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 |
|
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; |
|
|
} |
|
|
|
|
250 |
} |
} |
251 |
|
|
252 |
|
|
|
/* 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) |
|
|
{ |
|
253 |
|
|
|
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; |
|
|
} |
|
254 |
|
|
|
} |
|
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 |
if (pParam->global_flags & XVID_INTER4V) |
// 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 |
pMB->sad8[0] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
// levels. |
|
2 * j, 2 * i, mv16.x, mv16.y, pParam->motion_flags, |
|
|
pParam, pMBs, &pMB->mvs[0], &pMB->pmvs[0]); |
|
333 |
|
|
|
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]); |
|
334 |
|
|
|
pMB->sad8[2] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
|
|
2 * j, 2 * i + 1, mv16.x, mv16.y, pParam->motion_flags, |
|
|
pParam, pMBs, &pMB->mvs[2], &pMB->pmvs[2]); |
|
335 |
|
|
336 |
pMB->sad8[3] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
if (current->global_flags & XVID_INTER4V) |
337 |
2 * j + 1, 2 * i + 1, mv16.x, mv16.y, pParam->motion_flags, |
for (y = 0; y < iHcount; y++) |
338 |
pParam, pMBs, &pMB->mvs[3], &pMB->pmvs[3]); |
for (x = 0; x < iWcount; x++) |
339 |
|
{ |
340 |
|
MACROBLOCK* const pMB = &pMBs[x + y * iWcount]; |
341 |
|
|
342 |
sad8 = pMB->sad8[0] + pMB->sad8[1] + pMB->sad8[2] + pMB->sad8[3]; |
if (pMB->mode == MODE_INTRA) |
343 |
} |
continue; |
344 |
|
|
345 |
|
|
346 |
|
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; |
|
751 |
|
|
752 |
/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
int32_t Full16_MainSearch( |
753 |
Otherwise select large Diamond Search. |
const uint8_t * const pRef, |
754 |
*/ |
const uint8_t * const pRefH, |
755 |
|
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 |
|
for (dx = min_dx; dx<=max_dx; dx+=iDiamondSize) |
778 |
|
for (dy = min_dy; dy<= max_dy; dy+=iDiamondSize) |
779 |
|
NOCHECK_MV16_CANDIDATE(dx,dy); |
780 |
|
|
781 |
|
return iMinSAD; |
782 |
|
} |
783 |
|
|
784 |
|
int32_t AdvDiamond16_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 iDirection) |
802 |
|
{ |
803 |
|
|
804 |
|
int32_t iSAD; |
805 |
|
|
806 |
|
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
807 |
|
|
808 |
|
if (iDirection) |
809 |
|
{ |
810 |
|
CHECK_MV16_CANDIDATE(startx-iDiamondSize, starty); |
811 |
|
CHECK_MV16_CANDIDATE(startx+iDiamondSize, starty); |
812 |
|
CHECK_MV16_CANDIDATE(startx, starty-iDiamondSize); |
813 |
|
CHECK_MV16_CANDIDATE(startx, starty+iDiamondSize); |
814 |
|
} |
815 |
|
else |
816 |
|
{ |
817 |
|
int bDirection = 1+2+4+8; |
818 |
|
do |
819 |
|
{ |
820 |
|
iDirection = 0; |
821 |
|
if (bDirection&1) //we only want to check left if we came from the right (our last motion was to the left, up-left or down-left) |
822 |
|
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize,starty,1); |
823 |
|
|
824 |
|
if (bDirection&2) |
825 |
|
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize,starty,2); |
826 |
|
|
827 |
|
if (bDirection&4) |
828 |
|
CHECK_MV16_CANDIDATE_DIR(startx,starty-iDiamondSize,4); |
829 |
|
|
830 |
|
if (bDirection&8) |
831 |
|
CHECK_MV16_CANDIDATE_DIR(startx,starty+iDiamondSize,8); |
832 |
|
|
833 |
|
/* now we're doing diagonal checks near our candidate */ |
834 |
|
|
835 |
|
if (iDirection) //checking if anything found |
836 |
|
{ |
837 |
|
bDirection = iDirection; |
838 |
|
iDirection = 0; |
839 |
|
startx=currMV->x; starty=currMV->y; |
840 |
|
if (bDirection & 3) //our candidate is left or right |
841 |
|
{ |
842 |
|
CHECK_MV16_CANDIDATE_DIR(startx,starty+iDiamondSize, 8); |
843 |
|
CHECK_MV16_CANDIDATE_DIR(startx,starty-iDiamondSize, 4); |
844 |
|
} |
845 |
|
else // what remains here is up or down |
846 |
|
{ |
847 |
|
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty, 2); |
848 |
|
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty, 1); |
849 |
|
} |
850 |
|
|
851 |
|
if (iDirection) |
852 |
|
{ bDirection+=iDirection; |
853 |
|
startx=currMV->x; starty=currMV->y; |
854 |
|
} |
855 |
|
} |
856 |
|
else //about to quit, eh? not so fast.... |
857 |
|
{ |
858 |
|
switch (bDirection) |
859 |
|
{ |
860 |
|
case 2: |
861 |
|
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
862 |
|
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
863 |
|
break; |
864 |
|
case 1: |
865 |
|
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
866 |
|
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
867 |
|
break; |
868 |
|
case 2+4: |
869 |
|
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
870 |
|
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
871 |
|
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
872 |
|
break; |
873 |
|
case 4: |
874 |
|
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
875 |
|
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
876 |
|
break; |
877 |
|
case 8: |
878 |
|
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
879 |
|
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
880 |
|
break; |
881 |
|
case 1+4: |
882 |
|
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
883 |
|
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
884 |
|
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
885 |
|
break; |
886 |
|
case 2+8: |
887 |
|
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
888 |
|
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
889 |
|
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
890 |
|
break; |
891 |
|
case 1+8: |
892 |
|
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
893 |
|
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
894 |
|
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
895 |
|
break; |
896 |
|
default: //1+2+4+8 == we didn't find anything at all |
897 |
|
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
898 |
|
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
899 |
|
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
900 |
|
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
901 |
|
break; |
902 |
|
} |
903 |
|
if (!iDirection) break; //ok, the end. really |
904 |
|
else |
905 |
|
{ bDirection=iDirection; |
906 |
|
startx=currMV->x; starty=currMV->y; |
907 |
|
} |
908 |
|
} |
909 |
|
} |
910 |
|
while (1); //forever |
911 |
|
} |
912 |
|
return iMinSAD; |
913 |
|
} |
914 |
|
|
915 |
|
int32_t AdvDiamond8_MainSearch( |
916 |
|
const uint8_t * const pRef, |
917 |
|
const uint8_t * const pRefH, |
918 |
|
const uint8_t * const pRefV, |
919 |
|
const uint8_t * const pRefHV, |
920 |
|
const uint8_t * const cur, |
921 |
|
const int x, const int y, |
922 |
|
int32_t startx, int32_t starty, |
923 |
|
int32_t iMinSAD, |
924 |
|
VECTOR * const currMV, |
925 |
|
const VECTOR * const pmv, |
926 |
|
const int32_t min_dx, const int32_t max_dx, |
927 |
|
const int32_t min_dy, const int32_t max_dy, |
928 |
|
const int32_t iEdgedWidth, |
929 |
|
const int32_t iDiamondSize, |
930 |
|
const int32_t iFcode, |
931 |
|
const int32_t iQuant, |
932 |
|
int iDirection) |
933 |
|
{ |
934 |
|
|
935 |
|
int32_t iSAD; |
936 |
|
|
937 |
|
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
938 |
|
|
939 |
|
if (iDirection) |
940 |
|
{ |
941 |
|
CHECK_MV8_CANDIDATE(startx-iDiamondSize, starty); |
942 |
|
CHECK_MV8_CANDIDATE(startx+iDiamondSize, starty); |
943 |
|
CHECK_MV8_CANDIDATE(startx, starty-iDiamondSize); |
944 |
|
CHECK_MV8_CANDIDATE(startx, starty+iDiamondSize); |
945 |
|
} |
946 |
|
else |
947 |
|
{ |
948 |
|
int bDirection = 1+2+4+8; |
949 |
|
do |
950 |
|
{ |
951 |
|
iDirection = 0; |
952 |
|
if (bDirection&1) //we only want to check left if we came from the right (our last motion was to the left, up-left or down-left) |
953 |
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize,starty,1); |
954 |
|
|
955 |
|
if (bDirection&2) |
956 |
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize,starty,2); |
957 |
|
|
958 |
|
if (bDirection&4) |
959 |
|
CHECK_MV8_CANDIDATE_DIR(startx,starty-iDiamondSize,4); |
960 |
|
|
961 |
|
if (bDirection&8) |
962 |
|
CHECK_MV8_CANDIDATE_DIR(startx,starty+iDiamondSize,8); |
963 |
|
|
964 |
|
/* now we're doing diagonal checks near our candidate */ |
965 |
|
|
966 |
|
if (iDirection) //checking if anything found |
967 |
|
{ |
968 |
|
bDirection = iDirection; |
969 |
|
iDirection = 0; |
970 |
|
startx=currMV->x; starty=currMV->y; |
971 |
|
if (bDirection & 3) //our candidate is left or right |
972 |
|
{ |
973 |
|
CHECK_MV8_CANDIDATE_DIR(startx,starty+iDiamondSize, 8); |
974 |
|
CHECK_MV8_CANDIDATE_DIR(startx,starty-iDiamondSize, 4); |
975 |
|
} |
976 |
|
else // what remains here is up or down |
977 |
|
{ |
978 |
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty, 2); |
979 |
