Parent Directory | Revision Log
Revision 78 - (view) (download)
1 : | Isibaar | 3 | /************************************************************************** |
2 : | * | ||
3 : | * Modifications: | ||
4 : | * | ||
5 : | * 08.02.2002 split up PMVfast into three routines: PMVFast, PMVFast_MainLoop | ||
6 : | * PMVFast_Refine to support multiple searches with different start points | ||
7 : | edgomez | 78 | * 07.01.2002 uv-block-based interpolation |
8 : | Isibaar | 3 | * 06.01.2002 INTER/INTRA-decision is now done before any SEARCH8 (speedup) |
9 : | edgomez | 78 | * changed INTER_BIAS to 150 (as suggested by suxen_drol) |
10 : | * removed halfpel refinement step in PMVfastSearch8 + quality=5 | ||
11 : | * added new quality mode = 6 which performs halfpel refinement | ||
12 : | * filesize difference between quality 5 and 6 is smaller than 1% | ||
13 : | Isibaar | 3 | * (Isibaar) |
14 : | * 31.12.2001 PMVfastSearch16 and PMVfastSearch8 (gruel) | ||
15 : | edgomez | 78 | * 30.12.2001 get_range/MotionSearchX simplified; blue/green bug fix |
16 : | * 22.12.2001 commented best_point==99 check | ||
17 : | * 19.12.2001 modified get_range (purple bug fix) | ||
18 : | Isibaar | 3 | * 15.12.2001 moved pmv displacement from mbprediction |
19 : | * 02.12.2001 motion estimation/compensation split (Isibaar) | ||
20 : | edgomez | 78 | * 16.11.2001 rewrote/tweaked search algorithms; pross@cs.rmit.edu.au |
21 : | Isibaar | 3 | * 10.11.2001 support for sad16/sad8 functions |
22 : | * 28.08.2001 reactivated MODE_INTER4V for EXT_MODE | ||
23 : | * 24.08.2001 removed MODE_INTER4V_Q, disabled MODE_INTER4V for EXT_MODE | ||
24 : | edgomez | 78 | * 22.08.2001 added MODE_INTER4V_Q |
25 : | Isibaar | 3 | * 20.08.2001 added pragma to get rid of internal compiler error with VC6 |
26 : | * idea by Cyril. Thanks. | ||
27 : | * | ||
28 : | * Michael Militzer <isibaar@videocoding.de> | ||
29 : | * | ||
30 : | **************************************************************************/ | ||
31 : | |||
32 : | #include <assert.h> | ||
33 : | #include <stdio.h> | ||
34 : | |||
35 : | #include "../encoder.h" | ||
36 : | #include "../utils/mbfunctions.h" | ||
37 : | #include "../prediction/mbprediction.h" | ||
38 : | #include "../global.h" | ||
39 : | #include "../utils/timer.h" | ||
40 : | #include "sad.h" | ||
41 : | |||
42 : | // very large value | ||
43 : | #define MV_MAX_ERROR (4096 * 256) | ||
44 : | |||
45 : | // stop search if sdelta < THRESHOLD | ||
46 : | #define MV16_THRESHOLD 192 | ||
47 : | #define MV8_THRESHOLD 56 | ||
48 : | |||
49 : | /* sad16(0,0) bias; mpeg4 spec suggests nb/2+1 */ | ||
50 : | /* nb = vop pixels * 2^(bpp-8) */ | ||
51 : | #define MV16_00_BIAS (128+1) | ||
52 : | |||
53 : | /* INTER bias for INTER/INTRA decision; mpeg4 spec suggests 2*nb */ | ||
54 : | #define INTER_BIAS 512 | ||
55 : | |||
56 : | /* Parameters which control inter/inter4v decision */ | ||
57 : | #define IMV16X16 5 | ||
58 : | |||
59 : | /* vector map (vlc delta size) smoother parameters */ | ||
60 : | #define NEIGH_TEND_16X16 2 | ||
61 : | #define NEIGH_TEND_8X8 2 | ||
62 : | |||
63 : | |||
64 : | // fast ((A)/2)*2 | ||
65 : | #define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) | ||
66 : | |||
67 : | |||
68 : | #define MIN(X, Y) ((X)<(Y)?(X):(Y)) | ||
69 : | #define MAX(X, Y) ((X)>(Y)?(X):(Y)) | ||
70 : | edgomez | 78 | #define ABS(X) (((X)>0)?(X):-(X)) |
71 : | #define SIGN(X) (((X)>0)?1:-1) | ||
72 : | Isibaar | 3 | |
73 : | |||
74 : | int32_t PMVfastSearch8( | ||
75 : | edgomez | 78 | const uint8_t * const pRef, |
76 : | const uint8_t * const pRefH, | ||
77 : | const uint8_t * const pRefV, | ||
78 : | const uint8_t * const pRefHV, | ||
79 : | const IMAGE * const pCur, | ||
80 : | const int x, const int y, | ||
81 : | const int start_x, int start_y, | ||
82 : | const uint32_t iQuality, | ||
83 : | MBParam * const pParam, | ||
84 : | MACROBLOCK * const pMBs, | ||
85 : | VECTOR * const currMV, | ||
86 : | VECTOR * const currPMV); | ||
87 : | Isibaar | 3 | |
88 : | int32_t PMVfastSearch16( | ||
89 : | edgomez | 78 | const uint8_t * const pRef, |
90 : | const uint8_t * const pRefH, | ||
91 : | const uint8_t * const pRefV, | ||
92 : | const uint8_t * const pRefHV, | ||
93 : | const IMAGE * const pCur, | ||
94 : | const int x, const int y, | ||
95 : | const uint32_t iQuality, | ||
96 : | MBParam * const pParam, | ||
97 : | MACROBLOCK * const pMBs, | ||
98 : | VECTOR * const currMV, | ||
99 : | VECTOR * const currPMV); | ||
100 : | Isibaar | 3 | |
101 : | |||
102 : | |||
103 : | edgomez | 78 | /* |
104 : | * diamond search stuff | ||
105 : | * keep the the sequence in circular order (so optimization works) | ||
106 : | */ | ||
107 : | Isibaar | 3 | |
108 : | typedef struct | ||
109 : | { | ||
110 : | int32_t dx; | ||
111 : | int32_t dy; | ||
112 : | } | ||
113 : | DPOINT; | ||
114 : | |||
115 : | |||
116 : | static const DPOINT diamond_small[4] = | ||
117 : | { | ||
118 : | {0, 1}, {1, 0}, {0, -1}, {-1, 0} | ||
119 : | }; | ||
120 : | |||
121 : | |||
122 : | static const DPOINT diamond_large[8] = | ||
123 : | { | ||
124 : | {0, 2}, {1, 1}, {2, 0}, {1, -1}, {0, -2}, {-1, -1}, {-2, 0}, {-1, 1} | ||
125 : | }; | ||
126 : | |||
127 : | |||
128 : | // mv.length table | ||
129 : | static const uint32_t mvtab[33] = { | ||
130 : | 1, 2, 3, 4, 6, 7, 7, 7, | ||
131 : | 9, 9, 9, 10, 10, 10, 10, 10, | ||
