37 |
#include "../prediction/mbprediction.h" |
#include "../prediction/mbprediction.h" |
38 |
#include "../global.h" |
#include "../global.h" |
39 |
#include "../utils/timer.h" |
#include "../utils/timer.h" |
40 |
|
#include "../image/interpolate8x8.h" |
41 |
#include "motion_est.h" |
#include "motion_est.h" |
42 |
#include "motion.h" |
#include "motion.h" |
43 |
#include "sad.h" |
#include "sad.h" |
44 |
|
#include "../utils/emms.h" |
45 |
|
|
46 |
#define INITIAL_SKIP_THRESH (10) |
#define INITIAL_SKIP_THRESH (10) |
47 |
#define FINAL_SKIP_THRESH (50) |
#define FINAL_SKIP_THRESH (50) |
48 |
#define MAX_SAD00_FOR_SKIP (20) |
#define MAX_SAD00_FOR_SKIP (20) |
49 |
#define MAX_CHROMA_SAD_FOR_SKIP (18) |
#define MAX_CHROMA_SAD_FOR_SKIP (22) |
50 |
#define SKIP_THRESH_B (10) |
#define SKIP_THRESH_B (25) |
51 |
|
|
52 |
#define CHECK_CANDIDATE(X,Y,D) { \ |
#define CHECK_CANDIDATE(X,Y,D) { \ |
53 |
(*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } |
(*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } |
54 |
|
|
55 |
#define iDiamondSize 2 |
#define GET_REFERENCE(X, Y, REF) { \ |
56 |
|
switch ( (((X)&1)<<1) + ((Y)&1) ) \ |
57 |
|
{ \ |
58 |
|
case 0 : REF = (uint8_t *)data->Ref + (X)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ |
59 |
|
case 1 : REF = (uint8_t *)data->RefV + (X)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ |
60 |
|
case 2 : REF = (uint8_t *)data->RefH + ((X)-1)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ |
61 |
|
default : REF = (uint8_t *)data->RefHV + ((X)-1)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ |
62 |
|
} \ |
63 |
|
} |
64 |
|
// I hate those macros :/ |
65 |
|
#define GET_REFERENCE2(X, Y, REF) { \ |
66 |
|
switch ( (((X)&1)<<1) + ((Y)&1) ) \ |
67 |
|
{ \ |
68 |
|
case 0 : REF = (uint8_t *)data->bRef + (X)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ |
69 |
|
case 1 : REF = (uint8_t *)data->bRefV + (X)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ |
70 |
|
case 2 : REF = (uint8_t *)data->bRefH + ((X)-1)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ |
71 |
|
default : REF = (uint8_t *)data->bRefHV + ((X)-1)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ |
72 |
|
} \ |
73 |
|
} |
74 |
|
|
75 |
|
|
76 |
//FILE * debug; |
#define iDiamondSize 2 |
77 |
|
|
78 |
static __inline int |
static __inline int |
79 |
d_mv_bits(int x, int y, const uint32_t iFcode) |
d_mv_bits(int x, int y, const uint32_t iFcode) |
100 |
return xb + yb; |
return xb + yb; |
101 |
} |
} |
102 |
|
|
103 |
/* CHACK_CANDIATE FUNCTIONS START */ |
static int32_t |
104 |
|
ChromaSAD(int dx, int dy, const SearchData * const data) |
105 |
|
{ |
106 |
|
int sad; |
107 |
|
dx = (dx >> 1) + roundtab_79[dx & 0x3]; |
108 |
|
dy = (dy >> 1) + roundtab_79[dy & 0x3]; |
109 |
|
|
110 |
|
switch (((dx & 1) << 1) + (dy & 1)) { // ((dx%2)?2:0)+((dy%2)?1:0) |
111 |
|
case 0: |
112 |
|
sad = sad8(data->CurU, data->RefCU + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); |
113 |
|
sad += sad8(data->CurV, data->RefCV + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); |
114 |
|
break; |
115 |
|
case 1: |
116 |
|
dx = dx / 2; dy = (dy - 1) / 2; |
117 |
|
sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); |
118 |
|
sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); |
119 |
|
break; |
120 |
|
case 2: |
121 |
|
dx = (dx - 1) / 2; dy = dy / 2; |
122 |
|
sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); |
123 |
|
sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); |
124 |
|
break; |
125 |
|
default: |
126 |
|
dx = (dx - 1) / 2; dy = (dy - 1) / 2; |
127 |
|
interpolate8x8_halfpel_hv(data->RefQ, |
128 |
|
data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, |
129 |
|
data->rounding); |
130 |
|
sad = sad8(data->CurU, data->RefQ, data->iEdgedWidth/2); |
131 |
|
interpolate8x8_halfpel_hv(data->RefQ, |
132 |
|
data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, |
133 |
|
data->rounding); |
134 |
|
sad += sad8(data->CurV, data->RefQ, data->iEdgedWidth/2); |
135 |
|
break; |
136 |
|
} |
137 |
|
return sad; |
138 |
|
} |
139 |
|
|
140 |
|
|
141 |
|
/* CHECK_CANDIATE FUNCTIONS START */ |
142 |
|
|
143 |
|
|
144 |
static void |
static void |
145 |
CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
146 |
{ |
{ |
|
int32_t * const sad = data->temp; |
|
|
// static int32_t sad[5]; |
|
147 |
int t; |
int t; |
148 |
const uint8_t * Reference; |
const uint8_t * Reference; |
149 |
|
|
157 |
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
158 |
} |
} |
159 |
|
|
160 |
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, sad+1); |
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
161 |
|
|
162 |
t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
if (data->qpel) t = d_mv_bits(2*x - data->predQMV.x, 2*y - data->predQMV.y, data->iFcode); |
163 |
data->temp[0] += lambda_vec16[data->iQuant] * t; |
else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
164 |
data->temp[1] += lambda_vec8[data->iQuant] * t; |
|
165 |
|
data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; |
166 |
|
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; |
167 |
|
|
168 |
|
if (data->chroma) data->temp[0] += ChromaSAD(x, y, data); |
169 |
|
|
170 |
if (data->temp[0] < data->iMinSAD[0]) { |
if (data->temp[0] < data->iMinSAD[0]) { |
171 |
data->iMinSAD[0] = data->temp[0]; |
data->iMinSAD[0] = data->temp[0]; |
200 |
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
201 |
} |
} |
202 |
|
|
203 |
sad = lambda_vec16[data->iQuant] * |
sad = sad16(data->Cur, Reference, data->iEdgedWidth, MV_MAX_ERROR); |
204 |
d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
if (data->qpel) //only to be used in b-frames' ME |
205 |
sad += sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
sad += (data->lambda16 * d_mv_bits(2*x - data->predMV.x, 2*y - data->predMV.y, data->iFcode) * sad)/1000; |
206 |
|
else |
207 |
|
sad += (data->lambda16 * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode) * sad)/1000; |
208 |
|
|
209 |
|
if (sad < *(data->iMinSAD)) { |
210 |
|
*(data->iMinSAD) = sad; |
211 |
|
data->currentMV[0].x = x; data->currentMV[0].y = y; |
212 |
|
*dir = Direction; } |
213 |
|
} |
214 |
|
|
215 |
|
static void |
216 |
|
CheckCandidate16_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
217 |
|
|
218 |
|
// CheckCandidate16 variant which expects x and y in quarter pixel resolution |
219 |
|
// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!) |
220 |
|
// around currentMV! |
221 |
|
{ |
222 |
|
int t; |
223 |
|
uint8_t * Reference = (uint8_t *)data->RefQ; |
224 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
225 |
|
VECTOR halfpelMV = *(data->currentMV); |
226 |
|
|
227 |
|
int32_t iEdgedWidth = data->iEdgedWidth; |
228 |
|
uint32_t rounding = data->rounding; |
229 |
|
|
230 |
|
if (( x > data->max_dx) || ( x < data->min_dx) |
231 |
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
232 |
|
|
233 |
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this refenrence is used in all cases |
234 |
|
switch( ((x&1)<<1) + (y&1) ) |
235 |
|
{ |
236 |
|
case 0: // pure halfpel position - shouldn't happen during a refinement step |
237 |
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, Reference); |
238 |
|
break; |
239 |
|
|
240 |
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
241 |
|
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
242 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
243 |
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); |
244 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); |
245 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); |
246 |
|
break; |
247 |
|
|
248 |
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
249 |
|
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); |
250 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
251 |
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); |
252 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); |
253 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); |
254 |
|
break; |
255 |
|
|
256 |
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
257 |
|
// bottom left/right) during qpel refinement |
258 |
|
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
259 |
|
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); |
260 |
|
GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); |
261 |
|
|
262 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
263 |
|
interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); |
264 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); |
265 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); |
266 |
|
break; |
267 |
|
} |
268 |
|
|
269 |
|
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp+1); |
270 |
|
|
271 |
|
t = d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode); |
272 |
|
data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; |
273 |
|
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; |
274 |
|
|
275 |
|
if (data->chroma) |
276 |
|
data->temp[0] += ChromaSAD(x/2, y/2, data); |
277 |
|
|
278 |
|
if (data->temp[0] < data->iMinSAD[0]) { |
279 |
|
data->iMinSAD[0] = data->temp[0]; |
280 |
|
data->currentQMV[0].x = x; data->currentQMV[0].y = y; |
281 |
|
/* *dir = Direction;*/ } |
282 |
|
|
283 |
|
if (data->temp[1] < data->iMinSAD[1]) { |
284 |
|
data->iMinSAD[1] = data->temp[1]; data->currentQMV[1].x = x; data->currentQMV[1].y = y; } |
285 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
286 |
|
data->iMinSAD[2] = data->temp[2]; data->currentQMV[2].x = x; data->currentQMV[2].y = y; } |
287 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
288 |
|
data->iMinSAD[3] = data->temp[3]; data->currentQMV[3].x = x; data->currentQMV[3].y = y; } |
289 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
290 |
|
data->iMinSAD[4] = data->temp[4]; data->currentQMV[4].x = x; data->currentQMV[4].y = y; } |
291 |
|
} |
292 |
|
|
293 |
|
static void |
294 |
|
CheckCandidate16no4v_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
295 |
|
|
296 |
|
// CheckCandidate16no4v variant which expects x and y in quarter pixel resolution |
297 |
|
// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!) |
298 |
|
// around currentMV! |
299 |
|
// this function is for B-frames' search only |
300 |
|
{ |
301 |
|
uint8_t * Reference = (uint8_t *)data->RefQ; |
302 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
303 |
|
VECTOR halfpelMV = *(data->currentMV); |
304 |
|
|
305 |
|
int32_t iEdgedWidth = data->iEdgedWidth; |
306 |
|
int32_t sad; |
307 |
|
|
308 |
|
if (( x > data->max_dx) || ( x < data->min_dx) |
309 |
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
310 |
|
|
311 |
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this refenrence is used in all cases |
312 |
|
switch( ((x&1)<<1) + (y&1) ) |
313 |
|
{ |
314 |
|
case 0: // pure halfpel position - shouldn't happen during a refinement step |
315 |
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, Reference); |
316 |
|
break; |
317 |
|
|
318 |
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
319 |
|
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
320 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, 0); |
321 |
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, 0); |
322 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); |
323 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); |
324 |
|
break; |
325 |
|
|
326 |
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
327 |
|
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); |
328 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, 0); |
329 |
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, 0); |
330 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); |
331 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); |
332 |
|
break; |
333 |
|
|
334 |
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
335 |
|
// bottom left/right) during qpel refinement |
336 |
|
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
337 |
|
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); |
338 |
|
GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); |
339 |
|
|
340 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, 0); |
341 |
|
interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, 0); |
342 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, 0); |
343 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, 0); |
344 |
|
break; |
345 |
|
} |
346 |
|
|
347 |
|
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
348 |
|
sad += (data->lambda16 * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode) * sad)/1000; |
349 |
|
|
350 |
|
if (sad < data->iMinSAD[0]) { |
351 |
|
data->iMinSAD[0] = sad; |
352 |
|
data->currentQMV[0].x = x; data->currentQMV[0].y = y; |
353 |
|
/* *dir = Direction;*/ } |
354 |
|
} |
355 |
|
|
356 |
|
static void |
357 |
|
CheckCandidate16no4vI(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
358 |
|
{ |
359 |
|
// maximum speed - for P/B/I decision |
360 |
|
int32_t sad; |
361 |
|
|
362 |
|
if (( x > data->max_dx) || ( x < data->min_dx) |
363 |
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
364 |
|
|
365 |
|
sad = sad16(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), |
366 |
|
data->iEdgedWidth, 256*4096); |
367 |
|
|
368 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
369 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
371 |
*dir = Direction; } |
*dir = Direction; } |
372 |
} |
} |
373 |
|
|
374 |
|
|
375 |
static void |
static void |
376 |
CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) |
377 |
{ |
{ |
397 |
default : ReferenceB = data->bRefHV + (xb-1)/2 + ((yb-1)/2)*(data->iEdgedWidth); break; |
default : ReferenceB = data->bRefHV + (xb-1)/2 + ((yb-1)/2)*(data->iEdgedWidth); break; |
398 |
} |
} |
399 |
|
|
400 |
sad = lambda_vec16[data->iQuant] * |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
|
( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) + |
|
|
d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode) ); |
|
401 |
|
|
402 |
sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
if (data->qpel) |
403 |
|
sad += (data->lambda16 * |
404 |
|
( d_mv_bits(2*xf - data->predMV.x, 2*yf - data->predMV.y, data->iFcode) + |
405 |
|
d_mv_bits(2*xb - data->bpredMV.x, 2*yb - data->bpredMV.y, data->iFcode)) * sad)/1000; |
406 |
|
else |
407 |
|
sad += (data->lambda16 * |
408 |
|
( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) + |
409 |
|
d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode)) * sad)/1000; |
410 |
|
|
411 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
412 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
414 |
*dir = Direction; } |
*dir = Direction; } |
415 |
} |
} |
416 |
|
|
417 |
|
|
418 |
static void |
static void |
419 |
CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidateInt_qpel(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) |
420 |
{ |
{ |
421 |
|
// CheckCandidateInt variant which expects x and y in quarter pixel resolution |
422 |
|
|
423 |
int32_t sad; |
int32_t sad; |
424 |
|
const int xb = data->currentQMV[1].x; |
425 |
|
const int yb = data->currentQMV[1].y; |
426 |
|
uint8_t * ReferenceF = (uint8_t *)data->RefQ; |
427 |
|
uint8_t * ReferenceB = (uint8_t *)data->RefQ + 16; |
428 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
429 |
|
VECTOR halfpelMV; |
430 |
|
const int32_t iEdgedWidth = data->iEdgedWidth; |
431 |
|
|
432 |
|
if (( xf > data->max_dx) || ( xf < data->min_dx) |
433 |
|
|| ( yf > data->max_dy) || (yf < data->min_dy)) return; |
