2 |
* |
* |
3 |
* Modifications: |
* Modifications: |
4 |
* |
* |
5 |
|
* 25.04.2002 partial prevMB conversion |
6 |
|
* 22.04.2002 remove some compile warning by chenm001 <chenm001@163.com> |
7 |
|
* 14.04.2002 added MotionEstimationBVOP() |
8 |
|
* 02.04.2002 add EPZS(^2) as ME algorithm, use PMV_USESQUARES to choose between |
9 |
|
* EPZS and EPZS^2 |
10 |
* 08.02.2002 split up PMVfast into three routines: PMVFast, PMVFast_MainLoop |
* 08.02.2002 split up PMVfast into three routines: PMVFast, PMVFast_MainLoop |
11 |
* PMVFast_Refine to support multiple searches with different start points |
* PMVFast_Refine to support multiple searches with different start points |
12 |
* 07.01.2002 uv-block-based interpolation |
* 07.01.2002 uv-block-based interpolation |
36 |
|
|
37 |
#include <assert.h> |
#include <assert.h> |
38 |
#include <stdio.h> |
#include <stdio.h> |
39 |
|
#include <stdlib.h> |
40 |
|
|
41 |
#include "../encoder.h" |
#include "../encoder.h" |
42 |
#include "../utils/mbfunctions.h" |
#include "../utils/mbfunctions.h" |
43 |
#include "../prediction/mbprediction.h" |
#include "../prediction/mbprediction.h" |
44 |
#include "../global.h" |
#include "../global.h" |
45 |
#include "../utils/timer.h" |
#include "../utils/timer.h" |
46 |
|
#include "motion.h" |
47 |
#include "sad.h" |
#include "sad.h" |
48 |
|
|
49 |
// very large value |
// very large value |
72 |
#define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) |
#define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) |
73 |
|
|
74 |
|
|
75 |
#define MIN(X, Y) ((X)<(Y)?(X):(Y)) |
int32_t PMVfastSearch16( |
|
#define MAX(X, Y) ((X)>(Y)?(X):(Y)) |
|
|
#define ABS(X) (((X)>0)?(X):-(X)) |
|
|
#define SIGN(X) (((X)>0)?1:-1) |
|
|
|
|
|
|
|
|
int32_t PMVfastSearch8( |
|
76 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
77 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
78 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
79 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
80 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
81 |
const int x, const int y, |
const int x, const int y, |
82 |
const int start_x, int start_y, |
const uint32_t MotionFlags, |
83 |
const uint32_t iQuality, |
const uint32_t iQuant, |
84 |
MBParam * const pParam, |
const uint32_t iFcode, |
85 |
MACROBLOCK * const pMBs, |
const MBParam * const pParam, |
86 |
|
const MACROBLOCK * const pMBs, |
87 |
|
const MACROBLOCK * const prevMBs, |
88 |
VECTOR * const currMV, |
VECTOR * const currMV, |
89 |
VECTOR * const currPMV); |
VECTOR * const currPMV); |
90 |
|
|
91 |
int32_t PMVfastSearch16( |
int32_t EPZSSearch16( |
92 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
93 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
94 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
95 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
96 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
97 |
const int x, const int y, |
const int x, const int y, |
98 |
const uint32_t iQuality, |
const uint32_t MotionFlags, |
99 |
MBParam * const pParam, |
const uint32_t iQuant, |
100 |
MACROBLOCK * const pMBs, |
const uint32_t iFcode, |
101 |
|
const MBParam * const pParam, |
102 |
|
const MACROBLOCK * const pMBs, |
103 |
|
const MACROBLOCK * const prevMBs, |
104 |
VECTOR * const currMV, |
VECTOR * const currMV, |
105 |
VECTOR * const currPMV); |
VECTOR * const currPMV); |
106 |
|
|
107 |
|
|
108 |
|
int32_t PMVfastSearch8( |
109 |
|
const uint8_t * const pRef, |
110 |
|
const uint8_t * const pRefH, |
111 |
|
const uint8_t * const pRefV, |
112 |
|
const uint8_t * const pRefHV, |
113 |
|
const IMAGE * const pCur, |
114 |
|
const int x, const int y, |
115 |
|
const int start_x, const int start_y, |
116 |
|
const uint32_t MotionFlags, |
117 |
|
const uint32_t iQuant, |
118 |
|
const uint32_t iFcode, |
119 |
|
const MBParam * const pParam, |
120 |
|
const MACROBLOCK * const pMBs, |
121 |
|
const MACROBLOCK * const prevMBs, |
122 |
|
VECTOR * const currMV, |
123 |
|
VECTOR * const currPMV); |
124 |
|
|
125 |
/* diamond search stuff |
int32_t EPZSSearch8( |
126 |
keep the the sequence in circular order (so optimization works) |
const uint8_t * const pRef, |
127 |
*/ |
const uint8_t * const pRefH, |
128 |
|
const uint8_t * const pRefV, |
129 |
|
const uint8_t * const pRefHV, |
130 |
|
const IMAGE * const pCur, |
131 |
|
const int x, const int y, |
132 |
|
const int start_x, const int start_y, |
133 |
|
const uint32_t MotionFlags, |
134 |
|
const uint32_t iQuant, |
135 |
|
const uint32_t iFcode, |
136 |
|
const MBParam * const pParam, |
137 |
|
const MACROBLOCK * const pMBs, |
138 |
|
const MACROBLOCK * const prevMBs, |
139 |
|
VECTOR * const currMV, |
140 |
|
VECTOR * const currPMV); |
141 |
|
|
|
typedef struct |
|
|
{ |
|
|
int32_t dx; |
|
|
int32_t dy; |
|
|
} |
|
|
DPOINT; |
|
142 |
|
|
143 |
|
typedef int32_t (MainSearch16Func)( |
144 |
|
const uint8_t * const pRef, |
145 |
|
const uint8_t * const pRefH, |
146 |
|
const uint8_t * const pRefV, |
147 |
|
const uint8_t * const pRefHV, |
148 |
|
const uint8_t * const cur, |
149 |
|
const int x, const int y, |
150 |
|
int32_t startx, int32_t starty, |
151 |
|
int32_t iMinSAD, |
152 |
|
VECTOR * const currMV, |
153 |
|
const VECTOR * const pmv, |
154 |
|
const int32_t min_dx, const int32_t max_dx, |
155 |
|
const int32_t min_dy, const int32_t max_dy, |
156 |
|
const int32_t iEdgedWidth, |
157 |
|
const int32_t iDiamondSize, |
158 |
|
const int32_t iFcode, |
159 |
|
const int32_t iQuant, |
160 |
|
int iFound); |
161 |
|
|
162 |
static const DPOINT diamond_small[4] = |
typedef MainSearch16Func* MainSearch16FuncPtr; |
|
{ |
|
|
{0, 1}, {1, 0}, {0, -1}, {-1, 0} |
|
|
}; |
|
163 |
|
|
164 |
|
|
165 |
static const DPOINT diamond_large[8] = |
typedef int32_t (MainSearch8Func)( |
166 |
{ |
const uint8_t * const pRef, |
167 |
{0, 2}, {1, 1}, {2, 0}, {1, -1}, {0, -2}, {-1, -1}, {-2, 0}, {-1, 1} |
const uint8_t * const pRefH, |
168 |
}; |
const uint8_t * const pRefV, |
169 |
|
const uint8_t * const pRefHV, |
170 |
|
const uint8_t * const cur, |
171 |
|
const int x, const int y, |
172 |
|
int32_t startx, int32_t starty, |
173 |
|
int32_t iMinSAD, |
174 |
|
VECTOR * const currMV, |
175 |
|
const VECTOR * const pmv, |
176 |
|
const int32_t min_dx, const int32_t max_dx, |
177 |
|
const int32_t min_dy, const int32_t max_dy, |
178 |
|
const int32_t iEdgedWidth, |
179 |
|
const int32_t iDiamondSize, |
180 |
|
const int32_t iFcode, |
181 |
|
const int32_t iQuant, |
182 |
|
int iFound); |
183 |
|
|
184 |
|
typedef MainSearch8Func* MainSearch8FuncPtr; |
185 |
|
|
186 |
// mv.length table |
// mv.length table |
187 |
static const uint32_t mvtab[33] = { |
static const uint32_t mvtab[33] = { |
232 |
|
|
233 |
|
|
234 |
|
|
|
/* 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) |
|
|
{ |
|
|
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; |
|
|
} |
|
|
} |
|
|
|
|
|
|
|
|
/* 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) |
|
|
{ |
|
|
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; |
|
|
} |
|