132 : | 10, 10, 10, 10, 10, 10, 10, 10, | ||
133 : | 10, 11, 11, 11, 11, 11, 11, 12, 12 | ||
134 : | }; | ||
135 : | |||
136 : | |||
137 : | static __inline uint32_t mv_bits(int32_t component, const uint32_t iFcode) | ||
138 : | { | ||
139 : | if (component == 0) | ||
140 : | return 1; | ||
141 : | |||
142 : | if (component < 0) | ||
143 : | component = -component; | ||
144 : | |||
145 : | if (iFcode == 1) | ||
146 : | { | ||
147 : | if (component > 32) | ||
148 : | component = 32; | ||
149 : | |||
150 : | return mvtab[component] + 1; | ||
151 : | } | ||
152 : | |||
153 : | component += (1 << (iFcode - 1)) - 1; | ||
154 : | component >>= (iFcode - 1); | ||
155 : | |||
156 : | if (component > 32) | ||
157 : | component = 32; | ||
158 : | |||
159 : | return mvtab[component] + 1 + iFcode - 1; | ||
160 : | } | ||
161 : | |||
162 : | |||
163 : | static __inline uint32_t calc_delta_16(const int32_t dx, const int32_t dy, const uint32_t iFcode) | ||
164 : | { | ||
165 : | return NEIGH_TEND_16X16 * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); | ||
166 : | } | ||
167 : | |||
168 : | static __inline uint32_t calc_delta_8(const int32_t dx, const int32_t dy, const uint32_t iFcode) | ||
169 : | |||
170 : | { | ||
171 : | return NEIGH_TEND_8X8 * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); | ||
172 : | } | ||
173 : | |||
174 : | |||
175 : | |||
176 : | |||
177 : | /* calculate the min/max range (in halfpixels) | ||
178 : | relative to the _MACROBLOCK_ position | ||
179 : | */ | ||
180 : | |||
181 : | static void __inline get_range( | ||
182 : | edgomez | 78 | int32_t * const min_dx, int32_t * const max_dx, |
183 : | int32_t * const min_dy, int32_t * const max_dy, | ||
184 : | const uint32_t x, const uint32_t y, | ||
185 : | const uint32_t block_sz, // block dimension, 8 or 16 | ||
186 : | const uint32_t width, const uint32_t height, | ||
187 : | const uint32_t fcode) | ||
188 : | Isibaar | 3 | { |
189 : | edgomez | 78 | |
190 : | Isibaar | 3 | const int search_range = 32 << (fcode - 1); |
191 : | edgomez | 78 | const int high = search_range - 1; |
192 : | const int low = -search_range; | ||
193 : | Isibaar | 3 | |
194 : | // convert full-pixel measurements to half pixel | ||
195 : | const int hp_width = 2 * width; | ||
196 : | const int hp_height = 2 * height; | ||
197 : | const int hp_edge = 2 * block_sz; | ||
198 : | const int hp_x = 2 * (x) * block_sz; // we need _right end_ of block, not x-coordinate | ||
199 : | const int hp_y = 2 * (y) * block_sz; // same for _bottom end_ | ||
200 : | |||
201 : | edgomez | 78 | *max_dx = MIN(high, hp_width - hp_x); |
202 : | *max_dy = MIN(high, hp_height - hp_y); | ||
203 : | *min_dx = MAX(low, -(hp_edge + hp_x)); | ||
204 : | *min_dy = MAX(low, -(hp_edge + hp_y)); | ||
205 : | |||
206 : | Isibaar | 3 | } |
207 : | |||
208 : | |||
209 : | edgomez | 78 | /* |
210 : | * getref: calculate reference image pointer | ||
211 : | * the decision to use interpolation h/v/hv or the normal image is | ||
212 : | * based on dx & dy. | ||
213 : | */ | ||
214 : | Isibaar | 3 | |
215 : | static __inline const uint8_t * get_ref( | ||
216 : | edgomez | 78 | const uint8_t * const refn, |
217 : | const uint8_t * const refh, | ||
218 : | const uint8_t * const refv, | ||
219 : | const uint8_t * const refhv, | ||
220 : | const uint32_t x, const uint32_t y, | ||
221 : | const uint32_t block, // block dimension, 8 or 16 | ||
222 : | const int32_t dx, const int32_t dy, | ||
223 : | const uint32_t stride) | ||
224 : | Isibaar | 3 | { |
225 : | edgomez | 78 | |
226 : | Isibaar | 3 | switch ( ((dx&1)<<1) + (dy&1) ) // ((dx%2)?2:0)+((dy%2)?1:0) |
227 : | edgomez | 78 | { |
228 : | case 0 : return refn + (x*block+dx/2) + (y*block+dy/2)*stride; | ||
229 : | case 1 : return refv + (x*block+dx/2) + (y*block+(dy-1)/2)*stride; | ||
230 : | case 2 : return refh + (x*block+(dx-1)/2) + (y*block+dy/2)*stride; | ||
231 : | Isibaar | 3 | default : |
232 : | edgomez | 78 | case 3 : return refhv + (x*block+(dx-1)/2) + (y*block+(dy-1)/2)*stride; |
233 : | Isibaar | 3 | } |
234 : | edgomez | 78 | |
235 : | Isibaar | 3 | } |
236 : | |||
237 : | |||
238 : | /* This is somehow a copy of get_ref, but with MV instead of X,Y */ | ||
239 : | |||
240 : | static __inline const uint8_t * get_ref_mv( | ||
241 : | edgomez | 78 | const uint8_t * const refn, |
242 : | const uint8_t * const refh, | ||
243 : | const uint8_t * const refv, | ||
244 : | const uint8_t * const refhv, | ||
245 : | const uint32_t x, const uint32_t y, | ||
246 : | const uint32_t block, // block dimension, 8 or 16 | ||
247 : | const VECTOR* mv, // measured in half-pel! | ||
248 : | const uint32_t stride) | ||
249 : | Isibaar | 3 | { |
250 : | edgomez | 78 | |
251 : | Isibaar | 3 | switch ( (((mv->x)&1)<<1) + ((mv->y)&1) ) |
252 : | edgomez | 78 | { |
253 : | case 0 : return refn + (x*block+(mv->x)/2) + (y*block+(mv->y)/2)*stride; | ||
254 : | case 1 : return refv + (x*block+(mv->x)/2) + (y*block+((mv->y)-1)/2)*stride; | ||
255 : | case 2 : return refh + (x*block+((mv->x)-1)/2) + (y*block+(mv->y)/2)*stride; | ||
256 : | Isibaar | 3 | default : |
257 : | edgomez | 78 | case 3 : return refhv + (x*block+((mv->x)-1)/2) + (y*block+((mv->y)-1)/2)*stride; |
258 : | Isibaar | 3 | } |
259 : | edgomez | 78 | |
260 : | Isibaar | 3 | } |
261 : | |||
262 : | #ifndef SEARCH16 | ||
263 : | #define SEARCH16 PMVfastSearch16 | ||