434 |
|
|
435 |
|
halfpelMV.x = xf/2; //forward first |
436 |
|
halfpelMV.y = yf/2; |
437 |
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this reference is used in all cases |
438 |
|
switch( ((xf&1)<<1) + (yf&1) ) |
439 |
|
{ |
440 |
|
case 0: // pure halfpel position - shouldn't happen during a refinement step |
441 |
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ReferenceF); |
442 |
|
break; |
443 |
|
|
444 |
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
445 |
|
GET_REFERENCE(halfpelMV.x, yf - halfpelMV.y, ref2); |
446 |
|
interpolate8x8_avg2(ReferenceF, ref1, ref2, iEdgedWidth, 0); |
447 |
|
interpolate8x8_avg2(ReferenceF+8, ref1+8, ref2+8, iEdgedWidth, 0); |
448 |
|
interpolate8x8_avg2(ReferenceF+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); |
449 |
|
interpolate8x8_avg2(ReferenceF+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); |
450 |
|
break; |
451 |
|
|
452 |
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
453 |
|
GET_REFERENCE(xf - halfpelMV.x, halfpelMV.y, ref2); |
454 |
|
interpolate8x8_avg2(ReferenceF, ref1, ref2, iEdgedWidth, 0); |
455 |
|
interpolate8x8_avg2(ReferenceF+8, ref1+8, ref2+8, iEdgedWidth, 0); |
456 |
|
interpolate8x8_avg2(ReferenceF+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); |
457 |
|
interpolate8x8_avg2(ReferenceF+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); |
458 |
|
break; |
459 |
|
|
460 |
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
461 |
|
// bottom left/right) during qpel refinement |
462 |
|
GET_REFERENCE(halfpelMV.x, yf - halfpelMV.y, ref2); |
463 |
|
GET_REFERENCE(xf - halfpelMV.x, halfpelMV.y, ref3); |
464 |
|
GET_REFERENCE(xf - halfpelMV.x, yf - halfpelMV.y, ref4); |
465 |
|
|
466 |
|
interpolate8x8_avg4(ReferenceF, ref1, ref2, ref3, ref4, iEdgedWidth, 0); |
467 |
|
interpolate8x8_avg4(ReferenceF+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, 0); |
468 |
|
interpolate8x8_avg4(ReferenceF+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, 0); |
469 |
|
interpolate8x8_avg4(ReferenceF+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, 0); |
470 |
|
break; |
471 |
|
} |
472 |
|
|
473 |
|
halfpelMV.x = xb/2; //backward |
474 |
|
halfpelMV.y = yb/2; |
475 |
|
GET_REFERENCE2(halfpelMV.x, halfpelMV.y, ref1); // this reference is used in all cases |
476 |
|
switch( ((xb&1)<<1) + (yb&1) ) |
477 |
|
{ |
478 |
|
case 0: // pure halfpel position - shouldn't happen during a refinement step |
479 |
|
GET_REFERENCE2(halfpelMV.x, halfpelMV.y, ReferenceB); |
480 |
|
break; |
481 |
|
|
482 |
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
483 |
|
GET_REFERENCE2(halfpelMV.x, yb - halfpelMV.y, ref2); |
484 |
|
interpolate8x8_avg2(ReferenceB, ref1, ref2, iEdgedWidth, 0); |
485 |
|
interpolate8x8_avg2(ReferenceB+8, ref1+8, ref2+8, iEdgedWidth, 0); |
486 |
|
interpolate8x8_avg2(ReferenceB+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); |
487 |
|
interpolate8x8_avg2(ReferenceB+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); |
488 |
|
break; |
489 |
|
|
490 |
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
491 |
|
GET_REFERENCE2(xb - halfpelMV.x, halfpelMV.y, ref2); |
492 |
|
interpolate8x8_avg2(ReferenceB, ref1, ref2, iEdgedWidth, 0); |
493 |
|
interpolate8x8_avg2(ReferenceB+8, ref1+8, ref2+8, iEdgedWidth, 0); |
494 |
|
interpolate8x8_avg2(ReferenceB+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); |
495 |
|
interpolate8x8_avg2(ReferenceB+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); |
496 |
|
break; |
497 |
|
|
498 |
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
499 |
|
// bottom left/right) during qpel refinement |
500 |
|
GET_REFERENCE2(halfpelMV.x, yb - halfpelMV.y, ref2); |
501 |
|
GET_REFERENCE2(xb - halfpelMV.x, halfpelMV.y, ref3); |
502 |
|
GET_REFERENCE2(xb - halfpelMV.x, yb - halfpelMV.y, ref4); |
503 |
|
|
504 |
|
interpolate8x8_avg4(ReferenceB, ref1, ref2, ref3, ref4, iEdgedWidth, 0); |
505 |
|
interpolate8x8_avg4(ReferenceB+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, 0); |
506 |
|
interpolate8x8_avg4(ReferenceB+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, 0); |
507 |
|
interpolate8x8_avg4(ReferenceB+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, 0); |
508 |
|
break; |
509 |
|
} |
510 |
|
|
511 |
|
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
512 |
|
|
513 |
|
sad += (data->lambda16 * |
514 |
|
( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) + |
515 |
|
d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode)) * sad)/1000; |
516 |
|
|
517 |
|
if (sad < *(data->iMinSAD)) { |
518 |
|
*(data->iMinSAD) = sad; |
519 |
|
data->currentQMV->x = xf; data->currentQMV->y = yf; |
520 |
|
*dir = Direction; } |
521 |
|
} |
522 |
|
|
523 |
|
static void |
524 |
|
CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
525 |
|
{ |
526 |
|
int32_t sad = 0; |
527 |
int k; |
int k; |
528 |
const uint8_t *ReferenceF; |
const uint8_t *ReferenceF; |
529 |
const uint8_t *ReferenceB; |
const uint8_t *ReferenceB; |
531 |
|
|
532 |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
533 |
|
|
|
sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1); |
|
|
|
|
534 |
for (k = 0; k < 4; k++) { |
for (k = 0; k < 4; k++) { |
535 |
mvs.x = data->directmvF[k].x + x; |
mvs.x = data->directmvF[k].x + x; |
536 |
b_mvs.x = ((x == 0) ? |
b_mvs.x = ((x == 0) ? |
568 |
if (sad > *(data->iMinSAD)) return; |
if (sad > *(data->iMinSAD)) return; |
569 |
} |
} |
570 |
|
|
571 |
|
sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000; |
572 |
|
|
573 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
574 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
575 |
data->currentMV->x = x; data->currentMV->y = y; |
data->currentMV->x = x; data->currentMV->y = y; |
576 |
*dir = Direction; } |
*dir = Direction; } |
577 |
} |
} |
578 |
|
|
579 |
|
|
580 |
|
static void |
581 |
|
CheckCandidateDirect_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
582 |
|
{ |
583 |
|
int32_t sad = 0; |
584 |
|
int k; |
585 |
|
VECTOR mvs, b_mvs, halfpelMV; |
586 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
587 |
|
uint8_t *ReferenceF, *ReferenceB; |
588 |
|
const uint32_t iEdgedWidth = data->iEdgedWidth; |
589 |
|
|
590 |
|
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
591 |
|
|
592 |
|
for (k = 0; k < 4; k++) { |
593 |
|
ReferenceF = (uint8_t *)data->RefQ; |
594 |
|
ReferenceB = (uint8_t *)data->RefQ + 64; |
595 |
|
|
596 |
|
mvs.x = data->directmvF[k].x + x; |
597 |
|
b_mvs.x = ((x == 0) ? |
598 |
|
data->directmvB[k].x |
599 |
|
: mvs.x - data->referencemv[k].x); |
600 |
|
|
601 |
|
mvs.y = data->directmvF[k].y + y; |
602 |
|
b_mvs.y = ((y == 0) ? |
603 |
|
data->directmvB[k].y |
604 |
|
: mvs.y - data->referencemv[k].y); |
605 |
|
|
606 |
|
if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
607 |
|
|| ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) |
608 |
|
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
609 |
|
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
610 |
|
|
611 |
|
halfpelMV.x = mvs.x/2; //forward first |
612 |
|
halfpelMV.y = mvs.y/2; |
613 |
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this reference is used in all cases |
614 |
|
switch( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
615 |
|
case 0: // pure halfpel position |
616 |
|
GET_REFERENCE(halfpelMV.x + 16*(k&1), halfpelMV.y + 16*(k>>1), ReferenceF); |
617 |
|
break; |
618 |
|
|
619 |
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
620 |
|
GET_REFERENCE(halfpelMV.x, mvs.y - halfpelMV.y, ref2); |
621 |
|
interpolate8x8_avg2(ReferenceF, ref1+8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
622 |
|
ref2+ 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0); |
623 |
|
break; |
624 |
|
|
625 |
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
626 |
|
GET_REFERENCE(mvs.x - halfpelMV.x, halfpelMV.y, ref2); |
627 |
|
interpolate8x8_avg2(ReferenceF, ref1 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
628 |
|
ref2 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0); |
629 |
|
break; |
630 |
|
|
631 |
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
632 |
|
// bottom left/right) during qpel refinement |
633 |
|
GET_REFERENCE(halfpelMV.x, mvs.y - halfpelMV.y, ref2); |
634 |
|
GET_REFERENCE(mvs.x - halfpelMV.x, halfpelMV.y, ref3); |
635 |
|
GET_REFERENCE(mvs.x - halfpelMV.x, mvs.y - halfpelMV.y, ref4); |
636 |
|
interpolate8x8_avg4(ReferenceF, ref1 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
637 |
|
ref2 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
638 |
|
ref3 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
639 |
|
ref4 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0); |
640 |
|
break; |
641 |
|
} |
642 |
|
|
643 |
|
halfpelMV.x = b_mvs.x/2; |
644 |
|
halfpelMV.y = b_mvs.y/2; |
645 |
|
GET_REFERENCE2(halfpelMV.x, halfpelMV.y, ref1); // this reference is used in most cases |
646 |
|
switch( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { |
647 |
|
case 0: // pure halfpel position |
648 |
|
GET_REFERENCE2(halfpelMV.x + 16*(k&1), halfpelMV.y + 16*(k>>1), ReferenceB); |
649 |
|
break; |
650 |
|
|
651 |
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
652 |
|
GET_REFERENCE2(halfpelMV.x, b_mvs.y - halfpelMV.y, ref2); |
653 |
|
interpolate8x8_avg2(ReferenceB, ref1+8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
654 |
|
ref2+ 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0); |
655 |
|
break; |
656 |
|
|
657 |
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
658 |
|
GET_REFERENCE2(b_mvs.x - halfpelMV.x, halfpelMV.y, ref2); |
659 |
|
interpolate8x8_avg2(ReferenceB, ref1 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
660 |
|
ref2 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0); |
661 |
|
break; |
662 |
|
|
663 |
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
664 |
|
// bottom left/right) during qpel refinement |
665 |
|
GET_REFERENCE2(halfpelMV.x, b_mvs.y - halfpelMV.y, ref2); |
666 |
|
GET_REFERENCE2(b_mvs.x - halfpelMV.x, halfpelMV.y, ref3); |
667 |
|
GET_REFERENCE2(b_mvs.x - halfpelMV.x, b_mvs.y - halfpelMV.y, ref4); |
668 |
|
interpolate8x8_avg4(ReferenceB, ref1 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
669 |
|
ref2 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
670 |
|
ref3 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
671 |
|
ref4 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0); |
672 |
|
break; |
673 |
|
} |
674 |
|
|
675 |
|
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
676 |
|
ReferenceF, |
677 |
|
ReferenceB, |
678 |
|
data->iEdgedWidth); |
679 |
|
if (sad > *(data->iMinSAD)) return; |
680 |
|
} |
681 |
|
|
682 |
|
sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000; |
683 |
|
|
684 |
|
if (sad < *(data->iMinSAD)) { |
685 |
|
*(data->iMinSAD) = sad; |
686 |
|
data->currentMV->x = x; data->currentMV->y = y; |
687 |
|
*dir = Direction; } |
688 |
|
} |
689 |
|
|
690 |
|
static void |
691 |
|
CheckCandidateDirectno4v_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
692 |
|
{ |
693 |
|
int32_t sad = 0; |
694 |
|
VECTOR mvs, b_mvs, halfpelMV; |
695 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
696 |
|
const uint32_t iEdgedWidth = data->iEdgedWidth; |
697 |
|
uint8_t * ReferenceF = (uint8_t *)data->RefQ; |
698 |
|
uint8_t * ReferenceB = (uint8_t *)data->RefQ + 64; |
699 |
|
|
700 |
|
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
701 |
|
|
702 |
|
mvs.x = data->directmvF[0].x + x; |
703 |
|
b_mvs.x = ((x == 0) ? |
704 |
|
data->directmvB[0].x |
705 |
|
: mvs.x - data->referencemv[0].x); |
706 |
|
|
707 |
|
mvs.y = data->directmvF[0].y + y; |
708 |
|
b_mvs.y = ((y == 0) ? |
709 |
|
data->directmvB[0].y |
710 |
|
: mvs.y - data->referencemv[0].y); |
711 |
|
|
712 |
|
if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
713 |
|
|| ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) |
714 |
|
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
715 |
|
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
716 |
|
|
717 |
|
halfpelMV.x = mvs.x/2; //forward first |
718 |
|
halfpelMV.y = mvs.y/2; |
719 |
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this reference is used in all cases |
720 |
|
switch( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
721 |
|
case 0: // pure halfpel position |
722 |
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ReferenceF); |
723 |
|
break; |
724 |
|
|
725 |
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
726 |
|
GET_REFERENCE(halfpelMV.x, mvs.y - halfpelMV.y, ref2); |
727 |
|
interpolate8x8_avg2(ReferenceF, ref1, ref2, iEdgedWidth, 0); |
728 |
|
interpolate8x8_avg2(ReferenceF+8, ref1+8, ref2+8, iEdgedWidth, 0); |
729 |
|
interpolate8x8_avg2(ReferenceF+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); |
730 |
|
interpolate8x8_avg2(ReferenceF+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); |
731 |
|
break; |
732 |
|
|
733 |
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
734 |
|
GET_REFERENCE(mvs.x - halfpelMV.x, halfpelMV.y, ref2); |
735 |
|
interpolate8x8_avg2(ReferenceF, ref1, ref2, iEdgedWidth, 0); |
736 |
|
interpolate8x8_avg2(ReferenceF+8, ref1+8, ref2+8, iEdgedWidth, 0); |
737 |
|
interpolate8x8_avg2(ReferenceF+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); |
738 |
|
interpolate8x8_avg2(ReferenceF+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); |
739 |
|
break; |
740 |
|
|
741 |
|
default: // x and y in qpel resolution |
742 |
|
GET_REFERENCE(halfpelMV.x, mvs.y - halfpelMV.y, ref2); |
743 |
|
GET_REFERENCE(mvs.x - halfpelMV.x, halfpelMV.y, ref3); |
744 |
|
GET_REFERENCE(mvs.x - halfpelMV.x, mvs.y - halfpelMV.y, ref4); |
745 |
|
|
746 |
|
interpolate8x8_avg4(ReferenceF, ref1, ref2, ref3, ref4, iEdgedWidth, 0); |
747 |
|
interpolate8x8_avg4(ReferenceF+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, 0); |
748 |
|
interpolate8x8_avg4(ReferenceF+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, 0); |
749 |
|
interpolate8x8_avg4(ReferenceF+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, 0); |
750 |
|
break; |
751 |
|
} |
752 |
|
|
753 |
|
halfpelMV.x = b_mvs.x/2; //backward |
754 |
|
halfpelMV.y = b_mvs.y/2; |
755 |
|
GET_REFERENCE2(halfpelMV.x, halfpelMV.y, ref1); |
756 |
|
switch( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) |
757 |
|
{ |
758 |
|
case 0: // pure halfpel position |
759 |
|
GET_REFERENCE2(halfpelMV.x, halfpelMV.y, ReferenceB); |
760 |
|
break; |
761 |
|
|
762 |
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
763 |
|
GET_REFERENCE2(halfpelMV.x, b_mvs.y - halfpelMV.y, ref2); |
764 |
|
interpolate8x8_avg2(ReferenceB, ref1, ref2, iEdgedWidth, 0); |
765 |
|
interpolate8x8_avg2(ReferenceB+8, ref1+8, ref2+8, iEdgedWidth, 0); |
766 |
|
interpolate8x8_avg2(ReferenceB+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); |
767 |
|
interpolate8x8_avg2(ReferenceB+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); |
768 |
|
break; |
769 |
|
|
770 |
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
771 |
|
GET_REFERENCE2(b_mvs.x - halfpelMV.x, halfpelMV.y, ref2); |
772 |
|
interpolate8x8_avg2(ReferenceB, ref1, ref2, iEdgedWidth, 0); |
773 |
|
interpolate8x8_avg2(ReferenceB+8, ref1+8, ref2+8, iEdgedWidth, 0); |
774 |
|
interpolate8x8_avg2(ReferenceB+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); |
775 |
|
interpolate8x8_avg2(ReferenceB+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); |
776 |
|
break; |
777 |
|
|
778 |
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
779 |
|