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty, 1); |
980 |
|
} |
981 |
|
|
982 |
|
if (iDirection) |
983 |
|
{ bDirection+=iDirection; |
984 |
|
startx=currMV->x; starty=currMV->y; |
985 |
|
} |
986 |
|
} |
987 |
|
else //about to quit, eh? not so fast.... |
988 |
|
{ |
989 |
|
switch (bDirection) |
990 |
|
{ |
991 |
|
case 2: |
992 |
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
993 |
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
994 |
|
break; |
995 |
|
case 1: |
996 |
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
997 |
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
998 |
|
break; |
999 |
|
case 2+4: |
1000 |
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
1001 |
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
1002 |
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
1003 |
|
break; |
1004 |
|
case 4: |
1005 |
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
1006 |
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
1007 |
|
break; |
1008 |
|
case 8: |
1009 |
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
1010 |
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
1011 |
|
break; |
1012 |
|
case 1+4: |
1013 |
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
1014 |
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
1015 |
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
1016 |
|
break; |
1017 |
|
case 2+8: |
1018 |
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
1019 |
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
1020 |
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
1021 |
|
break; |
1022 |
|
case 1+8: |
1023 |
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
1024 |
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
1025 |
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
1026 |
|
break; |
1027 |
|
default: //1+2+4+8 == we didn't find anything at all |
1028 |
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
1029 |
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
1030 |
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
1031 |
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
1032 |
|
break; |
1033 |
|
} |
1034 |
|
if (!(iDirection)) break; //ok, the end. really |
1035 |
|
else |
1036 |
|
{ bDirection=iDirection; |
1037 |
|
startx=currMV->x; starty=currMV->y; |
1038 |
|
} |
1039 |
|
} |
1040 |
|
} |
1041 |
|
while (1); //forever |
1042 |
|
} |
1043 |
|
return iMinSAD; |
1044 |
|
} |
1045 |
|
|
1046 |
|
|
1047 |
|
int32_t Full8_MainSearch( |
1048 |
|
const uint8_t * const pRef, |
1049 |
|
const uint8_t * const pRefH, |
1050 |
|
const uint8_t * const pRefV, |
1051 |
|
const uint8_t * const pRefHV, |
1052 |
|
const uint8_t * const cur, |
1053 |
|
const int x, const int y, |
1054 |
|
int32_t startx, int32_t starty, |
1055 |
|
int32_t iMinSAD, |
1056 |
|
VECTOR * const currMV, |
1057 |
|
const VECTOR * const pmv, |
1058 |
|
const int32_t min_dx, const int32_t max_dx, |
1059 |
|
const int32_t min_dy, const int32_t max_dy, |
1060 |
|
const int32_t iEdgedWidth, |
1061 |
|
const int32_t iDiamondSize, |
1062 |
|
const int32_t iFcode, |
1063 |
|
const int32_t iQuant, |
1064 |
|
int iFound) |
1065 |
|
{ |
1066 |
|
int32_t iSAD; |
1067 |
|
int32_t dx,dy; |
1068 |
|
VECTOR backupMV; |
1069 |
|
backupMV.x = startx; |
1070 |
|
backupMV.y = starty; |
1071 |
|
|
1072 |
|
for (dx = min_dx; dx<=max_dx; dx+=iDiamondSize) |
1073 |
|
for (dy = min_dy; dy<= max_dy; dy+=iDiamondSize) |
1074 |
|
NOCHECK_MV8_CANDIDATE(dx,dy); |
1075 |
|
|
1076 |
|
return iMinSAD; |
1077 |
|
} |
1078 |
|
|
1079 |
|
|
1080 |
|
|
1081 |
|
int32_t Halfpel16_Refine( |
1082 |
|
const uint8_t * const pRef, |
1083 |
|
const uint8_t * const pRefH, |
1084 |
|
const uint8_t * const pRefV, |
1085 |
|
const uint8_t * const pRefHV, |
1086 |
|
const uint8_t * const cur, |
1087 |
|
const int x, const int y, |
1088 |
|
VECTOR * const currMV, |
1089 |
|
int32_t iMinSAD, |
1090 |
|
const VECTOR * const pmv, |
1091 |
|
const int32_t min_dx, const int32_t max_dx, |
1092 |
|
const int32_t min_dy, const int32_t max_dy, |
1093 |
|
const int32_t iFcode, |
1094 |
|
const int32_t iQuant, |
1095 |
|
const int32_t iEdgedWidth) |
1096 |
|
{ |
1097 |
|
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
1098 |
|
|
1099 |
|
int32_t iSAD; |
1100 |
|
VECTOR backupMV = *currMV; |
1101 |
|
|
1102 |
|
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y-1); |
1103 |
|
CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y-1); |
1104 |
|
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y-1); |
1105 |
|
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y); |
1106 |
|
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y); |
1107 |
|
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y+1); |
1108 |
|
CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y+1); |
1109 |
|
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y+1); |
1110 |
|
|
1111 |
|
return iMinSAD; |
1112 |
|
} |
1113 |
|
|
1114 |
|
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
1115 |
|
|
1116 |
|
|
1117 |
|
int32_t PMVfastSearch16( |
1118 |
|
const uint8_t * const pRef, |
1119 |
|
const uint8_t * const pRefH, |
1120 |
|
const uint8_t * const pRefV, |
1121 |
|
const uint8_t * const pRefHV, |
1122 |
|
const IMAGE * const pCur, |
1123 |
|
const int x, const int y, |
1124 |
|
const uint32_t MotionFlags, |
1125 |
|
const uint32_t iQuant, |
1126 |
|
const uint32_t iFcode, |
1127 |
|
const MBParam * const pParam, |
1128 |
|
const MACROBLOCK * const pMBs, |
1129 |
|
const MACROBLOCK * const prevMBs, |
1130 |
|
VECTOR * const currMV, |
1131 |
|
VECTOR * const currPMV) |
1132 |
|
{ |
1133 |
|
const uint32_t iWcount = pParam->mb_width; |
1134 |
|
const int32_t iWidth = pParam->width; |
1135 |
|
const int32_t iHeight = pParam->height; |
1136 |
|
const int32_t iEdgedWidth = pParam->edged_width; |
1137 |
|
|
1138 |
|
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
1139 |
|
|
1140 |
|
int32_t iDiamondSize; |
1141 |
|
|
1142 |
|
int32_t min_dx; |
1143 |
|
int32_t max_dx; |
1144 |
|
int32_t min_dy; |
1145 |
|
int32_t max_dy; |
1146 |
|
|
1147 |
|
int32_t iFound; |
1148 |
|
|
1149 |
|
VECTOR newMV; |
1150 |
|
VECTOR backupMV; /* just for PMVFAST */ |
1151 |
|
|
1152 |
|
VECTOR pmv[4]; |
1153 |
|
int32_t psad[4]; |
1154 |
|
|
1155 |
|
MainSearch16FuncPtr MainSearchPtr; |
1156 |
|
|
1157 |
|
// const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
1158 |
|
const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; |
1159 |
|
|
1160 |
|
static int32_t threshA,threshB; |
1161 |
|
int32_t bPredEq; |
1162 |
|
int32_t iMinSAD,iSAD; |
1163 |
|
|
1164 |
|
/* Get maximum range */ |
1165 |
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
1166 |
|
x, y, 16, iWidth, iHeight, iFcode); |
1167 |
|
|
1168 |
|
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
1169 |
|
|
1170 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1171 |
|
{ min_dx = EVEN(min_dx); |
1172 |
|
max_dx = EVEN(max_dx); |
1173 |
|
min_dy = EVEN(min_dy); |
1174 |
|
max_dy = EVEN(max_dy); |
1175 |
|
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
1176 |
|
|
1177 |
|
|
1178 |
|
bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
1179 |
|
|
1180 |
|
if ((x==0) && (y==0) ) |
1181 |
|
{ |
1182 |
|
threshA = 512; |
1183 |
|
threshB = 1024; |
1184 |
|
|
1185 |
|
} |
1186 |
|
else |
1187 |
|
{ |
1188 |
|
threshA = psad[0]; |
1189 |
|
threshB = threshA+256; |
1190 |
|
if (threshA< 512) threshA = 512; |
1191 |
|
if (threshA>1024) threshA = 1024; |
1192 |
|
if (threshB>1792) threshB = 1792; |
1193 |
|
} |
1194 |
|
|
1195 |
|
iFound=0; |
1196 |
|
|
1197 |
|
/* Step 4: Calculate SAD around the Median prediction. |
1198 |
|
MinSAD=SAD |
1199 |
|
If Motion Vector equal to Previous frame motion vector |
1200 |
|
and MinSAD<PrevFrmSAD goto Step 10. |
1201 |
|
If SAD<=256 goto Step 10. |
1202 |
|
*/ |
1203 |
|
|
1204 |
|
*currMV=pmv[0]; /* current best := prediction */ |
1205 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1206 |
|
{ /* This should NOT be necessary! */ |
1207 |
|
currMV->x = EVEN(currMV->x); |
1208 |
|
currMV->y = EVEN(currMV->y); |
1209 |
|
} |
1210 |
|
|
1211 |
|
if (currMV->x > max_dx) |
1212 |
|
{ |
1213 |
|
currMV->x=max_dx; |
1214 |
|
} |
1215 |
|
if (currMV->x < min_dx) |
1216 |
|
{ |
1217 |
|
currMV->x=min_dx; |
1218 |
|
} |
1219 |
|
if (currMV->y > max_dy) |
1220 |
|
{ |
1221 |
|
currMV->y=max_dy; |
1222 |
|
} |
1223 |
|
if (currMV->y < min_dy) |
1224 |
|
{ |
1225 |
|
currMV->y=min_dy; |
1226 |
|
} |
1227 |
|
|
1228 |
|
iMinSAD = sad16( cur, |
1229 |
|
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
1230 |
|
iEdgedWidth, MV_MAX_ERROR); |
1231 |
|
iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode, iQuant); |
1232 |
|
|
1233 |
|