264 : | #endif | ||
265 : | |||
266 : | #ifndef SEARCH8 | ||
267 : | #define SEARCH8 PMVfastSearch8 | ||
268 : | #endif | ||
269 : | |||
270 : | bool MotionEstimation( | ||
271 : | edgomez | 78 | MACROBLOCK * const pMBs, |
272 : | MBParam * const pParam, | ||
273 : | const IMAGE * const pRef, | ||
274 : | const IMAGE * const pRefH, | ||
275 : | const IMAGE * const pRefV, | ||
276 : | const IMAGE * const pRefHV, | ||
277 : | IMAGE * const pCurrent, | ||
278 : | const uint32_t iLimit) | ||
279 : | Isibaar | 3 | |
280 : | { | ||
281 : | edgomez | 78 | const uint32_t iWcount = pParam->mb_width; |
282 : | const uint32_t iHcount = pParam->mb_height; | ||
283 : | Isibaar | 3 | |
284 : | uint32_t i, j, iIntra = 0; | ||
285 : | |||
286 : | edgomez | 78 | VECTOR mv16; |
287 : | VECTOR pmv16; | ||
288 : | Isibaar | 3 | |
289 : | edgomez | 78 | int32_t sad8 = 0; |
290 : | int32_t sad16; | ||
291 : | int32_t deviation; | ||
292 : | Isibaar | 3 | |
293 : | // note: i==horizontal, j==vertical | ||
294 : | edgomez | 78 | for (i = 0; i < iHcount; i++) |
295 : | Isibaar | 3 | for (j = 0; j < iWcount; j++) |
296 : | { | ||
297 : | MACROBLOCK *pMB = &pMBs[j + i * iWcount]; | ||
298 : | |||
299 : | sad16 = SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, | ||
300 : | edgomez | 78 | j, i, pParam->motion_flags, |
301 : | pParam, pMBs, &mv16, &pmv16); | ||
302 : | Isibaar | 3 | pMB->sad16=sad16; |
303 : | |||
304 : | |||
305 : | edgomez | 78 | /* decide: MODE_INTER or MODE_INTRA |
306 : | if (dev_intra < sad_inter - 2 * nb) use_intra | ||
307 : | */ | ||
308 : | Isibaar | 3 | |
309 : | edgomez | 78 | deviation = dev16(pCurrent->y + j*16 + i*16*pParam->edged_width, pParam->edged_width); |
310 : | Isibaar | 3 | |
311 : | edgomez | 78 | if (deviation < (sad16 - INTER_BIAS)) |
312 : | { | ||
313 : | pMB->mode = MODE_INTRA; | ||
314 : | pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; | ||
315 : | pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; | ||
316 : | Isibaar | 3 | |
317 : | edgomez | 78 | iIntra++; |
318 : | if(iIntra >= iLimit) | ||
319 : | return 1; | ||
320 : | Isibaar | 3 | |
321 : | edgomez | 78 | continue; |
322 : | } | ||
323 : | Isibaar | 3 | |
324 : | edgomez | 78 | if (pParam->global_flags & XVID_INTER4V) |
325 : | { | ||
326 : | pMB->sad8[0] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, | ||
327 : | 2 * j, 2 * i, mv16.x, mv16.y, pParam->motion_flags, | ||
328 : | pParam, pMBs, &pMB->mvs[0], &pMB->pmvs[0]); | ||
329 : | Isibaar | 3 | |
330 : | edgomez | 78 | pMB->sad8[1] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
331 : | 2 * j + 1, 2 * i, mv16.x, mv16.y, pParam->motion_flags, | ||
332 : | pParam, pMBs, &pMB->mvs[1], &pMB->pmvs[1]); | ||
333 : | Isibaar | 3 | |
334 : | edgomez | 78 | pMB->sad8[2] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
335 : | 2 * j, 2 * i + 1, mv16.x, mv16.y, pParam->motion_flags, | ||
336 : | pParam, pMBs, &pMB->mvs[2], &pMB->pmvs[2]); | ||
337 : | Isibaar | 3 | |
338 : | edgomez | 78 | pMB->sad8[3] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
339 : | 2 * j + 1, 2 * i + 1, mv16.x, mv16.y, pParam->motion_flags, | ||
340 : | pParam, pMBs, &pMB->mvs[3], &pMB->pmvs[3]); | ||
341 : | Isibaar | 3 | |
342 : | edgomez | 78 | sad8 = pMB->sad8[0] + pMB->sad8[1] + pMB->sad8[2] + pMB->sad8[3]; |
343 : | } | ||
344 : | Isibaar | 3 | |
345 : | |||
346 : | edgomez | 78 | /* decide: MODE_INTER or MODE_INTER4V |
347 : | mpeg4: if (sad8 < sad16 - nb/2+1) use_inter4v | ||
348 : | */ | ||
349 : | Isibaar | 3 | |
350 : | edgomez | 78 | if (pMB->dquant == NO_CHANGE) { |
351 : | if (((pParam->global_flags & XVID_INTER4V)==0) || | ||
352 : | (sad16 < (sad8 + (int32_t)(IMV16X16 * pParam->quant)))) { | ||
353 : | Isibaar | 3 | |
354 : | edgomez | 78 | sad8 = sad16; |
355 : | pMB->mode = MODE_INTER; | ||
356 : | pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; | ||
357 : | pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; | ||
358 : | pMB->pmvs[0].x = pmv16.x; | ||
359 : | pMB->pmvs[0].y = pmv16.y; | ||
360 : | } | ||
361 : | else | ||
362 : | pMB->mode = MODE_INTER4V; | ||
363 : | } | ||
364 : | else | ||
365 : | { | ||
366 : | Isibaar | 3 | sad8 = sad16; |
367 : | pMB->mode = MODE_INTER; | ||
368 : | pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; | ||
369 : | pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; | ||
370 : | pMB->pmvs[0].x = pmv16.x; | ||
371 : | pMB->pmvs[0].y = pmv16.y; | ||
372 : | } | ||
373 : | } | ||
374 : | |||
375 : | return 0; | ||
376 : | } | ||
377 : | |||
378 : | #define MVzero(A) ( ((A).x)==(0) && ((A).y)==(0) ) | ||
379 : | |||
380 : | #define MVequal(A,B) ( ((A).x)==((B).x) && ((A).y)==((B).y) ) | ||
381 : | |||
382 : | |||
383 : | #define CHECK_MV16_ZERO {\ | ||
384 : | if ( (0 <= max_dx) && (0 >= min_dx) \ | ||
385 : | && (0 <= max_dy) && (0 >= min_dy) ) \ | ||
386 : | { \ | ||
387 : | iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0, 0 , iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); \ | ||
388 : | iSAD += calc_delta_16(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode) * iQuant;\ | ||
389 : | if (iSAD <= iQuant * 96) \ | ||
390 : | iSAD -= MV16_00_BIAS; \ | ||
391 : | if (iSAD < iMinSAD) \ | ||
392 : | { iMinSAD=iSAD; currMV->x=0; currMV->y=0; } } \ | ||
393 : | } | ||
394 : | |||
395 : | |||