|
} |
|
235 |
|
|
236 |
#ifndef SEARCH16 |
#ifndef SEARCH16 |
237 |
#define SEARCH16 PMVfastSearch16 |
#define SEARCH16 PMVfastSearch16 |
238 |
|
//#define SEARCH16 FullSearch16 |
239 |
|
//#define SEARCH16 EPZSSearch16 |
240 |
#endif |
#endif |
241 |
|
|
242 |
#ifndef SEARCH8 |
#ifndef SEARCH8 |
243 |
#define SEARCH8 PMVfastSearch8 |
#define SEARCH8 PMVfastSearch8 |
244 |
|
//#define SEARCH8 EPZSSearch8 |
245 |
#endif |
#endif |
246 |
|
|
247 |
bool MotionEstimation( |
bool MotionEstimation( |
|
MACROBLOCK * const pMBs, |
|
248 |
MBParam * const pParam, |
MBParam * const pParam, |
249 |
const IMAGE * const pRef, |
FRAMEINFO * const current, |
250 |
|
FRAMEINFO * const reference, |
251 |
const IMAGE * const pRefH, |
const IMAGE * const pRefH, |
252 |
const IMAGE * const pRefV, |
const IMAGE * const pRefV, |
253 |
const IMAGE * const pRefHV, |
const IMAGE * const pRefHV, |
|
IMAGE * const pCurrent, |
|
254 |
const uint32_t iLimit) |
const uint32_t iLimit) |
255 |
|
|
256 |
{ |
{ |
257 |
const uint32_t iWcount = pParam->mb_width; |
const uint32_t iWcount = pParam->mb_width; |
258 |
const uint32_t iHcount = pParam->mb_height; |
const uint32_t iHcount = pParam->mb_height; |
259 |
|
MACROBLOCK * pMBs = current->mbs; |
260 |
|
IMAGE * pCurrent = ¤t->image; |
261 |
|
|
262 |
|
MACROBLOCK * prevMBs = reference->mbs; // previous frame |
263 |
|
IMAGE * pRef = &reference->image; |
264 |
|
|
265 |
|
|
266 |
uint32_t i, j, iIntra = 0; |
uint32_t i, j, iIntra = 0; |
267 |
|
|
272 |
int32_t sad16; |
int32_t sad16; |
273 |
int32_t deviation; |
int32_t deviation; |
274 |
|
|
275 |
|
if (sadInit) |
276 |
|
(*sadInit)(); |
277 |
|
|
278 |
|
|
279 |
|
/* eventhough we have a seperate prevMBs, |
280 |
|
pmvfast/epsz does something "funny" with the previous frames data */ |
281 |
|
|
282 |
|
for (i = 0; i < iHcount; i++) |
283 |
|
for (j = 0; j < iWcount; j++) |
284 |
|
{ |
285 |
|
pMBs[j + i * iWcount].mvs[0] = prevMBs[j + i * iWcount].mvs[0]; |
286 |
|
pMBs[j + i * iWcount].mvs[1] = prevMBs[j + i * iWcount].mvs[1]; |
287 |
|
pMBs[j + i * iWcount].mvs[2] = prevMBs[j + i * iWcount].mvs[2]; |
288 |
|
pMBs[j + i * iWcount].mvs[3] = prevMBs[j + i * iWcount].mvs[3]; |
289 |
|
} |
290 |
|
|
291 |
|
/*dprintf("*** BEFORE ***"); |
292 |
|
for (i = 0; i < iHcount; i++) |
293 |
|
for (j = 0; j < iWcount; j++) |
294 |
|
{ |
295 |
|
dprintf(" [%i,%i] mode=%i dquant=%i mvs=(%i %i %i %i) sad8=(%i %i %i %i) sad16=(%i)", j,i, |
296 |
|
pMBs[j + i * iWcount].mode, |
297 |
|
pMBs[j + i * iWcount].dquant, |
298 |
|
pMBs[j + i * iWcount].mvs[0], |
299 |
|
pMBs[j + i * iWcount].mvs[1], |
300 |
|
pMBs[j + i * iWcount].mvs[2], |
301 |
|
pMBs[j + i * iWcount].mvs[3], |
302 |
|
prevMBs[j + i * iWcount].sad8[0], |
303 |
|
prevMBs[j + i * iWcount].sad8[1], |
304 |
|
prevMBs[j + i * iWcount].sad8[2], |
305 |
|
prevMBs[j + i * iWcount].sad8[3], |
306 |
|
prevMBs[j + i * iWcount].sad16); |
307 |
|
} |
308 |
|
*/ |
309 |
|
|
310 |
// note: i==horizontal, j==vertical |
// note: i==horizontal, j==vertical |
311 |
for (i = 0; i < iHcount; i++) |
for (i = 0; i < iHcount; i++) |
312 |
for (j = 0; j < iWcount; j++) |
for (j = 0; j < iWcount; j++) |
313 |
{ |
{ |
314 |
MACROBLOCK *pMB = &pMBs[j + i * iWcount]; |
MACROBLOCK *pMB = &pMBs[j + i * iWcount]; |
315 |
|
MACROBLOCK *prevMB = &prevMBs[j + i * iWcount]; |
316 |
|
|
317 |
sad16 = SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
sad16 = SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
318 |
j, i, pParam->motion_flags, |
j, i, current->motion_flags, current->quant, current->fcode, |
319 |
pParam, pMBs, &mv16, &pmv16); |
pParam, pMBs, prevMBs, &mv16, &pmv16); |
320 |
pMB->sad16=sad16; |
pMB->sad16=sad16; |
321 |
|
|
322 |
|
|
332 |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
333 |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; |
334 |
|
|
335 |
|
pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = 0; |
336 |
|
|
337 |
iIntra++; |
iIntra++; |
338 |
if(iIntra >= iLimit) |
if(iIntra >= iLimit) |
339 |
return 1; |
return 1; |
341 |
continue; |
continue; |
342 |
} |
} |
343 |
|
|
344 |
if (pParam->global_flags & XVID_INTER4V) |
if (current->global_flags & XVID_INTER4V) |
345 |
{ |
{ |
346 |
pMB->sad8[0] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
pMB->sad8[0] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
347 |
2 * j, 2 * i, mv16.x, mv16.y, pParam->motion_flags, |
2 * j, 2 * i, mv16.x, mv16.y, |
348 |
pParam, pMBs, &pMB->mvs[0], &pMB->pmvs[0]); |
current->motion_flags, current->quant, current->fcode, |
349 |
|
pParam, pMBs, prevMBs, &pMB->mvs[0], &pMB->pmvs[0]); |
350 |
|
|
351 |
pMB->sad8[1] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
pMB->sad8[1] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
352 |
2 * j + 1, 2 * i, mv16.x, mv16.y, pParam->motion_flags, |
2 * j + 1, 2 * i, mv16.x, mv16.y, |
353 |
pParam, pMBs, &pMB->mvs[1], &pMB->pmvs[1]); |
current->motion_flags, current->quant, current->fcode, |
354 |
|
pParam, pMBs, prevMBs, &pMB->mvs[1], &pMB->pmvs[1]); |
355 |
|
|
356 |
pMB->sad8[2] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
pMB->sad8[2] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
357 |
2 * j, 2 * i + 1, mv16.x, mv16.y, pParam->motion_flags, |
2 * j, 2 * i + 1, mv16.x, mv16.y, |
358 |
pParam, pMBs, &pMB->mvs[2], &pMB->pmvs[2]); |
current->motion_flags, current->quant, current->fcode, |
359 |
|
pParam, pMBs, prevMBs, &pMB->mvs[2], &pMB->pmvs[2]); |
360 |
|
|
361 |
pMB->sad8[3] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
pMB->sad8[3] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
362 |
2 * j + 1, 2 * i + 1, mv16.x, mv16.y, pParam->motion_flags, |
2 * j + 1, 2 * i + 1, mv16.x, mv16.y, |
363 |
pParam, pMBs, &pMB->mvs[3], &pMB->pmvs[3]); |
current->motion_flags, current->quant, current->fcode, |
364 |
|
pParam, pMBs, prevMBs, &pMB->mvs[3], &pMB->pmvs[3]); |
365 |
|
|
366 |
sad8 = pMB->sad8[0] + pMB->sad8[1] + pMB->sad8[2] + pMB->sad8[3]; |
sad8 = pMB->sad8[0] + pMB->sad8[1] + pMB->sad8[2] + pMB->sad8[3]; |
367 |
} |
} |
371 |
mpeg4: if (sad8 < sad16 - nb/2+1) use_inter4v |
mpeg4: if (sad8 < sad16 - nb/2+1) use_inter4v |
372 |
*/ |
*/ |
373 |
|
|
374 |
if (pMB->dquant == NO_CHANGE) { |
if (!(current->global_flags & XVID_LUMIMASKING) || pMB->dquant == NO_CHANGE) |
375 |
if (((pParam->global_flags & XVID_INTER4V)==0) || |
{ |
376 |
(sad16 < (sad8 + (int32_t)(IMV16X16 * pParam->quant)))) { |
if (((current->global_flags & XVID_INTER4V)==0) || |
377 |
|
(sad16 < (sad8 + (int32_t)(IMV16X16 * current->quant)))) |
378 |
|
{ |
379 |
|
|
380 |
sad8 = sad16; |
sad8 = sad16; |
381 |
pMB->mode = MODE_INTER; |
pMB->mode = MODE_INTER; |
398 |
} |
} |
399 |
} |
} |
400 |
|
|
401 |
|
/* dprintf("*** AFTER ***", pMBs[0].b_mvs[0].x); |
402 |
|
for (i = 0; i < iHcount; i++) |
403 |
|
for (j = 0; j < iWcount; j++) |
404 |
|
{ |
405 |
|
dprintf(" [%i,%i] mode=%i dquant=%i mvs=(%i %i %i %i) sad8=(%i %i %i %i) sad16=(%i)", j,i, |
406 |
|
pMBs[j + i * iWcount].mode, |
407 |
|
pMBs[j + i * iWcount].dquant, |
408 |
|