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,prevMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
1234 |
|
{ |
1235 |
|
if (iMinSAD < 2*iQuant) // high chances for SKIP-mode |
1236 |
|
{ |
1237 |
|
if (!MVzero(*currMV)) |
1238 |
|
{ |
1239 |
|
iMinSAD += MV16_00_BIAS; |
1240 |
|
CHECK_MV16_ZERO; // (0,0) saves space for letterboxed pictures |
1241 |
|
iMinSAD -= MV16_00_BIAS; |
1242 |
|
} |
1243 |
|
} |
1244 |
|
|
1245 |
|
if (MotionFlags & PMV_QUICKSTOP16) |
1246 |
|
goto PMVfast16_Terminate_without_Refine; |
1247 |
|
if (MotionFlags & PMV_EARLYSTOP16) |
1248 |
|
goto PMVfast16_Terminate_with_Refine; |
1249 |
|
} |
1250 |
|
|
1251 |
|
|
1252 |
|
/* Step 2 (lazy eval): Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
1253 |
|
vector of the median. |
1254 |
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
1255 |
|
*/ |
1256 |
|
|
1257 |
|
if ((bPredEq) && (MVequal(pmv[0],prevMB->mvs[0]) ) ) |
1258 |
|
iFound=2; |
1259 |
|
|
1260 |
|
/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
1261 |
|
Otherwise select large Diamond Search. |
1262 |
|
*/ |
1263 |
|
|
1264 |
|
if ( (!MVzero(pmv[0])) || (threshB<1536) || (bPredEq) ) |
1265 |
|
iDiamondSize=1; // halfpel! |
1266 |
|
else |
1267 |
|
iDiamondSize=2; // halfpel! |
1268 |
|
|
1269 |
|
if (!(MotionFlags & PMV_HALFPELDIAMOND16) ) |
1270 |
|
iDiamondSize*=2; |
1271 |
|
|
1272 |
|
/* |
1273 |
|
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
1274 |
|
Also calculate (0,0) but do not subtract offset. |
1275 |
|
Let MinSAD be the smallest SAD up to this point. |
1276 |
|
If MV is (0,0) subtract offset. |
1277 |
|
*/ |
1278 |
|
|
1279 |
|
// (0,0) is always possible |
1280 |
|
|
1281 |
|
if (!MVzero(pmv[0])) |
1282 |
|
CHECK_MV16_ZERO; |
1283 |
|
|
1284 |
|
// previous frame MV is always possible |
1285 |
|
|
1286 |
|
if (!MVzero(prevMB->mvs[0])) |
1287 |
|
if (!MVequal(prevMB->mvs[0],pmv[0])) |
1288 |
|
CHECK_MV16_CANDIDATE(prevMB->mvs[0].x,prevMB->mvs[0].y); |
1289 |
|
|
1290 |
|
// left neighbour, if allowed |
1291 |
|
|
1292 |
|
if (!MVzero(pmv[1])) |
1293 |
|
if (!MVequal(pmv[1],prevMB->mvs[0])) |
1294 |
|
if (!MVequal(pmv[1],pmv[0])) |
1295 |
|
{ |
1296 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1297 |
|
{ pmv[1].x = EVEN(pmv[1].x); |
1298 |
|
pmv[1].y = EVEN(pmv[1].y); |
1299 |
|
} |
1300 |
|
|
1301 |
|
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
1302 |
|
} |
1303 |
|
|
1304 |
|
// top neighbour, if allowed |
1305 |
|
if (!MVzero(pmv[2])) |
1306 |
|
if (!MVequal(pmv[2],prevMB->mvs[0])) |
1307 |
|
if (!MVequal(pmv[2],pmv[0])) |
1308 |
|
if (!MVequal(pmv[2],pmv[1])) |
1309 |
|
{ |
1310 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1311 |
|
{ pmv[2].x = EVEN(pmv[2].x); |
1312 |
|
pmv[2].y = EVEN(pmv[2].y); |
1313 |
|
} |
1314 |
|
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
1315 |
|
|
1316 |
|
// top right neighbour, if allowed |
1317 |
|
if (!MVzero(pmv[3])) |
1318 |
|
if (!MVequal(pmv[3],prevMB->mvs[0])) |
1319 |
|
if (!MVequal(pmv[3],pmv[0])) |
1320 |
|
if (!MVequal(pmv[3],pmv[1])) |
1321 |
|
if (!MVequal(pmv[3],pmv[2])) |
1322 |
|
{ |
1323 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1324 |
|
{ pmv[3].x = EVEN(pmv[3].x); |
1325 |
|
pmv[3].y = EVEN(pmv[3].y); |
1326 |
|
} |
1327 |
|
CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); |
1328 |
|
} |
1329 |
|
} |
1330 |
|
|
1331 |
|
if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96)*/ ) |
1332 |
|
iMinSAD -= MV16_00_BIAS; |
1333 |
|
|
1334 |
|
|
1335 |
|
/* Step 6: If MinSAD <= thresa goto Step 10. |
1336 |
|
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
1337 |
|
*/ |
1338 |
|
|
1339 |
|
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,prevMB->mvs[0]) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
1340 |
|
{ |
1341 |
|
if (MotionFlags & PMV_QUICKSTOP16) |
1342 |
|
goto PMVfast16_Terminate_without_Refine; |
1343 |
|
if (MotionFlags & PMV_EARLYSTOP16) |
1344 |
|
goto PMVfast16_Terminate_with_Refine; |
1345 |
|
} |
1346 |
|
|
1347 |
|
|
1348 |
|
/************ (Diamond Search) **************/ |
1349 |
|
/* |
1350 |
|
Step 7: Perform Diamond search, with either the small or large diamond. |
1351 |
|
If Found=2 only examine one Diamond pattern, and afterwards goto step 10 |
1352 |
|
Step 8: If small diamond, iterate small diamond search pattern until motion vector lies in the center of the diamond. |
1353 |
|
If center then goto step 10. |
1354 |
|
Step 9: If large diamond, iterate large diamond search pattern until motion vector lies in the center. |
1355 |
|
Refine by using small diamond and goto step 10. |
1356 |
|
*/ |
1357 |
|
|
1358 |
|
if (MotionFlags & PMV_USESQUARES16) |
1359 |
|
MainSearchPtr = Square16_MainSearch; |
1360 |
|
else |
1361 |
|
if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
1362 |
|
MainSearchPtr = AdvDiamond16_MainSearch; |
1363 |
|
else |
1364 |
|
MainSearchPtr = Diamond16_MainSearch; |
1365 |
|
|
1366 |
|
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
1367 |
|
|
1368 |
|
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
1369 |
|
iSAD = (*MainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1370 |
|
x, y, |
1371 |
|
currMV->x, currMV->y, iMinSAD, &newMV, |
1372 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1373 |
|
|
1374 |
|
if (iSAD < iMinSAD) |
1375 |
|
{ |
1376 |
|
*currMV = newMV; |
1377 |
|
iMinSAD = iSAD; |
1378 |
|
} |
1379 |
|
|
1380 |
|
if (MotionFlags & PMV_EXTSEARCH16) |
1381 |
|
{ |
1382 |
|
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
1383 |
|
|
1384 |
|
if (!(MVequal(pmv[0],backupMV)) ) |
1385 |
|
{ iSAD = (*MainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1386 |
|
x, y, |
1387 |
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
1388 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1389 |
|
|
1390 |
|
if (iSAD < iMinSAD) |
1391 |
|
{ |
1392 |
|
*currMV = newMV; |
1393 |
|
iMinSAD = iSAD; |
1394 |
|
} |
1395 |
|
} |
1396 |
|
|
1397 |
|
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
1398 |
|
{ iSAD = (*MainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1399 |
|
x, y, |
1400 |
|
0, 0, iMinSAD, &newMV, |
1401 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1402 |
|
|
1403 |
|
if (iSAD < iMinSAD) |
1404 |
|
{ |
1405 |
|
*currMV = newMV; |
1406 |
|
iMinSAD = iSAD; |
1407 |
|
} |
1408 |
|
} |
1409 |
|
} |
1410 |
|
|
1411 |
|
/* |
1412 |
|
Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
1413 |
|
*/ |
1414 |
|
|
1415 |
|
PMVfast16_Terminate_with_Refine: |
1416 |
|
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
1417 |
|
iMinSAD = Halfpel16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
1418 |
|
x, y, |
1419 |
|
currMV, iMinSAD, |
1420 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
1421 |
|
|
1422 |
|
PMVfast16_Terminate_without_Refine: |
1423 |
|
currPMV->x = currMV->x - pmv[0].x; |
1424 |
|
currPMV->y = currMV->y - pmv[0].y; |
1425 |
|
return iMinSAD; |
1426 |
|
} |
1427 |
|
|
1428 |
|
|
1429 |
|
|
1430 |
|
|
1431 |
|
|
1432 |
|
|
1433 |
|
int32_t Diamond8_MainSearch( |
1434 |
|
const uint8_t * const pRef, |
1435 |
|
const uint8_t * const pRefH, |
1436 |
|
const uint8_t * const pRefV, |
1437 |
|
const uint8_t * const pRefHV, |
1438 |
|
const uint8_t * const cur, |
1439 |
|
const int x, const int y, |
1440 |
|
int32_t startx, int32_t starty, |
1441 |
|
int32_t iMinSAD, |
1442 |
|
VECTOR * const currMV, |
1443 |
|
const VECTOR * const pmv, |
1444 |
|
const int32_t min_dx, const int32_t max_dx, |
1445 |
|
const int32_t min_dy, const int32_t max_dy, |
1446 |
|
const int32_t iEdgedWidth, |
1447 |
|
const int32_t iDiamondSize, |
1448 |
|
const int32_t iFcode, |
1449 |
|
const int32_t iQuant, |
1450 |
|
int iFound) |
1451 |
|
{ |
1452 |
|
/* Do a diamond search around given starting point, return SAD of best */ |
1453 |
|
|
1454 |
|
int32_t iDirection=0; |
1455 |
|
int32_t iSAD; |
1456 |
|
VECTOR backupMV; |
1457 |
|
backupMV.x = startx; |
1458 |
|
backupMV.y = starty; |
1459 |
|
|
1460 |
|
/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ |
1461 |
|
|
1462 |
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
1463 |
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
1464 |
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
1465 |
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
1466 |
|
|
1467 |
|
if (iDirection) |
1468 |
|
while (!iFound) |
1469 |
|
{ |
1470 |
|
iFound = 1; |
1471 |
|
backupMV=*currMV; // since iDirection!=0, this is well defined! |
1472 |
|
|
1473 |
|
if ( iDirection != 2) |
1474 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
1475 |
|
if ( iDirection != 1) |
1476 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x+iDiamondSize,backupMV.y,2); |
1477 |
|
if ( iDirection != 4) |