396 : | #define CHECK_MV16_CANDIDATE(X,Y) { \ | ||
397 : | if ( ((X) <= max_dx) && ((X) >= min_dx) \ | ||
398 : | && ((Y) <= max_dy) && ((Y) >= min_dy) ) \ | ||
399 : | { \ | ||
400 : | iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ | ||
401 : | iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ | ||
402 : | if (iSAD < iMinSAD) \ | ||
403 : | { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ | ||
404 : | } | ||
405 : | |||
406 : | #define CHECK_MV16_CANDIDATE_DIR(X,Y,D) { \ | ||
407 : | if ( ((X) <= max_dx) && ((X) >= min_dx) \ | ||
408 : | && ((Y) <= max_dy) && ((Y) >= min_dy) ) \ | ||
409 : | { \ | ||
410 : | iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ | ||
411 : | iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ | ||
412 : | if (iSAD < iMinSAD) \ | ||
413 : | { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ | ||
414 : | } | ||
415 : | |||
416 : | #define CHECK_MV16_CANDIDATE_FOUND(X,Y,D) { \ | ||
417 : | if ( ((X) <= max_dx) && ((X) >= min_dx) \ | ||
418 : | && ((Y) <= max_dy) && ((Y) >= min_dy) ) \ | ||
419 : | { \ | ||
420 : | iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ | ||
421 : | iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ | ||
422 : | if (iSAD < iMinSAD) \ | ||
423 : | { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ | ||
424 : | } | ||
425 : | |||
426 : | |||
427 : | #define CHECK_MV8_ZERO {\ | ||
428 : | iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, 0, 0 , iEdgedWidth), iEdgedWidth); \ | ||
429 : | iSAD += calc_delta_8(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode) * iQuant;\ | ||
430 : | if (iSAD < iMinSAD) \ | ||
431 : | { iMinSAD=iSAD; currMV->x=0; currMV->y=0; } \ | ||
432 : | } | ||
433 : | |||
434 : | |||
435 : | #define CHECK_MV8_CANDIDATE(X,Y) { \ | ||
436 : | if ( ((X) <= max_dx) && ((X) >= min_dx) \ | ||
437 : | && ((Y) <= max_dy) && ((Y) >= min_dy) ) \ | ||
438 : | { \ | ||
439 : | iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ | ||
440 : | iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ | ||
441 : | if (iSAD < iMinSAD) \ | ||
442 : | { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ | ||
443 : | } | ||
444 : | |||
445 : | #define CHECK_MV8_CANDIDATE_DIR(X,Y,D) { \ | ||
446 : | if ( ((X) <= max_dx) && ((X) >= min_dx) \ | ||
447 : | && ((Y) <= max_dy) && ((Y) >= min_dy) ) \ | ||
448 : | { \ | ||
449 : | iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ | ||
450 : | iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ | ||
451 : | if (iSAD < iMinSAD) \ | ||
452 : | { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ | ||
453 : | } | ||
454 : | |||
455 : | #define CHECK_MV8_CANDIDATE_FOUND(X,Y,D) { \ | ||
456 : | if ( ((X) <= max_dx) && ((X) >= min_dx) \ | ||
457 : | && ((Y) <= max_dy) && ((Y) >= min_dy) ) \ | ||
458 : | { \ | ||
459 : | iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ | ||
460 : | iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ | ||
461 : | if (iSAD < iMinSAD) \ | ||
462 : | { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ | ||
463 : | } | ||
464 : | |||
465 : | /* too slow and not fully functional at the moment */ | ||
466 : | /* | ||
467 : | int32_t ZeroSearch16( | ||
468 : | const uint8_t * const pRef, | ||
469 : | const uint8_t * const pRefH, | ||
470 : | const uint8_t * const pRefV, | ||
471 : | const uint8_t * const pRefHV, | ||
472 : | const IMAGE * const pCur, | ||
473 : | const int x, const int y, | ||
474 : | const uint32_t MotionFlags, | ||
475 : | MBParam * const pParam, | ||
476 : | MACROBLOCK * const pMBs, | ||
477 : | VECTOR * const currMV, | ||
478 : | VECTOR * const currPMV) | ||
479 : | { | ||
480 : | const int32_t iEdgedWidth = pParam->edged_width; | ||
481 : | const int32_t iQuant = pParam->quant; | ||
482 : | const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; | ||
483 : | int32_t iSAD; | ||
484 : | int32_t pred_x,pred_y; | ||
485 : | |||
486 : | get_pmv(pMBs, x, y, pParam->mb_width, 0, &pred_x, &pred_y); | ||
487 : | |||
488 : | iSAD = sad16( cur, | ||
489 : | get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0,0, iEdgedWidth), | ||
490 : | iEdgedWidth, MV_MAX_ERROR); | ||
491 : | if (iSAD <= iQuant * 96) | ||
492 : | iSAD -= MV16_00_BIAS; | ||
493 : | |||
494 : | currMV->x = 0; | ||
495 : | currMV->y = 0; | ||
496 : | currPMV->x = -pred_x; | ||
497 : | currPMV->y = -pred_y; | ||
498 : | |||
499 : | return iSAD; | ||
500 : | |||
501 : | } | ||
502 : | */ | ||
503 : | |||
504 : | int32_t PMVfastSearch16_MainSearch( | ||
505 : | edgomez | 78 | const uint8_t * const pRef, |
506 : | const uint8_t * const pRefH, | ||
507 : | const uint8_t * const pRefV, | ||
508 : | const uint8_t * const pRefHV, | ||
509 : | const uint8_t * const cur, | ||
510 : | const int x, const int y, | ||
511 : | int32_t startx, int32_t starty, | ||
512 : | int32_t iMinSAD, | ||
513 : | VECTOR * const currMV, | ||
514 : | const VECTOR * const pmv, | ||
515 : | const int32_t min_dx, const int32_t max_dx, | ||
516 : | const int32_t min_dy, const int32_t max_dy, | ||
517 : | const int32_t iEdgedWidth, | ||
518 : | const int32_t iDiamondSize, | ||
519 : | const int32_t iFcode, | ||
520 : | const int32_t iQuant, | ||
521 : | int iFound) | ||
522 : | Isibaar | 3 | { |
523 : | /* Do a diamond search around given starting point, return SAD of best */ | ||
524 : | |||
525 : | int32_t iDirection=0; | ||