pMBs[j + i * iWcount].mvs[0], |
409 |
|
pMBs[j + i * iWcount].mvs[1], |
410 |
|
pMBs[j + i * iWcount].mvs[2], |
411 |
|
pMBs[j + i * iWcount].mvs[3], |
412 |
|
pMBs[j + i * iWcount].sad8[0], |
413 |
|
pMBs[j + i * iWcount].sad8[1], |
414 |
|
pMBs[j + i * iWcount].sad8[2], |
415 |
|
pMBs[j + i * iWcount].sad8[3], |
416 |
|
pMBs[j + i * iWcount].sad16); |
417 |
|
} |
418 |
|
*/ |
419 |
|
|
420 |
return 0; |
return 0; |
421 |
} |
} |
422 |
|
|
437 |
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } } \ |
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } } \ |
438 |
} |
} |
439 |
|
|
440 |
|
#define NOCHECK_MV16_CANDIDATE(X,Y) { \ |
441 |
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
442 |
|
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
443 |
|
if (iSAD < iMinSAD) \ |
444 |
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } \ |
445 |
|
} |
446 |
|
|
447 |
#define CHECK_MV16_CANDIDATE(X,Y) { \ |
#define CHECK_MV16_CANDIDATE(X,Y) { \ |
448 |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
482 |
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } \ |
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } \ |
483 |
} |
} |
484 |
|
|
485 |
|
#define NOCHECK_MV8_CANDIDATE(X,Y) \ |
486 |
|
{ \ |
487 |
|
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
488 |
|
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
489 |
|
if (iSAD < iMinSAD) \ |
490 |
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } \ |
491 |
|
} |
492 |
|
|
493 |
#define CHECK_MV8_CANDIDATE(X,Y) { \ |
#define CHECK_MV8_CANDIDATE(X,Y) { \ |
494 |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
530 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
531 |
const int x, const int y, |
const int x, const int y, |
532 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
533 |
|
const uint32_t iQuant, |
534 |
|
const uint32_t iFcode, |
535 |
MBParam * const pParam, |
MBParam * const pParam, |
536 |
MACROBLOCK * const pMBs, |
const MACROBLOCK * const pMBs, |
537 |
|
const MACROBLOCK * const prevMBs, |
538 |
VECTOR * const currMV, |
VECTOR * const currMV, |
539 |
VECTOR * const currPMV) |
VECTOR * const currPMV) |
540 |
{ |
{ |
541 |
const int32_t iEdgedWidth = pParam->edged_width; |
const int32_t iEdgedWidth = pParam->edged_width; |
|
const int32_t iQuant = pParam->quant; |
|
542 |
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
543 |
int32_t iSAD; |
int32_t iSAD; |
544 |
int32_t pred_x,pred_y; |
int32_t pred_x,pred_y; |
561 |
} |
} |
562 |
*/ |
*/ |
563 |
|
|
564 |
int32_t PMVfastSearch16_MainSearch( |
int32_t Diamond16_MainSearch( |
565 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
566 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
567 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
618 |
return iMinSAD; |
return iMinSAD; |
619 |
} |
} |
620 |
|
|
621 |
int32_t PMVfastSearch16_Refine( |
int32_t Square16_MainSearch( |
622 |
|
const uint8_t * const pRef, |
623 |
|
const uint8_t * const pRefH, |
624 |
|
const uint8_t * const pRefV, |
625 |
|
const uint8_t * const pRefHV, |
626 |
|
const uint8_t * const cur, |
627 |
|
const int x, const int y, |
628 |
|
int32_t startx, int32_t starty, |
629 |
|
int32_t iMinSAD, |
630 |
|
VECTOR * const currMV, |
631 |
|
const VECTOR * const pmv, |
632 |
|
const int32_t min_dx, const int32_t max_dx, |
633 |
|
const int32_t min_dy, const int32_t max_dy, |
634 |
|
const int32_t iEdgedWidth, |
635 |
|
const int32_t iDiamondSize, |
636 |
|
const int32_t iFcode, |
637 |
|
const int32_t iQuant, |
638 |
|
int iFound) |
639 |
|
{ |
640 |
|
/* Do a square search around given starting point, return SAD of best */ |
641 |
|
|
642 |
|
int32_t iDirection=0; |
643 |
|
int32_t iSAD; |
644 |
|
VECTOR backupMV; |
645 |
|
backupMV.x = startx; |
646 |
|
backupMV.y = starty; |
647 |
|
|
648 |
|
/* It's one search with full square pattern, and new parts for all following diamonds */ |
649 |
|
|
650 |
|
/* new direction are extra, so 1-4 is normal diamond |
651 |
|
537 |
652 |
|
1*2 |
653 |
|
648 |
654 |
|
*/ |
655 |
|
|
656 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
657 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
658 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
659 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
660 |
|
|
661 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
662 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
663 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
664 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
665 |
|
|
666 |
|
|
667 |
|
if (iDirection) |
668 |
|
while (!iFound) |
669 |
|
{ |
670 |
|
iFound = 1; |
671 |
|
backupMV=*currMV; |
672 |
|
|
673 |
|
switch (iDirection) |
674 |
|
{ |
675 |
|
case 1: |
676 |
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
677 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
678 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
679 |
|
break; |
680 |
|
case 2: |
681 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
682 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
683 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
684 |
|
break; |
685 |
|
|
686 |
|
case 3: |
687 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
688 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
689 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
690 |
|
break; |
691 |
|
|
692 |
|
case 4: |
693 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
694 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
695 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
696 |
|
break; |
697 |
|
|
698 |
|
case 5: |
699 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
700 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
701 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
702 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
703 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
704 |
|
break; |
705 |
|
|
706 |
|
case 6: |
707 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
708 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
709 |
|
|
710 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
711 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
712 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
713 |
|
|
714 |
|
break; |
715 |
|
|
716 |
|
case 7: |
717 |
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
718 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
719 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
720 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
721 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
722 |
|
break; |
723 |
|
|
724 |
|
case 8: |
725 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
726 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
727 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
728 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
729 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
730 |
|
break; |
731 |
|
default: |
732 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
733 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