// bottom left/right) during qpel refinement |
780 |
|
GET_REFERENCE2(halfpelMV.x, b_mvs.y - halfpelMV.y, ref2); |
781 |
|
GET_REFERENCE2(b_mvs.x - halfpelMV.x, halfpelMV.y, ref3); |
782 |
|
GET_REFERENCE2(b_mvs.x - halfpelMV.x, b_mvs.y - halfpelMV.y, ref4); |
783 |
|
|
784 |
|
interpolate8x8_avg4(ReferenceB, ref1, ref2, ref3, ref4, iEdgedWidth, 0); |
785 |
|
interpolate8x8_avg4(ReferenceB+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, 0); |
786 |
|
interpolate8x8_avg4(ReferenceB+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, 0); |
787 |
|
interpolate8x8_avg4(ReferenceB+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, 0); |
788 |
|
break; |
789 |
|
} |
790 |
|
|
791 |
|
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
792 |
|
sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000; |
793 |
|
|
794 |
|
if (sad < *(data->iMinSAD)) { |
795 |
|
*(data->iMinSAD) = sad; |
796 |
|
data->currentMV->x = x; data->currentMV->y = y; |
797 |
|
*dir = Direction; } |
798 |
|
} |
799 |
|
|
800 |
|
|
801 |
static void |
static void |
802 |
CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
803 |
{ |
{ |
806 |
const uint8_t *ReferenceB; |
const uint8_t *ReferenceB; |
807 |
VECTOR mvs, b_mvs; |
VECTOR mvs, b_mvs; |
808 |
|
|
809 |
if (( x > 31) || ( x < -31) || ( y > 31) || (y < -31)) return; |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
|
|
|
|
sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1); |
|
810 |
|
|
811 |
mvs.x = data->directmvF[0].x + x; |
mvs.x = data->directmvF[0].x + x; |
812 |
b_mvs.x = ((x == 0) ? |
b_mvs.x = ((x == 0) ? |
837 |
default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
838 |
} |
} |
839 |
|
|
840 |
sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
841 |
|
sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000; |
842 |
|
|
843 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
844 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
849 |
static void |
static void |
850 |
CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
851 |
{ |
{ |
852 |
int32_t sad; |
int32_t sad; int t; |
853 |
const uint8_t * Reference; |
const uint8_t * Reference; |
854 |
|
|
855 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if (( x > data->max_dx) || ( x < data->min_dx) |
864 |
} |
} |
865 |
|
|
866 |
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
867 |
sad += lambda_vec8[data->iQuant] * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
if (data->qpel) t = d_mv_bits(2 * x - data->predQMV.x, 2 * y - data->predQMV.y, data->iFcode); |
868 |
|
else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
869 |
|
|
870 |
|
sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))/100; |
871 |
|
|
872 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
873 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
875 |
*dir = Direction; } |
*dir = Direction; } |
876 |
} |
} |
877 |
|
|
878 |
/* CHACK_CANDIATE FUNCTIONS END */ |
static void |
879 |
|
CheckCandidate8_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
880 |
|
// CheckCandidate8 variant which expects x and y in quarter pixel resolution |
881 |
|
// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!) |
882 |
|
// around currentMV! |
883 |
|
|
884 |
|
{ |
885 |
|
int32_t sad; |
886 |
|
uint8_t *Reference = (uint8_t *) data->RefQ; |
887 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
888 |
|
VECTOR halfpelMV = *(data->currentMV); |
889 |
|
|
890 |
|
int32_t iEdgedWidth = data->iEdgedWidth; |
891 |
|
uint32_t rounding = data->rounding; |
892 |
|
|
893 |
|
if (( x > data->max_dx) || ( x < data->min_dx) |
894 |
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
895 |
|
|
896 |
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); |
897 |
|
switch( ((x&1)<<1) + (y&1) ) |
898 |
|
{ |
899 |
|
case 0: // pure halfpel position - shouldn't happen during a refinement step |
900 |
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, Reference); |
901 |
|
break; |
902 |
|
|
903 |
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
904 |
|
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
905 |
|
|
906 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
907 |
|
break; |
908 |
|
|
909 |
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
910 |
|
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); |
911 |
|
|
912 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
913 |
|
break; |
914 |
|
|
915 |
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
916 |
|
// bottom left/right) during qpel refinement |
917 |
|
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
918 |
|
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); |
919 |
|
GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); |
920 |
|
|
921 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
922 |
|
break; |
923 |
|
} |
924 |
|
|
925 |
|
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
926 |
|
sad += (data->lambda8 * d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode) * (sad+NEIGH_8X8_BIAS))/100; |
927 |
|
|
928 |
|
if (sad < *(data->iMinSAD)) { |
929 |
|
*(data->iMinSAD) = sad; |
930 |
|
data->currentQMV->x = x; data->currentQMV->y = y; |
931 |
|
*dir = Direction; } |
932 |
|
} |
933 |
|
|
934 |
|
/* CHECK_CANDIATE FUNCTIONS END */ |
935 |
|
|
936 |
/* MAINSEARCH FUNCTIONS START */ |
/* MAINSEARCH FUNCTIONS START */ |
937 |
|
|
1099 |
CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); |
CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); |
1100 |
} |
} |
1101 |
|
|
1102 |
|
|
1103 |
|
static void |
1104 |
|
QuarterpelRefine(const SearchData * const data) |
1105 |
|
{ |
1106 |
|
/* Perform quarter pixel refinement*/ |
1107 |
|
|
1108 |
|
VECTOR backupMV = *(data->currentQMV); |
1109 |
|
int iDirection; //not needed |
1110 |
|
|
1111 |
|
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); |
1112 |
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); |
1113 |
|
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); |
1114 |
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0); |
1115 |
|
|
1116 |
|
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); |
1117 |
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); |
1118 |
|
|
1119 |
|
CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); |
1120 |
|
CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); |
1121 |
|
|
1122 |
|
} |
1123 |
|
|
1124 |
static __inline int |
static __inline int |
1125 |
SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
1126 |
const int x, const int y, |
const int x, const int y, |
1134 |
uint32_t sadC = sad8(current->u + x*8 + y*(iEdgedWidth/2)*8, |
uint32_t sadC = sad8(current->u + x*8 + y*(iEdgedWidth/2)*8, |
1135 |
reference->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); |
reference->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); |
1136 |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
1137 |
sadC += sad8(current->v + x*8 + y*(iEdgedWidth/2)*8, |
sadC += sad8(current->v + (x + y*(iEdgedWidth/2))*8, |
1138 |
reference->v + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); |
reference->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); |
1139 |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
1140 |
|
|
1141 |
return 1; |
return 1; |
1145 |
SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
1146 |
{ |
{ |
1147 |
pMB->mode = MODE_NOT_CODED; |
pMB->mode = MODE_NOT_CODED; |
1148 |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = pMB->mv16.x = 0; |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
1149 |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = pMB->mv16.y = 0; |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; |
1150 |
|
|
1151 |
|
pMB->qmvs[0].x = pMB->qmvs[1].x = pMB->qmvs[2].x = pMB->qmvs[3].x = 0; |
1152 |
|
pMB->qmvs[0].y = pMB->qmvs[1].y = pMB->qmvs[2].y = pMB->qmvs[3].y = 0; |
1153 |
|
|
1154 |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
1155 |
} |
} |
1156 |
|
|
1171 |
|
|
1172 |
uint32_t x, y; |
uint32_t x, y; |
1173 |
uint32_t iIntra = 0; |
uint32_t iIntra = 0; |
1174 |
int32_t InterBias; |
int32_t InterBias, quant = current->quant, sad00; |
1175 |
|
uint8_t *qimage; |
1176 |
|
|
1177 |
|
// some pre-initialized thingies for SearchP |
1178 |
|
int32_t temp[5]; |
1179 |
|
VECTOR currentMV[5]; |
1180 |
|
VECTOR currentQMV[5]; |
1181 |
|
int32_t iMinSAD[5]; |
1182 |
|
SearchData Data; |
1183 |
|
Data.iEdgedWidth = pParam->edged_width; |
1184 |
|
Data.currentMV = currentMV; |
1185 |
|
Data.currentQMV = currentQMV; |
1186 |
|
Data.iMinSAD = iMinSAD; |
1187 |
|
Data.temp = temp; |
1188 |
|
Data.iFcode = current->fcode; |
1189 |
|
Data.rounding = pParam->m_rounding_type; |
1190 |
|
Data.qpel = pParam->m_quarterpel; |
1191 |
|
Data.chroma = current->global_flags & XVID_ME_COLOUR; |
1192 |
|
|
1193 |
|
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
1194 |
|
return 1; // allocate some mem for qpel interpolated blocks |
1195 |
|
// somehow this is dirty since I think we shouldn't use malloc outside |
1196 |
|
// encoder_create() - so please fix me! |
1197 |
|
Data.RefQ = qimage; |
1198 |
if (sadInit) (*sadInit) (); |
if (sadInit) (*sadInit) (); |
1199 |
|
|
1200 |
for (y = 0; y < pParam->mb_height; y++) { |
for (y = 0; y < pParam->mb_height; y++) { |
1201 |
for (x = 0; x < pParam->mb_width; x++) { |
for (x = 0; x < pParam->mb_width; x++) { |
|
|
|
1202 |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
1203 |
int32_t sad00 = pMB->sad16 |
|
1204 |
|
pMB->sad16 |
1205 |
= sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, |
= sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, |
1206 |
pRef->y + (x + y * pParam->edged_width) * 16, |
pRef->y + (x + y * pParam->edged_width) * 16, |
1207 |
pParam->edged_width, pMB->sad8 ); |
pParam->edged_width, pMB->sad8 ); |
1208 |
|
|
1209 |
|
if (Data.chroma) { |
1210 |
|
pMB->sad16 += sad8(pCurrent->u + x*8 + y*(pParam->edged_width/2)*8, |
1211 |
|
pRef->u + x*8 + y*(pParam->edged_width/2)*8, pParam->edged_width/2); |
1212 |
|
|
1213 |
|
pMB->sad16 += sad8(pCurrent->v + (x + y*(pParam->edged_width/2))*8, |
1214 |
|
pRef->v + (x + y*(pParam->edged_width/2))*8, pParam->edged_width/2); |
1215 |
|
} |
1216 |
|
|
1217 |
|
sad00 = pMB->sad16; //if no gmc; else sad00 = (..) |
1218 |
|
|
1219 |
if (!(current->global_flags & XVID_LUMIMASKING)) { |
if (!(current->global_flags & XVID_LUMIMASKING)) { |
1220 |
pMB->dquant = NO_CHANGE; |
pMB->dquant = NO_CHANGE; |
1221 |
pMB->quant = current->quant; } |
pMB->quant = current->quant; |
1222 |
|
} else { |
1223 |
|
if (pMB->dquant != NO_CHANGE) { |
1224 |
|
quant += DQtab[pMB->dquant]; |
1225 |
|
if (quant > 31) quant = 31; |
1226 |
|
else if (quant < 1) quant = 1; |
1227 |
|
} |
1228 |
|
pMB->quant = quant; |
1229 |
|
} |
1230 |
|
|
1231 |
//initial skip decision |
//initial skip decision |
1232 |
|
/* no early skip for GMC (global vector = skip vector is unknown!) */ |
1233 |
if ((pMB->dquant == NO_CHANGE) && (sad00 <= MAX_SAD00_FOR_SKIP * pMB->quant) |
if (current->coding_type == P_VOP) { /* no fast SKIP for S(GMC)-VOPs */ |
1234 |
&& (SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) ) { |
if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH) |
1235 |
if (pMB->sad16 < pMB->quant * INITIAL_SKIP_THRESH) { |
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) { |
1236 |
SkipMacroblockP(pMB, sad00); |
SkipMacroblockP(pMB, sad00); |
1237 |
continue; |
continue; |
1238 |
} |
} |
1239 |
} else sad00 = 256*4096; // skip not allowed - for final skip decision |
} |
1240 |
|
|
1241 |
SearchP(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
1242 |
y, current->motion_flags, pMB->quant, |
y, current->motion_flags, pMB->quant, |
1243 |
current->fcode, pParam, pMBs, reference->mbs, |
&Data, pParam, pMBs, reference->mbs, |
1244 |
current->global_flags & XVID_INTER4V, pMB); |
current->global_flags & XVID_INTER4V, pMB); |
1245 |
|
|
1246 |
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
1247 |
if (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
if (current->coding_type == P_VOP) { |
1248 |
if ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) |
if ( (pMB->dquant == NO_CHANGE) && (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
1249 |
{ SkipMacroblockP(pMB, sad00); continue; } |
&& ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) ) |
1250 |
|
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) { |
1251 |
|
SkipMacroblockP(pMB, sad00); |
1252 |
|
continue; |
1253 |
|
} |
1254 |
|
} |
1255 |
|
|
1256 |
/* finally, intra decision */ |
/* finally, intra decision */ |
1257 |
|
|
1258 |
InterBias = MV16_INTER_BIAS; |
InterBias = MV16_INTER_BIAS; |
1259 |
if (pMB->quant > 8) InterBias += 50 * (pMB->quant - 8); // to make high quants work |
if (pMB->quant > 8) InterBias += 100 * (pMB->quant - 8); // to make high quants work |
1260 |
if (y != 0) |
if (y != 0) |
1261 |
if ((pMB - pParam->mb_width)->mode == MODE_INTER ) InterBias -= 50; |
if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; |
1262 |
if (x != 0) |
if (x != 0) |
1263 |
if ((pMB - 1)->mode == MODE_INTER ) InterBias -= 50; |
if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; |
1264 |
|
|
1265 |
|
if (Data.chroma) InterBias += 50; // to compensate bigger SAD |
1266 |
|
|
1267 |
if (InterBias < pMB->sad16) { |
if (InterBias < pMB->sad16) { |
1268 |
const int32_t deviation = |
const int32_t deviation = |
1270 |
pParam->edged_width); |
pParam->edged_width); |
1271 |
|
|
1272 |
if (deviation < (pMB->sad16 - InterBias)) { |
if (deviation < (pMB->sad16 - InterBias)) { |
1273 |
if (++iIntra >= iLimit) return 1; |
if (++iIntra >= iLimit) { free(qimage); return 1; } |
1274 |
pMB->mode = MODE_INTRA; |
pMB->mode = MODE_INTRA; |
1275 |
pMB->mv16 = pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = |
1276 |
pMB->mvs[3] = zeroMV; |
pMB->mvs[3] = zeroMV; |
1277 |
|
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = |
1278 |
|
pMB->qmvs[3] = zeroMV; |
1279 |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = |
1280 |
pMB->sad8[3] = 0; |
pMB->sad8[3] = 0; |
1281 |
} |
} |
1282 |
} |
} |
1283 |
} |
} |
1284 |
} |
} |
1285 |
|
free(qimage); |
1286 |
|
|
1287 |
|
if (current->coding_type == S_VOP) /* first GMC step only for S(GMC)-VOPs */ |
1288 |
|
current->GMC_MV = GlobalMotionEst( pMBs, pParam, current->fcode ); |
1289 |
|
else |
1290 |
|
current->GMC_MV = zeroMV; |
1291 |
|
|
1292 |
return 0; |
return 0; |
1293 |
} |
} |
1294 |
|
|
1298 |
static __inline int |
static __inline int |
1299 |
make_mask(const VECTOR * const pmv, const int i) |
make_mask(const VECTOR * const pmv, const int i) |
1300 |
{ |
{ |
1301 |
int mask = 0xFF, j; |
int mask = 255, j; |
1302 |
for (j = 0; j < i; j++) { |
for (j = 0; j < i; j++) { |
1303 |
if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already |
if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already |
1304 |
if (pmv[i].x == pmv[j].x) { |
if (pmv[i].x == pmv[j].x) { |
1322 |
|
|
1323 |
if ( (y != 0) && (x != (iWcount-1)) ) { // [5] top-right neighbour |
if ( (y != 0) && (x != (iWcount-1)) ) { // [5] top-right neighbour |
1324 |
pmv[5].x = EVEN(pmv[3].x); |
pmv[5].x = EVEN(pmv[3].x); |
1325 |
pmv[5].y = EVEN(pmv[3].y); } |
pmv[5].y = EVEN(pmv[3].y); |
1326 |
else pmv[5].x = pmv[5].y = 0; |
} else pmv[5].x = pmv[5].y = 0; |
1327 |
|
|
1328 |
if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour |
if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour |
1329 |
else pmv[3].x = pmv[3].y = 0; |
else pmv[3].x = pmv[3].y = 0; |
1341 |
|
|
1342 |
if ((x != iWcount-1) && (y != iHcount-1)) { |