526 : | int32_t iSAD; | ||
527 : | VECTOR backupMV; | ||
528 : | backupMV.x = startx; | ||
529 : | backupMV.y = starty; | ||
530 : | |||
531 : | /* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ | ||
532 : | |||
533 : | CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); | ||
534 : | CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); | ||
535 : | CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); | ||
536 : | CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); | ||
537 : | |||
538 : | if (iDirection) | ||
539 : | while (!iFound) | ||
540 : | { | ||
541 : | iFound = 1; | ||
542 : | backupMV=*currMV; | ||
543 : | |||
544 : | if ( iDirection != 2) | ||
545 : | CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); | ||
546 : | if ( iDirection != 1) | ||
547 : | CHECK_MV16_CANDIDATE_FOUND(backupMV.x+iDiamondSize,backupMV.y,2); | ||
548 : | if ( iDirection != 4) | ||
549 : | CHECK_MV16_CANDIDATE_FOUND(backupMV.x,backupMV.y-iDiamondSize,3); | ||
550 : | if ( iDirection != 3) | ||
551 : | CHECK_MV16_CANDIDATE_FOUND(backupMV.x,backupMV.y+iDiamondSize,4); | ||
552 : | } | ||
553 : | else | ||
554 : | edgomez | 78 | { |
555 : | currMV->x = startx; | ||
556 : | currMV->y = starty; | ||
557 : | } | ||
558 : | Isibaar | 3 | return iMinSAD; |
559 : | } | ||
560 : | |||
561 : | int32_t PMVfastSearch16_Refine( | ||
562 : | edgomez | 78 | const uint8_t * const pRef, |
563 : | const uint8_t * const pRefH, | ||
564 : | const uint8_t * const pRefV, | ||
565 : | const uint8_t * const pRefHV, | ||
566 : | const uint8_t * const cur, | ||
567 : | const int x, const int y, | ||
568 : | VECTOR * const currMV, | ||
569 : | int32_t iMinSAD, | ||
570 : | const VECTOR * const pmv, | ||
571 : | const int32_t min_dx, const int32_t max_dx, | ||
572 : | const int32_t min_dy, const int32_t max_dy, | ||
573 : | const int32_t iFcode, | ||
574 : | const int32_t iQuant, | ||
575 : | const int32_t iEdgedWidth) | ||
576 : | Isibaar | 3 | { |
577 : | /* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ | ||
578 : | |||
579 : | int32_t iSAD; | ||
580 : | VECTOR backupMV = *currMV; | ||
581 : | |||
582 : | CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y-1); | ||
583 : | CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y-1); | ||
584 : | CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y-1); | ||
585 : | CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y); | ||
586 : | CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y); | ||
587 : | CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y+1); | ||
588 : | CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y+1); | ||
589 : | CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y+1); | ||
590 : | |||
591 : | return iMinSAD; | ||
592 : | } | ||
593 : | |||
594 : | #define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) | ||
595 : | |||
596 : | int32_t PMVfastSearch16( | ||
597 : | edgomez | 78 | const uint8_t * const pRef, |
598 : | const uint8_t * const pRefH, | ||
599 : | const uint8_t * const pRefV, | ||
600 : | const uint8_t * const pRefHV, | ||
601 : | const IMAGE * const pCur, | ||
602 : | const int x, const int y, | ||
603 : | const uint32_t MotionFlags, | ||
604 : | MBParam * const pParam, | ||
605 : | MACROBLOCK * const pMBs, | ||
606 : | VECTOR * const currMV, | ||
607 : | VECTOR * const currPMV) | ||
608 : | Isibaar | 3 | { |
609 : | const uint32_t iWcount = pParam->mb_width; | ||
610 : | const int32_t iFcode = pParam->fixed_code; | ||
611 : | const int32_t iQuant = pParam->quant; | ||
612 : | const int32_t iWidth = pParam->width; | ||
613 : | const int32_t iHeight = pParam->height; | ||
614 : | const int32_t iEdgedWidth = pParam->edged_width; | ||
615 : | |||
616 : | const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; | ||
617 : | |||
618 : | int32_t iDiamondSize; | ||
619 : | |||
620 : | int32_t min_dx; | ||
621 : | int32_t max_dx; | ||
622 : | int32_t min_dy; | ||
623 : | int32_t max_dy; | ||
624 : | |||
625 : | int32_t iFound; | ||
626 : | |||
627 : | VECTOR newMV; | ||
628 : | VECTOR backupMV; /* just for PMVFAST */ | ||
629 : | |||
630 : | VECTOR pmv[4]; | ||
631 : | int32_t psad[4]; | ||
632 : | |||
633 : | MACROBLOCK * const pMB = pMBs + x + y * iWcount; | ||
634 : | |||
635 : | static int32_t threshA,threshB; | ||
636 : | int32_t bPredEq; | ||
637 : | int32_t iMinSAD,iSAD; | ||
638 : | |||
639 : | /* Get maximum range */ | ||
640 : | get_range(&min_dx, &max_dx, &min_dy, &max_dy, | ||
641 : | edgomez | 78 | x, y, 16, iWidth, iHeight, iFcode); |
642 : | Isibaar | 3 | |
643 : | /* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ | ||
644 : | |||
645 : | if (!(MotionFlags & PMV_HALFPEL16 )) | ||
646 : | { min_dx = EVEN(min_dx); | ||
647 : | edgomez | 78 | max_dx = EVEN(max_dx); |
648 : | min_dy = EVEN(min_dy); | ||
649 : | max_dy = EVEN(max_dy); | ||
650 : | Isibaar | 3 | } /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
651 : | |||
652 : | |||
653 : | bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); | ||
654 : | |||
655 : | if ((x==0) && (y==0) ) | ||
656 : | { | ||
657 : | threshA = 512; | ||
658 : | threshB = 1024; | ||
659 : | |||
660 : | } | ||
661 : | else | ||
662 : | { | ||
663 : | threshA = psad[0]; | ||
664 : | threshB = threshA+256; | ||
665 : | if (threshA< 512) threshA = 512; | ||
666 : | if (threshA>1024) threshA = 1024; | ||