1478 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x,backupMV.y-iDiamondSize,3); |
1479 |
|
if ( iDirection != 3) |
1480 |
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x,backupMV.y+iDiamondSize,4); |
1481 |
|
} |
1482 |
|
else |
1483 |
|
{ |
1484 |
|
currMV->x = startx; |
1485 |
|
currMV->y = starty; |
1486 |
|
} |
1487 |
|
return iMinSAD; |
1488 |
|
} |
1489 |
|
|
1490 |
|
int32_t Halfpel8_Refine( |
1491 |
|
const uint8_t * const pRef, |
1492 |
|
const uint8_t * const pRefH, |
1493 |
|
const uint8_t * const pRefV, |
1494 |
|
const uint8_t * const pRefHV, |
1495 |
|
const uint8_t * const cur, |
1496 |
|
const int x, const int y, |
1497 |
|
VECTOR * const currMV, |
1498 |
|
int32_t iMinSAD, |
1499 |
|
const VECTOR * const pmv, |
1500 |
|
const int32_t min_dx, const int32_t max_dx, |
1501 |
|
const int32_t min_dy, const int32_t max_dy, |
1502 |
|
const int32_t iFcode, |
1503 |
|
const int32_t iQuant, |
1504 |
|
const int32_t iEdgedWidth) |
1505 |
|
{ |
1506 |
|
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
1507 |
|
|
1508 |
|
int32_t iSAD; |
1509 |
|
VECTOR backupMV = *currMV; |
1510 |
|
|
1511 |
|
CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y-1); |
1512 |
|
CHECK_MV8_CANDIDATE(backupMV.x ,backupMV.y-1); |
1513 |
|
CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y-1); |
1514 |
|
CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y); |
1515 |
|
CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y); |
1516 |
|
CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y+1); |
1517 |
|
CHECK_MV8_CANDIDATE(backupMV.x ,backupMV.y+1); |
1518 |
|
CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y+1); |
1519 |
|
|
1520 |
|
return iMinSAD; |
1521 |
|
} |
1522 |
|
|
1523 |
|
|
1524 |
|
#define PMV_HALFPEL8 (PMV_HALFPELDIAMOND8|PMV_HALFPELREFINE8) |
1525 |
|
|
1526 |
|
int32_t PMVfastSearch8( |
1527 |
|
const uint8_t * const pRef, |
1528 |
|
const uint8_t * const pRefH, |
1529 |
|
const uint8_t * const pRefV, |
1530 |
|
const uint8_t * const pRefHV, |
1531 |
|
const IMAGE * const pCur, |
1532 |
|
const int x, const int y, |
1533 |
|
const int start_x, const int start_y, |
1534 |
|
const uint32_t MotionFlags, |
1535 |
|
const uint32_t iQuant, |
1536 |
|
const uint32_t iFcode, |
1537 |
|
const MBParam * const pParam, |
1538 |
|
const MACROBLOCK * const pMBs, |
1539 |
|
const MACROBLOCK * const prevMBs, |
1540 |
|
VECTOR * const currMV, |
1541 |
|
VECTOR * const currPMV) |
1542 |
|
{ |
1543 |
|
const uint32_t iWcount = pParam->mb_width; |
1544 |
|
const int32_t iWidth = pParam->width; |
1545 |
|
const int32_t iHeight = pParam->height; |
1546 |
|
const int32_t iEdgedWidth = pParam->edged_width; |
1547 |
|
|
1548 |
|
const uint8_t * cur = pCur->y + x*8 + y*8*iEdgedWidth; |
1549 |
|
|
1550 |
|
int32_t iDiamondSize; |
1551 |
|
|
1552 |
|
int32_t min_dx; |
1553 |
|
int32_t max_dx; |
1554 |
|
int32_t min_dy; |
1555 |
|
int32_t max_dy; |
1556 |
|
|
1557 |
|
VECTOR pmv[4]; |
1558 |
|
int32_t psad[4]; |
1559 |
|
VECTOR newMV; |
1560 |
|
VECTOR backupMV; |
1561 |
|
VECTOR startMV; |
1562 |
|
|
1563 |
|
// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
1564 |
|
const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; |
1565 |
|
|
1566 |
|
static int32_t threshA,threshB; |
1567 |
|
int32_t iFound,bPredEq; |
1568 |
|
int32_t iMinSAD,iSAD; |
1569 |
|
|
1570 |
|
int32_t iSubBlock = (y&1)+(y&1) + (x&1); |
1571 |
|
|
1572 |
|
MainSearch8FuncPtr MainSearchPtr; |
1573 |
|
|
1574 |
|
/* Init variables */ |
1575 |
|
startMV.x = start_x; |
1576 |
|
startMV.y = start_y; |
1577 |
|
|
1578 |
|
/* Get maximum range */ |
1579 |
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
1580 |
|
x, y, 8, iWidth, iHeight, iFcode); |
1581 |
|
|
1582 |
|
if (!(MotionFlags & PMV_HALFPELDIAMOND8 )) |
1583 |
|
{ min_dx = EVEN(min_dx); |
1584 |
|
max_dx = EVEN(max_dx); |
1585 |
|
min_dy = EVEN(min_dy); |
1586 |
|
max_dy = EVEN(max_dy); |
1587 |
|
} /* because we might use IF (dx>max_dx) THEN dx=max_dx; */ |
1588 |
|
|
1589 |
|
|
1590 |
|
bPredEq = get_pmvdata(pMBs, (x>>1), (y>>1), iWcount, iSubBlock, pmv, psad); |
1591 |
|
|
1592 |
|
if ((x==0) && (y==0) ) |
1593 |
|
{ |
1594 |
|
threshA = 512/4; |
1595 |
|
threshB = 1024/4; |
1596 |
|
|
1597 |
|
} |
1598 |
|
else |
1599 |
|
{ |
1600 |
|
threshA = psad[0]/4; /* good estimate */ |
1601 |
|
threshB = threshA+256/4; |
1602 |
|
if (threshA< 512/4) threshA = 512/4; |
1603 |
|
if (threshA>1024/4) threshA = 1024/4; |
1604 |
|
if (threshB>1792/4) threshB = 1792/4; |
1605 |
|
} |
1606 |
|
|
1607 |
|
iFound=0; |
1608 |
|
|
1609 |
|
/* Step 4: Calculate SAD around the Median prediction. |
1610 |
|
MinSAD=SAD |
1611 |
|
If Motion Vector equal to Previous frame motion vector |
1612 |
|
and MinSAD<PrevFrmSAD goto Step 10. |
1613 |
|
If SAD<=256 goto Step 10. |
1614 |
|
*/ |
1615 |
|
|
1616 |
|
|
1617 |
|
// Prepare for main loop |
1618 |
|
|
1619 |
|
// if (MotionFlags & PMV_USESQUARES8) |
1620 |
|
// MainSearchPtr = Square8_MainSearch; |
1621 |
|
// else |
1622 |
|
|
1623 |
|
if (MotionFlags & PMV_ADVANCEDDIAMOND8) |
1624 |
|
MainSearchPtr = AdvDiamond8_MainSearch; |
1625 |
|
else |
1626 |
|
MainSearchPtr = Diamond8_MainSearch; |
1627 |
|
|
1628 |
|
|
1629 |
|
*currMV = startMV; |
1630 |
|
|
1631 |
|
iMinSAD = sad8( cur, |
1632 |
|
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
1633 |
|
iEdgedWidth); |
1634 |
|
iMinSAD += calc_delta_8(currMV->x - pmv[0].x, currMV->y - pmv[0].y, (uint8_t)iFcode, iQuant); |
1635 |
|
|
1636 |
if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536) || (bPredEq) ) |
if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,prevMB->mvs[iSubBlock])) |
1637 |
iDiamondSize=1; // halfpel! |
&& ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
1638 |
|
{ |
1639 |
|
if (MotionFlags & PMV_QUICKSTOP16) |
1640 |
|
goto PMVfast8_Terminate_without_Refine; |
1641 |
|
if (MotionFlags & PMV_EARLYSTOP16) |
1642 |
|
goto PMVfast8_Terminate_with_Refine; |
1643 |
|
} |
1644 |
|
|
1645 |
|
/* Step 2 (lazy eval): Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
1646 |
|
vector of the median. |
1647 |
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
1648 |
|
*/ |
1649 |
|
|
1650 |
|
if ((bPredEq) && (MVequal(pmv[0],prevMB->mvs[iSubBlock]) ) ) |
1651 |
|
iFound=2; |
1652 |
|
|
1653 |
|
/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
1654 |
|
Otherwise select large Diamond Search. |
1655 |
|
*/ |
1656 |
|
|
1657 |
|
if ( (!MVzero(pmv[0])) || (threshB<1536/4) || (bPredEq) ) |
1658 |
|
iDiamondSize=1; // 1 halfpel! |
1659 |
else |
else |
1660 |
iDiamondSize=2; // halfpel! |
iDiamondSize=2; // 2 halfpel = 1 full pixel! |
1661 |
|
|
1662 |
if (!(MotionFlags & PMV_HALFPELDIAMOND16) ) |
if (!(MotionFlags & PMV_HALFPELDIAMOND8) ) |
1663 |
iDiamondSize*=2; |
iDiamondSize*=2; |
1664 |
|
|
1665 |
|
|
1666 |
|
/* |
1667 |
|
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
1668 |
|
Also calculate (0,0) but do not subtract offset. |
1669 |
|
Let MinSAD be the smallest SAD up to this point. |
1670 |
|
If MV is (0,0) subtract offset. |
1671 |
|
*/ |
1672 |
|
|
1673 |
|
// the median prediction might be even better than mv16 |
1674 |
|
|
1675 |
|
if (!MVequal(pmv[0],startMV)) |
1676 |
|
CHECK_MV8_CANDIDATE(pmv[0].x,pmv[0].y); |
1677 |
|
|
1678 |
|
// (0,0) if needed |
1679 |
|
if (!MVzero(pmv[0])) |
1680 |
|
if (!MVzero(startMV)) |
1681 |
|
CHECK_MV8_ZERO; |
1682 |
|
|
1683 |
|
// previous frame MV if needed |
1684 |
|
if (!MVzero(prevMB->mvs[iSubBlock])) |
1685 |
|
if (!MVequal(prevMB->mvs[iSubBlock],startMV)) |
1686 |
|
if (!MVequal(prevMB->mvs[iSubBlock],pmv[0])) |
1687 |
|
CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x,prevMB->mvs[iSubBlock].y); |
1688 |
|
|
1689 |
|
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,prevMB->mvs[iSubBlock]) && ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
1690 |
|
{ |
1691 |
|
if (MotionFlags & PMV_QUICKSTOP16) |
1692 |
|
goto PMVfast8_Terminate_without_Refine; |
1693 |
|
if (MotionFlags & PMV_EARLYSTOP16) |
1694 |
|
goto PMVfast8_Terminate_with_Refine; |
1695 |
|
} |
1696 |
|
|
1697 |
|
|
1698 |
|
// left neighbour, if allowed and needed |
1699 |
|
if (!MVzero(pmv[1])) |
1700 |
|
if (!MVequal(pmv[1],startMV)) |
1701 |
|
if (!MVequal(pmv[1],prevMB->mvs[iSubBlock])) |
1702 |
|
if (!MVequal(pmv[1],pmv[0])) |
1703 |
|
{ |
1704 |
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
1705 |
|
{ pmv[1].x = EVEN(pmv[1].x); |
1706 |
|
pmv[1].y = EVEN(pmv[1].y); |
1707 |
|
} |
1708 |
|
CHECK_MV8_CANDIDATE(pmv[1].x,pmv[1].y); |
1709 |
|
} |
1710 |
|
|
1711 |
|
// top neighbour, if allowed and needed |
1712 |
|
if (!MVzero(pmv[2])) |
1713 |
|
if (!MVequal(pmv[2],startMV)) |
1714 |
|
if (!MVequal(pmv[2],prevMB->mvs[iSubBlock])) |