734 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
735 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
736 |
|
|
737 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
738 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
739 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
740 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
741 |
|
break; |
742 |
|
} |
743 |
|
} |
744 |
|
else |
745 |
|
{ |
746 |
|
currMV->x = startx; |
747 |
|
currMV->y = starty; |
748 |
|
} |
749 |
|
return iMinSAD; |
750 |
|
} |
751 |
|
|
752 |
|
|
753 |
|
int32_t Full16_MainSearch( |
754 |
|
const uint8_t * const pRef, |
755 |
|
const uint8_t * const pRefH, |
756 |
|
const uint8_t * const pRefV, |
757 |
|
const uint8_t * const pRefHV, |
758 |
|
const uint8_t * const cur, |
759 |
|
const int x, const int y, |
760 |
|
int32_t startx, int32_t starty, |
761 |
|
int32_t iMinSAD, |
762 |
|
VECTOR * const currMV, |
763 |
|
const VECTOR * const pmv, |
764 |
|
const int32_t min_dx, const int32_t max_dx, |
765 |
|
const int32_t min_dy, const int32_t max_dy, |
766 |
|
const int32_t iEdgedWidth, |
767 |
|
const int32_t iDiamondSize, |
768 |
|
const int32_t iFcode, |
769 |
|
const int32_t iQuant, |
770 |
|
int iFound) |
771 |
|
{ |
772 |
|
int32_t iSAD; |
773 |
|
int32_t dx,dy; |
774 |
|
VECTOR backupMV; |
775 |
|
backupMV.x = startx; |
776 |
|
backupMV.y = starty; |
777 |
|
|
778 |
|
for (dx = min_dx; dx<=max_dx; dx+=iDiamondSize) |
779 |
|
for (dy = min_dy; dy<= max_dy; dy+=iDiamondSize) |
780 |
|
NOCHECK_MV16_CANDIDATE(dx,dy); |
781 |
|
|
782 |
|
return iMinSAD; |
783 |
|
} |
784 |
|
|
785 |
|
int32_t Full8_MainSearch( |
786 |
|
const uint8_t * const pRef, |
787 |
|
const uint8_t * const pRefH, |
788 |
|
const uint8_t * const pRefV, |
789 |
|
const uint8_t * const pRefHV, |
790 |
|
const uint8_t * const cur, |
791 |
|
const int x, const int y, |
792 |
|
int32_t startx, int32_t starty, |
793 |
|
int32_t iMinSAD, |
794 |
|
VECTOR * const currMV, |
795 |
|
const VECTOR * const pmv, |
796 |
|
const int32_t min_dx, const int32_t max_dx, |
797 |
|
const int32_t min_dy, const int32_t max_dy, |
798 |
|
const int32_t iEdgedWidth, |
799 |
|
const int32_t iDiamondSize, |
800 |
|
const int32_t iFcode, |
801 |
|
const int32_t iQuant, |
802 |
|
int iFound) |
803 |
|
{ |
804 |
|
int32_t iSAD; |
805 |
|
int32_t dx,dy; |
806 |
|
VECTOR backupMV; |
807 |
|
backupMV.x = startx; |
808 |
|
backupMV.y = starty; |
809 |
|
|
810 |
|
for (dx = min_dx; dx<=max_dx; dx+=iDiamondSize) |
811 |
|
for (dy = min_dy; dy<= max_dy; dy+=iDiamondSize) |
812 |
|
NOCHECK_MV8_CANDIDATE(dx,dy); |
813 |
|
|
814 |
|
return iMinSAD; |
815 |
|
} |
816 |
|
|
817 |
|
|
818 |
|
|
819 |
|
int32_t Halfpel16_Refine( |
820 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
821 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
822 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
851 |
|
|
852 |
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
853 |
|
|
854 |
|
|
855 |
int32_t PMVfastSearch16( |
int32_t PMVfastSearch16( |
856 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
857 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
860 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
861 |
const int x, const int y, |
const int x, const int y, |
862 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
863 |
MBParam * const pParam, |
const uint32_t iQuant, |
864 |
MACROBLOCK * const pMBs, |
const uint32_t iFcode, |
865 |
|
const MBParam * const pParam, |
866 |
|
const MACROBLOCK * const pMBs, |
867 |
|
const MACROBLOCK * const prevMBs, |
868 |
VECTOR * const currMV, |
VECTOR * const currMV, |
869 |
VECTOR * const currPMV) |
VECTOR * const currPMV) |
870 |
{ |
{ |
871 |
const uint32_t iWcount = pParam->mb_width; |
const uint32_t iWcount = pParam->mb_width; |
|
const int32_t iFcode = pParam->fixed_code; |
|
|
const int32_t iQuant = pParam->quant; |
|
872 |
const int32_t iWidth = pParam->width; |
const int32_t iWidth = pParam->width; |
873 |
const int32_t iHeight = pParam->height; |
const int32_t iHeight = pParam->height; |
874 |
const int32_t iEdgedWidth = pParam->edged_width; |
const int32_t iEdgedWidth = pParam->edged_width; |
890 |
VECTOR pmv[4]; |
VECTOR pmv[4]; |
891 |
int32_t psad[4]; |
int32_t psad[4]; |
892 |
|
|
893 |
MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
894 |
|
const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; |
895 |
|
|
896 |
static int32_t threshA,threshB; |
static int32_t threshA,threshB; |
897 |
int32_t bPredEq; |
int32_t bPredEq; |
935 |
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
936 |
*/ |
*/ |
937 |
|
|
938 |
if ((bPredEq) && (MVequal(pmv[0],pMB->mvs[0]) ) ) |
if ((bPredEq) && (MVequal(pmv[0],prevMB->mvs[0]) ) ) |
939 |
iFound=2; |
iFound=2; |
940 |
|
|
941 |
/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
989 |
iEdgedWidth, MV_MAX_ERROR); |
iEdgedWidth, MV_MAX_ERROR); |
990 |
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; |
991 |
|
|
992 |
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,pMB->mvs[0])) && (iMinSAD < pMB->sad16) ) ) |
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,prevMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
993 |
{ |
{ |
994 |
|
|
995 |
if (MotionFlags & PMV_QUICKSTOP16) |
if (MotionFlags & PMV_QUICKSTOP16) |
996 |
goto step10b; |
goto PMVfast16_Terminate_without_Refine; |
997 |
if (MotionFlags & PMV_EARLYSTOP16) |
if (MotionFlags & PMV_EARLYSTOP16) |
998 |
goto step10; |
goto PMVfast16_Terminate_with_Refine; |
999 |
} |
} |
1000 |
|
|
1001 |
/* |
/* |
1010 |
CHECK_MV16_ZERO; |
CHECK_MV16_ZERO; |
1011 |
|
|
1012 |
// previous frame MV is always possible |
// previous frame MV is always possible |
1013 |
CHECK_MV16_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); |
CHECK_MV16_CANDIDATE(prevMB->mvs[0].x,prevMB->mvs[0].y); |
1014 |
|
|
1015 |
// left neighbour, if allowed |
// left neighbour, if allowed |
1016 |
if (x != 0) |
if (x != 0) |
1032 |
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
1033 |
|
|
1034 |
// top right neighbour, if allowed |
// top right neighbour, if allowed |
1035 |
if (x != (iWcount-1)) |
if ((uint32_t)x != (iWcount-1)) |
1036 |
{ |
{ |
1037 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
if (!(MotionFlags & PMV_HALFPEL16 )) |
1038 |
{ pmv[3].x = EVEN(pmv[3].x); |
{ pmv[3].x = EVEN(pmv[3].x); |
1046 |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
1047 |
*/ |
*/ |
1048 |
|
|
1049 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,pMB->mvs[0]) && (iMinSAD < pMB->sad16) ) ) |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,prevMB->mvs[0]) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
1050 |
{ |
{ |
1051 |
if (MotionFlags & PMV_QUICKSTOP16) |
if (MotionFlags & PMV_QUICKSTOP16) |
1052 |
goto step10b; |
goto PMVfast16_Terminate_without_Refine; |
1053 |
if (MotionFlags & PMV_EARLYSTOP16) |
if (MotionFlags & PMV_EARLYSTOP16) |
1054 |
goto step10; |
goto PMVfast16_Terminate_with_Refine; |
1055 |
} |
} |
1056 |
|
|
1057 |
|
|