if ((x != iWcount-1) && (y != iHcount-1)) { |
1343 |
pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame |
pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame |
1344 |
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); } |
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
1345 |
else pmv[6].x = pmv[6].y = 0; |
} else pmv[6].x = pmv[6].y = 0; |
1346 |
} |
} |
1347 |
|
|
1348 |
static void |
static void |
1349 |
SearchP(const uint8_t * const pRef, |
SearchP(const IMAGE * const pRef, |
1350 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1351 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
1352 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
1355 |
const int y, |
const int y, |
1356 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
1357 |
const uint32_t iQuant, |
const uint32_t iQuant, |
1358 |
const uint32_t iFcode, |
SearchData * const Data, |
1359 |
const MBParam * const pParam, |
const MBParam * const pParam, |
1360 |
const MACROBLOCK * const pMBs, |
const MACROBLOCK * const pMBs, |
1361 |
const MACROBLOCK * const prevMBs, |
const MACROBLOCK * const prevMBs, |
1363 |
MACROBLOCK * const pMB) |
MACROBLOCK * const pMB) |
1364 |
{ |
{ |
1365 |
|
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
|
|
|
1366 |
int i, iDirection = 255, mask, threshA; |
int i, iDirection = 255, mask, threshA; |
1367 |
int32_t temp[5]; |
VECTOR pmv[7]; |
|
VECTOR currentMV[5], pmv[7]; |
|
|
int32_t psad[4], iMinSAD[5]; |
|
|
MainSearchFunc * MainSearchPtr; |
|
|
SearchData Data; |
|
1368 |
|
|
1369 |
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, psad); //has to be changed to get_pmv(2)() |
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); //has to be changed to get_pmv(2)() |
1370 |
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 16, |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1371 |
pParam->width, pParam->height, iFcode); |
pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); |
1372 |
|
|
1373 |
Data.predMV = pmv[0]; |
Data->predMV = pmv[0]; |
1374 |
Data.Cur = pCur->y + (x + y * iEdgedWidth) * 16; |
|
1375 |
Data.iEdgedWidth = iEdgedWidth; |
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16; |
1376 |
Data.currentMV = currentMV; |
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
1377 |
Data.iMinSAD = iMinSAD; |
Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
1378 |
Data.Ref = pRef + (x + iEdgedWidth*y)*16; |
|
1379 |
Data.RefH = pRefH + (x + iEdgedWidth*y) * 16; |
Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16; |
1380 |
Data.RefV = pRefV + (x + iEdgedWidth*y) * 16; |
Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16; |
1381 |
Data.RefHV = pRefHV + (x + iEdgedWidth*y) * 16; |
Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16; |
1382 |
Data.temp = temp; |
Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16; |
1383 |
|
Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
1384 |
|
Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
1385 |
|
|
1386 |
Data.iQuant = iQuant; |
Data->lambda16 = lambda_vec16[iQuant]; |
1387 |
Data.iFcode = iFcode; |
Data->lambda8 = lambda_vec8[iQuant]; |
1388 |
|
|
1389 |
if (!(MotionFlags & PMV_HALFPEL16)) { |
if (!(MotionFlags & PMV_HALFPEL16)) { |
1390 |
Data.min_dx = EVEN(Data.min_dx); |
Data->min_dx = EVEN(Data->min_dx); |
1391 |
Data.max_dx = EVEN(Data.max_dx); |
Data->max_dx = EVEN(Data->max_dx); |
1392 |
Data.min_dy = EVEN(Data.min_dy); |
Data->min_dy = EVEN(Data->min_dy); |
1393 |
Data.max_dy = EVEN(Data.max_dy); } |
Data->max_dy = EVEN(Data->max_dy); } |
|
|
|
|
for(i = 0; i < 5; i++) currentMV[i].x = currentMV[i].y = 0; |
|
|
|
|
|
i = d_mv_bits(pmv[0].x, pmv[0].y, iFcode); |
|
|
|
|
|
iMinSAD[0] = pMB->sad16 + lambda_vec16[iQuant] * i; |
|
|
iMinSAD[1] = pMB->sad8[0] + lambda_vec8[iQuant] * i; |
|
|
iMinSAD[2] = pMB->sad8[1]; |
|
|
iMinSAD[3] = pMB->sad8[2]; |
|
|
iMinSAD[4] = pMB->sad8[3]; |
|
1394 |
|
|
1395 |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
1396 |
|
|
1397 |
|
for(i = 0; i < 5; i++) |
1398 |
|
Data->currentMV[i].x = Data->currentMV[i].y = 0; |
1399 |
|
|
1400 |
|
if (pParam->m_quarterpel) { |
1401 |
|
Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
1402 |
|
i = d_mv_bits(Data->predQMV.x, Data->predQMV.y, Data->iFcode); |
1403 |
|
} else i = d_mv_bits(Data->predMV.x, Data->predMV.y, Data->iFcode); |
1404 |
|
|
1405 |
|
Data->iMinSAD[0] = pMB->sad16 + (Data->lambda16 * i * pMB->sad16)/1000; |
1406 |
|
Data->iMinSAD[1] = pMB->sad8[0] + (Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS))/100; |
1407 |
|
Data->iMinSAD[2] = pMB->sad8[1]; |
1408 |
|
Data->iMinSAD[3] = pMB->sad8[2]; |
1409 |
|
Data->iMinSAD[4] = pMB->sad8[3]; |
1410 |
|
|
1411 |
if ((x == 0) && (y == 0)) threshA = 512; |
if ((x == 0) && (y == 0)) threshA = 512; |
1412 |
else { |
else { |
1413 |
threshA = psad[0] + 20; |
threshA = Data->temp[0]; // that's when we keep this SAD atm |
1414 |
if (threshA < 512) threshA = 512; |
if (threshA < 512) threshA = 512; |
1415 |
if (threshA > 1024) threshA = 1024; } |
if (threshA > 1024) threshA = 1024; } |
1416 |
|
|
1417 |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
1418 |
prevMBs + x + y * pParam->mb_width); |
prevMBs + x + y * pParam->mb_width); |
1419 |
|
|
1420 |
if (inter4v) CheckCandidate = CheckCandidate16; |
if (inter4v || pParam->m_quarterpel || Data->chroma) CheckCandidate = CheckCandidate16; |
1421 |
else CheckCandidate = CheckCandidate16no4v; |
else CheckCandidate = CheckCandidate16no4v; |
1422 |
|
|
1423 |
/* main loop. checking all predictions */ |
/* main loop. checking all predictions */ |
1424 |
|
|
1425 |
for (i = 1; i < 7; i++) { |
for (i = 1; i < 7; i++) { |
1426 |
if (!(mask = make_mask(pmv, i)) ) continue; |
if (!(mask = make_mask(pmv, i)) ) continue; |
1427 |
CheckCandidate16(pmv[i].x, pmv[i].y, mask, &iDirection, &Data); |
(*CheckCandidate)(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1428 |
if (iMinSAD[0] < threshA) break; |
if (Data->iMinSAD[0] <= threshA) break; |
1429 |
} |
} |
1430 |
|
|
1431 |
if ((iMinSAD[0] <= threshA) || |
if ((Data->iMinSAD[0] <= threshA) || |
1432 |
(MVequal(currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
1433 |
(iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { |
(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { |
1434 |
inter4v = 0; |
inter4v = 0; |
1435 |
if (MotionFlags & PMV_QUICKSTOP16) goto PMVfast16_Terminate_without_Refine; |
} else { |
|
if (MotionFlags & PMV_EARLYSTOP16) { |
|
|
CheckCandidate = CheckCandidate16no4v; // I sure hope it's faster |
|
|
goto PMVfast16_Terminate_with_Refine; |
|
|
} |
|
|
} |
|
1436 |
|
|
1437 |
if (MotionFlags & PMV_USESQUARES16) |
MainSearchFunc * MainSearchPtr; |
1438 |
MainSearchPtr = SquareSearch; |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1439 |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
|
MainSearchPtr = AdvDiamondSearch; |
|
1440 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1441 |
|
|
1442 |
(*MainSearchPtr)(currentMV->x, currentMV->y, &Data, iDirection); |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
1443 |
|
|
1444 |
/* extended search, diamond starting in 0,0 and in prediction. |
/* extended search, diamond starting in 0,0 and in prediction. |
1445 |
note that this search is/might be done in halfpel positions, |
note that this search is/might be done in halfpel positions, |
1447 |
|
|
1448 |
if (MotionFlags & PMV_EXTSEARCH16) { |
if (MotionFlags & PMV_EXTSEARCH16) { |
1449 |
int32_t bSAD; |
int32_t bSAD; |
1450 |
VECTOR startMV = Data.predMV, backupMV = currentMV[0]; |
VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; |
1451 |
if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? |
if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? |
1452 |
startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); |
startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); |
1453 |
if (!(MVequal(startMV, backupMV))) { |
if (!(MVequal(startMV, backupMV))) { |
1454 |
bSAD = iMinSAD[0]; iMinSAD[0] = MV_MAX_ERROR; |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1455 |
|
|
1456 |
CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, &Data); |
(*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); |
1457 |
(*MainSearchPtr)(startMV.x, startMV.y, &Data, 255); |
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
1458 |
if (bSAD < iMinSAD[0]) { |
if (bSAD < Data->iMinSAD[0]) { |
1459 |
currentMV[0] = backupMV; |
Data->currentMV[0] = backupMV; |
1460 |
iMinSAD[0] = bSAD; } |
Data->iMinSAD[0] = bSAD; } |
1461 |
} |
} |
1462 |
|
|
1463 |
backupMV = currentMV[0]; |
backupMV = Data->currentMV[0]; |
1464 |
if (MotionFlags & PMV_HALFPELREFINE16) startMV.x = startMV.y = 1; |
if (MotionFlags & PMV_HALFPELREFINE16) startMV.x = startMV.y = 1; |
1465 |
else startMV.x = startMV.y = 0; |
else startMV.x = startMV.y = 0; |
1466 |
if (!(MVequal(startMV, backupMV))) { |
if (!(MVequal(startMV, backupMV))) { |
1467 |
bSAD = iMinSAD[0]; iMinSAD[0] = MV_MAX_ERROR; |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1468 |
|
|
1469 |
|
(*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); |
1470 |
|
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
1471 |
|
if (bSAD < Data->iMinSAD[0]) { |
1472 |
|
Data->currentMV[0] = backupMV; |
1473 |
|
Data->iMinSAD[0] = bSAD; } |
1474 |
|
} |
1475 |
|
} |
1476 |
|
} |
1477 |
|
|
1478 |
|
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data); |
1479 |
|
|
1480 |
CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, &Data); |
for(i = 0; i < 5; i++) { |
1481 |
(*MainSearchPtr)(startMV.x, startMV.y, &Data, 255); |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors |
1482 |
if (bSAD < iMinSAD[0]) { |
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
|
currentMV[0] = backupMV; |
|
|
iMinSAD[0] = bSAD; } |
|
1483 |
} |
} |
1484 |
|
|
1485 |
|
if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { |
1486 |
|
|
1487 |
|
CheckCandidate = CheckCandidate16_qpel; |
1488 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1489 |
|
pParam->width, pParam->height, Data->iFcode, 0); |
1490 |
|
|
1491 |
|
QuarterpelRefine(Data); |
1492 |
} |
} |
1493 |
|
|
1494 |
PMVfast16_Terminate_with_Refine: |
if (Data->iMinSAD[0] < (int32_t)iQuant * 30 ) inter4v = 0; |
1495 |
|
if (inter4v) { |
1496 |
|
SearchData Data8; |
1497 |
|
Data8.iFcode = Data->iFcode; |
1498 |
|
Data8.lambda8 = Data->lambda8; |
1499 |
|
Data8.iEdgedWidth = Data->iEdgedWidth; |
1500 |
|
Data8.RefQ = Data->RefQ; |
1501 |
|
Data8.qpel = Data->qpel; |
1502 |
|
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
1503 |
|
Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); |
1504 |
|
Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); |
1505 |
|
Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); |
1506 |
|
|
1507 |
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(&Data); |
if (Data->chroma) { |
1508 |
|
int sum, dx, dy; |
1509 |
|
|
1510 |
PMVfast16_Terminate_without_Refine: |
if(pParam->m_quarterpel) { |
1511 |
|
sum = pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2; |
1512 |
|
} else sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
1513 |
|
dy = (sum >> 3) + roundtab_76[sum & 0xf]; |
1514 |
|
|
1515 |
if (inter4v) |
if(pParam->m_quarterpel) { |
1516 |
for(i = 0; i < 4; i++) |
sum = pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2; |
1517 |
Search8(&Data, 2*x+(i&1), 2*y+(i>>1), MotionFlags, pParam, pMB, pMBs, i); |
} else sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
1518 |
|
dx = (sum >> 3) + roundtab_76[sum & 0xf]; |
1519 |
|
|
1520 |
|
Data->iMinSAD[1] += ChromaSAD(dx, dy, Data); |
1521 |
|
} |
1522 |
|
} |
1523 |
|
|
1524 |
if (!(inter4v) || |
if (!(inter4v) || |
1525 |
(iMinSAD[0] < iMinSAD[1] + iMinSAD[2] + iMinSAD[3] + iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
1526 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
1527 |
// INTER MODE |
// INTER MODE |
1528 |
pMB->mode = MODE_INTER; |
pMB->mode = MODE_INTER; |
1529 |
pMB->mv16 = pMB->mvs[0] = pMB->mvs[1] |
pMB->mvs[0] = pMB->mvs[1] |
1530 |
= pMB->mvs[2] = pMB->mvs[3] = currentMV[0]; |
= pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
1531 |
|
|
1532 |
|
pMB->qmvs[0] = pMB->qmvs[1] |
1533 |
|
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
1534 |
|
|
1535 |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
1536 |
pMB->sad8[2] = pMB->sad8[3] = iMinSAD[0]; |
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
1537 |
|
|
1538 |
pMB->pmvs[0].x = currentMV[0].x - Data.predMV.x; |
if(pParam->m_quarterpel) { |
1539 |
pMB->pmvs[0].y = currentMV[0].y - Data.predMV.y; |
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predQMV.x; |
1540 |
|
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predQMV.y; |
1541 |
|
} else { |
1542 |
|
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
1543 |
|
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
1544 |
|
} |
1545 |
} else { |
} else { |
1546 |
// INTER4V MODE; all other things are already set in Search8 |
// INTER4V MODE; all other things are already set in Search8 |
1547 |
pMB->mode = MODE_INTER4V; |
pMB->mode = MODE_INTER4V; |
1548 |
pMB->sad16 = iMinSAD[1] + iMinSAD[2] + iMinSAD[3] + iMinSAD[4] + IMV16X16 * iQuant; |
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + |
1549 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant; |
1550 |
} |
} |
|
|
|
1551 |
} |
} |
1552 |
|
|
1553 |
static void |
static void |
1557 |
const MBParam * const pParam, |
const MBParam * const pParam, |
1558 |
MACROBLOCK * const pMB, |
MACROBLOCK * const pMB, |
1559 |
const MACROBLOCK * const pMBs, |
const MACROBLOCK * const pMBs, |
1560 |
const int block) |
const int block, |
1561 |
|
SearchData * const Data) |
1562 |
{ |
{ |
1563 |
SearchData Data; |
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
1564 |
|
Data->currentMV = OldData->currentMV + 1 + block; |
1565 |
Data.predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
Data->currentQMV = OldData->currentQMV + 1 + block; |
1566 |
Data.iMinSAD = OldData->iMinSAD + 1 + block; |
|
1567 |
Data.currentMV = OldData->currentMV+1+block; |
if(pParam->m_quarterpel) { |
1568 |
Data.iFcode = OldData->iFcode; |
Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
1569 |
Data.iQuant = OldData->iQuant; |
if (block != 0) *(Data->iMinSAD) += (Data->lambda8 * |
1570 |
|
d_mv_bits( Data->currentQMV->x - Data->predQMV.x, |
1571 |
if (block != 0) |
Data->currentQMV->y - Data->predQMV.y, |
1572 |
*(Data.iMinSAD) += lambda_vec8[Data.iQuant] * |
Data->iFcode) * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100; |
1573 |
d_mv_bits( Data.currentMV->x - Data.predMV.x, |
} else { |
1574 |
Data.currentMV->y - Data.predMV.y, |
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
1575 |
Data.iFcode); |
if (block != 0) *(Data->iMinSAD) += (Data->lambda8 * |
1576 |
|
d_mv_bits( Data->currentMV->x - Data->predMV.x, |
1577 |
|
Data->currentMV->y - Data->predMV.y, |
1578 |
|
Data->iFcode) * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100; |
1579 |
|
} |
1580 |
|
|
1581 |
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { |
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { |
1582 |
|
|
1583 |
Data.Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1584 |
Data.RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1585 |
Data.RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1586 |
Data.RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
|
|
|
Data.iEdgedWidth = pParam->edged_width; |
|
|
|
|
|
Data.Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
1587 |
|
|
1588 |