1715 |
|
if (!MVequal(pmv[2],pmv[0])) |
1716 |
|
if (!MVequal(pmv[2],pmv[1])) |
1717 |
|
{ |
1718 |
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
1719 |
|
{ pmv[2].x = EVEN(pmv[2].x); |
1720 |
|
pmv[2].y = EVEN(pmv[2].y); |
1721 |
|
} |
1722 |
|
CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); |
1723 |
|
|
1724 |
|
// top right neighbour, if allowed and needed |
1725 |
|
if (!MVzero(pmv[3])) |
1726 |
|
if (!MVequal(pmv[3],startMV)) |
1727 |
|
if (!MVequal(pmv[3],prevMB->mvs[iSubBlock])) |
1728 |
|
if (!MVequal(pmv[3],pmv[0])) |
1729 |
|
if (!MVequal(pmv[3],pmv[1])) |
1730 |
|
if (!MVequal(pmv[3],pmv[2])) |
1731 |
|
{ |
1732 |
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
1733 |
|
{ pmv[3].x = EVEN(pmv[3].x); |
1734 |
|
pmv[3].y = EVEN(pmv[3].y); |
1735 |
|
} |
1736 |
|
CHECK_MV8_CANDIDATE(pmv[3].x,pmv[3].y); |
1737 |
|
} |
1738 |
|
} |
1739 |
|
|
1740 |
|
if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96) */ ) |
1741 |
|
iMinSAD -= MV8_00_BIAS; |
1742 |
|
|
1743 |
|
|
1744 |
|
/* Step 6: If MinSAD <= thresa goto Step 10. |
1745 |
|
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
1746 |
|
*/ |
1747 |
|
|
1748 |
|
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,prevMB->mvs[iSubBlock]) && ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
1749 |
|
{ |
1750 |
|
if (MotionFlags & PMV_QUICKSTOP16) |
1751 |
|
goto PMVfast8_Terminate_without_Refine; |
1752 |
|
if (MotionFlags & PMV_EARLYSTOP16) |
1753 |
|
goto PMVfast8_Terminate_with_Refine; |
1754 |
|
} |
1755 |
|
|
1756 |
|
/************ (Diamond Search) **************/ |
1757 |
|
/* |
1758 |
|
Step 7: Perform Diamond search, with either the small or large diamond. |
1759 |
|
If Found=2 only examine one Diamond pattern, and afterwards goto step 10 |
1760 |
|
Step 8: If small diamond, iterate small diamond search pattern until motion vector lies in the center of the diamond. |
1761 |
|
If center then goto step 10. |
1762 |
|
Step 9: If large diamond, iterate large diamond search pattern until motion vector lies in the center. |
1763 |
|
Refine by using small diamond and goto step 10. |
1764 |
|
*/ |
1765 |
|
|
1766 |
|
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
1767 |
|
|
1768 |
|
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
1769 |
|
iSAD = (*MainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1770 |
|
x, y, |
1771 |
|
currMV->x, currMV->y, iMinSAD, &newMV, |
1772 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1773 |
|
|
1774 |
|
if (iSAD < iMinSAD) |
1775 |
|
{ |
1776 |
|
*currMV = newMV; |
1777 |
|
iMinSAD = iSAD; |
1778 |
|
} |
1779 |
|
|
1780 |
|
if (MotionFlags & PMV_EXTSEARCH8) |
1781 |
|
{ |
1782 |
|
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
1783 |
|
|
1784 |
|
if (!(MVequal(pmv[0],backupMV)) ) |
1785 |
|
{ iSAD = (*MainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1786 |
|
x, y, |
1787 |
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
1788 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1789 |
|
|
1790 |
|
if (iSAD < iMinSAD) |
1791 |
|
{ |
1792 |
|
*currMV = newMV; |
1793 |
|
iMinSAD = iSAD; |
1794 |
|
} |
1795 |
|
} |
1796 |
|
|
1797 |
|
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
1798 |
|
{ iSAD = (*MainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1799 |
|
x, y, |
1800 |
|
0, 0, iMinSAD, &newMV, |
1801 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1802 |
|
|
1803 |
|
if (iSAD < iMinSAD) |
1804 |
|
{ |
1805 |
|
*currMV = newMV; |
1806 |
|
iMinSAD = iSAD; |
1807 |
|
} |
1808 |
|
} |
1809 |
|
} |
1810 |
|
|
1811 |
|
/* Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
1812 |
|
By performing an optional local half-pixel search, we can refine this result even further. |
1813 |
|
*/ |
1814 |
|
|
1815 |
|
PMVfast8_Terminate_with_Refine: |
1816 |
|
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
1817 |
|
iMinSAD = Halfpel8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
1818 |
|
x, y, |
1819 |
|
currMV, iMinSAD, |
1820 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
1821 |
|
|
1822 |
|
|
1823 |
|
PMVfast8_Terminate_without_Refine: |
1824 |
|
currPMV->x = currMV->x - pmv[0].x; |
1825 |
|
currPMV->y = currMV->y - pmv[0].y; |
1826 |
|
|
1827 |
|
return iMinSAD; |
1828 |
|
} |
1829 |
|
|
1830 |
|
int32_t EPZSSearch16( |
1831 |
|
const uint8_t * const pRef, |
1832 |
|
const uint8_t * const pRefH, |
1833 |
|
const uint8_t * const pRefV, |
1834 |
|
const uint8_t * const pRefHV, |
1835 |
|
const IMAGE * const pCur, |
1836 |
|
const int x, const int y, |
1837 |
|
const uint32_t MotionFlags, |
1838 |
|
const uint32_t iQuant, |
1839 |
|
const uint32_t iFcode, |
1840 |
|
const MBParam * const pParam, |
1841 |
|
const MACROBLOCK * const pMBs, |
1842 |
|
const MACROBLOCK * const prevMBs, |
1843 |
|
VECTOR * const currMV, |
1844 |
|
VECTOR * const currPMV) |
1845 |
|
{ |
1846 |
|
const uint32_t iWcount = pParam->mb_width; |
1847 |
|
const uint32_t iHcount = pParam->mb_height; |
1848 |
|
|
1849 |
|
const int32_t iWidth = pParam->width; |
1850 |
|
const int32_t iHeight = pParam->height; |
1851 |
|
const int32_t iEdgedWidth = pParam->edged_width; |
1852 |
|
|
1853 |
|
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
1854 |
|
|
1855 |
|
int32_t min_dx; |
1856 |
|
int32_t max_dx; |
1857 |
|
int32_t min_dy; |
1858 |
|
int32_t max_dy; |
1859 |
|
|
1860 |
|
VECTOR newMV; |
1861 |
|
VECTOR backupMV; |
1862 |
|
|
1863 |
|
VECTOR pmv[4]; |
1864 |
|
int32_t psad[8]; |
1865 |
|
|
1866 |
|
static MACROBLOCK * oldMBs = NULL; |
1867 |
|
// const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
1868 |
|
const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; |
1869 |
|
MACROBLOCK * oldMB = NULL; |
1870 |
|
|
1871 |
|
static int32_t thresh2; |
1872 |
|
int32_t bPredEq; |
1873 |
|
int32_t iMinSAD,iSAD=9999; |
1874 |
|
|
1875 |
|
MainSearch16FuncPtr MainSearchPtr; |
1876 |
|
|
1877 |
|
if (oldMBs == NULL) |
1878 |
|
{ oldMBs = (MACROBLOCK*) calloc(iWcount*iHcount,sizeof(MACROBLOCK)); |
1879 |
|
// fprintf(stderr,"allocated %d bytes for oldMBs\n",iWcount*iHcount*sizeof(MACROBLOCK)); |
1880 |
|
} |
1881 |
|
oldMB = oldMBs + x + y * iWcount; |
1882 |
|
|
1883 |
|
/* Get maximum range */ |
1884 |
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
1885 |
|
x, y, 16, iWidth, iHeight, iFcode); |
1886 |
|
|
1887 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1888 |
|
{ min_dx = EVEN(min_dx); |
1889 |
|
max_dx = EVEN(max_dx); |
1890 |
|
min_dy = EVEN(min_dy); |
1891 |
|
max_dy = EVEN(max_dy); |
1892 |
|
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
1893 |
|
|
1894 |
|
bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
1895 |
|
|
1896 |
/* Step 4: Calculate SAD around the Median prediction. |
/* Step 4: Calculate SAD around the Median prediction. |
1897 |
MinSAD=SAD |
MinSAD=SAD |
1898 |
If Motion Vector equal to Previous frame motion vector |
If Motion Vector equal to Previous frame motion vector |
1900 |
If SAD<=256 goto Step 10. |
If SAD<=256 goto Step 10. |
1901 |
*/ |
*/ |
1902 |
|
|
|
|
|
1903 |
// Prepare for main loop |
// Prepare for main loop |
1904 |
|
|
1905 |
*currMV=pmv[0]; /* current best := prediction */ |
*currMV=pmv[0]; /* current best := median prediction */ |
1906 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
if (!(MotionFlags & PMV_HALFPEL16 )) |
1907 |
{ /* This should NOT be necessary! */ |
{ |
1908 |
currMV->x = EVEN(currMV->x); |
currMV->x = EVEN(currMV->x); |
1909 |
currMV->y = EVEN(currMV->y); |
currMV->y = EVEN(currMV->y); |
1910 |
} |
} |
1911 |
|
|
1912 |
if (currMV->x > max_dx) |
if (currMV->x > max_dx) |
|
{ |
|
1913 |
currMV->x=max_dx; |
currMV->x=max_dx; |
|
} |
|
1914 |
if (currMV->x < min_dx) |
if (currMV->x < min_dx) |
|
{ |
|
1915 |
currMV->x=min_dx; |
currMV->x=min_dx; |
|
} |
|
1916 |
if (currMV->y > max_dy) |
if (currMV->y > max_dy) |
|
{ |
|
1917 |
currMV->y=max_dy; |
currMV->y=max_dy; |
|
} |
|
1918 |
if (currMV->y < min_dy) |
if (currMV->y < min_dy) |
|
{ |
|
1919 |
currMV->y=min_dy; |
currMV->y=min_dy; |
1920 |
} |
|
1921 |
|
/***************** This is predictor SET A: only median prediction ******************/ |
1922 |
|
|
1923 |
iMinSAD = sad16( cur, |
iMinSAD = sad16( cur, |
1924 |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
1925 |
iEdgedWidth, MV_MAX_ERROR); |
iEdgedWidth, MV_MAX_ERROR); |
1926 |
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); |
1927 |
|
|
1928 |
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,pMB->mvs[0])) && (iMinSAD < pMB->sad16) ) ) |
// thresh1 is fixed to 256 |
1929 |
|
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV, prevMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
1930 |
{ |
{ |
|
|
|
1931 |
if (MotionFlags & PMV_QUICKSTOP16) |