1068 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
1069 |
|
|
1070 |
/* 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 */ |
1071 |
iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1072 |
x, y, |
x, y, |
1073 |
currMV->x, currMV->y, iMinSAD, &newMV, |
currMV->x, currMV->y, iMinSAD, &newMV, |
1074 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1084 |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
1085 |
|
|
1086 |
if (!(MVequal(pmv[0],backupMV)) ) |
if (!(MVequal(pmv[0],backupMV)) ) |
1087 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1088 |
x, y, |
x, y, |
1089 |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
1090 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1097 |
} |
} |
1098 |
|
|
1099 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
1100 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1101 |
x, y, |
x, y, |
1102 |
0, 0, iMinSAD, &newMV, |
0, 0, iMinSAD, &newMV, |
1103 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1114 |
Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
1115 |
*/ |
*/ |
1116 |
|
|
1117 |
step10: |
PMVfast16_Terminate_with_Refine: |
1118 |
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
1119 |
iMinSAD = PMVfastSearch16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
iMinSAD = Halfpel16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
1120 |
x, y, |
x, y, |
1121 |
currMV, iMinSAD, |
currMV, iMinSAD, |
1122 |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
1123 |
|
|
1124 |
step10b: |
PMVfast16_Terminate_without_Refine: |
1125 |
currPMV->x = currMV->x - pmv[0].x; |
currPMV->x = currMV->x - pmv[0].x; |
1126 |
currPMV->y = currMV->y - pmv[0].y; |
currPMV->y = currMV->y - pmv[0].y; |
1127 |
return iMinSAD; |
return iMinSAD; |
1132 |
|
|
1133 |
|
|
1134 |
|
|
1135 |
int32_t PMVfastSearch8_MainSearch( |
int32_t Diamond8_MainSearch( |
1136 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
1137 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1138 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
1189 |
return iMinSAD; |
return iMinSAD; |
1190 |
} |
} |
1191 |
|
|
1192 |
int32_t PMVfastSearch8_Refine( |
int32_t Halfpel8_Refine( |
1193 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
1194 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1195 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
1232 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
1233 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
1234 |
const int x, const int y, |
const int x, const int y, |
1235 |
const int start_x, int start_y, |
const int start_x, const int start_y, |
1236 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
1237 |
MBParam * const pParam, |
const uint32_t iQuant, |
1238 |
MACROBLOCK * const pMBs, |
const uint32_t iFcode, |
1239 |
|
const MBParam * const pParam, |
1240 |
|
const MACROBLOCK * const pMBs, |
1241 |
|
const MACROBLOCK * const prevMBs, |
1242 |
VECTOR * const currMV, |
VECTOR * const currMV, |
1243 |
VECTOR * const currPMV) |
VECTOR * const currPMV) |
1244 |
{ |
{ |
1245 |
const uint32_t iWcount = pParam->mb_width; |
const uint32_t iWcount = pParam->mb_width; |
|
|
|
|
const int32_t iFcode = pParam->fixed_code; |
|
|
const int32_t iQuant = pParam->quant; |
|
1246 |
const int32_t iWidth = pParam->width; |
const int32_t iWidth = pParam->width; |
1247 |
const int32_t iHeight = pParam->height; |
const int32_t iHeight = pParam->height; |
1248 |
const int32_t iEdgedWidth = pParam->edged_width; |
const int32_t iEdgedWidth = pParam->edged_width; |
1261 |
VECTOR newMV; |
VECTOR newMV; |
1262 |
VECTOR backupMV; |
VECTOR backupMV; |
1263 |
|
|
1264 |
MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
1265 |
|
const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; |
1266 |
|
|
1267 |
static int32_t threshA,threshB; |
static int32_t threshA,threshB; |
1268 |
int32_t iFound,bPredEq; |
int32_t iFound,bPredEq; |
1341 |
iEdgedWidth); |
iEdgedWidth); |
1342 |
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; |
1343 |
|
|
1344 |
if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,pMB->mvs[iSubBlock])) && (iMinSAD < pMB->sad8[iSubBlock]) ) ) |
if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,pMB->mvs[iSubBlock])) && ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
1345 |
{ |
{ |
1346 |
if (MotionFlags & PMV_QUICKSTOP8) |
if (MotionFlags & PMV_QUICKSTOP16) |
1347 |
goto step10_8b; |
goto PMVfast8_Terminate_without_Refine; |
1348 |
if (MotionFlags & PMV_EARLYSTOP8) |
if (MotionFlags & PMV_EARLYSTOP16) |
1349 |
goto step10_8; |
goto PMVfast8_Terminate_with_Refine; |
1350 |
} |
} |
1351 |
|
|
1352 |
|
|
1353 |
/* |
/* |
1354 |
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
1355 |
Also calculate (0,0) but do not subtract offset. |
Also calculate (0,0) but do not subtract offset. |
1400 |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
1401 |
*/ |
*/ |
1402 |
|
|
1403 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,pMB->mvs[iSubBlock]) && (iMinSAD < pMB->sad8[iSubBlock]) ) ) |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,pMB->mvs[iSubBlock]) && ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
1404 |
{ |
{ |
1405 |
if (MotionFlags & PMV_QUICKSTOP8) |
if (MotionFlags & PMV_QUICKSTOP16) |
1406 |
goto step10_8b; |
goto PMVfast8_Terminate_without_Refine; |
1407 |
if (MotionFlags & PMV_EARLYSTOP8) |
if (MotionFlags & PMV_EARLYSTOP16) |
1408 |
goto step10_8; |
goto PMVfast8_Terminate_with_Refine; |
1409 |
} |
} |
1410 |
|
|
1411 |
/************ (Diamond Search) **************/ |
/************ (Diamond Search) **************/ |
1421 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
1422 |
|
|
1423 |
/* 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 */ |
1424 |
iSAD = PMVfastSearch8_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
iSAD = Diamond8_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1425 |
x, y, |
x, y, |
1426 |
currMV->x, currMV->y, iMinSAD, &newMV, |
currMV->x, currMV->y, iMinSAD, &newMV, |
1427 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1437 |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
1438 |
|
|
1439 |
if (!(MVequal(pmv[0],backupMV)) ) |
if (!(MVequal(pmv[0],backupMV)) ) |
1440 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1441 |
x, y, |
x, y, |
1442 |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
1443 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1450 |
} |
} |
1451 |
|
|
1452 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
1453 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1454 |
x, y, |
x, y, |
1455 |
0, 0, iMinSAD, &newMV, |
0, 0, iMinSAD, &newMV, |
1456 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1467 |
By performing an optional local half-pixel search, we can refine this result even further. |
By performing an optional local half-pixel search, we can refine this result even further. |
1468 |
*/ |
*/ |
1469 |
|
|
1470 |
step10_8: |
PMVfast8_Terminate_with_Refine: |
1471 |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
1472 |