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 8, |
Data->Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
pParam->width, pParam->height, OldData->iFcode); |
|
1589 |
|
|
1590 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
1591 |
|
pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel); |
1592 |
CheckCandidate = CheckCandidate8; |
CheckCandidate = CheckCandidate8; |
1593 |
|
|
1594 |
if (MotionFlags & PMV_EXTSEARCH8) { |
if (MotionFlags & PMV_EXTSEARCH8) { |
1595 |
|
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
1596 |
|
|
1597 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1598 |
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
1599 |
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
1600 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1601 |
|
|
1602 |
(*MainSearchPtr)(Data.currentMV->x, Data.currentMV->y, &Data, 255); } |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1603 |
|
|
1604 |
|
if(*(Data->iMinSAD) < temp_sad) { |
1605 |
|
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
1606 |
|
Data->currentQMV->y = 2 * Data->currentMV->y; |
1607 |
|
} |
1608 |
|
} |
1609 |
|
|
1610 |
|
if (MotionFlags & PMV_HALFPELREFINE8) { |
1611 |
|
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
1612 |
|
|
1613 |
|
HalfpelRefine(Data); // perform halfpel refine of current best vector |
1614 |
|
|
1615 |
|
if(*(Data->iMinSAD) < temp_sad) { // we have found a better match |
1616 |
|
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
1617 |
|
Data->currentQMV->y = 2 * Data->currentMV->y; |
1618 |
|
} |
1619 |
|
} |
1620 |
|
|
1621 |
|
if(pParam->m_quarterpel) { |
1622 |
|
if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && |
1623 |
|
(MotionFlags & PMV_QUARTERPELREFINE8)) { |
1624 |
|
CheckCandidate = CheckCandidate8_qpel; |
1625 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
1626 |
|
pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel); |
1627 |
|
QuarterpelRefine(Data); |
1628 |
|
} |
1629 |
|
} |
1630 |
|
} |
1631 |
|
|
1632 |
if (MotionFlags & PMV_HALFPELREFINE8) HalfpelRefine(&Data); |
if(pParam->m_quarterpel) { |
1633 |
|
pMB->pmvs[block].x = Data->currentQMV->x - Data->predQMV.x; |
1634 |
|
pMB->pmvs[block].y = Data->currentQMV->y - Data->predQMV.y; |
1635 |
|
} |
1636 |
|
else { |
1637 |
|
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
1638 |
|
pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; |
1639 |
} |
} |
1640 |
|
|
1641 |
pMB->pmvs[block].x = Data.currentMV->x - Data.predMV.x; |
pMB->mvs[block] = *(Data->currentMV); |
1642 |
pMB->pmvs[block].y = Data.currentMV->y - Data.predMV.y; |
pMB->qmvs[block] = *(Data->currentQMV); |
1643 |
pMB->mvs[block] = *(Data.currentMV); |
|
1644 |
pMB->sad8[block] = 4 * (*(Data.iMinSAD)); |
pMB->sad8[block] = 4 * (*Data->iMinSAD); |
1645 |
} |
} |
1646 |
|
|
1647 |
/* B-frames code starts here */ |
/* B-frames code starts here */ |
1669 |
pmv[2] = ChoosePred(pMB, mode_curr); |
pmv[2] = ChoosePred(pMB, mode_curr); |
1670 |
pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); |
pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); |
1671 |
|
|
|
pmv[3].x = pmv[3].y = 0; |
|
1672 |
if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour |
if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour |
1673 |
pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); |
pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); |
1674 |
pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); } |
pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); |
1675 |
|
} else pmv[3].x = pmv[3].y = 0; |
1676 |
|
|
1677 |
if (y != 0) { |
if (y != 0) { |
1678 |
pmv[4] = ChoosePred(pMB-iWcount, mode_curr); |
pmv[4] = ChoosePred(pMB-iWcount, mode_curr); |
1702 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
1703 |
const int x, const int y, |
const int x, const int y, |
1704 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
|
const uint32_t iQuant, |
|
1705 |
const uint32_t iFcode, |
const uint32_t iFcode, |
1706 |
const MBParam * const pParam, |
const MBParam * const pParam, |
1707 |
MACROBLOCK * const pMB, |
MACROBLOCK * const pMB, |
1708 |
const VECTOR * const predMV, |
const VECTOR * const predMV, |
1709 |
int32_t * const best_sad, |
int32_t * const best_sad, |
1710 |
const int32_t mode_current) |
const int32_t mode_current, |
1711 |
|
SearchData * const Data) |
1712 |
{ |
{ |
1713 |
|
|
1714 |
const int32_t iEdgedWidth = pParam->edged_width; |
const int32_t iEdgedWidth = pParam->edged_width; |
1715 |
|
|
1716 |
int i, iDirection, mask; |
int i, iDirection, mask; |
1717 |
VECTOR currentMV, pmv[7]; |
VECTOR pmv[7]; |
1718 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1719 |
int32_t iMinSAD = MV_MAX_ERROR; |
*Data->iMinSAD = MV_MAX_ERROR; |
1720 |
SearchData Data; |
Data->iFcode = iFcode; |
|
|
|
|
Data.iMinSAD = &iMinSAD; |
|
|
Data.Cur = pCur->y + (x + y * iEdgedWidth) * 16; |
|
|
Data.iEdgedWidth = iEdgedWidth; |
|
|
Data.currentMV = ¤tMV; |
|
|
Data.iMinSAD = &iMinSAD; |
|
|
Data.Ref = pRef + (x + y * iEdgedWidth) * 16; |
|
|
Data.RefH = pRefH + (x + y * iEdgedWidth) * 16; |
|
|
Data.RefV = pRefV + (x + y * iEdgedWidth) * 16; |
|
|
Data.RefHV = pRefHV + (x + y * iEdgedWidth) * 16; |
|
|
|
|
|
Data.iQuant = iQuant; |
|
|
Data.iFcode = iFcode; |
|
|
Data.predMV = *predMV; |
|
1721 |
|
|
1722 |
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 16, |
Data->Ref = pRef + (x + y * iEdgedWidth) * 16; |
1723 |
pParam->width, pParam->height, iFcode); |
Data->RefH = pRefH + (x + y * iEdgedWidth) * 16; |
1724 |
|
Data->RefV = pRefV + (x + y * iEdgedWidth) * 16; |
1725 |
|
Data->RefHV = pRefHV + (x + y * iEdgedWidth) * 16; |
1726 |
|
|
1727 |
if (!(MotionFlags & PMV_HALFPEL16)) { |
Data->predMV = *predMV; |
|
Data.min_dx = EVEN(Data.min_dx); |
|
|
Data.max_dx = EVEN(Data.max_dx); |
|
|
Data.min_dy = EVEN(Data.min_dy); |
|
|
Data.max_dy = EVEN(Data.max_dy); } // no-halpel and b-frames. do we need it? |
|
1728 |
|
|
1729 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1730 |
|
pParam->width, pParam->height, iFcode, pParam->m_quarterpel); |
1731 |
|
|
1732 |
pmv[0] = Data.predMV; |
pmv[0] = Data->predMV; |
1733 |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, |
if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } |
1734 |
pMB, mode_current); |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
1735 |
|
|
1736 |
currentMV.x = currentMV.y = 0; |
Data->currentMV->x = Data->currentMV->y = 0; |
1737 |
|
|
1738 |
CheckCandidate = CheckCandidate16no4v; |
CheckCandidate = CheckCandidate16no4v; |
1739 |
|
|
1740 |
// main loop. checking all predictions |
// main loop. checking all predictions |
1741 |
for (i = 0; i < 8; i++) { |
for (i = 0; i < 8; i++) { |
1742 |
if (!(mask = make_mask(pmv, i)) ) continue; |
if (!(mask = make_mask(pmv, i)) ) continue; |
1743 |
CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, &Data); |
CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1744 |
} |
} |
1745 |
|
|
1746 |
if (MotionFlags & PMV_USESQUARES16) |
if (MotionFlags & PMV_USESQUARES16) |
1749 |
MainSearchPtr = AdvDiamondSearch; |
MainSearchPtr = AdvDiamondSearch; |
1750 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1751 |
|
|
1752 |
(*MainSearchPtr)(currentMV.x, currentMV.y, &Data, 255); |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1753 |
|
|
1754 |
|
HalfpelRefine(Data); |
1755 |
|
|
1756 |
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(&Data); |
if (Data->qpel) { |
1757 |
|
Data->currentQMV->x = 2*Data->currentMV->x; |
1758 |
|
Data->currentQMV->y = 2*Data->currentMV->y; |
1759 |
|
CheckCandidate = CheckCandidate16no4v_qpel; |
1760 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1761 |
|
pParam->width, pParam->height, iFcode, pParam->m_quarterpel); |
1762 |
|
QuarterpelRefine(Data); |
1763 |
|
} |
1764 |
|
|
1765 |
// three bits are needed to code backward mode. four for forward |
// three bits are needed to code backward mode. four for forward |
1766 |
// we treat the bits just like they were vector's |
// we treat the bits just like they were vector's |
1767 |
if (mode_current == MODE_FORWARD) iMinSAD += 4 * lambda_vec16[iQuant]; |
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
1768 |
else iMinSAD += 3 * lambda_vec16[iQuant]; |
else *Data->iMinSAD += 3 * Data->lambda16; |
|
|
|
1769 |
|
|
1770 |
if (iMinSAD < *best_sad) { |
if (*Data->iMinSAD < *best_sad) { |
1771 |
*best_sad = iMinSAD; |
*best_sad = *Data->iMinSAD; |
1772 |
pMB->mode = mode_current; |
pMB->mode = mode_current; |
1773 |
pMB->pmvs[0].x = currentMV.x - predMV->x; |
if (Data->qpel) { |
1774 |
pMB->pmvs[0].y = currentMV.y - predMV->y; |
pMB->pmvs[0].x = Data->currentQMV->x - predMV->x; |
1775 |
if (mode_current == MODE_FORWARD) pMB->mvs[0] = currentMV; |
pMB->pmvs[0].y = Data->currentQMV->y - predMV->y; |
1776 |
else pMB->b_mvs[0] = currentMV; |
if (mode_current == MODE_FORWARD) |
1777 |
|
pMB->qmvs[0] = *Data->currentQMV; |
1778 |
|
else |
1779 |
|
pMB->b_qmvs[0] = *Data->currentQMV; |
1780 |
|
} else { |
1781 |
|
pMB->pmvs[0].x = Data->currentMV->x - predMV->x; |
1782 |
|
pMB->pmvs[0].y = Data->currentMV->y - predMV->y; |
1783 |
|
} |
1784 |
|
if (mode_current == MODE_FORWARD) |
1785 |
|
pMB->mvs[0] = *(Data->currentMV+2) = *Data->currentMV; |
1786 |
|
else |
1787 |
|
pMB->b_mvs[0] = *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search |
1788 |
|
|
1789 |
} |
} |
1790 |
|
|
1791 |
} |
} |
1792 |
|
|
1793 |
static int32_t |
static int32_t |
1794 |
SearchDirect(const uint8_t * const f_Ref, |
SearchDirect(const IMAGE * const f_Ref, |
1795 |
const uint8_t * const f_RefH, |
const uint8_t * const f_RefH, |
1796 |
const uint8_t * const f_RefV, |
const uint8_t * const f_RefV, |
1797 |
const uint8_t * const f_RefHV, |
const uint8_t * const f_RefHV, |
1798 |
const uint8_t * const b_Ref, |
const IMAGE * const b_Ref, |
1799 |
const uint8_t * const b_RefH, |
const uint8_t * const b_RefH, |
1800 |
const uint8_t * const b_RefV, |
const uint8_t * const b_RefV, |
1801 |
const uint8_t * const b_RefHV, |
const uint8_t * const b_RefHV, |
1802 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
1803 |
const int x, const int y, |
const int x, const int y, |
1804 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
|
const uint32_t iQuant, |
|
1805 |
const int32_t TRB, const int32_t TRD, |
const int32_t TRB, const int32_t TRD, |
1806 |
const MBParam * const pParam, |
const MBParam * const pParam, |
1807 |
MACROBLOCK * const pMB, |
MACROBLOCK * const pMB, |
1808 |
const MACROBLOCK * const b_mb, |
const MACROBLOCK * const b_mb, |
1809 |
int32_t * const best_sad) |
int32_t * const best_sad, |
1810 |
|
SearchData * const Data) |
1811 |
|
|
1812 |
{ |
{ |
1813 |
const uint32_t iEdgedWidth = pParam->edged_width; |
int32_t skip_sad; |
|
int32_t iMinSAD = 0, skip_sad; |
|
1814 |
int k; |
int k; |
1815 |
VECTOR currentMV; |
|
1816 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
|
SearchData Data; |
|
1817 |
|
|
1818 |
Data.iMinSAD = &iMinSAD; |
*Data->iMinSAD = 256*4096; |
|
Data.Cur = pCur->y + x * 16 + y * 16 * iEdgedWidth; |
|
|
Data.iEdgedWidth = iEdgedWidth; |
|
|
Data.currentMV = ¤tMV; |
|
|
Data.iQuant = iQuant; |
|
|
Data.referencemv = b_mb->mvs; |
|
1819 |
|
|
1820 |
Data.Ref= f_Ref + (x + iEdgedWidth*y) * 16; |
Data->Ref = f_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
1821 |
Data.RefH = f_RefH + (x + iEdgedWidth*y) * 16; |
Data->RefH = f_RefH + (x + Data->iEdgedWidth*y) * 16; |
1822 |
Data.RefV = f_RefV + (x + iEdgedWidth*y) * 16; |
Data->RefV = f_RefV + (x + Data->iEdgedWidth*y) * 16; |
1823 |
Data.RefHV = f_RefHV + (x + iEdgedWidth*y) * 16; |
Data->RefHV = f_RefHV + (x + Data->iEdgedWidth*y) * 16; |
1824 |
Data.bRef = b_Ref + (x + iEdgedWidth*y) * 16; |
Data->bRef = b_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
1825 |
Data.bRefH = b_RefH + (x + iEdgedWidth*y) * 16; |
Data->bRefH = b_RefH + (x + Data->iEdgedWidth*y) * 16; |
1826 |
Data.bRefV = b_RefV + (x + iEdgedWidth*y) * 16; |
Data->bRefV = b_RefV + (x + Data->iEdgedWidth*y) * 16; |
1827 |
Data.bRefHV = b_RefHV + (x + iEdgedWidth*y) * 16; |
Data->bRefHV = b_RefHV + (x + Data->iEdgedWidth*y) * 16; |
1828 |
/* |
|
1829 |
//What we do here is a complicated version of CheckCandidateDirect(0,0); |
Data->max_dx = 2 * pParam->width - 2 * (x) * 16; |
1830 |
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 16, pParam->width, pParam->height, 19); |
Data->max_dy = 2 * pParam->height - 2 * (y) * 16; |
1831 |
|
Data->min_dx = -(2 * 16 + 2 * (x) * 16); |
1832 |
*/ |
Data->min_dy = -(2 * 16 + 2 * (y) * 16); |
1833 |
Data.max_dx = 2 * pParam->width - 2 * (x) * 16; |
if (Data->qpel) { //we measure in qpixels |
1834 |
Data.max_dy = 2 * pParam->height - 2 * (y) * 16; |
Data->max_dx *= 2; |
1835 |
Data.min_dx = -(2 * 16 + 2 * (x) * 16); |
Data->max_dy *= 2; |
1836 |
Data.min_dy = -(2 * 16 + 2 * (y) * 16); |
Data->min_dx *= 2; |
1837 |
|
Data->min_dy *= 2; |
1838 |
|
Data->referencemv = b_mb->qmvs; |
1839 |
|
} else Data->referencemv = b_mb->mvs; |
1840 |
|
|
1841 |
for (k = 0; k < 4; k++) { |
for (k = 0; k < 4; k++) { |
1842 |
pMB->mvs[k].x = Data.directmvF[k].x = ((TRB * Data.referencemv[k].x) / TRD); |
pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); |
1843 |
pMB->b_mvs[k].x = Data.directmvB[k].x = ((TRB - TRD) * Data.referencemv[k].x) / TRD; |
pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD; |
1844 |
pMB->mvs[k].y = Data.directmvF[k].y = ((TRB * Data.referencemv[k].y) / TRD); |
pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); |
1845 |
pMB->b_mvs[k].y = Data.directmvB[k].y = ((TRB - TRD) * Data.referencemv[k].y) / TRD; |
pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; |
|
|
|
|
if (( pMB->mvs[k].x > Data.max_dx ) || ( pMB->mvs[k].x < Data.min_dx ) |
|
|
|| ( pMB->mvs[k].y > Data.max_dy ) || ( pMB->mvs[k].y < Data.min_dy ) |
|
|
|| ( pMB->b_mvs[k].x > Data.max_dx ) || ( pMB->b_mvs[k].x < Data.min_dx ) |
|
|
|| ( pMB->b_mvs[k].y > Data.max_dy ) || ( pMB->b_mvs[k].y < Data.min_dy )) { |
|
|
/* |
|
|
fprintf(debug, "\nERROR - out of range : vector %d,%d and %d,%d\n", pMB->mvs[k].x, pMB->mvs[k].y,pMB->b_mvs[k].x,pMB->b_mvs[k].y ); |
|
|
fprintf(debug, " range is x: %d..%d y: %d..%d \n", Data.min_dx,Data.max_dx,Data.min_dy,Data.max_dy); |
|
|
fprintf(debug,"macroblock %d, %d \n", x, y); |
|
|
fprintf(debug, "direct MV is %d,%d \n", directmv[k].x, directmv[k].y); |
|
|
*/ |
|
|
*best_sad = 256*4096; // in that case, we won't use direct mode |
|
|
pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV" |
|
|
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; /* because backwards and interpol might rely on this */ |
|
|
return 0; } |
|
1846 |
|
|
1847 |
|
if ( ( pMB->b_mvs[k].x > Data->max_dx ) || ( pMB->b_mvs[k].x < Data->min_dx ) |
1848 |
|
|| ( pMB->b_mvs[k].y > Data->max_dy ) || ( pMB->b_mvs[k].y < Data->min_dy )) { |
1849 |
|
|
1850 |
|
*best_sad = 256*4096; // in that case, we won't use direct mode |
1851 |
|
pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV" |
1852 |
|
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
1853 |
|
return 0; |
1854 |
|
} |
1855 |
if (b_mb->mode != MODE_INTER4V) { |
if (b_mb->mode != MODE_INTER4V) { |
1856 |
iMinSAD = sad16bi(Data.Cur, |
pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; |
1857 |
get_ref_mv(f_Ref, f_RefH, f_RefV, f_RefHV, |
pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0]; |
1858 |
x, y, 16, &pMB->mvs[0], iEdgedWidth), |
Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0]; |
1859 |
get_ref_mv(b_Ref, b_RefH, b_RefV, b_RefHV, |
Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0]; |
1860 |
x, y, 16, &pMB->b_mvs[0], iEdgedWidth), iEdgedWidth); |
break; |
1861 |
|
} |
1862 |
Data.directmvF[1] = Data.directmvF[2] = Data.directmvF[3] = Data.directmvF[0]; |
} |
1863 |
Data.directmvB[1] = Data.directmvB[2] = Data.directmvB[3] = Data.directmvB[0]; |
|
1864 |
break; |
if (Data->qpel) { |
1865 |
} |
if (b_mb->mode == MODE_INTER4V) |
1866 |
iMinSAD += sad8bi(Data.Cur + (k&1)*8 + (k>>1)* 8 * iEdgedWidth, |
CheckCandidate = CheckCandidateDirect_qpel; |
1867 |
get_ref_mv(f_Ref, f_RefH, f_RefV, f_RefHV, |
else CheckCandidate = CheckCandidateDirectno4v_qpel; |
1868 |
(2*x+(k&1)), (2*y+(k>>1)), 8, &pMB->mvs[k], iEdgedWidth), |
} else { |
1869 |
get_ref_mv(b_Ref, b_RefH, b_RefV, b_RefHV, |
if (b_mb->mode == MODE_INTER4V) CheckCandidate = CheckCandidateDirect; |
1870 |
(2*x+(k&1)), (2*y+(k>>1)), 8, &pMB->b_mvs[k], iEdgedWidth), |
else CheckCandidate = CheckCandidateDirectno4v; |
|
iEdgedWidth); |
|
1871 |
} |
} |
1872 |
|
|
1873 |
|
(*CheckCandidate)(0, 0, 255, &k, Data); |
1874 |
|
|
1875 |
// skip decision |
// skip decision |
1876 |
if (iMinSAD < (int32_t)iQuant * SKIP_THRESH_B) { |
if (*Data->iMinSAD < pMB->quant * SKIP_THRESH_B) { |
1877 |
|
//possible skip - checking chroma. everything copied from MC |
1878 |
|
//this is not full chroma compensation, only it's fullpel approximation. should work though |
1879 |
|
int sum, dx, dy, b_dx, b_dy; |
1880 |
|
|
1881 |
|
if (Data->qpel) { |
1882 |
|
sum = pMB->mvs[0].y/2 + pMB->mvs[1].y/2 + pMB->mvs[2].y/2 + pMB->mvs[3].y/2; |
1883 |
|
dy = (sum >> 3) + roundtab_76[sum & 0xf]; |
1884 |
|
sum = pMB->mvs[0].x/2 + pMB->mvs[1].x/2 + pMB->mvs[2].x/2 + pMB->mvs[3].x/2; |
1885 |
|
dx = (sum >> 3) + roundtab_76[sum & 0xf]; |
1886 |
|
|
1887 |
|
sum = pMB->b_mvs[0].y/2 + pMB->b_mvs[1].y/2 + pMB->b_mvs[2].y/2 + pMB->b_mvs[3].y/2; |
1888 |
|
b_dy = (sum >> 3) + roundtab_76[sum & 0xf]; |
1889 |
|
sum = pMB->b_mvs[0].x/2 + pMB->b_mvs[1].x/2 + pMB->b_mvs[2].x/2 + pMB->b_mvs[3].x/2; |
1890 |
|
b_dx = (sum >> 3) + roundtab_76[sum & 0xf]; |
1891 |
|
|
1892 |
|
} else { |
1893 |
|
sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
1894 |
|
dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
1895 |
|
sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
1896 |
|
dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
1897 |
|
|
1898 |
|
sum = pMB->b_mvs[0].x + pMB->b_mvs[1].x + pMB->b_mvs[2].x + pMB->b_mvs[3].x; |
1899 |
|
b_dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
1900 |
|
sum = pMB->b_mvs[0].y + pMB->b_mvs[1].y + pMB->b_mvs[2].y + pMB->b_mvs[3].y; |
1901 |
|
b_dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
1902 |
|
} |
1903 |
|
sum = sad8bi(pCur->u + 8*x + 8*y*(Data->iEdgedWidth/2), |
1904 |
|
f_Ref->u + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
1905 |
|
b_Ref->u + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
1906 |
|
Data->iEdgedWidth/2); |
1907 |
|
sum += sad8bi(pCur->v + 8*x + 8*y*(Data->iEdgedWidth/2), |
1908 |
|
f_Ref->v + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
1909 |
|
b_Ref->v + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
1910 |
|
Data->iEdgedWidth/2); |
1911 |
|
|
1912 |
|
if (sum < MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) { |
1913 |
pMB->mode = MODE_DIRECT_NONE_MV; |
pMB->mode = MODE_DIRECT_NONE_MV; |
1914 |
return iMinSAD; } |
return *Data->iMinSAD; |
1915 |
|
} |
1916 |
|
} |
1917 |
|
|
1918 |
skip_sad = iMinSAD; |
skip_sad = *Data->iMinSAD; |
|
iMinSAD += 2 * lambda_vec16[iQuant]; // 2 bits needed to code vector 0,0 |
|
|
currentMV.x = currentMV.y = 0; |
|
|
if (b_mb->mode == MODE_INTER4V) |
|
|
CheckCandidate = CheckCandidateDirect; |
|
|
else CheckCandidate = CheckCandidateDirectno4v; |
|
1919 |
|
|
1920 |
// DIRECT MODE DELTA VECTOR SEARCH. |
// DIRECT MODE DELTA VECTOR SEARCH. |
1921 |
// This has to be made more effective, but at the moment I'm happy it's running at all |
// This has to be made more effective, but at the moment I'm happy it's running at all |
1924 |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1925 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1926 |
|
|
1927 |
(*MainSearchPtr)(0, 0, &Data, 255); |
(*MainSearchPtr)(0, 0, Data, 255); |
1928 |
|
|
1929 |
HalfpelRefine(&Data); |
HalfpelRefine(Data); //or qpel refine, if we're in qpel mode |
1930 |
|
|
1931 |
iMinSAD += 1 * lambda_vec16[iQuant]; // one bit is needed to code direct mode. we treat this bit just like it was vector's |
*Data->iMinSAD += 1 * Data->lambda16; // one bit is needed to code direct mode |
1932 |
*best_sad = iMinSAD; |
*best_sad = *Data->iMinSAD; |
1933 |
|
|
1934 |
if (b_mb->mode == MODE_INTER4V) |
if (b_mb->mode == MODE_INTER4V) |
1935 |
pMB->mode = MODE_DIRECT; |
pMB->mode = MODE_DIRECT; |
1936 |
else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
1937 |
|
|
1938 |
pMB->pmvs[3] = currentMV; |
pMB->pmvs[3] = *Data->currentMV; |
1939 |
|
|
1940 |
for (k = 0; k < 4; k++) { |
for (k = 0; k < 4; k++) { |
1941 |
pMB->mvs[k].x = Data.directmvF[k].x + currentMV.x; |
pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x; |
1942 |
pMB->b_mvs[k].x = ((currentMV.x == 0) |
pMB->b_mvs[k].x = ( (Data->currentMV->x == 0) |
1943 |
? Data.directmvB[k].x |
? Data->directmvB[k].x |
1944 |
: pMB->mvs[k].x - Data.referencemv[k].x); |
:pMB->mvs[k].x - Data->referencemv[k].x); |
1945 |
pMB->mvs[k].y = (Data.directmvF[k].y + currentMV.y); |
pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y); |
1946 |
pMB->b_mvs[k].y = ((currentMV.y == 0) |
pMB->b_mvs[k].y = ((Data->currentMV->y == 0) |
1947 |
? Data.directmvB[k].y |
? Data->directmvB[k].y |
1948 |
: pMB->mvs[k].y - Data.referencemv[k].y); |
: pMB->mvs[k].y - Data->referencemv[k].y); |
1949 |
|
if (Data->qpel) { |
1950 |
|
pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2; |
1951 |
|
pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2; |
1952 |
|
pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2; |
1953 |
|
pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2; |
1954 |
|
} |
1955 |
|
|
1956 |
if (b_mb->mode != MODE_INTER4V) { |
if (b_mb->mode != MODE_INTER4V) { |
1957 |
pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; |
pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; |
1958 |
pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; |
pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; |
1959 |
|
pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0]; |
1960 |
|
pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0]; |
1961 |
break; |
break; |
1962 |
} |
} |
1963 |
} |
} |
1964 |
return 0;//skip_sad; |
return skip_sad; |
1965 |
} |
} |
1966 |
|
|
1967 |
|
|
1968 |
static __inline void |
static __inline void |
1969 |
SearchInterpolate(const uint8_t * const f_Ref, |
SearchInterpolate(const uint8_t * const f_Ref, |
1970 |
const uint8_t * const f_RefH, |
const uint8_t * const f_RefH, |
1979 |
const uint32_t fcode, |
const uint32_t fcode, |
1980 |
const uint32_t bcode, |
const uint32_t bcode, |
1981 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
|
const uint32_t iQuant, |
|
1982 |
const MBParam * const pParam, |
const MBParam * const pParam, |
1983 |
const VECTOR * const f_predMV, |
const VECTOR * const f_predMV, |
1984 |
const VECTOR * const b_predMV, |
const VECTOR * const b_predMV, |
1985 |
MACROBLOCK * const pMB, |
MACROBLOCK * const pMB, |
1986 |
int32_t * const best_sad) |
int32_t * const best_sad, |
1987 |
|
SearchData * const fData) |
1988 |
|
|
1989 |
{ |
{ |
|
/* Interpolated MC motion vector search, this is tedious and more complicated because there are |
|
|
two values for everything, always one for backward and one for forward ME. Still, we don't gain |
|
|
much from this search, maybe it should simply be skipped and simply current i_sad16 value used |
|
|
as "optimal". */ |
|
1990 |
|
|
1991 |
const int32_t iEdgedWidth = pParam->edged_width; |
const int32_t iEdgedWidth = pParam->edged_width; |
|
|
|
1992 |
int iDirection, i, j; |
int iDirection, i, j; |
1993 |
int32_t iMinSAD = 256*4096; |
SearchData bData; |
|
VECTOR currentMV[3]; |
|
|
SearchData fData, bData; |
|
|
|
|
|
|
|
|
fData.iMinSAD = bData.iMinSAD = &iMinSAD; |
|
|
|
|
|
fData.Cur = bData.Cur = pCur->y + (x + y * iEdgedWidth) * 16; |
|
|
fData.iEdgedWidth = bData.iEdgedWidth = iEdgedWidth; |
|
|
fData.currentMV = currentMV; bData.currentMV = currentMV + 1; |
|
|
fData.iQuant = bData.iQuant = iQuant; |
|
|
fData.iFcode = bData.bFcode = fcode; fData.bFcode = bData.iFcode = bcode; |
|
|
|
|
|
|
|
|
bData.bRef = fData.Ref = f_Ref + (x + y * iEdgedWidth) * 16; |
|
|
bData.bRefH = fData.RefH = f_RefH + (x + y * iEdgedWidth) * 16; |
|
|
bData.bRefV = fData.RefV = f_RefV + (x + y * iEdgedWidth) * 16; |
|
|
bData.bRefHV = fData.RefHV = f_RefHV + (x + y * iEdgedWidth) * 16; |
|
|
bData.Ref = fData.bRef = b_Ref + (x + y * iEdgedWidth) * 16; |
|
|
bData.RefH = fData.bRefH = b_RefH + (x + y * iEdgedWidth) * 16; |
|
|
bData.RefV = fData.bRefV = b_RefV + (x + y * iEdgedWidth) * 16; |
|
|
bData.RefHV = fData.bRefHV = b_RefHV + (x + y * iEdgedWidth) * 16; |
|
1994 |
|
|
1995 |
bData.bpredMV = fData.predMV = *f_predMV; |
*(bData.iMinSAD = fData->iMinSAD) = 4096*256; |
1996 |
fData.bpredMV = bData.predMV = *b_predMV; |
bData.Cur = fData->Cur; |
1997 |
|
fData->iEdgedWidth = bData.iEdgedWidth = iEdgedWidth; |
1998 |
|
bData.currentMV = fData->currentMV + 1; bData.currentQMV = fData->currentQMV + 1; |
1999 |
|
bData.lambda16 = fData->lambda16; |
2000 |
|
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
2001 |
|
|
2002 |
|
bData.bRef = fData->Ref = f_Ref + (x + y * iEdgedWidth) * 16; |
2003 |
|
bData.bRefH = fData->RefH = f_RefH + (x + y * iEdgedWidth) * 16; |
2004 |
|
bData.bRefV = fData->RefV = f_RefV + (x + y * iEdgedWidth) * 16; |
2005 |
|
bData.bRefHV = fData->RefHV = f_RefHV + (x + y * iEdgedWidth) * 16; |
2006 |
|
bData.Ref = fData->bRef = b_Ref + (x + y * iEdgedWidth) * 16; |
2007 |
|
bData.RefH = fData->bRefH = b_RefH + (x + y * iEdgedWidth) * 16; |
2008 |
|
bData.RefV = fData->bRefV = b_RefV + (x + y * iEdgedWidth) * 16; |
2009 |
|
bData.RefHV = fData->bRefHV = b_RefHV + (x + y * iEdgedWidth) * 16; |
2010 |
|
bData.RefQ = fData->RefQ; |
2011 |
|
|
2012 |
|
bData.bpredMV = fData->predMV = *f_predMV; |
2013 |
|
fData->bpredMV = bData.predMV = *b_predMV; |
2014 |
|
|
2015 |
|
fData->currentMV[0] = fData->currentMV[2]; |
2016 |
|
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, pParam->m_quarterpel); |
2017 |
|
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, pParam->m_quarterpel); |
2018 |
|
|
2019 |
|
if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; |
2020 |
|
if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dy; |
2021 |
|
if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dx; |
2022 |
|
if (fData->currentMV[0].y > fData->min_dy) fData->currentMV[0].y = fData->min_dy; |
2023 |
|
|
2024 |
|
if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx; |
2025 |
|
if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dy; |
2026 |
|
if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dx; |
2027 |
|
if (fData->currentMV[1].y > bData.min_dy) fData->currentMV[1].y = bData.min_dy; |
2028 |
|
|
2029 |
|
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
|
currentMV[0] = pMB->mvs[0]; |
|
|
currentMV[1] = pMB->b_mvs[0]; |
|
|
get_range(&fData.min_dx, &fData.max_dx, &fData.min_dy, &fData.max_dy, x, y, 16, pParam->width, pParam->height, fcode); |
|
|
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode); |
|
|
|
|
|
CheckCandidateInt(currentMV[0].x, currentMV[0].y, 255, &iDirection, &fData); |
|
2030 |
|
|
2031 |
//diamond. I wish we could use normal mainsearch functions (square, advdiamond) |
//diamond. I wish we could use normal mainsearch functions (square, advdiamond) |
2032 |
|
|
2033 |
do { |
do { |
2034 |
iDirection = 255; |
iDirection = 255; |
2035 |
// forward MV moves |
// forward MV moves |
2036 |
i = currentMV[0].x; j = currentMV[0].y; |
i = fData->currentMV[0].x; j = fData->currentMV[0].y; |
2037 |
|
|
2038 |
CheckCandidateInt(i + 2, j, 0, &iDirection, &fData); |
CheckCandidateInt(i + 1, j, 0, &iDirection, fData); |
2039 |
CheckCandidateInt(i, j + 2, 0, &iDirection, &fData); |
CheckCandidateInt(i, j + 1, 0, &iDirection, fData); |
2040 |
CheckCandidateInt(i - 2, j, 0, &iDirection, &fData); |
CheckCandidateInt(i - 1, j, 0, &iDirection, fData); |
2041 |
CheckCandidateInt(i, j - 2, 0, &iDirection, &fData); |
CheckCandidateInt(i, j - 1, 0, &iDirection, fData); |
2042 |
|
|
2043 |
// backward MV moves |
// backward MV moves |
2044 |
i = currentMV[1].x; j = currentMV[1].y; |
i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
2045 |
currentMV[2] = currentMV[0]; |
fData->currentMV[2] = fData->currentMV[0]; |
2046 |
|
CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); |
2047 |
CheckCandidateInt(i + 2, j, 0, &iDirection, &bData); |
CheckCandidateInt(i, j + 1, 0, &iDirection, &bData); |
2048 |
CheckCandidateInt(i, j + 2, 0, &iDirection, &bData); |
CheckCandidateInt(i - 1, j, 0, &iDirection, &bData); |
2049 |
CheckCandidateInt(i - 2, j, 0, &iDirection, &bData); |
CheckCandidateInt(i, j - 1, 0, &iDirection, &bData); |
|
CheckCandidateInt(i, j - 2, 0, &iDirection, &bData); |
|
2050 |
|
|
2051 |
} while (!(iDirection)); |
} while (!(iDirection)); |
2052 |
|
|
2053 |
/* halfpel refinement. luckly we can use normal halfpel function for it */ |
*fData->iMinSAD += 2 * fData->lambda16; // two bits are needed to code interpolate mode. |
|
|
|
|
if (MotionFlags & PMV_HALFPELREFINE16) { |
|
|
CheckCandidate = CheckCandidateInt; |
|
|
HalfpelRefine(&fData); |
|
|
currentMV[2] = currentMV[0]; |
|
|
HalfpelRefine(&bData); |
|
|
} |
|
2054 |
|
|
2055 |
// two bits are needed to code interpolate mode. we treat the bits just like they were vector's |
if (fData->qpel) { |
2056 |
iMinSAD += 2 * lambda_vec16[iQuant]; |
CheckCandidate = CheckCandidateInt_qpel; |
2057 |
if (iMinSAD < *best_sad) { |
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 0); |
2058 |
*best_sad = iMinSAD; |
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 0); |
2059 |
pMB->mvs[0] = currentMV[0]; |
fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x; |
2060 |
pMB->b_mvs[0] = currentMV[1]; |
fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y; |
2061 |
|
fData->currentQMV[1].x = 2 * fData->currentMV[1].x; |
2062 |
|
fData->currentQMV[1].y = 2 * fData->currentMV[1].y; |
2063 |
|
// QuarterpelRefine(fData); |
2064 |
|
fData->currentQMV[2] = fData->currentQMV[0]; |
2065 |
|
// QuarterpelRefine(&bData); |
2066 |
|
} |
2067 |
|
|
2068 |
|
if (*fData->iMinSAD < *best_sad) { |
2069 |
|
*best_sad = *fData->iMinSAD; |
2070 |
|
pMB->mvs[0] = fData->currentMV[0]; |
2071 |
|
pMB->b_mvs[0] = fData->currentMV[1]; |
2072 |
pMB->mode = MODE_INTERPOLATE; |
pMB->mode = MODE_INTERPOLATE; |
2073 |
|
if (fData->qpel) { |
2074 |
|
pMB->qmvs[0] = fData->currentQMV[0]; |
2075 |
|
pMB->b_qmvs[0] = fData->currentQMV[1]; |
2076 |
|
pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x; |
2077 |
|
pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y; |
2078 |
|
pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x; |
2079 |
|
pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y; |
2080 |
|
} else { |
2081 |
pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; |
pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; |
2082 |
pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; |
pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; |
2083 |
pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; |
pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; |
2084 |
pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; |
pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; |
2085 |
} |
} |
2086 |
} |
} |
2087 |
|
} |
2088 |
|
|
2089 |
void |
void |
2090 |
MotionEstimationBVOP(MBParam * const pParam, |
MotionEstimationBVOP(MBParam * const pParam, |
2098 |
const IMAGE * const f_refV, |
const IMAGE * const f_refV, |
2099 |
const IMAGE * const f_refHV, |
const IMAGE * const f_refHV, |
2100 |
// backward (future) reference |
// backward (future) reference |
2101 |
const MACROBLOCK * const b_mbs, |
const FRAMEINFO * const b_reference, |
2102 |
const IMAGE * const b_ref, |
const IMAGE * const b_ref, |
2103 |
const IMAGE * const b_refH, |
const IMAGE * const b_refH, |
2104 |
const IMAGE * const b_refV, |
const IMAGE * const b_refV, |
2108 |
int32_t best_sad, skip_sad; |
int32_t best_sad, skip_sad; |
2109 |
int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; |
int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; |
2110 |
static const VECTOR zeroMV={0,0}; |
static const VECTOR zeroMV={0,0}; |
2111 |
|
const MACROBLOCK * const b_mbs = b_reference->mbs; |
2112 |
|
|
2113 |
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
2114 |
|
|
2115 |
const int32_t TRB = time_pp - time_bp; |
const int32_t TRB = time_pp - time_bp; |
2116 |
const int32_t TRD = time_pp; |
const int32_t TRD = time_pp; |
2117 |
|
uint8_t * qimage; |
2118 |
|
|
2119 |
// note: i==horizontal, j==vertical |
// some pre-inintialized data for the rest of the search |
2120 |
|
|
2121 |
|
SearchData Data; |
2122 |
|
int32_t iMinSAD; |
2123 |
|
VECTOR currentMV[3]; |
2124 |
|
VECTOR currentQMV[3]; |
2125 |
|
Data.iEdgedWidth = pParam->edged_width; |
2126 |
|
Data.currentMV = currentMV; Data.currentQMV = currentQMV; |
2127 |
|
Data.iMinSAD = &iMinSAD; |
2128 |
|
Data.lambda16 = lambda_vec16[frame->quant]; |
2129 |
|
Data.qpel = pParam->m_quarterpel; |
2130 |
|
|
2131 |
|
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
2132 |
|
return; // allocate some mem for qpel interpolated blocks |
2133 |
|
// somehow this is dirty since I think we shouldn't use malloc outside |
2134 |
|
// encoder_create() - so please fix me! |
2135 |
|
Data.RefQ = qimage; |
2136 |
|
|
2137 |
|
// note: i==horizontal, j==vertical |
2138 |
for (j = 0; j < pParam->mb_height; j++) { |
for (j = 0; j < pParam->mb_height; j++) { |
2139 |
|
|
2140 |
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
2143 |
MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; |
MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; |
2144 |
const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; |
const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; |
2145 |
|
|
2146 |
/* special case, if collocated block is SKIPed: encoding is forward (0,0), cpb=0 without further ado */ |
/* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */ |
2147 |
|
if (b_reference->coding_type != S_VOP) |
2148 |
if (b_mb->mode == MODE_NOT_CODED) { |
if (b_mb->mode == MODE_NOT_CODED) { |
2149 |
pMB->mode = MODE_NOT_CODED; |
pMB->mode = MODE_NOT_CODED; |
2150 |
continue; |
continue; |
2151 |
} |
} |
2152 |
|
|
2153 |
|
Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
2154 |
|
pMB->quant = frame->quant; |
2155 |
|
|
2156 |
/* direct search comes first, because it (1) checks for SKIP-mode |
/* direct search comes first, because it (1) checks for SKIP-mode |
2157 |
and (2) sets very good predictions for forward and backward search */ |
and (2) sets very good predictions for forward and backward search */ |
2158 |
|
skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
2159 |
skip_sad = SearchDirect(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
|
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
2160 |
&frame->image, |
&frame->image, |
2161 |
i, j, |
i, j, |
2162 |
frame->motion_flags, |
frame->motion_flags, |
|
frame->quant, |
|
2163 |
TRB, TRD, |
TRB, TRD, |
2164 |
pParam, |
pParam, |
2165 |
pMB, b_mb, |
pMB, b_mb, |
2166 |
&best_sad); |
&best_sad, |
2167 |
|
&Data); |
2168 |
|
|
|
if (!(frame->global_flags & XVID_HALFPEL)) best_sad = skip_sad = 256*4096; |
|
|
else |
|
2169 |
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
2170 |
|
|
|
// best_sad = 256*4096; //uncomment to disable Directsearch. |
|
|
// To disable any other mode, just comment the function call |
|
|
|
|
2171 |
// forward search |
// forward search |
2172 |
SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
2173 |
&frame->image, i, j, |
&frame->image, i, j, |
2174 |
frame->motion_flags, |
frame->motion_flags, |
2175 |
frame->quant, frame->fcode, pParam, |
frame->fcode, pParam, |
2176 |
pMB, &f_predMV, &best_sad, |
pMB, &f_predMV, &best_sad, |
2177 |
MODE_FORWARD); |
MODE_FORWARD, &Data); |
2178 |
|
|
2179 |
// backward search |
// backward search |
2180 |
SearchBF(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
SearchBF(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
2181 |
&frame->image, i, j, |
&frame->image, i, j, |
2182 |
frame->motion_flags, |
frame->motion_flags, |
2183 |
frame->quant, frame->bcode, pParam, |
frame->bcode, pParam, |
2184 |
pMB, &b_predMV, &best_sad, |
pMB, &b_predMV, &best_sad, |
2185 |
MODE_BACKWARD); |
MODE_BACKWARD, &Data); |
2186 |
|
|
2187 |
// interpolate search comes last, because it uses data from forward and backward as prediction |
// interpolate search comes last, because it uses data from forward and backward as prediction |
2188 |
|
|
2192 |
i, j, |
i, j, |
2193 |
frame->fcode, frame->bcode, |
frame->fcode, frame->bcode, |
2194 |
frame->motion_flags, |
frame->motion_flags, |
2195 |
frame->quant, pParam, |
pParam, |
2196 |
&f_predMV, &b_predMV, |
&f_predMV, &b_predMV, |
2197 |
pMB, &best_sad); |
pMB, &best_sad, |
2198 |
|
&Data); |
2199 |
|
|
2200 |
switch (pMB->mode) { |
switch (pMB->mode) { |
2201 |
case MODE_FORWARD: |
case MODE_FORWARD: |
2202 |
f_count++; |
f_count++; |
2203 |
f_predMV = pMB->mvs[0]; |
if (pParam->m_quarterpel) f_predMV = pMB->qmvs[0]; |
2204 |
|
else f_predMV = pMB->mvs[0]; |
2205 |
break; |
break; |
2206 |
case MODE_BACKWARD: |
case MODE_BACKWARD: |
2207 |
b_count++; |
b_count++; |
2208 |
b_predMV = pMB->b_mvs[0]; |
if (pParam->m_quarterpel) b_predMV = pMB->b_qmvs[0]; |
2209 |
|
else b_predMV = pMB->b_mvs[0]; |
2210 |
break; |
break; |
2211 |
case MODE_INTERPOLATE: |
case MODE_INTERPOLATE: |
2212 |
i_count++; |
i_count++; |
2213 |
|
if (pParam->m_quarterpel) { |
2214 |
|
f_predMV = pMB->qmvs[0]; |
2215 |
|
b_predMV = pMB->b_qmvs[0]; |
2216 |
|
} else { |
2217 |
f_predMV = pMB->mvs[0]; |
f_predMV = pMB->mvs[0]; |
2218 |
b_predMV = pMB->b_mvs[0]; |
b_predMV = pMB->b_mvs[0]; |
2219 |
|
} |
2220 |
break; |
break; |
2221 |
case MODE_DIRECT: |
case MODE_DIRECT: |
2222 |
case MODE_DIRECT_NO4V: |
case MODE_DIRECT_NO4V: |
2227 |
} |
} |
2228 |
} |
} |
2229 |
} |
} |
2230 |
|
free(qimage); |
|
// fprintf(debug,"B-Stat: F: %04d B: %04d I: %04d D: %04d, N: %04d\n", |
|
|
// f_count,b_count,i_count,d_count,n_count); |
|
|
|
|
2231 |
} |
} |
2232 |
|
|
2233 |
/* Hinted ME starts here */ |
/* Hinted ME starts here */ |
2234 |
|
|
|
static __inline void |
|
|
Search8hinted( const SearchData * const OldData, |
|
|
const int x, const int y, |
|
|
const uint32_t MotionFlags, |
|
|
const MBParam * const pParam, |
|
|
MACROBLOCK * const pMB, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const int block) |
|
|
{ |
|
|
SearchData Data; |
|
|
MainSearchFunc *MainSearchPtr; |
|
|
|
|
|
Data.predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
|
|
Data.iMinSAD = OldData->iMinSAD + 1 + block; |
|
|
Data.currentMV = OldData->currentMV+1+block; |
|
|
Data.iFcode = OldData->iFcode; |
|
|
Data.iQuant = OldData->iQuant; |
|
|
|
|
|
Data.Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
|
Data.RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
|
Data.RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
|
Data.RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
|
Data.iEdgedWidth = pParam->edged_width; |
|
|
Data.Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
|
|
|
|
CheckCandidate = CheckCandidate8; |
|
|
|
|
|
if (block != 0) |
|
|
*(Data.iMinSAD) += lambda_vec8[Data.iQuant] * |
|
|
d_mv_bits( Data.currentMV->x - Data.predMV.x, |
|
|
Data.currentMV->y - Data.predMV.y, |
|
|
Data.iFcode); |
|
|
|
|
|
|
|
|
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 8, |
|
|
pParam->width, pParam->height, OldData->iFcode); |
|
|
|
|
|
if (pMB->mode == MODE_INTER4V) { |
|
|
int dummy; |
|
|
CheckCandidate8(pMB->mvs[block].x, pMB->mvs[block].y, 0, &dummy, &Data); } |
|
|
|
|
|
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
|
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
|
|
else MainSearchPtr = DiamondSearch; |
|
|
|
|
|
(*MainSearchPtr)(Data.currentMV->x, Data.currentMV->y, &Data, 255); |
|
|
|
|
|
if (MotionFlags & PMV_HALFPELREFINE8) HalfpelRefine(&Data); |
|
|
|
|
|
pMB->pmvs[block].x = Data.currentMV->x - Data.predMV.x; |
|
|
pMB->pmvs[block].y = Data.currentMV->y - Data.predMV.y; |
|
|
pMB->mvs[block] = *(Data.currentMV); |
|
|
pMB->sad8[block] = 4 * (*(Data.iMinSAD)); |
|
|
} |
|
|
|
|
|
|
|
2235 |
static void |
static void |
2236 |
SearchPhinted ( const uint8_t * const pRef, |
SearchPhinted ( const IMAGE * const pRef, |
2237 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
2238 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
2239 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
2242 |
const int y, |
const int y, |
2243 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
2244 |
const uint32_t iQuant, |
const uint32_t iQuant, |
|
const uint32_t iFcode, |
|
2245 |
const MBParam * const pParam, |
const MBParam * const pParam, |
2246 |
const MACROBLOCK * const pMBs, |
const MACROBLOCK * const pMBs, |
2247 |
int inter4v, |
int inter4v, |
2248 |
MACROBLOCK * const pMB) |
MACROBLOCK * const pMB, |
2249 |
|
SearchData * const Data) |
2250 |
{ |
{ |
2251 |
|
|
2252 |
const int32_t iEdgedWidth = pParam->edged_width; |
int i, t; |
|
|
|
|
int i; |
|
|
VECTOR currentMV[5]; |
|
|
int32_t iMinSAD[5]; |
|
|
int32_t temp[5]; |
|
2253 |
MainSearchFunc * MainSearchPtr; |
MainSearchFunc * MainSearchPtr; |
|
SearchData Data; |
|
|
|
|
|
Data.predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
|
|
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 16, |
|
|
pParam->width, pParam->height, iFcode); |
|
2254 |
|
|
2255 |
Data.Cur = pCur->y + (x + y * iEdgedWidth) * 16; |
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
2256 |
Data.iEdgedWidth = iEdgedWidth; |
Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
2257 |
Data.currentMV = currentMV; |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2258 |
Data.iMinSAD = iMinSAD; |
pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); |
2259 |
Data.Ref = pRef + (x + iEdgedWidth*y)*16; |
|
2260 |
Data.RefH = pRefH + (x + iEdgedWidth*y) * 16; |
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16; |
2261 |
Data.RefV = pRefV + (x + iEdgedWidth*y) * 16; |
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
2262 |
Data.RefHV = pRefHV + (x + iEdgedWidth*y) * 16; |
Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
2263 |
Data.temp = temp; |
|
2264 |
Data.iQuant = iQuant; |
Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16; |
2265 |
Data.iFcode = iFcode; |
Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16; |
2266 |
|
Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16; |
2267 |
|
Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16; |
2268 |
|
Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
2269 |
|
Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
2270 |
|
|
2271 |
if (!(MotionFlags & PMV_HALFPEL16)) { |
if (!(MotionFlags & PMV_HALFPEL16)) { |
2272 |
Data.min_dx = EVEN(Data.min_dx); |
Data->min_dx = EVEN(Data->min_dx); |
2273 |
Data.max_dx = EVEN(Data.max_dx); |
Data->max_dx = EVEN(Data->max_dx); |
2274 |
Data.min_dy = EVEN(Data.min_dy); |
Data->min_dy = EVEN(Data->min_dy); |
2275 |
Data.max_dy = EVEN(Data.max_dy); |
Data->max_dy = EVEN(Data->max_dy); |
2276 |
} |
} |
2277 |
|
|
2278 |
for(i = 0; i < 5; i++) iMinSAD[i] = MV_MAX_ERROR; |
for(i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
2279 |
|
|
2280 |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
2281 |
|
|
2282 |
if (inter4v) |
if (inter4v || pParam->m_quarterpel || Data->chroma) CheckCandidate = CheckCandidate16; |
|
CheckCandidate = CheckCandidate16; |
|
2283 |
else CheckCandidate = CheckCandidate16no4v; |
else CheckCandidate = CheckCandidate16no4v; |
2284 |
|
|
|
|
|
2285 |
pMB->mvs[0].x = EVEN(pMB->mvs[0].x); |
pMB->mvs[0].x = EVEN(pMB->mvs[0].x); |
2286 |
pMB->mvs[0].y = EVEN(pMB->mvs[0].y); |
pMB->mvs[0].y = EVEN(pMB->mvs[0].y); |
2287 |
if (pMB->mvs[0].x > Data.max_dx) pMB->mvs[0].x = Data.max_dx; // this is in case iFcode changed |
if (pMB->mvs[0].x > Data->max_dx) pMB->mvs[0].x = Data->max_dx; // this is in case iFcode changed |
2288 |
if (pMB->mvs[0].x < Data.min_dx) pMB->mvs[0].x = Data.min_dx; |
if (pMB->mvs[0].x < Data->min_dx) pMB->mvs[0].x = Data->min_dx; |
2289 |
if (pMB->mvs[0].y > Data.max_dy) pMB->mvs[0].y = Data.max_dy; |
if (pMB->mvs[0].y > Data->max_dy) pMB->mvs[0].y = Data->max_dy; |
2290 |
if (pMB->mvs[0].y < Data.min_dy) pMB->mvs[0].y = Data.min_dy; |
if (pMB->mvs[0].y < Data->min_dy) pMB->mvs[0].y = Data->min_dy; |
2291 |
|
|
2292 |
CheckCandidate16(pMB->mvs[0].x, pMB->mvs[0].y, 0, &i, &Data); |
(*CheckCandidate)(pMB->mvs[0].x, pMB->mvs[0].y, 0, &t, Data); |
2293 |
|
|
2294 |
if (pMB->mode == MODE_INTER4V) |
if (pMB->mode == MODE_INTER4V) |
2295 |
for (i = 1; i < 4; i++) { // all four vectors will be used as four predictions for 16x16 search |
for (i = 1; i < 4; i++) { // all four vectors will be used as four predictions for 16x16 search |
2296 |
pMB->mvs[i].x = EVEN(pMB->mvs[i].x); |
pMB->mvs[i].x = EVEN(pMB->mvs[i].x); |
2297 |
pMB->mvs[i].y = EVEN(pMB->mvs[i].y); |
pMB->mvs[i].y = EVEN(pMB->mvs[i].y); |
2298 |
if (!(make_mask(pMB->mvs, i))) |
if (!(make_mask(pMB->mvs, i))) |
2299 |
CheckCandidate16(pMB->mvs[i].x, pMB->mvs[i].y, 0, &i, &Data); |
(*CheckCandidate)(pMB->mvs[i].x, pMB->mvs[i].y, 0, &t, Data); |
2300 |
} |
} |
2301 |
|
|
2302 |
if (MotionFlags & PMV_USESQUARES16) |
if (MotionFlags & PMV_USESQUARES16) |
2305 |
MainSearchPtr = AdvDiamondSearch; |
MainSearchPtr = AdvDiamondSearch; |
2306 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
2307 |
|
|
2308 |
(*MainSearchPtr)(currentMV->x, currentMV->y, &Data, 255); |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
2309 |
|
|
2310 |
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(&Data); |
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data); |
2311 |
|
|
2312 |
if (inter4v) |
for(i = 0; i < 5; i++) { |
2313 |
for(i = 0; i < 4; i++) |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors |
2314 |
Search8hinted(&Data, 2*x+(i&1), 2*y+(i>>1), MotionFlags, pParam, pMB, pMBs, i); |
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
2315 |
|
} |
2316 |
|
|
2317 |
|
if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { |
2318 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2319 |
|
pParam->width, pParam->height, Data->iFcode, 0); |
2320 |
|
CheckCandidate = CheckCandidate16_qpel; |
2321 |
|
QuarterpelRefine(Data); |
2322 |
|
} |
2323 |
|
|
2324 |
|
if (inter4v) { |
2325 |
|
SearchData Data8; |
2326 |
|
Data8.iFcode = Data->iFcode; |
2327 |
|
Data8.lambda8 = Data->lambda8; |
2328 |
|
Data8.iEdgedWidth = Data->iEdgedWidth; |
2329 |
|
