |
|
iMinSAD = iSAD; |
2040 |
|
} |
2041 |
|
} |
2042 |
|
|
2043 |
|
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
2044 |
|
{ |
2045 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
2046 |
|
x, y, |
2047 |
|
0, 0, iMinSAD, &newMV, |
2048 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); |
2049 |
|
|
2050 |
|
if (iSAD < iMinSAD) |
2051 |
|
{ |
2052 |
|
*currMV = newMV; |
2053 |
|
iMinSAD = iSAD; |
2054 |
|
} |
2055 |
|
} |
2056 |
|
} |
2057 |
|
|
2058 |
|
/*************** Choose best MV found **************/ |
2059 |
|
|
2060 |
|
EPZS8_Terminate_with_Refine: |
2061 |
|
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
2062 |
|
iMinSAD = Halfpel8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
2063 |
|
x, y, |
2064 |
|
currMV, iMinSAD, |
2065 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
2066 |
|
|
2067 |
|
EPZS8_Terminate_without_Refine: |
2068 |
|
|
2069 |
|
currPMV->x = currMV->x - pmv[0].x; |
2070 |
|
currPMV->y = currMV->y - pmv[0].y; |
2071 |
|
return iMinSAD; |
2072 |
|
} |
2073 |
|
|
2074 |
|
|
2075 |
|
|
2076 |
|
|
2077 |
|
|
2078 |
|
/* *********************************************************** |
2079 |
|
bvop motion estimation |
2080 |
|
// TODO: need to incorporate prediction here (eg. sad += calc_delta_16) |
2081 |
|
***************************************************************/ |
2082 |
|
|
2083 |
|
|
2084 |
|
void MotionEstimationBVOP( |
2085 |
|
MBParam * const pParam, |
2086 |
|
FRAMEINFO * const frame, |
2087 |
|
|
2088 |
|
// forward (past) reference |
2089 |
|
const MACROBLOCK * const f_mbs, |
2090 |
|
const IMAGE * const f_ref, |
2091 |
|
const IMAGE * const f_refH, |
2092 |
|
const IMAGE * const f_refV, |
2093 |
|
const IMAGE * const f_refHV, |
2094 |
|
// backward (future) reference |
2095 |
|
const MACROBLOCK * const b_mbs, |
2096 |
|
const IMAGE * const b_ref, |
2097 |
|
const IMAGE * const b_refH, |
2098 |
|
const IMAGE * const b_refV, |
2099 |
|
const IMAGE * const b_refHV) |
2100 |
|
{ |
2101 |
|
const uint32_t mb_width = pParam->mb_width; |
2102 |
|
const uint32_t mb_height = pParam->mb_height; |
2103 |
|
const int32_t edged_width = pParam->edged_width; |
2104 |
|
|
2105 |
|
uint32_t i,j; |
2106 |
|
|
2107 |
|
int32_t f_sad16; |
2108 |
|
int32_t b_sad16; |
2109 |
|
int32_t i_sad16; |
2110 |
|
int32_t d_sad16; |
2111 |
|
int32_t best_sad; |
2112 |
|
|
2113 |
|
VECTOR pmv_dontcare; |
2114 |
|
|
2115 |
|
// note: i==horizontal, j==vertical |
2116 |
|
for (j = 0; j < mb_height; j++) |
2117 |
|
{ |
2118 |
|
for (i = 0; i < mb_width; i++) |
2119 |
|
{ |
2120 |
|
MACROBLOCK *mb = &frame->mbs[i + j*mb_width]; |
2121 |
|
const MACROBLOCK *f_mb = &f_mbs[i + j*mb_width]; |
2122 |
|
const MACROBLOCK *b_mb = &b_mbs[i + j*mb_width]; |
2123 |
|
|
2124 |
|
if (b_mb->mode == MODE_INTER |
2125 |
|
&& b_mb->cbp == 0 |
2126 |
|
&& b_mb->mvs[0].x == 0 |
2127 |
|
&& b_mb->mvs[0].y == 0) |
2128 |
|
{ |
2129 |
|
mb->mode = MODE_NOT_CODED; |
2130 |
|
mb->mvs[0].x = 0; |
2131 |
|
mb->mvs[0].y = 0; |
2132 |
|
mb->b_mvs[0].x = 0; |
2133 |
|
mb->b_mvs[0].y = 0; |
2134 |
|
continue; |
2135 |
|
} |
2136 |
|
|
2137 |
|
|
2138 |
|
// forward search |
2139 |
|
f_sad16 = SEARCH16(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
2140 |
|
&frame->image, |
2141 |
|
i, j, |
2142 |
|
frame->motion_flags, frame->quant, frame->fcode, |
2143 |
|
pParam, |
2144 |
|
f_mbs, f_mbs /* todo */, |
2145 |
|
&mb->mvs[0], &pmv_dontcare); // ignore pmv |
2146 |
|
|
2147 |
|
// backward search |
2148 |
|
b_sad16 = SEARCH16(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
2149 |
|
&frame->image, |
2150 |
|
i, j, |
2151 |
|
frame->motion_flags, frame->quant, frame->bcode, |
2152 |
|
pParam, |
2153 |
|
b_mbs, b_mbs, /* todo */ |
2154 |
|
&mb->b_mvs[0], &pmv_dontcare); // ignore pmv |
2155 |
|
|
2156 |
|
// interpolate search (simple, but effective) |
2157 |
|
i_sad16 = sad16bi_c( |
2158 |
|
frame->image.y + i*16 + j*16*edged_width, |
2159 |
|
get_ref(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
2160 |
|
i, j, 16, mb->mvs[0].x, mb->mvs[0].y, edged_width), |
2161 |
|
get_ref(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
2162 |
|
i, j, 16, mb->b_mvs[0].x, mb->b_mvs[0].x, edged_width), |
2163 |
|
edged_width); |
2164 |
|
|
2165 |
|
// TODO: direct search |
2166 |
|
// predictor + range of [-32,32] |
2167 |
|
d_sad16 = 65535; |
2168 |
|
|
2169 |
|
|
2170 |
|
if (f_sad16 < b_sad16) |
2171 |
|
{ |
2172 |
|
best_sad = f_sad16; |
2173 |
|
mb->mode = MODE_FORWARD; |
2174 |
|
} |
2175 |
|
else |
2176 |
|
{ |
2177 |
|
best_sad = b_sad16; |
2178 |
|
mb->mode = MODE_BACKWARD; |
2179 |
|
} |
2180 |
|
|
2181 |
|
if (i_sad16 < best_sad) |
2182 |
|
{ |
2183 |
|
best_sad = i_sad16; |
2184 |
|
mb->mode = MODE_INTERPOLATE; |
2185 |
|
} |
2186 |
|
|
2187 |
|
if (d_sad16 < best_sad) |
2188 |
|
{ |
2189 |
|
best_sad = d_sad16; |
2190 |
|
mb->mode = MODE_DIRECT; |
2191 |
|
} |
2192 |
|
|
2193 |
|
} |
2194 |
|
} |
2195 |
|
} |