if (MotionFlags & PMV_QUICKSTOP16) |
1932 |
goto step10b; |
goto EPZS16_Terminate_without_Refine; |
1933 |
if (MotionFlags & PMV_EARLYSTOP16) |
if (MotionFlags & PMV_EARLYSTOP16) |
1934 |
goto step10; |
goto EPZS16_Terminate_with_Refine; |
1935 |
} |
} |
1936 |
|
|
1937 |
/* |
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
|
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
|
|
Also calculate (0,0) but do not subtract offset. |
|
|
Let MinSAD be the smallest SAD up to this point. |
|
|
If MV is (0,0) subtract offset. ******** WHAT'S THIS 'OFFSET' ??? *********** |
|
|
*/ |
|
1938 |
|
|
1939 |
// (0,0) is always possible |
// previous frame MV |
1940 |
|
CHECK_MV16_CANDIDATE(prevMB->mvs[0].x,prevMB->mvs[0].y); |
1941 |
|
|
1942 |
|
// set threshhold based on Min of Prediction and SAD of collocated block |
1943 |
|
// CHECK_MV16 always uses iSAD for the SAD of last vector to check, so now iSAD is what we want |
1944 |
|
|
1945 |
|
if ((x==0) && (y==0) ) |
1946 |
|
{ |
1947 |
|
thresh2 = 512; |
1948 |
|
} |
1949 |
|
else |
1950 |
|
{ |
1951 |
|
/* T_k = 1.2 * MIN(SAD_top,SAD_left,SAD_topleft,SAD_coll) +128; [Tourapis, 2002] */ |
1952 |
|
|
1953 |
|
thresh2 = MIN(psad[0],iSAD)*6/5 + 128; |
1954 |
|
} |
1955 |
|
|
1956 |
|
// MV=(0,0) is often a good choice |
1957 |
|
|
1958 |
CHECK_MV16_ZERO; |
CHECK_MV16_ZERO; |
1959 |
|
|
|
// previous frame MV is always possible |
|
|
CHECK_MV16_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); |
|
1960 |
|
|
1961 |
// left neighbour, if allowed |
// left neighbour, if allowed |
1962 |
if (x != 0) |
if (x != 0) |
1978 |
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
1979 |
|
|
1980 |
// top right neighbour, if allowed |
// top right neighbour, if allowed |
1981 |
if (x != (iWcount-1)) |
if ((uint32_t)x != (iWcount-1)) |
1982 |
{ |
{ |
1983 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
if (!(MotionFlags & PMV_HALFPEL16 )) |
1984 |
{ pmv[3].x = EVEN(pmv[3].x); |
{ pmv[3].x = EVEN(pmv[3].x); |
1988 |
} |
} |
1989 |
} |
} |
1990 |
|
|
1991 |
/* Step 6: If MinSAD <= thresa goto Step 10. |
/* Terminate if MinSAD <= T_2 |
1992 |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
1993 |
*/ |
*/ |
1994 |
|
|
1995 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,pMB->mvs[0]) && (iMinSAD < pMB->sad16) ) ) |
if ( (iMinSAD <= thresh2) |
1996 |
|
|| ( MVequal(*currMV,prevMB->mvs[0]) && ((uint32_t)iMinSAD <= prevMB->sad16) ) ) |
1997 |
{ |
{ |
1998 |
if (MotionFlags & PMV_QUICKSTOP16) |
if (MotionFlags & PMV_QUICKSTOP16) |
1999 |
goto step10b; |
goto EPZS16_Terminate_without_Refine; |
2000 |
if (MotionFlags & PMV_EARLYSTOP16) |
if (MotionFlags & PMV_EARLYSTOP16) |
2001 |
goto step10; |
goto EPZS16_Terminate_with_Refine; |
2002 |
} |
} |
2003 |
|
|
2004 |
|
/***** predictor SET C: acceleration MV (new!), neighbours in prev. frame(new!) ****/ |
2005 |
|
|
2006 |
/************ (Diamond Search) **************/ |
backupMV = prevMB->mvs[0]; // collocated MV |
2007 |
/* |
backupMV.x += (prevMB->mvs[0].x - oldMB->mvs[0].x ); // acceleration X |
2008 |
Step 7: Perform Diamond search, with either the small or large diamond. |
backupMV.y += (prevMB->mvs[0].y - oldMB->mvs[0].y ); // acceleration Y |
2009 |
If Found=2 only examine one Diamond pattern, and afterwards goto step 10 |
|
2010 |
Step 8: If small diamond, iterate small diamond search pattern until motion vector lies in the center of the diamond. |
CHECK_MV16_CANDIDATE(backupMV.x,backupMV.y); |
2011 |
If center then goto step 10. |
|
2012 |
Step 9: If large diamond, iterate large diamond search pattern until motion vector lies in the center. |
// left neighbour |
2013 |
Refine by using small diamond and goto step 10. |
if (x != 0) |
2014 |
*/ |
CHECK_MV16_CANDIDATE((prevMB-1)->mvs[0].x,(prevMB-1)->mvs[0].y); |
2015 |
|
|
2016 |
|
// top neighbour |
2017 |
|
if (y != 0) |
2018 |
|
CHECK_MV16_CANDIDATE((prevMB-iWcount)->mvs[0].x,(prevMB-iWcount)->mvs[0].y); |
2019 |
|
|
2020 |
|
// right neighbour, if allowed (this value is not written yet, so take it from pMB->mvs |
2021 |
|
|
2022 |
|
if ((uint32_t)x != iWcount-1) |
2023 |
|
CHECK_MV16_CANDIDATE((prevMB+1)->mvs[0].x,(prevMB+1)->mvs[0].y); |
2024 |
|
|
2025 |
|
// bottom neighbour, dito |
2026 |
|
if ((uint32_t)y != iHcount-1) |
2027 |
|
CHECK_MV16_CANDIDATE((prevMB+iWcount)->mvs[0].x,(prevMB+iWcount)->mvs[0].y); |
2028 |
|
|
2029 |
|
/* Terminate if MinSAD <= T_3 (here T_3 = T_2) */ |
2030 |
|
if (iMinSAD <= thresh2) |
2031 |
|
{ |
2032 |
|
if (MotionFlags & PMV_QUICKSTOP16) |
2033 |
|
goto EPZS16_Terminate_without_Refine; |
2034 |
|
if (MotionFlags & PMV_EARLYSTOP16) |
2035 |
|
goto EPZS16_Terminate_with_Refine; |
2036 |
|
} |
2037 |
|
|
2038 |
|
/************ (if Diamond Search) **************/ |
2039 |
|
|
2040 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
2041 |
|
|
2042 |
/* 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 */ |
2043 |
iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
|
2044 |
|
iSAD = (*MainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
2045 |
x, y, |
x, y, |
2046 |
currMV->x, currMV->y, iMinSAD, &newMV, |
currMV->x, currMV->y, iMinSAD, &newMV, pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
2047 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
2, iFcode, iQuant, 0); |
2048 |
|
|
2049 |
if (iSAD < iMinSAD) |
if (iSAD < iMinSAD) |
2050 |
{ |
{ |
2052 |
iMinSAD = iSAD; |
iMinSAD = iSAD; |
2053 |
} |
} |
2054 |
|
|
2055 |
|
|
2056 |
if (MotionFlags & PMV_EXTSEARCH16) |
if (MotionFlags & PMV_EXTSEARCH16) |
2057 |
{ |
{ |
2058 |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
2059 |
|
|
2060 |
if (!(MVequal(pmv[0],backupMV)) ) |
if (!(MVequal(pmv[0],backupMV)) ) |
2061 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
{ |
2062 |
|
iSAD = (*MainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
2063 |
x, y, |
x, y, |
2064 |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
2065 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); |
2066 |
|
} |
2067 |
|
|
2068 |
if (iSAD < iMinSAD) |
if (iSAD < iMinSAD) |
2069 |
{ |
{ |
2070 |
*currMV = newMV; |
*currMV = newMV; |
2071 |
iMinSAD = iSAD; |
iMinSAD = iSAD; |
2072 |
} |
} |
|
} |
|
2073 |
|
|
2074 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
2075 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
{ |
2076 |
|
iSAD = (*MainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
2077 |
x, y, |
x, y, |
2078 |
0, 0, iMinSAD, &newMV, |
0, 0, iMinSAD, &newMV, |
2079 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); |
2080 |
|
|
2081 |
if (iSAD < iMinSAD) |
if (iSAD < iMinSAD) |
2082 |
{ |
{ |
2086 |
} |
} |
2087 |
} |
} |
2088 |
|
|
2089 |
/* |
/*************** Choose best MV found **************/ |
|
Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
|
|
*/ |
|
2090 |
|
|
2091 |
step10: |
EPZS16_Terminate_with_Refine: |
2092 |
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
2093 |
iMinSAD = PMVfastSearch16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
iMinSAD = Halfpel16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
2094 |
x, y, |
x, y, |
2095 |
currMV, iMinSAD, |
currMV, iMinSAD, |
2096 |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
2097 |
|
|
2098 |
step10b: |
EPZS16_Terminate_without_Refine: |
|
currPMV->x = currMV->x - pmv[0].x; |
|
|
currPMV->y = currMV->y - pmv[0].y; |
|
|
return iMinSAD; |
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
int32_t PMVfastSearch8_MainSearch( |
|
|
const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, const int y, |
|
|
int32_t startx, int32_t starty, |
|
|
int32_t iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const VECTOR * const pmv, |
|
|
const int32_t min_dx, const int32_t max_dx, |
|
|
const int32_t min_dy, const int32_t max_dy, |
|
|
const int32_t iEdgedWidth, |
|
|
const int32_t iDiamondSize, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
int iFound) |
|
|
{ |
|
|
/* Do a diamond search around given starting point, return SAD of best */ |
|
|
|
|
|
int32_t iDirection=0; |
|
|
int32_t iSAD; |
|
|
VECTOR backupMV; |
|
|
backupMV.x = startx; |
|
|
backupMV.y = starty; |
|
|
|
|
|
/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ |
|
|
|
|
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
|
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
|
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
|
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
|
|
|
|
|
if (iDirection) |
|
|
while (!iFound) |
|
|
{ |
|
|
iFound = 1; |
|
|
backupMV=*currMV; // since iDirection!=0, this is well defined! |
|
|
|
|
|
if ( iDirection != 2) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
|
|
if ( iDirection != 1) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x+iDiamondSize,backupMV.y,2); |