iMinSAD = PMVfastSearch8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
iMinSAD = Halfpel8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
1473 |
|
x, y, |
1474 |
|
currMV, iMinSAD, |
1475 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
1476 |
|
|
1477 |
|
|
1478 |
|
PMVfast8_Terminate_without_Refine: |
1479 |
|
currPMV->x = currMV->x - pmv[0].x; |
1480 |
|
currPMV->y = currMV->y - pmv[0].y; |
1481 |
|
|
1482 |
|
return iMinSAD; |
1483 |
|
} |
1484 |
|
|
1485 |
|
int32_t EPZSSearch16( |
1486 |
|
const uint8_t * const pRef, |
1487 |
|
const uint8_t * const pRefH, |
1488 |
|
const uint8_t * const pRefV, |
1489 |
|
const uint8_t * const pRefHV, |
1490 |
|
const IMAGE * const pCur, |
1491 |
|
const int x, const int y, |
1492 |
|
const uint32_t MotionFlags, |
1493 |
|
const uint32_t iQuant, |
1494 |
|
const uint32_t iFcode, |
1495 |
|
const MBParam * const pParam, |
1496 |
|
const MACROBLOCK * const pMBs, |
1497 |
|
const MACROBLOCK * const prevMBs, |
1498 |
|
VECTOR * const currMV, |
1499 |
|
VECTOR * const currPMV) |
1500 |
|
{ |
1501 |
|
const uint32_t iWcount = pParam->mb_width; |
1502 |
|
const uint32_t iHcount = pParam->mb_height; |
1503 |
|
|
1504 |
|
const int32_t iWidth = pParam->width; |
1505 |
|
const int32_t iHeight = pParam->height; |
1506 |
|
const int32_t iEdgedWidth = pParam->edged_width; |
1507 |
|
|
1508 |
|
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
1509 |
|
|
1510 |
|
int32_t min_dx; |
1511 |
|
int32_t max_dx; |
1512 |
|
int32_t min_dy; |
1513 |
|
int32_t max_dy; |
1514 |
|
|
1515 |
|
VECTOR newMV; |
1516 |
|
VECTOR backupMV; |
1517 |
|
|
1518 |
|
VECTOR pmv[4]; |
1519 |
|
int32_t psad[8]; |
1520 |
|
|
1521 |
|
static MACROBLOCK * oldMBs = NULL; |
1522 |
|
const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
1523 |
|
const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; |
1524 |
|
MACROBLOCK * oldMB = NULL; |
1525 |
|
|
1526 |
|
static int32_t thresh2; |
1527 |
|
int32_t bPredEq; |
1528 |
|
int32_t iMinSAD,iSAD=9999; |
1529 |
|
|
1530 |
|
MainSearch16FuncPtr EPZSMainSearchPtr; |
1531 |
|
|
1532 |
|
if (oldMBs == NULL) |
1533 |
|
{ oldMBs = (MACROBLOCK*) calloc(1,iWcount*iHcount*sizeof(MACROBLOCK)); |
1534 |
|
fprintf(stderr,"allocated %d bytes for oldMBs\n",iWcount*iHcount*sizeof(MACROBLOCK)); |
1535 |
|
} |
1536 |
|
oldMB = oldMBs + x + y * iWcount; |
1537 |
|
|
1538 |
|
/* Get maximum range */ |
1539 |
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
1540 |
|
x, y, 16, iWidth, iHeight, iFcode); |
1541 |
|
|
1542 |
|
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
1543 |
|
|
1544 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1545 |
|
{ min_dx = EVEN(min_dx); |
1546 |
|
max_dx = EVEN(max_dx); |
1547 |
|
min_dy = EVEN(min_dy); |
1548 |
|
max_dy = EVEN(max_dy); |
1549 |
|
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
1550 |
|
|
1551 |
|
bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
1552 |
|
|
1553 |
|
/* Step 4: Calculate SAD around the Median prediction. |
1554 |
|
MinSAD=SAD |
1555 |
|
If Motion Vector equal to Previous frame motion vector |
1556 |
|
and MinSAD<PrevFrmSAD goto Step 10. |
1557 |
|
If SAD<=256 goto Step 10. |
1558 |
|
*/ |
1559 |
|
|
1560 |
|
// Prepare for main loop |
1561 |
|
|
1562 |
|
*currMV=pmv[0]; /* current best := median prediction */ |
1563 |
|
if (!(MotionFlags & PMV_HALFPEL16)) |
1564 |
|
{ |
1565 |
|
currMV->x = EVEN(currMV->x); |
1566 |
|
currMV->y = EVEN(currMV->y); |
1567 |
|
} |
1568 |
|
|
1569 |
|
if (currMV->x > max_dx) |
1570 |
|
currMV->x=max_dx; |
1571 |
|
if (currMV->x < min_dx) |
1572 |
|
currMV->x=min_dx; |
1573 |
|
if (currMV->y > max_dy) |
1574 |
|
currMV->y=max_dy; |
1575 |
|
if (currMV->y < min_dy) |
1576 |
|
currMV->y=min_dy; |
1577 |
|
|
1578 |
|
/***************** This is predictor SET A: only median prediction ******************/ |
1579 |
|
|
1580 |
|
iMinSAD = sad16( cur, |
1581 |
|
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
1582 |
|
iEdgedWidth, MV_MAX_ERROR); |
1583 |
|
iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; |
1584 |
|
|
1585 |
|
// thresh1 is fixed to 256 |
1586 |
|
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,pMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
1587 |
|
{ |
1588 |
|
if (MotionFlags & PMV_QUICKSTOP16) |
1589 |
|
goto EPZS16_Terminate_without_Refine; |
1590 |
|
if (MotionFlags & PMV_EARLYSTOP16) |
1591 |
|
goto EPZS16_Terminate_with_Refine; |
1592 |
|
} |
1593 |
|
|
1594 |
|
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
1595 |
|
|
1596 |
|
// previous frame MV |
1597 |
|
CHECK_MV16_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); |
1598 |
|
|
1599 |
|
// set threshhold based on Min of Prediction and SAD of collocated block |
1600 |
|
// CHECK_MV16 always uses iSAD for the SAD of last vector to check, so now iSAD is what we want |
1601 |
|
|
1602 |
|
if ((x==0) && (y==0) ) |
1603 |
|
{ |
1604 |
|
thresh2 = 512; |
1605 |
|
} |
1606 |
|
else |
1607 |
|
{ |
1608 |
|
/* T_k = 1.2 * MIN(SAD_top,SAD_left,SAD_topleft,SAD_coll) +128; [Tourapis, 2002] */ |
1609 |
|
|
1610 |
|
thresh2 = MIN(psad[0],iSAD)*6/5 + 128; |
1611 |
|
} |
1612 |
|
|
1613 |
|
// MV=(0,0) is often a good choice |
1614 |
|
|
1615 |
|
CHECK_MV16_ZERO; |
1616 |
|
|
1617 |
|
|
1618 |
|
// left neighbour, if allowed |
1619 |
|
if (x != 0) |
1620 |
|
{ |
1621 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1622 |
|
{ pmv[1].x = EVEN(pmv[1].x); |
1623 |
|
pmv[1].y = EVEN(pmv[1].y); |
1624 |
|
} |
1625 |
|
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
1626 |
|
} |
1627 |
|
|
1628 |
|
// top neighbour, if allowed |
1629 |
|
if (y != 0) |
1630 |
|
{ |
1631 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1632 |
|
{ pmv[2].x = EVEN(pmv[2].x); |
1633 |
|
pmv[2].y = EVEN(pmv[2].y); |
1634 |
|
} |
1635 |
|
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
1636 |
|
|
1637 |
|
// top right neighbour, if allowed |
1638 |
|
if ((uint32_t)x != (iWcount-1)) |
1639 |
|
{ |
1640 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1641 |
|
{ pmv[3].x = EVEN(pmv[3].x); |
1642 |
|
pmv[3].y = EVEN(pmv[3].y); |
1643 |
|
} |
1644 |
|
CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); |
1645 |
|
} |
1646 |
|
} |
1647 |
|
|
1648 |
|
/* Terminate if MinSAD <= T_2 |
1649 |
|
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
1650 |
|
*/ |
1651 |
|
|
1652 |
|
if ( (iMinSAD <= thresh2) |
1653 |
|
|| ( MVequal(*currMV,pMB->mvs[0]) && ((uint32_t)iMinSAD <= prevMB->sad16) ) ) |
1654 |
|
{ |
1655 |
|
if (MotionFlags & PMV_QUICKSTOP16) |
1656 |
|
goto EPZS16_Terminate_without_Refine; |
1657 |
|
if (MotionFlags & PMV_EARLYSTOP16) |
1658 |
|
goto EPZS16_Terminate_with_Refine; |
1659 |
|
} |
1660 |
|
|
1661 |
|
/***** predictor SET C: acceleration MV (new!), neighbours in prev. frame(new!) ****/ |
1662 |
|
|
1663 |
|
backupMV = pMB->mvs[0]; // last MV |
1664 |
|
backupMV.x += (pMB->mvs[0].x - oldMB->mvs[0].x ); // acceleration X |
1665 |
|
backupMV.y += (pMB->mvs[0].y - oldMB->mvs[0].y ); // acceleration Y |
1666 |
|
|
1667 |
|