Data8.RefQ = Data->RefQ; |
2330 |
|
Data8.qpel = Data->qpel; |
2331 |
|
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
2332 |
|
Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); |
2333 |
|
Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); |
2334 |
|
Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); |
2335 |
|
|
2336 |
|
if (Data->chroma) { |
2337 |
|
int sum, dx, dy; |
2338 |
|
|
2339 |
|
if(pParam->m_quarterpel) |
2340 |
|
sum = (pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2); |
2341 |
|
else sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
2342 |
|
dy = (sum ? SIGN(sum) * |
2343 |
|
(roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2) : 0); |
2344 |
|
|
2345 |
|
if(pParam->m_quarterpel) |
2346 |
|
sum = (pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2); |
2347 |
|
else sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
2348 |
|
dx = (sum ? SIGN(sum) * |
2349 |
|
(roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2) : 0); |
2350 |
|
Data->iMinSAD[1] += ChromaSAD(dx, dy, Data); |
2351 |
|
} |
2352 |
|
} |
2353 |
|
|
2354 |
if (!(inter4v) || |
if (!(inter4v) || |
2355 |
(iMinSAD[0] < iMinSAD[1] + iMinSAD[2] + iMinSAD[3] + iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] + |
2356 |
|
Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
2357 |
// INTER MODE |
// INTER MODE |
|
|
|
2358 |
pMB->mode = MODE_INTER; |
pMB->mode = MODE_INTER; |
2359 |
pMB->mv16 = pMB->mvs[0] = pMB->mvs[1] |
pMB->mvs[0] = pMB->mvs[1] |
2360 |
= pMB->mvs[2] = pMB->mvs[3] = currentMV[0]; |
= pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
2361 |
|
|
2362 |
|
pMB->qmvs[0] = pMB->qmvs[1] |
2363 |
|
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
2364 |
|
|
2365 |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
2366 |
pMB->sad8[2] = pMB->sad8[3] = iMinSAD[0]; |
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
2367 |
|
|
2368 |
pMB->pmvs[0].x = currentMV[0].x - Data.predMV.x; |
if(pParam->m_quarterpel) { |
2369 |
pMB->pmvs[0].y = currentMV[0].y - Data.predMV.y; |
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predQMV.x; |
2370 |
|
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predQMV.y; |
2371 |
} else { |
} else { |
2372 |
// INTER4V MODE; all other things are already set in Search8hinted |
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
2373 |
|
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
2374 |
|
} |
2375 |
|
} else { |
2376 |
|
// INTER4V MODE; all other things are already set in Search8 |
2377 |
pMB->mode = MODE_INTER4V; |
pMB->mode = MODE_INTER4V; |
2378 |
pMB->sad16 = iMinSAD[1] + iMinSAD[2] + iMinSAD[3] + iMinSAD[4] + IMV16X16 * iQuant; |
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] |
2379 |
|
+ Data->iMinSAD[4] + IMV16X16 * iQuant; |
2380 |
} |
} |
2381 |
|
|
2382 |
} |
} |
2394 |
const IMAGE *const pRef = &reference->image; |
const IMAGE *const pRef = &reference->image; |
2395 |
|
|
2396 |
uint32_t x, y; |
uint32_t x, y; |
2397 |
|
uint8_t * qimage; |
2398 |
|
int32_t temp[5], quant = current->quant; |
2399 |
|
int32_t iMinSAD[5]; |
2400 |
|
VECTOR currentMV[5], currentQMV[5]; |
2401 |
|
SearchData Data; |
2402 |
|
Data.iEdgedWidth = pParam->edged_width; |
2403 |
|
Data.currentMV = currentMV; |
2404 |
|
Data.currentQMV = currentQMV; |
2405 |
|
Data.iMinSAD = iMinSAD; |
2406 |
|
Data.temp = temp; |
2407 |
|
Data.iFcode = current->fcode; |
2408 |
|
Data.rounding = pParam->m_rounding_type; |
2409 |
|
Data.qpel = pParam->m_quarterpel; |
2410 |
|
Data.chroma = current->global_flags & XVID_ME_COLOUR; |
2411 |
|
|
2412 |
|
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
2413 |
|
return; // allocate some mem for qpel interpolated blocks |
2414 |
|
// somehow this is dirty since I think we shouldn't use malloc outside |
2415 |
|
// encoder_create() - so please fix me! |
2416 |
|
|
2417 |
|
Data.RefQ = qimage; |
2418 |
|
|
2419 |
if (sadInit) (*sadInit) (); |
if (sadInit) (*sadInit) (); |
2420 |
|
|
2421 |
for (y = 0; y < pParam->mb_height; y++) { |
for (y = 0; y < pParam->mb_height; y++) { |
2422 |
for (x = 0; x < pParam->mb_width; x++) { |
for (x = 0; x < pParam->mb_width; x++) { |
|
int32_t sad00; |
|
2423 |
|
|
2424 |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
2425 |
|
|
2429 |
if (!(current->global_flags & XVID_LUMIMASKING)) { |
if (!(current->global_flags & XVID_LUMIMASKING)) { |
2430 |
pMB->dquant = NO_CHANGE; |
pMB->dquant = NO_CHANGE; |
2431 |
pMB->quant = current->quant; } |
pMB->quant = current->quant; } |
2432 |
|
else { |
2433 |
|
if (pMB->dquant != NO_CHANGE) { |
2434 |
|
quant += DQtab[pMB->dquant]; |
2435 |
|
if (quant > 31) quant = 31; |
2436 |
|
else if (quant < 1) quant = 1; |
2437 |
|
} |
2438 |
|
pMB->quant = quant; |
2439 |
|
} |
2440 |
|
|
2441 |
if (pMB->dquant == NO_CHANGE) //no skip otherwise, anyway |
SearchPhinted(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
2442 |
sad00 = pMB->sad16 |
y, current->motion_flags, pMB->quant, |
2443 |
= sad16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
pParam, pMBs, current->global_flags & XVID_INTER4V, pMB, |
2444 |
pRef->y + (x + y * pParam->edged_width) * 16, |
&Data); |
|
pParam->edged_width, 256*4096 ); |
|
|
else sad00 = 256*4096; |
|
2445 |
|
|
2446 |
|
} |
2447 |
|
} |
2448 |
|
free(qimage); |
2449 |
|
} |
2450 |
|
|
2451 |
//initial skip decision |
static __inline int |
2452 |
|
MEanalyzeMB ( const uint8_t * const pRef, |
2453 |
|
const uint8_t * const pCur, |
2454 |
|
const int x, |
2455 |
|
const int y, |
2456 |
|
const MBParam * const pParam, |
2457 |
|
const MACROBLOCK * const pMBs, |
2458 |
|
MACROBLOCK * const pMB, |
2459 |
|
SearchData * const Data) |
2460 |
|
{ |
2461 |
|
|
2462 |
if ( (pMB->dquant == NO_CHANGE) && (sad00 <= MAX_SAD00_FOR_SKIP * pMB->quant) |
int i = 255, mask; |
2463 |
&& ( //(pMB->mode == MODE_NOT_CODED) || |
VECTOR pmv[3]; |
2464 |
(SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant) )) ) { |
*(Data->iMinSAD) = MV_MAX_ERROR; |
2465 |
if (sad00 < pMB->quant * INITIAL_SKIP_THRESH) { |
|
2466 |
SkipMacroblockP(pMB, sad00); |
//median is only used as prediction. it doesn't have to be real |
2467 |
continue; } //skipped |
if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; |
2468 |
|
else |
2469 |
|
if (x == 1) //left macroblock does not have any vector now |
2470 |
|
Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median |
2471 |
|
else if (y == 1) // top macroblock don't have it's vector |
2472 |
|
Data->predMV = (pMB - 1)->mvs[0]; // left instead of median |
2473 |
|
else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median |
2474 |
|
|
2475 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2476 |
|
pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); |
2477 |
|
|
2478 |
|
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
2479 |
|
Data->Ref = pRef + (x + y * pParam->edged_width) * 16; |
2480 |
|
|
2481 |
|
pmv[1].x = EVEN(pMB->mvs[0].x); |
2482 |
|
pmv[1].y = EVEN(pMB->mvs[0].y); |
2483 |
|
pmv[2].x = EVEN(Data->predMV.x); |
2484 |
|
pmv[2].y = EVEN(Data->predMV.y); |
2485 |
|
pmv[0].x = pmv[0].y = 0; |
2486 |
|
|
2487 |
|
(*CheckCandidate)(0, 0, 255, &i, Data); |
2488 |
|
|
2489 |
|
//early skip for 0,0 |
2490 |
|
if (*Data->iMinSAD < MAX_SAD00_FOR_SKIP * 4) { |
2491 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
2492 |
|
pMB->mode = MODE_NOT_CODED; |
2493 |
|
return 0; |
2494 |
} |
} |
|
else sad00 = 256*4096; |
|
2495 |
|
|
2496 |
if (pMB->mode == MODE_NOT_CODED) |
if (!(mask = make_mask(pmv, 1))) |
2497 |
SearchP( pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
(*CheckCandidate)(pmv[1].x, pmv[1].y, mask, &i, Data); |
2498 |
y, current->motion_flags, pMB->quant, |
if (!(mask = make_mask(pmv, 2))) |
2499 |
current->fcode, pParam, pMBs, reference->mbs, |
(*CheckCandidate)(pmv[2].x, pmv[2].y, mask, &i, Data); |
2500 |
current->global_flags & XVID_INTER4V, pMB); |
|
2501 |
|
if (*Data->iMinSAD > MAX_SAD00_FOR_SKIP * 4) // diamond only if needed |
2502 |
|
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
2503 |
|
|
2504 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
2505 |
|
pMB->mode = MODE_INTER; |
2506 |
|
return *(Data->iMinSAD); |
2507 |
|
} |
2508 |
|
|
2509 |
|
#define INTRA_THRESH 1350 |
2510 |
|
#define INTER_THRESH 900 |
2511 |
|
|
2512 |
|
|
2513 |
|
int |
2514 |
|
MEanalysis( const IMAGE * const pRef, |
2515 |
|
FRAMEINFO * const Current, |
2516 |
|
MBParam * const pParam, |
2517 |
|
int maxIntra, //maximum number if non-I frames |
2518 |
|
int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame |
2519 |
|
int bCount) // number if B frames in a row |
2520 |
|
{ |
2521 |
|
uint32_t x, y, intra = 0; |
2522 |
|
int sSAD = 0; |
2523 |
|
MACROBLOCK * const pMBs = Current->mbs; |
2524 |
|
const IMAGE * const pCurrent = &Current->image; |
2525 |
|
int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH; |
2526 |
|
|
2527 |
|
VECTOR currentMV; |
2528 |
|
int32_t iMinSAD; |
2529 |
|
SearchData Data; |
2530 |
|
Data.iEdgedWidth = pParam->edged_width; |
2531 |
|
Data.currentMV = ¤tMV; |
2532 |
|
Data.iMinSAD = &iMinSAD; |
2533 |
|
Data.iFcode = Current->fcode; |
2534 |
|
CheckCandidate = CheckCandidate16no4vI; |
2535 |
|
|
2536 |
|
if (intraCount < 10) // we're right after an I frame |
2537 |
|
IntraThresh += 4 * (intraCount - 10) * (intraCount - 10); |
2538 |
else |
else |
2539 |
SearchPhinted(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec |
2540 |
y, current->motion_flags, pMB->quant, |
IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; |
|
current->fcode, pParam, pMBs, |
|
|
current->global_flags & XVID_INTER4V, pMB); |
|
2541 |
|
|
|
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
|
|
if (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
|
|
if ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) |
|
|
SkipMacroblockP(pMB, sad00); |
|
2542 |
|
|
2543 |
|
InterThresh += 300 * (1 - bCount); |
2544 |
|
if (InterThresh < 200) InterThresh = 200; |
2545 |
|
|
2546 |
|
if (sadInit) (*sadInit) (); |
2547 |
|
|
2548 |
|
for (y = 1; y < pParam->mb_height-1; y++) { |
2549 |
|
for (x = 1; x < pParam->mb_width-1; x++) { |
2550 |
|
int sad, dev; |
2551 |
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
2552 |
|
|
2553 |
|
sad = MEanalyzeMB(pRef->y, pCurrent->y, x, y, |
2554 |
|
pParam, pMBs, pMB, &Data); |
2555 |
|
|
2556 |
|
if (sad > IntraThresh) { |
2557 |
|
dev = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
2558 |
|
pParam->edged_width); |
2559 |
|
if (dev + IntraThresh < sad) { |
2560 |
|
pMB->mode = MODE_INTRA; |
2561 |
|
if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return 2; // I frame |
2562 |
|
} |
2563 |
|
} |
2564 |
|
sSAD += sad; |
2565 |
|
} |
2566 |
|
} |
2567 |
|
sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); |
2568 |
|
if (sSAD > InterThresh ) return 1; //P frame |
2569 |
|
emms(); |
2570 |
|
return 0; // B frame |
2571 |
|
|
2572 |
|
} |
2573 |
|
|
2574 |
|
int |
2575 |
|
FindFcode( const MBParam * const pParam, |
2576 |
|
const FRAMEINFO * const current) |
2577 |
|
{ |
2578 |
|
uint32_t x, y; |
2579 |
|
int max = 0, min = 0, i; |
2580 |
|
|
2581 |
|
for (y = 0; y < pParam->mb_height; y++) { |
2582 |
|
for (x = 0; x < pParam->mb_width; x++) { |
2583 |
|
|
2584 |
|
MACROBLOCK *pMB = ¤t->mbs[x + y * pParam->mb_width]; |
2585 |
|
for(i = 0; i < (pMB->mode == MODE_INTER4V ? 4:1); i++) { |
2586 |
|
if (pMB->mvs[i].x > max) max = pMB->mvs[i].x; |
2587 |
|
if (pMB->mvs[i].y > max) max = pMB->mvs[i].y; |
2588 |
|
|
2589 |
|
if (pMB->mvs[i].x < min) min = pMB->mvs[i].x; |
2590 |
|
if (pMB->mvs[i].y < min) min = pMB->mvs[i].y; |
2591 |
} |
} |
2592 |
} |
} |
2593 |
} |
} |
2594 |
|
|
2595 |
|
min = -min; |
2596 |
|
max += 1; |
2597 |
|
if (min > max) max = min; |
2598 |
|
if (pParam->m_quarterpel) max *= 2; |
2599 |
|
|
2600 |
|
for (i = 1; (max > 32 << (i - 1)); i++); |
2601 |
|
return i; |
2602 |
|
} |
2603 |
|
|
2604 |
|
static void |
2605 |
|
CheckGMC(int x, int y, const int dir, int * iDirection, |
2606 |
|
const MACROBLOCK * const pMBs, uint32_t * bestcount, VECTOR * GMC, |
2607 |
|
const MBParam * const pParam) |
2608 |
|
{ |
2609 |
|
uint32_t mx, my, a, count = 0; |
2610 |
|
|
2611 |
|
for (my = 1; my < pParam->mb_height-1; my++) |
2612 |
|
for (mx = 1; mx < pParam->mb_width-1; mx++) { |
2613 |
|
VECTOR mv; |
2614 |
|
const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; |
2615 |
|
if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) continue; |
2616 |
|
mv = pMB->mvs[0]; |
2617 |
|
a = ABS(mv.x - x) + ABS(mv.y - y); |
2618 |
|
if (a < 6) count += 6 - a; |
2619 |
|
} |
2620 |
|
|
2621 |
|
if (count > *bestcount) { |
2622 |
|
*bestcount = count; |
2623 |
|
*iDirection = dir; |
2624 |
|
GMC->x = x; GMC->y = y; |
2625 |
|
} |
2626 |
|
} |
2627 |
|
|
2628 |
|
|
2629 |
|
static VECTOR |
2630 |
|
GlobalMotionEst(const MACROBLOCK * const pMBs, const MBParam * const pParam, const uint32_t iFcode) |
2631 |
|
{ |
2632 |
|
|
2633 |
|
uint32_t count, bestcount = 0; |
2634 |
|
int x, y; |
2635 |
|
VECTOR gmc = {0,0}; |
2636 |
|
int step, min_x, max_x, min_y, max_y; |
2637 |
|
uint32_t mx, my; |
2638 |
|
int iDirection, bDirection; |
2639 |
|
|
2640 |
|
min_x = min_y = -32<<iFcode; |
2641 |
|
max_x = max_y = 32<<iFcode; |
2642 |
|
|
2643 |
|
//step1: let's find a rough camera panning |
2644 |
|
for (step = 32; step >= 2; step /= 2) { |
2645 |
|
bestcount = 0; |
2646 |
|
for (y = min_y; y <= max_y; y += step) |
2647 |
|
for (x = min_x ; x <= max_x; x += step) { |
2648 |
|
count = 0; |
2649 |
|
//for all macroblocks |
2650 |
|
for (my = 1; my < pParam->mb_height-1; my++) |
2651 |
|
for (mx = 1; mx < pParam->mb_width-1; mx++) { |
2652 |
|
const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; |
2653 |
|
VECTOR mv; |
2654 |
|
|
2655 |
|
if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) |
2656 |
|
continue; |
2657 |
|
|
2658 |
|
mv = pMB->mvs[0]; |
2659 |
|
if ( ABS(mv.x - x) <= step && ABS(mv.y - y) <= step ) /* GMC translation is always halfpel-res */ |
2660 |
|
count++; |
2661 |
|
} |
2662 |
|
if (count >= bestcount) { bestcount = count; gmc.x = x; gmc.y = y; } |
2663 |
|
} |
2664 |
|
min_x = gmc.x - step; |
2665 |
|
max_x = gmc.x + step; |
2666 |
|
min_y = gmc.y - step; |
2667 |
|
max_y = gmc.y + step; |
2668 |
|
|
2669 |
|
} |
2670 |
|
|
2671 |
|
if (bestcount < (pParam->mb_height-2)*(pParam->mb_width-2)/10) |
2672 |
|
gmc.x = gmc.y = 0; //no camara pan, no GMC |
2673 |
|
|
2674 |
|
// step2: let's refine camera panning using gradiend-descent approach. |
2675 |
|
// TODO: more warping points may be evaluated here (like in interpolate mode search - two vectors in one diamond) |
2676 |
|
bestcount = 0; |
2677 |
|
CheckGMC(gmc.x, gmc.y, 255, &iDirection, pMBs, &bestcount, &gmc, pParam); |
2678 |
|
do { |
2679 |
|
x = gmc.x; y = gmc.y; |
2680 |
|
bDirection = iDirection; iDirection = 0; |
2681 |
|
if (bDirection & 1) CheckGMC(x - 1, y, 1+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); |
2682 |
|
if (bDirection & 2) CheckGMC(x + 1, y, 2+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); |
2683 |
|
if (bDirection & 4) CheckGMC(x, y - 1, 1+2+4, &iDirection, pMBs, &bestcount, &gmc, pParam); |
2684 |
|
if (bDirection & 8) CheckGMC(x, y + 1, 1+2+8, &iDirection, pMBs, &bestcount, &gmc, pParam); |
2685 |
|
|
2686 |
|
} while (iDirection); |
2687 |
|
|
2688 |
|
if (pParam->m_quarterpel) { |
2689 |
|
gmc.x *= 2; |
2690 |
|
gmc.y *= 2; /* we store the halfpel value as pseudo-qpel to make comparison easier */ |
2691 |
|
} |
2692 |
|
|
2693 |
|
return gmc; |
2694 |
|
} |