667 : | if (threshB>1792) threshB = 1792; | ||
668 : | } | ||
669 : | |||
670 : | iFound=0; | ||
671 : | |||
672 : | /* Step 2: Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion | ||
673 : | edgomez | 78 | vector of the median. |
674 : | If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 | ||
675 : | Isibaar | 3 | */ |
676 : | |||
677 : | if ((bPredEq) && (MVequal(pmv[0],pMB->mvs[0]) ) ) | ||
678 : | iFound=2; | ||
679 : | |||
680 : | /* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. | ||
681 : | edgomez | 78 | Otherwise select large Diamond Search. |
682 : | Isibaar | 3 | */ |
683 : | |||
684 : | if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536) || (bPredEq) ) | ||
685 : | iDiamondSize=1; // halfpel! | ||
686 : | else | ||
687 : | iDiamondSize=2; // halfpel! | ||
688 : | |||
689 : | if (!(MotionFlags & PMV_HALFPELDIAMOND16) ) | ||
690 : | iDiamondSize*=2; | ||
691 : | |||
692 : | /* Step 4: Calculate SAD around the Median prediction. | ||
693 : | edgomez | 78 | MinSAD=SAD |
694 : | If Motion Vector equal to Previous frame motion vector | ||
695 : | and MinSAD<PrevFrmSAD goto Step 10. | ||
696 : | If SAD<=256 goto Step 10. | ||
697 : | Isibaar | 3 | */ |
698 : | |||
699 : | |||
700 : | // Prepare for main loop | ||
701 : | |||
702 : | *currMV=pmv[0]; /* current best := prediction */ | ||
703 : | if (!(MotionFlags & PMV_HALFPEL16 )) | ||
704 : | { /* This should NOT be necessary! */ | ||
705 : | currMV->x = EVEN(currMV->x); | ||
706 : | currMV->y = EVEN(currMV->y); | ||
707 : | } | ||
708 : | |||
709 : | if (currMV->x > max_dx) | ||
710 : | edgomez | 78 | { |
711 : | currMV->x=max_dx; | ||
712 : | } | ||
713 : | Isibaar | 3 | if (currMV->x < min_dx) |
714 : | edgomez | 78 | { |
715 : | currMV->x=min_dx; | ||
716 : | } | ||
717 : | Isibaar | 3 | if (currMV->y > max_dy) |
718 : | edgomez | 78 | { |
719 : | currMV->y=max_dy; | ||
720 : | } | ||
721 : | Isibaar | 3 | if (currMV->y < min_dy) |
722 : | edgomez | 78 | { |
723 : | currMV->y=min_dy; | ||
724 : | } | ||
725 : | Isibaar | 3 | |
726 : | iMinSAD = sad16( cur, | ||
727 : | edgomez | 78 | get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
728 : | iEdgedWidth, MV_MAX_ERROR); | ||
729 : | Isibaar | 3 | iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; |
730 : | |||
731 : | if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,pMB->mvs[0])) && (iMinSAD < pMB->sad16) ) ) | ||
732 : | edgomez | 78 | { |
733 : | Isibaar | 3 | |
734 : | edgomez | 78 | if (MotionFlags & PMV_QUICKSTOP16) |
735 : | goto step10b; | ||
736 : | if (MotionFlags & PMV_EARLYSTOP16) | ||
737 : | goto step10; | ||
738 : | } | ||
739 : | Isibaar | 3 | |
740 : | /* | ||
741 : | edgomez | 78 | Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
742 : | Also calculate (0,0) but do not subtract offset. | ||
743 : | Let MinSAD be the smallest SAD up to this point. | ||
744 : | If MV is (0,0) subtract offset. ******** WHAT'S THIS 'OFFSET' ??? *********** | ||
745 : | Isibaar | 3 | */ |
746 : | |||
747 : | // (0,0) is always possible | ||
748 : | |||
749 : | CHECK_MV16_ZERO; | ||
750 : | |||
751 : | // previous frame MV is always possible | ||
752 : | CHECK_MV16_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); | ||
753 : | |||
754 : | // left neighbour, if allowed | ||
755 : | if (x != 0) | ||
756 : | { | ||
757 : | if (!(MotionFlags & PMV_HALFPEL16 )) | ||
758 : | { pmv[1].x = EVEN(pmv[1].x); | ||
759 : | edgomez | 78 | pmv[1].y = EVEN(pmv[1].y); |
760 : | Isibaar | 3 | } |
761 : | CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); | ||
762 : | } | ||
763 : | |||
764 : | // top neighbour, if allowed | ||
765 : | if (y != 0) | ||
766 : | { | ||
767 : | if (!(MotionFlags & PMV_HALFPEL16 )) | ||
768 : | { pmv[2].x = EVEN(pmv[2].x); | ||
769 : | edgomez | 78 | pmv[2].y = EVEN(pmv[2].y); |
770 : | Isibaar | 3 | } |
771 : | CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); | ||
772 : | |||
773 : | // top right neighbour, if allowed | ||
774 : | if (x != (iWcount-1)) | ||
775 : | { | ||
776 : | if (!(MotionFlags & PMV_HALFPEL16 )) | ||
777 : | { pmv[3].x = EVEN(pmv[3].x); | ||
778 : | edgomez | 78 | pmv[3].y = EVEN(pmv[3].y); |
779 : | Isibaar | 3 | } |
780 : | CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); | ||
781 : | } | ||
782 : | } | ||
783 : | |||
784 : | /* Step 6: If MinSAD <= thresa goto Step 10. | ||
785 : | If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. | ||
786 : | */ | ||
787 : | |||
788 : | if ( (iMinSAD <= threshA) || ( MVequal(*currMV,pMB->mvs[0]) && (iMinSAD < pMB->sad16) ) ) | ||
789 : | edgomez | 78 | { |
790 : | if (MotionFlags & PMV_QUICKSTOP16) | ||
791 : | goto step10b; | ||
792 : | if (MotionFlags & PMV_EARLYSTOP16) | ||
793 : | goto step10; | ||
794 : | } | ||
795 : | Isibaar | 3 | |
796 : | |||
797 : | /************ (Diamond Search) **************/ | ||
798 : | /* | ||
799 : | edgomez | 78 | Step 7: Perform Diamond search, with either the small or large diamond. |
800 : | If Found=2 only examine one Diamond pattern, and afterwards goto step 10 | ||
801 : | Step 8: If small diamond, iterate small diamond search pattern until motion vector lies in the center of the diamond. | ||
802 : | If center then goto step 10. | ||
803 : | Step 9: If large diamond, iterate large diamond search pattern until motion vector lies in the center. | ||