CHECK_MV16_CANDIDATE(backupMV.x,backupMV.y); |
1668 |
|
|
1669 |
|
// left neighbour |
1670 |
|
if (x != 0) |
1671 |
|
CHECK_MV16_CANDIDATE((oldMB-1)->mvs[0].x,oldMB->mvs[0].y); |
1672 |
|
|
1673 |
|
// top neighbour |
1674 |
|
if (y != 0) |
1675 |
|
CHECK_MV16_CANDIDATE((oldMB-iWcount)->mvs[0].x,oldMB->mvs[0].y); |
1676 |
|
|
1677 |
|
// right neighbour, if allowed (this value is not written yet, so take it from pMB->mvs |
1678 |
|
|
1679 |
|
if ((uint32_t)x != iWcount-1) |
1680 |
|
CHECK_MV16_CANDIDATE((pMB+1)->mvs[0].x,oldMB->mvs[0].y); |
1681 |
|
|
1682 |
|
// bottom neighbour, dito |
1683 |
|
if ((uint32_t)y != iHcount-1) |
1684 |
|
CHECK_MV16_CANDIDATE((pMB+iWcount)->mvs[0].x,oldMB->mvs[0].y); |
1685 |
|
|
1686 |
|
/* Terminate if MinSAD <= T_3 (here T_3 = T_2) */ |
1687 |
|
if (iMinSAD <= thresh2) |
1688 |
|
{ |
1689 |
|
if (MotionFlags & PMV_QUICKSTOP16) |
1690 |
|
goto EPZS16_Terminate_without_Refine; |
1691 |
|
if (MotionFlags & PMV_EARLYSTOP16) |
1692 |
|
goto EPZS16_Terminate_with_Refine; |
1693 |
|
} |
1694 |
|
|
1695 |
|
/************ (if Diamond Search) **************/ |
1696 |
|
|
1697 |
|
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
1698 |
|
|
1699 |
|
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
1700 |
|
|
1701 |
|
if (MotionFlags & PMV_USESQUARES16) |
1702 |
|
EPZSMainSearchPtr = Square16_MainSearch; |
1703 |
|
else |
1704 |
|
EPZSMainSearchPtr = Diamond16_MainSearch; |
1705 |
|
|
1706 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1707 |
|
x, y, |
1708 |
|
currMV->x, currMV->y, iMinSAD, &newMV, pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
1709 |
|
2, iFcode, iQuant, 0); |
1710 |
|
|
1711 |
|
if (iSAD < iMinSAD) |
1712 |
|
{ |
1713 |
|
*currMV = newMV; |
1714 |
|
iMinSAD = iSAD; |
1715 |
|
} |
1716 |
|
|
1717 |
|
|
1718 |
|
if (MotionFlags & PMV_EXTSEARCH16) |
1719 |
|
{ |
1720 |
|
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
1721 |
|
|
1722 |
|
if (!(MVequal(pmv[0],backupMV)) ) |
1723 |
|
{ |
1724 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1725 |
|
x, y, |
1726 |
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
1727 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); |
1728 |
|
} |
1729 |
|
|
1730 |
|
if (iSAD < iMinSAD) |
1731 |
|
{ |
1732 |
|
*currMV = newMV; |
1733 |
|
iMinSAD = iSAD; |
1734 |
|
} |
1735 |
|
|
1736 |
|
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
1737 |
|
{ |
1738 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1739 |
|
x, y, |
1740 |
|
0, 0, iMinSAD, &newMV, |
1741 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, /*iDiamondSize*/ 2, iFcode, iQuant, 0); |
1742 |
|
|
1743 |
|
if (iSAD < iMinSAD) |
1744 |
|
{ |
1745 |
|
*currMV = newMV; |
1746 |
|
iMinSAD = iSAD; |
1747 |
|
} |
1748 |
|
} |
1749 |
|
} |
1750 |
|
|
1751 |
|
/*************** Choose best MV found **************/ |
1752 |
|
|
1753 |
|
EPZS16_Terminate_with_Refine: |
1754 |
|
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
1755 |
|
iMinSAD = Halfpel16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
1756 |
x, y, |
x, y, |
1757 |
currMV, iMinSAD, |
currMV, iMinSAD, |
1758 |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
1759 |
|
|
1760 |
step10_8b: |
EPZS16_Terminate_without_Refine: |
1761 |
|
|
1762 |
|
*oldMB = *pMB; |
1763 |
|
|
1764 |
currPMV->x = currMV->x - pmv[0].x; |
currPMV->x = currMV->x - pmv[0].x; |
1765 |
currPMV->y = currMV->y - pmv[0].y; |
currPMV->y = currMV->y - pmv[0].y; |
1766 |
|
return iMinSAD; |
1767 |
|
} |
1768 |
|
|
1769 |
|
|
1770 |
|
int32_t EPZSSearch8( |
1771 |
|
const uint8_t * const pRef, |
1772 |
|
const uint8_t * const pRefH, |
1773 |
|
const uint8_t * const pRefV, |
1774 |
|
const uint8_t * const pRefHV, |
1775 |
|
const IMAGE * const pCur, |
1776 |
|
const int x, const int y, |
1777 |
|
const int start_x, const int start_y, |
1778 |
|
const uint32_t MotionFlags, |
1779 |
|
const uint32_t iQuant, |
1780 |
|
const uint32_t iFcode, |
1781 |
|
const MBParam * const pParam, |
1782 |
|
const MACROBLOCK * const pMBs, |
1783 |
|
const MACROBLOCK * const prevMBs, |
1784 |
|
VECTOR * const currMV, |
1785 |
|
VECTOR * const currPMV) |
1786 |
|
{ |
1787 |
|
const uint32_t iWcount = pParam->mb_width; |
1788 |
|
const int32_t iWidth = pParam->width; |
1789 |
|
const int32_t iHeight = pParam->height; |
1790 |
|
const int32_t iEdgedWidth = pParam->edged_width; |
1791 |
|
|
1792 |
|
const uint8_t * cur = pCur->y + x*8 + y*8*iEdgedWidth; |
1793 |
|
|
1794 |
|
int32_t iDiamondSize=1; |
1795 |
|
|
1796 |
|
int32_t min_dx; |
1797 |
|
int32_t max_dx; |
1798 |
|
int32_t min_dy; |
1799 |
|
int32_t max_dy; |
1800 |
|
|
1801 |
|
VECTOR newMV; |
1802 |
|
VECTOR backupMV; |
1803 |
|
|
1804 |
|
VECTOR pmv[4]; |
1805 |
|
int32_t psad[8]; |
1806 |
|
|
1807 |
|
const int32_t iSubBlock = ((y&1)<<1) + (x&1); |
1808 |
|
|
1809 |
|
const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
1810 |
|
const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; |
1811 |
|
|
1812 |
|
int32_t bPredEq; |
1813 |
|
int32_t iMinSAD,iSAD=9999; |
1814 |
|
|
1815 |
|
MainSearch8FuncPtr EPZSMainSearchPtr; |
1816 |
|
|
1817 |
|
/* Get maximum range */ |
1818 |
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
1819 |
|
x, y, 8, iWidth, iHeight, iFcode); |
1820 |
|
|
1821 |
|
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
1822 |
|
|
1823 |
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
1824 |
|
{ min_dx = EVEN(min_dx); |
1825 |
|
max_dx = EVEN(max_dx); |
1826 |
|
min_dy = EVEN(min_dy); |
1827 |
|
max_dy = EVEN(max_dy); |
1828 |
|
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
1829 |
|
|
1830 |
|
bPredEq = get_pmvdata(pMBs, x>>1, y>>1, iWcount, iSubBlock, pmv, psad); |
1831 |
|
|
1832 |
|
|
1833 |
|
/* Step 4: Calculate SAD around the Median prediction. |
1834 |
|
MinSAD=SAD |
1835 |
|
If Motion Vector equal to Previous frame motion vector |
1836 |
|
and MinSAD<PrevFrmSAD goto Step 10. |
1837 |
|
If SAD<=256 goto Step 10. |
1838 |
|
*/ |
1839 |
|
|
1840 |
|
// Prepare for main loop |
1841 |
|
|
1842 |
|
|
1843 |
|
if (!(MotionFlags & PMV_HALFPEL8)) |
1844 |
|
{ |
1845 |
|
currMV->x = EVEN(currMV->x); |
1846 |
|
currMV->y = EVEN(currMV->y); |
1847 |
|
} |
1848 |
|
|
1849 |
|
if (currMV->x > max_dx) |
1850 |
|
currMV->x=max_dx; |
1851 |
|
if (currMV->x < min_dx) |
1852 |
|
currMV->x=min_dx; |
1853 |
|
if (currMV->y > max_dy) |
1854 |
|
currMV->y=max_dy; |
1855 |
|
if (currMV->y < min_dy) |
1856 |
|
currMV->y=min_dy; |
1857 |
|
|
1858 |
|
/***************** This is predictor SET A: only median prediction ******************/ |
1859 |
|
|
1860 |
|
|
1861 |
|
iMinSAD = sad8( cur, |
1862 |
|
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
1863 |
|
iEdgedWidth); |
1864 |
|
iMinSAD += calc_delta_8(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; |
1865 |
|
|
1866 |
|
|
1867 |
|
// thresh1 is fixed to 256 |
1868 |
|
if (iMinSAD < 256/4 ) |
1869 |
|
{ |
1870 |
|
if (MotionFlags & PMV_QUICKSTOP8) |
1871 |
|
goto EPZS8_Terminate_without_Refine; |
1872 |
|
if (MotionFlags & PMV_EARLYSTOP8) |
1873 |
|