|
|
if ( iDirection != 4) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x,backupMV.y-iDiamondSize,3); |
|
|
if ( iDirection != 3) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x,backupMV.y+iDiamondSize,4); |
|
|
} |
|
|
else |
|
|
{ |
|
|
currMV->x = startx; |
|
|
currMV->y = starty; |
|
|
} |
|
|
return iMinSAD; |
|
|
} |
|
|
|
|
|
int32_t PMVfastSearch8_Refine( |
|
|
const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, const int y, |
|
|
VECTOR * const currMV, |
|
|
int32_t iMinSAD, |
|
|
const VECTOR * const pmv, |
|
|
const int32_t min_dx, const int32_t max_dx, |
|
|
const int32_t min_dy, const int32_t max_dy, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
const int32_t iEdgedWidth) |
|
|
{ |
|
|
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
|
|
|
|
|
int32_t iSAD; |
|
|
VECTOR backupMV = *currMV; |
|
2099 |
|
|
2100 |
CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y-1); |
*oldMB = *prevMB; |
|
CHECK_MV8_CANDIDATE(backupMV.x ,backupMV.y-1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y-1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y+1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x ,backupMV.y+1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y+1); |
|
2101 |
|
|
2102 |
|
currPMV->x = currMV->x - pmv[0].x; |
2103 |
|
currPMV->y = currMV->y - pmv[0].y; |
2104 |
return iMinSAD; |
return iMinSAD; |
2105 |
} |
} |
2106 |
|
|
2107 |
|
|
2108 |
#define PMV_HALFPEL8 (PMV_HALFPELDIAMOND8|PMV_HALFPELREFINE8) |
int32_t EPZSSearch8( |
|
|
|
|
int32_t PMVfastSearch8( |
|
2109 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
2110 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
2111 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
2112 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
2113 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
2114 |
const int x, const int y, |
const int x, const int y, |
2115 |
const int start_x, int start_y, |
const int start_x, const int start_y, |
2116 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
2117 |
MBParam * const pParam, |
const uint32_t iQuant, |
2118 |
MACROBLOCK * const pMBs, |
const uint32_t iFcode, |
2119 |
|
const MBParam * const pParam, |
2120 |
|
const MACROBLOCK * const pMBs, |
2121 |
|
const MACROBLOCK * const prevMBs, |
2122 |
VECTOR * const currMV, |
VECTOR * const currMV, |
2123 |
VECTOR * const currPMV) |
VECTOR * const currPMV) |
2124 |
{ |
{ |
2125 |
const uint32_t iWcount = pParam->mb_width; |
/* Please not that EPZS might not be a good choice for 8x8-block motion search ! */ |
2126 |
|
|
2127 |
const int32_t iFcode = pParam->fixed_code; |
const uint32_t iWcount = pParam->mb_width; |
|
const int32_t iQuant = pParam->quant; |
|
2128 |
const int32_t iWidth = pParam->width; |
const int32_t iWidth = pParam->width; |
2129 |
const int32_t iHeight = pParam->height; |
const int32_t iHeight = pParam->height; |
2130 |
const int32_t iEdgedWidth = pParam->edged_width; |
const int32_t iEdgedWidth = pParam->edged_width; |
2131 |
|
|
2132 |
const uint8_t * cur = pCur->y + x*8 + y*8*iEdgedWidth; |
const uint8_t * cur = pCur->y + x*8 + y*8*iEdgedWidth; |
2133 |
|
|
2134 |
int32_t iDiamondSize; |
int32_t iDiamondSize=1; |
2135 |
|
|
2136 |
int32_t min_dx; |
int32_t min_dx; |
2137 |
int32_t max_dx; |
int32_t max_dx; |
2138 |
int32_t min_dy; |
int32_t min_dy; |
2139 |
int32_t max_dy; |
int32_t max_dy; |
2140 |
|
|
|
VECTOR pmv[4]; |
|
|
int32_t psad[4]; |
|
2141 |
VECTOR newMV; |
VECTOR newMV; |
2142 |
VECTOR backupMV; |
VECTOR backupMV; |
2143 |
|
|
2144 |
MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
VECTOR pmv[4]; |
2145 |
|
int32_t psad[8]; |
2146 |
|
|
2147 |
|
const int32_t iSubBlock = ((y&1)<<1) + (x&1); |
2148 |
|
|
2149 |
static int32_t threshA,threshB; |
// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
2150 |
int32_t iFound,bPredEq; |
const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; |
|
int32_t iMinSAD,iSAD; |
|
2151 |
|
|
2152 |
int32_t iSubBlock = ((y&1)<<1) + (x&1); |
int32_t bPredEq; |
2153 |
|
int32_t iMinSAD,iSAD=9999; |
2154 |
|
|
2155 |
|
MainSearch8FuncPtr MainSearchPtr; |
2156 |
|
|
2157 |
/* Get maximum range */ |
/* Get maximum range */ |
2158 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
2159 |
x, y, 8, iWidth, iHeight, iFcode); |
x, y, 8, iWidth, iHeight, iFcode); |
2160 |
|
|
2161 |
/* we work with abs. MVs, not relative to prediction, so range is relative to 0,0 */ |
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
2162 |
|
|
2163 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8 )) |
if (!(MotionFlags & PMV_HALFPEL8 )) |
2164 |
{ min_dx = EVEN(min_dx); |
{ min_dx = EVEN(min_dx); |
2165 |
max_dx = EVEN(max_dx); |
max_dx = EVEN(max_dx); |
2166 |
min_dy = EVEN(min_dy); |
min_dy = EVEN(min_dy); |
2167 |
max_dy = EVEN(max_dy); |
max_dy = EVEN(max_dy); |
2168 |
} /* because we might use IF (dx>max_dx) THEN dx=max_dx; */ |
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
|
|
|
|
|
|
|
bPredEq = get_pmvdata(pMBs, (x>>1), (y>>1), iWcount, iSubBlock, pmv, psad); |
|
|
|
|
|
if ((x==0) && (y==0) ) |
|
|
{ |
|
|
threshA = 512/4; |
|
|
threshB = 1024/4; |
|
|
|
|
|
} |
|
|
else |
|
|
{ |
|
|
threshA = psad[0]/4; /* good estimate */ |
|
|
threshB = threshA+256/4; |
|
|
if (threshA< 512/4) threshA = 512/4; |
|
|
if (threshA>1024/4) threshA = 1024/4; |
|
|
if (threshB>1792/4) threshB = 1792/4; |
|
|
} |
|
|
|
|
|
iFound=0; |
|
|
|
|
|
/* Step 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. |
|
|
*/ |
|
2169 |
|
|
2170 |
if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536/4) || (bPredEq) ) |
bPredEq = get_pmvdata(pMBs, x>>1, y>>1, iWcount, iSubBlock, pmv, psad); |
|
iDiamondSize=1; // 1 halfpel! |
|
|
else |
|
|
iDiamondSize=2; // 2 halfpel = 1 full pixel! |
|
2171 |
|
|
|
if (!(MotionFlags & PMV_HALFPELDIAMOND8) ) |
|
|
iDiamondSize*=2; |
|
2172 |
|
|
2173 |
/* Step 4: Calculate SAD around the Median prediction. |
/* Step 4: Calculate SAD around the Median prediction. |
2174 |
MinSAD=SAD |
MinSAD=SAD |
2177 |
If SAD<=256 goto Step 10. |
If SAD<=256 goto Step 10. |
2178 |
*/ |
*/ |
2179 |
|
|
|
|
|
2180 |
// Prepare for main loop |
// Prepare for main loop |
2181 |
|
|
2182 |
currMV->x=start_x; /* start with mv16 */ |
|
2183 |
currMV->y=start_y; |
if (!(MotionFlags & PMV_HALFPEL8)) |
2184 |
|
{ |
2185 |
|
currMV->x = EVEN(currMV->x); |
2186 |
|
currMV->y = EVEN(currMV->y); |
2187 |
|
} |
2188 |
|
|
2189 |
|
if (currMV->x > max_dx) |
2190 |
|
currMV->x=max_dx; |
2191 |
|
if (currMV->x < min_dx) |
2192 |
|
currMV->x=min_dx; |
2193 |
|
if (currMV->y > max_dy) |
2194 |
|
currMV->y=max_dy; |
2195 |
|
if (currMV->y < min_dy) |
2196 |
|
currMV->y=min_dy; |
2197 |
|
|
2198 |
|
/***************** This is predictor SET A: only median prediction ******************/ |
2199 |
|
|
2200 |
|
|
2201 |
iMinSAD = sad8( cur, |
iMinSAD = sad8( cur, |
2202 |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
2203 |
iEdgedWidth); |
iEdgedWidth); |
2204 |
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); |
2205 |
|
|
2206 |
|
|
2207 |
if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,pMB->mvs[iSubBlock])) && (iMinSAD < pMB->sad8[iSubBlock]) ) ) |
// thresh1 is fixed to 256 |
2208 |
|
if (iMinSAD < 256/4 ) |
2209 |
{ |
{ |
2210 |
if (MotionFlags & PMV_QUICKSTOP8) |
if (MotionFlags & PMV_QUICKSTOP8) |
2211 |
goto step10_8b; |
goto EPZS8_Terminate_without_Refine; |
2212 |
if (MotionFlags & PMV_EARLYSTOP8) |
if (MotionFlags & PMV_EARLYSTOP8) |
2213 |
goto step10_8; |
goto EPZS8_Terminate_with_Refine; |
2214 |
} |
} |
2215 |
|
|
2216 |
/* |
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
|
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
|
|
Also calculate (0,0) but do not subtract offset. |
|
|
Let MinSAD be the smallest SAD up to this point. |
|
|
If MV is (0,0) subtract offset. ******** WHAT'S THIS 'OFFSET' ??? *********** |
|
|
*/ |
|
2217 |
|
|
|
// the prediction might be even better than mv16 |
|
|
CHECK_MV8_CANDIDATE(pmv[0].x,pmv[0].y); |
|
2218 |
|
|
2219 |
// (0,0) is always possible |
// MV=(0,0) is often a good choice |
2220 |
CHECK_MV8_ZERO; |
CHECK_MV8_ZERO; |
2221 |
|
|
2222 |
// previous frame MV is always possible |
// previous frame MV |
2223 |
CHECK_MV8_CANDIDATE(pMB->mvs[iSubBlock].x,pMB->mvs[iSubBlock].y); |
CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x,prevMB->mvs[iSubBlock].y); |
2224 |
|
|
2225 |
// left neighbour, if allowed |
// left neighbour, if allowed |
2226 |
if (psad[1] != MV_MAX_ERROR) |
if (psad[1] != MV_MAX_ERROR) |
2252 |
} |
} |
2253 |
} |
} |
2254 |
|
|
2255 |
/* Step 6: If MinSAD <= thresa goto Step 10. |
/* // this bias is zero anyway, at the moment! |