804 : | Refine by using small diamond and goto step 10. | ||
805 : | Isibaar | 3 | */ |
806 : | |||
807 : | backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ | ||
808 : | |||
809 : | /* default: use best prediction as starting point for one call of PMVfast_MainSearch */ | ||
810 : | iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, | ||
811 : | edgomez | 78 | x, y, |
812 : | currMV->x, currMV->y, iMinSAD, &newMV, | ||
813 : | pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); | ||
814 : | Isibaar | 3 | |
815 : | if (iSAD < iMinSAD) | ||
816 : | { | ||
817 : | *currMV = newMV; | ||
818 : | iMinSAD = iSAD; | ||
819 : | } | ||
820 : | |||
821 : | if (MotionFlags & PMV_EXTSEARCH16) | ||
822 : | { | ||
823 : | /* extended: search (up to) two more times: orignal prediction and (0,0) */ | ||
824 : | |||
825 : | if (!(MVequal(pmv[0],backupMV)) ) | ||
826 : | { iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, | ||
827 : | edgomez | 78 | x, y, |
828 : | pmv[0].x, pmv[0].y, iMinSAD, &newMV, | ||
829 : | pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); | ||
830 : | Isibaar | 3 | |
831 : | edgomez | 78 | if (iSAD < iMinSAD) |
832 : | { | ||
833 : | *currMV = newMV; | ||
834 : | iMinSAD = iSAD; | ||
835 : | Isibaar | 3 | } |
836 : | edgomez | 78 | } |
837 : | Isibaar | 3 | |
838 : | if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) | ||
839 : | { iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, | ||
840 : | edgomez | 78 | x, y, |
841 : | 0, 0, iMinSAD, &newMV, | ||
842 : | pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); | ||
843 : | Isibaar | 3 | |
844 : | edgomez | 78 | if (iSAD < iMinSAD) |
845 : | { | ||
846 : | *currMV = newMV; | ||
847 : | iMinSAD = iSAD; | ||
848 : | Isibaar | 3 | } |
849 : | edgomez | 78 | } |
850 : | Isibaar | 3 | } |
851 : | |||
852 : | /* | ||
853 : | edgomez | 78 | Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
854 : | Isibaar | 3 | */ |
855 : | |||
856 : | edgomez | 78 | step10: |
857 : | Isibaar | 3 | if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
858 : | iMinSAD = PMVfastSearch16_Refine( pRef, pRefH, pRefV, pRefHV, cur, | ||
859 : | edgomez | 78 | x, y, |
860 : | currMV, iMinSAD, | ||
861 : | pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); | ||
862 : | Isibaar | 3 | |
863 : | edgomez | 78 | step10b: |
864 : | Isibaar | 3 | currPMV->x = currMV->x - pmv[0].x; |
865 : | currPMV->y = currMV->y - pmv[0].y; | ||
866 : | return iMinSAD; | ||
867 : | } | ||
868 : | |||
869 : | |||
870 : | |||
871 : | |||
872 : | |||
873 : | |||
874 : | int32_t PMVfastSearch8_MainSearch( | ||
875 : | edgomez | 78 | const uint8_t * const pRef, |
876 : | const uint8_t * const pRefH, | ||
877 : | const uint8_t * const pRefV, | ||
878 : | const uint8_t * const pRefHV, | ||
879 : | const uint8_t * const cur, | ||
880 : | const int x, const int y, | ||
881 : | int32_t startx, int32_t starty, | ||
882 : | int32_t iMinSAD, | ||
883 : | VECTOR * const currMV, | ||
884 : | const VECTOR * const pmv, | ||
885 : | const int32_t min_dx, const int32_t max_dx, | ||
886 : | const int32_t min_dy, const int32_t max_dy, | ||
887 : | const int32_t iEdgedWidth, | ||
888 : | const int32_t iDiamondSize, | ||
889 : | const int32_t iFcode, | ||
890 : | const int32_t iQuant, | ||
891 : | int iFound) | ||
892 : | Isibaar | 3 | { |
893 : | /* Do a diamond search around given starting point, return SAD of best */ | ||
894 : | |||
895 : | int32_t iDirection=0; | ||
896 : | int32_t iSAD; | ||
897 : | VECTOR backupMV; | ||
898 : | backupMV.x = startx; | ||
899 : | backupMV.y = starty; | ||
900 : | |||
901 : | /* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ | ||
902 : | |||
903 : | CHECK_MV8_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); | ||
904 : | CHECK_MV8_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); | ||
905 : | CHECK_MV8_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); | ||
906 : | CHECK_MV8_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); | ||
907 : | |||
908 : | if (iDirection) | ||
909 : | while (!iFound) | ||
910 : | { | ||
911 : | iFound = 1; | ||
912 : | backupMV=*currMV; // since iDirection!=0, this is well defined! | ||
913 : | |||
914 : | if ( iDirection != 2) | ||
915 : | CHECK_MV8_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); | ||
916 : | if ( iDirection != 1) | ||
917 : | CHECK_MV8_CANDIDATE_FOUND(backupMV.x+iDiamondSize,backupMV.y,2); | ||
918 : | if ( iDirection != 4) | ||
919 : | CHECK_MV8_CANDIDATE_FOUND(backupMV.x,backupMV.y-iDiamondSize,3); | ||
920 : | if ( iDirection != 3) | ||
921 : | CHECK_MV8_CANDIDATE_FOUND(backupMV.x,backupMV.y+iDiamondSize,4); | ||
922 : | } | ||
923 : | else | ||
924 : | edgomez | 78 | { |
925 : | currMV->x = startx; | ||
926 : | currMV->y = starty; | ||
927 : | } | ||
928 : | Isibaar | 3 | return iMinSAD; |
929 : | } | ||
930 : | |||
931 : | int32_t PMVfastSearch8_Refine( | ||
932 : | edgomez | 78 | const uint8_t * const pRef, |
933 : | const uint8_t * const pRefH, | ||
934 : | const uint8_t * const pRefV, | ||
935 : | const uint8_t * const pRefHV, | ||
936 : | const uint8_t * const cur, | ||
937 : | const int x, const int y, | ||
938 : | VECTOR * const currMV, | ||
939 : | int32_t iMinSAD, | ||
940 : | const VECTOR * const pmv, | ||
941 : | const int32_t min_dx, const int32_t max_dx, | ||
942 : | const int32_t min_dy, const int32_t max_dy, | ||