goto EPZS8_Terminate_with_Refine; |
1874 |
|
} |
1875 |
|
|
1876 |
|
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
1877 |
|
|
1878 |
|
// previous frame MV |
1879 |
|
CHECK_MV8_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); |
1880 |
|
|
1881 |
|
// MV=(0,0) is often a good choice |
1882 |
|
|
1883 |
|
CHECK_MV8_ZERO; |
1884 |
|
|
1885 |
|
/* Terminate if MinSAD <= T_2 |
1886 |
|
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
1887 |
|
*/ |
1888 |
|
|
1889 |
|
if (iMinSAD < 512/4) /* T_2 == 512/4 hardcoded */ |
1890 |
|
{ |
1891 |
|
if (MotionFlags & PMV_QUICKSTOP8) |
1892 |
|
goto EPZS8_Terminate_without_Refine; |
1893 |
|
if (MotionFlags & PMV_EARLYSTOP8) |
1894 |
|
goto EPZS8_Terminate_with_Refine; |
1895 |
|
} |
1896 |
|
|
1897 |
|
/************ (if Diamond Search) **************/ |
1898 |
|
|
1899 |
|
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
1900 |
|
|
1901 |
|
if (!(MotionFlags & PMV_HALFPELDIAMOND8)) |
1902 |
|
iDiamondSize *= 2; |
1903 |
|
|
1904 |
|
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
1905 |
|
|
1906 |
|
// if (MotionFlags & PMV_USESQUARES8) |
1907 |
|
// EPZSMainSearchPtr = Square8_MainSearch; |
1908 |
|
// else |
1909 |
|
EPZSMainSearchPtr = Diamond8_MainSearch; |
1910 |
|
|
1911 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1912 |
|
x, y, |
1913 |
|
currMV->x, currMV->y, iMinSAD, &newMV, |
1914 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
1915 |
|
iDiamondSize, iFcode, iQuant, 00); |
1916 |
|
|
1917 |
|
|
1918 |
|
if (iSAD < iMinSAD) |
1919 |
|
{ |
1920 |
|
*currMV = newMV; |
1921 |
|
iMinSAD = iSAD; |
1922 |
|
} |
1923 |
|
|
1924 |
|
if (MotionFlags & PMV_EXTSEARCH8) |
1925 |
|
{ |
1926 |
|
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
1927 |
|
|
1928 |
|
if (!(MVequal(pmv[0],backupMV)) ) |
1929 |
|
{ |
1930 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1931 |
|
x, y, |
1932 |
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
1933 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); |
1934 |
|
|
1935 |
|
if (iSAD < iMinSAD) |
1936 |
|
{ |
1937 |
|
*currMV = newMV; |
1938 |
|
iMinSAD = iSAD; |
1939 |
|
} |
1940 |
|
} |
1941 |
|
|
1942 |
|
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
1943 |
|
{ |
1944 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1945 |
|
x, y, |
1946 |
|
0, 0, iMinSAD, &newMV, |
1947 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); |
1948 |
|
|
1949 |
|
if (iSAD < iMinSAD) |
1950 |
|
{ |
1951 |
|
*currMV = newMV; |
1952 |
|
iMinSAD = iSAD; |
1953 |
|
} |
1954 |
|
} |
1955 |
|
} |
1956 |
|
|
1957 |
|
/*************** Choose best MV found **************/ |
1958 |
|
|
1959 |
|
EPZS8_Terminate_with_Refine: |
1960 |
|
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
1961 |
|
iMinSAD = Halfpel8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
1962 |
|
x, y, |
1963 |
|
currMV, iMinSAD, |
1964 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
1965 |
|
|
1966 |
|
EPZS8_Terminate_without_Refine: |
1967 |
|
|
1968 |
|
currPMV->x = currMV->x - pmv[0].x; |
1969 |
|
currPMV->y = currMV->y - pmv[0].y; |
1970 |
return iMinSAD; |
return iMinSAD; |
1971 |
} |
} |
1972 |
|
|
1973 |
|
|
1974 |
|
|
1975 |
|
|
1976 |
|
|
1977 |
|
/* *********************************************************** |
1978 |
|
bvop motion estimation |
1979 |
|
// TODO: need to incorporate prediction here (eg. sad += calc_delta_16) |
1980 |
|
***************************************************************/ |
1981 |
|
|
1982 |
|
/* |
1983 |
|
void MotionEstimationBVOP( |
1984 |
|
MBParam * const pParam, |
1985 |
|
FRAMEINFO * const frame, |
1986 |
|
|
1987 |
|
// forward (past) reference |
1988 |
|
const MACROBLOCK * const f_mbs, |
1989 |
|
const IMAGE * const f_ref, |
1990 |
|
const IMAGE * const f_refH, |
1991 |
|
const IMAGE * const f_refV, |
1992 |
|
const IMAGE * const f_refHV, |
1993 |
|
// backward (future) reference |
1994 |
|
const MACROBLOCK * const b_mbs, |
1995 |
|
const IMAGE * const b_ref, |
1996 |
|
const IMAGE * const b_refH, |
1997 |
|
const IMAGE * const b_refV, |
1998 |
|
const IMAGE * const b_refHV) |
1999 |
|
{ |
2000 |
|
const uint32_t mb_width = pParam->mb_width; |
2001 |
|
const uint32_t mb_height = pParam->mb_height; |
2002 |
|
const int32_t edged_width = pParam->edged_width; |
2003 |
|
|
2004 |
|
int32_t i,j; |
2005 |
|
|
2006 |
|
int32_t f_sad16; |
2007 |
|
int32_t b_sad16; |
2008 |
|
int32_t i_sad16; |
2009 |
|
int32_t d_sad16; |
2010 |
|
int32_t best_sad; |
2011 |
|
|
2012 |
|
VECTOR pmv_dontcare; |
2013 |
|
|
2014 |
|
// note: i==horizontal, j==vertical |
2015 |
|
for (j = 0; j < mb_height; j++) |
2016 |
|
{ |
2017 |
|
for (i = 0; i < mb_width; i++) |
2018 |
|
{ |
2019 |
|
MACROBLOCK *mb = &frame->mbs[i + j*mb_width]; |
2020 |
|
const MACROBLOCK *f_mb = &f_mbs[i + j*mb_width]; |
2021 |
|
const MACROBLOCK *b_mb = &b_mbs[i + j*mb_width]; |
2022 |
|
|
2023 |
|
if (b_mb->mode == MODE_INTER |
2024 |
|
&& b_mb->cbp == 0 |
2025 |
|
&& b_mb->mvs[0].x == 0 |
2026 |
|
&& b_mb->mvs[0].y == 0) |
2027 |
|
{ |
2028 |
|
mb->mode = MB_IGNORE; |
2029 |
|
mb->mvs[0].x = 0; |
2030 |
|
mb->mvs[0].y = 0; |
2031 |
|
mb->b_mvs[0].x = 0; |
2032 |
|
mb->b_mvs[0].y = 0; |
2033 |
|
continue; |
2034 |
|
} |
2035 |
|
|
2036 |
|
|
2037 |
|
// forward search |
2038 |
|
f_sad16 = SEARCH16(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
2039 |
|
&frame->image, |
2040 |
|
i, j, |
2041 |
|
frame->motion_flags, frame->quant, frame->fcode, |
2042 |
|
pParam, |
2043 |
|
f_mbs, |
2044 |
|
&mb->mvs[0], &pmv_dontcare); // ignore pmv |
2045 |
|
|
2046 |
|
// backward search |
2047 |
|
b_sad16 = SEARCH16(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
2048 |
|
&frame->image, |
2049 |
|
i, j, |
2050 |
|
frame->motion_flags, frame->quant, frame->bcode, |
2051 |
|
pParam, |
2052 |
|
b_mbs, |
2053 |
|
&mb->b_mvs[0], &pmv_dontcare); // ignore pmv |
2054 |
|
|
2055 |
|
// interpolate search (simple, but effective) |
2056 |
|
i_sad16 = sad16bi_c( |
2057 |
|
frame->image.y + i*16 + j*16*edged_width, |
2058 |
|
get_ref(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
2059 |
|
i, j, 16, mb->mvs[0].x, mb->mvs[0].y, edged_width), |
2060 |
|
get_ref(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
2061 |
|
i, j, 16, mb->b_mvs[0].x, mb->b_mvs[0].x, edged_width), |
2062 |
|
edged_width); |
2063 |
|
|
2064 |
|
// TODO: direct search |
2065 |
|
// predictor + range of [-32,32] |
2066 |
|
d_sad16 = 65535; |
2067 |
|
|
2068 |
|
|
2069 |
|
if (f_sad16 < b_sad16) |
2070 |
|
{ |
2071 |
|
best_sad = f_sad16; |
2072 |
|
mb->mode = MB_FORWARD; |
2073 |
|
} |
2074 |
|
else |
2075 |
|
{ |
2076 |
|
best_sad = b_sad16; |
2077 |
|
mb->mode = MB_BACKWARD; |
2078 |
|
} |
2079 |
|
|
2080 |
|
if (i_sad16 < best_sad) |
2081 |
|
{ |
2082 |
|
best_sad = i_sad16; |
2083 |
|
mb->mode = MB_INTERPOLATE; |
2084 |
|
} |
2085 |
|
|
2086 |
|
if (d_sad16 < best_sad) |
2087 |
|
{ |
2088 |
|
best_sad = d_sad16; |
2089 |
|
mb->mode = MB_DIRECT; |
2090 |
|
} |
2091 |
|
|
2092 |
|
} |
2093 |
|
} |
2094 |
|
} |
2095 |
|
|
2096 |
|
*/ |