2256 |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
|
2257 |
|
if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) ) // && (iMinSAD <= iQuant * 96) |
2258 |
|
iMinSAD -= MV8_00_BIAS; |
2259 |
|
|
2260 |
*/ |
*/ |
2261 |
|
|
2262 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,pMB->mvs[iSubBlock]) && (iMinSAD < pMB->sad8[iSubBlock]) ) ) |
/* Terminate if MinSAD <= T_2 |
2263 |
|
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
2264 |
|
*/ |
2265 |
|
|
2266 |
|
if (iMinSAD < 512/4) /* T_2 == 512/4 hardcoded */ |
2267 |
{ |
{ |
2268 |
if (MotionFlags & PMV_QUICKSTOP8) |
if (MotionFlags & PMV_QUICKSTOP8) |
2269 |
goto step10_8b; |
goto EPZS8_Terminate_without_Refine; |
2270 |
if (MotionFlags & PMV_EARLYSTOP8) |
if (MotionFlags & PMV_EARLYSTOP8) |
2271 |
goto step10_8; |
goto EPZS8_Terminate_with_Refine; |
2272 |
} |
} |
2273 |
|
|
2274 |
/************ (Diamond Search) **************/ |
/************ (Diamond Search) **************/ |
|
/* |
|
|
Step 7: Perform Diamond search, with either the small or large diamond. |
|
|
If Found=2 only examine one Diamond pattern, and afterwards goto step 10 |
|
|
Step 8: If small diamond, iterate small diamond search pattern until motion vector lies in the center of the diamond. |
|
|
If center then goto step 10. |
|
|
Step 9: If large diamond, iterate large diamond search pattern until motion vector lies in the center. |
|
|
Refine by using small diamond and goto step 10. |
|
|
*/ |
|
2275 |
|
|
2276 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
2277 |
|
|
2278 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
if (!(MotionFlags & PMV_HALFPELDIAMOND8)) |
2279 |
iSAD = PMVfastSearch8_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
iDiamondSize *= 2; |
2280 |
|
|
2281 |
|
/* default: use best prediction as starting point for one call of EPZS_MainSearch */ |
2282 |
|
|
2283 |
|
/* // there is no EPZS^2 for inter4v at the moment |
2284 |
|
|
2285 |
|
if (MotionFlags & PMV_USESQUARES8) |
2286 |
|
MainSearchPtr = Square8_MainSearch; |
2287 |
|
else |
2288 |
|
*/ |
2289 |
|
|
2290 |
|
// if (MotionFlags & PMV_USESQUARES8) |
2291 |
|
// MainSearchPtr = Square8_MainSearch; |
2292 |
|
// else |
2293 |
|
|
2294 |
|
if (MotionFlags & PMV_ADVANCEDDIAMOND8) |
2295 |
|
MainSearchPtr = AdvDiamond8_MainSearch; |
2296 |
|
else |
2297 |
|
MainSearchPtr = Diamond8_MainSearch; |
2298 |
|
|
2299 |
|
iSAD = (*MainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
2300 |
x, y, |
x, y, |
2301 |
currMV->x, currMV->y, iMinSAD, &newMV, |
currMV->x, currMV->y, iMinSAD, &newMV, |
2302 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
2303 |
|
iDiamondSize, iFcode, iQuant, 0); |
2304 |
|
|
2305 |
|
|
2306 |
if (iSAD < iMinSAD) |
if (iSAD < iMinSAD) |
2307 |
{ |
{ |
2311 |
|
|
2312 |
if (MotionFlags & PMV_EXTSEARCH8) |
if (MotionFlags & PMV_EXTSEARCH8) |
2313 |
{ |
{ |
2314 |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
2315 |
|
|
2316 |
if (!(MVequal(pmv[0],backupMV)) ) |
if (!(MVequal(pmv[0],backupMV)) ) |
2317 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
{ |
2318 |
|
iSAD = (*MainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
2319 |
x, y, |
x, y, |
2320 |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
2321 |
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, 0); |
2322 |
|
|
2323 |
if (iSAD < iMinSAD) |
if (iSAD < iMinSAD) |
2324 |
{ |
{ |
2328 |
} |
} |
2329 |
|
|
2330 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
2331 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
{ |
2332 |
|
iSAD = (*MainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
2333 |
x, y, |
x, y, |
2334 |
0, 0, iMinSAD, &newMV, |
0, 0, iMinSAD, &newMV, |
2335 |
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, 0); |
2336 |
|
|
2337 |
if (iSAD < iMinSAD) |
if (iSAD < iMinSAD) |
2338 |
{ |
{ |
2342 |
} |
} |
2343 |
} |
} |
2344 |
|
|
2345 |
/* Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
/*************** Choose best MV found **************/ |
|
By performing an optional local half-pixel search, we can refine this result even further. |
|
|
*/ |
|
2346 |
|
|
2347 |
step10_8: |
EPZS8_Terminate_with_Refine: |
2348 |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
2349 |
iMinSAD = PMVfastSearch8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
iMinSAD = Halfpel8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
2350 |
x, y, |
x, y, |
2351 |
currMV, iMinSAD, |
currMV, iMinSAD, |
2352 |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
2353 |
|
|
2354 |
step10_8b: |
EPZS8_Terminate_without_Refine: |
2355 |
|
|
2356 |
currPMV->x = currMV->x - pmv[0].x; |
currPMV->x = currMV->x - pmv[0].x; |
2357 |
currPMV->y = currMV->y - pmv[0].y; |
currPMV->y = currMV->y - pmv[0].y; |
|
|
|
2358 |
return iMinSAD; |
return iMinSAD; |
2359 |
} |
} |
2360 |
|
|
2361 |
|
|
2362 |
|
|
2363 |
|
|
2364 |
|
|
2365 |
|
/* *********************************************************** |
2366 |
|
bvop motion estimation |
2367 |
|
// TODO: need to incorporate prediction here (eg. sad += calc_delta_16) |
2368 |
|
***************************************************************/ |
2369 |
|
|
2370 |
|
|
2371 |
|
void MotionEstimationBVOP( |
2372 |
|
MBParam * const pParam, |
2373 |
|
FRAMEINFO * const frame, |
2374 |
|
|
2375 |
|
// forward (past) reference |
2376 |
|
const MACROBLOCK * const f_mbs, |
2377 |
|
const IMAGE * const f_ref, |
2378 |
|
const IMAGE * const f_refH, |
2379 |
|
const IMAGE * const f_refV, |
2380 |
|
const IMAGE * const f_refHV, |
2381 |
|
// backward (future) reference |
2382 |
|
const MACROBLOCK * const b_mbs, |
2383 |
|
const IMAGE * const b_ref, |
2384 |
|
const IMAGE * const b_refH, |
2385 |
|
const IMAGE * const b_refV, |
2386 |
|
const IMAGE * const b_refHV) |
2387 |
|
{ |
2388 |
|
const uint32_t mb_width = pParam->mb_width; |
2389 |
|
const uint32_t mb_height = pParam->mb_height; |
2390 |
|
const int32_t edged_width = pParam->edged_width; |
2391 |
|
|
2392 |
|
uint32_t i,j; |
2393 |
|
|
2394 |
|
int32_t f_sad16; |
2395 |
|
int32_t b_sad16; |
2396 |
|
int32_t i_sad16; |
2397 |
|
int32_t d_sad16; |
2398 |
|
int32_t best_sad; |
2399 |
|
|
2400 |
|
VECTOR pmv_dontcare; |
2401 |
|
|
2402 |
|
// note: i==horizontal, j==vertical |
2403 |
|
for (j = 0; j < mb_height; j++) |
2404 |
|
{ |
2405 |
|
for (i = 0; i < mb_width; i++) |
2406 |
|
{ |
2407 |
|
MACROBLOCK *mb = &frame->mbs[i + j*mb_width]; |
2408 |
|
const MACROBLOCK *f_mb = &f_mbs[i + j*mb_width]; |
2409 |
|
const MACROBLOCK *b_mb = &b_mbs[i + j*mb_width]; |
2410 |
|
|
2411 |
|
if (b_mb->mode == MODE_INTER |
2412 |
|
&& b_mb->cbp == 0 |
2413 |
|
&& b_mb->mvs[0].x == 0 |
2414 |
|
&& b_mb->mvs[0].y == 0) |
2415 |
|
{ |
2416 |
|
mb->mode = MODE_NOT_CODED; |
2417 |
|
mb->mvs[0].x = 0; |
2418 |
|
mb->mvs[0].y = 0; |
2419 |
|
mb->b_mvs[0].x = 0; |
2420 |
|
mb->b_mvs[0].y = 0; |
2421 |
|
continue; |
2422 |
|
} |
2423 |
|
|
2424 |
|
|
2425 |
|
// forward search |
2426 |
|
f_sad16 = SEARCH16(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
2427 |
|
&frame->image, |
2428 |
|
i, j, |
2429 |
|
frame->motion_flags, frame->quant, frame->fcode, |
2430 |
|
pParam, |
2431 |
|
f_mbs, f_mbs /* todo */, |
2432 |
|
&mb->mvs[0], &pmv_dontcare); // ignore pmv |
2433 |
|
|
2434 |
|
// backward search |
2435 |
|
b_sad16 = SEARCH16(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
2436 |
|
&frame->image, |
2437 |
|
i, j, |
2438 |
|
frame->motion_flags, frame->quant, frame->bcode, |
2439 |
|
pParam, |
2440 |
|
b_mbs, b_mbs, /* todo */ |
2441 |
|
&mb->b_mvs[0], &pmv_dontcare); // ignore pmv |
2442 |
|
|
2443 |
|
// interpolate search (simple, but effective) |
2444 |
|
i_sad16 = sad16bi_c( |
2445 |
|
frame->image.y + i*16 + j*16*edged_width, |
2446 |
|
get_ref(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
2447 |
|
i, j, 16, mb->mvs[0].x, mb->mvs[0].y, edged_width), |
2448 |
|
get_ref(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
2449 |
|
i, j, 16, mb->b_mvs[0].x, mb->b_mvs[0].x, edged_width), |
2450 |
|
edged_width); |
2451 |
|
|
2452 |
|
// TODO: direct search |
2453 |
|
// predictor + range of [-32,32] |
2454 |
|
d_sad16 = 65535; |
2455 |
|
|
2456 |
|
|
2457 |
|
if (f_sad16 < b_sad16) |
2458 |
|
{ |
2459 |
|
best_sad = f_sad16; |
2460 |
|
mb->mode = MODE_FORWARD; |
2461 |
|
} |
2462 |
|
else |
2463 |
|
{ |
2464 |
|
best_sad = b_sad16; |
2465 |
|
mb->mode = MODE_BACKWARD; |
2466 |
|
} |
2467 |
|
|
2468 |
|
if (i_sad16 < best_sad) |
2469 |
|
{ |
2470 |
|
best_sad = i_sad16; |
2471 |
|
mb->mode = MODE_INTERPOLATE; |
2472 |
|
} |
2473 |
|
|
2474 |
|
if (d_sad16 < best_sad) |
2475 |
|
{ |
2476 |
|
best_sad = d_sad16; |
2477 |
|
mb->mode = MODE_DIRECT; |
2478 |
|
} |
2479 |
|
|
2480 |
|
} |
2481 |
|
} |
2482 |
|
} |