943 : | const int32_t iFcode, | ||
944 : | const int32_t iQuant, | ||
945 : | const int32_t iEdgedWidth) | ||
946 : | Isibaar | 3 | { |
947 : | /* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ | ||
948 : | |||
949 : | int32_t iSAD; | ||
950 : | VECTOR backupMV = *currMV; | ||
951 : | |||
952 : | CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y-1); | ||
953 : | CHECK_MV8_CANDIDATE(backupMV.x ,backupMV.y-1); | ||
954 : | CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y-1); | ||
955 : | CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y); | ||
956 : | CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y); | ||
957 : | CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y+1); | ||
958 : | CHECK_MV8_CANDIDATE(backupMV.x ,backupMV.y+1); | ||
959 : | CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y+1); | ||
960 : | |||
961 : | return iMinSAD; | ||
962 : | } | ||
963 : | |||
964 : | |||
965 : | #define PMV_HALFPEL8 (PMV_HALFPELDIAMOND8|PMV_HALFPELREFINE8) | ||
966 : | |||
967 : | int32_t PMVfastSearch8( | ||
968 : | edgomez | 78 | const uint8_t * const pRef, |
969 : | const uint8_t * const pRefH, | ||
970 : | const uint8_t * const pRefV, | ||
971 : | const uint8_t * const pRefHV, | ||
972 : | const IMAGE * const pCur, | ||
973 : | const int x, const int y, | ||
974 : | const int start_x, int start_y, | ||
975 : | const uint32_t MotionFlags, | ||
976 : | MBParam * const pParam, | ||
977 : | MACROBLOCK * const pMBs, | ||
978 : | VECTOR * const currMV, | ||
979 : | VECTOR * const currPMV) | ||
980 : | Isibaar | 3 | { |
981 : | const uint32_t iWcount = pParam->mb_width; | ||
982 : | |||
983 : | const int32_t iFcode = pParam->fixed_code; | ||
984 : | const int32_t iQuant = pParam->quant; | ||
985 : | const int32_t iWidth = pParam->width; | ||
986 : | const int32_t iHeight = pParam->height; | ||
987 : | const int32_t iEdgedWidth = pParam->edged_width; | ||
988 : | |||
989 : | const uint8_t * cur = pCur->y + x*8 + y*8*iEdgedWidth; | ||
990 : | |||
991 : | int32_t iDiamondSize; | ||
992 : | |||
993 : | int32_t min_dx; | ||
994 : | int32_t max_dx; | ||
995 : | int32_t min_dy; | ||
996 : | int32_t max_dy; | ||
997 : | |||
998 : | VECTOR pmv[4]; | ||
999 : | int32_t psad[4]; | ||
1000 : | VECTOR newMV; | ||
1001 : | VECTOR backupMV; | ||
1002 : | |||
1003 : | MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; | ||
1004 : | |||
1005 : | static int32_t threshA,threshB; | ||
1006 : | int32_t iFound,bPredEq; | ||
1007 : | int32_t iMinSAD,iSAD; | ||
1008 : | |||
1009 : | int32_t iSubBlock = ((y&1)<<1) + (x&1); | ||
1010 : | |||
1011 : | /* Get maximum range */ | ||
1012 : | edgomez | 78 | get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
1013 : | x, y, 8, iWidth, iHeight, iFcode); | ||
1014 : | Isibaar | 3 | |
1015 : | /* we work with abs. MVs, not relative to prediction, so range is relative to 0,0 */ | ||
1016 : | |||
1017 : | if (!(MotionFlags & PMV_HALFPELDIAMOND8 )) | ||
1018 : | { min_dx = EVEN(min_dx); | ||
1019 : | edgomez | 78 | max_dx = EVEN(max_dx); |
1020 : | min_dy = EVEN(min_dy); | ||
1021 : | max_dy = EVEN(max_dy); | ||
1022 : | Isibaar | 3 | } /* because we might use IF (dx>max_dx) THEN dx=max_dx; */ |
1023 : | |||
1024 : | |||
1025 : | bPredEq = get_pmvdata(pMBs, (x>>1), (y>>1), iWcount, iSubBlock, pmv, psad); | ||
1026 : | |||
1027 : | if ((x==0) && (y==0) ) | ||
1028 : | { | ||
1029 : | threshA = 512/4; | ||
1030 : | threshB = 1024/4; | ||
1031 : | |||
1032 : | } | ||
1033 : | else | ||
1034 : | { | ||
1035 : | threshA = psad[0]/4; /* good estimate */ | ||
1036 : | threshB = threshA+256/4; | ||
1037 : | if (threshA< 512/4) threshA = 512/4; | ||
1038 : | if (threshA>1024/4) threshA = 1024/4; | ||
1039 : | if (threshB>1792/4) threshB = 1792/4; | ||
1040 : | } | ||
1041 : | |||
1042 : | iFound=0; | ||
1043 : | |||
1044 : | /* Step 2: Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion | ||
1045 : | edgomez | 78 | vector of the median. |
1046 : | If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 | ||
1047 : | Isibaar | 3 | */ |
1048 : | |||
1049 : | if ((bPredEq) && (MVequal(pmv[0],pMB->mvs[iSubBlock]) ) ) | ||
1050 : | iFound=2; | ||
1051 : | |||
1052 : | /* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. | ||
1053 : | edgomez | 78 | Otherwise select large Diamond Search. |
1054 : | Isibaar | 3 | */ |
1055 : | |||
1056 : | if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536/4) || (bPredEq) ) | ||
1057 : | iDiamondSize=1; // 1 halfpel! | ||
1058 : | else | ||
1059 : | iDiamondSize=2; // 2 halfpel = 1 full pixel! | ||
1060 : | |||
1061 : | if (!(MotionFlags & PMV_HALFPELDIAMOND8) ) | ||
1062 : | iDiamondSize*=2; | ||
1063 : | |||
1064 : | /* Step 4: Calculate SAD around the Median prediction. | ||
1065 : | edgomez | 78 | MinSAD=SAD |
1066 : | If Motion Vector equal to Previous frame motion vector | ||
1067 : | and MinSAD<PrevFrmSAD goto Step 10. | ||
1068 : | If SAD<=256 goto Step 10. | ||
1069 : | Isibaar | 3 | */ |
1070 : | |||
1071 : | |||
1072 : | // Prepare for main loop | ||
1073 : | |||
1074 : | currMV->x=start_x; /* start with mv16 */ | ||
1075 : | currMV->y=start_y; | ||
1076 : | |||
1077 : | iMinSAD = sad8( cur, | ||
1078 : | edgomez | 78 | get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
1079 : | iEdgedWidth); | ||
1080 : | Isibaar | 3 | iMinSAD += calc_delta_8(currMV->x - pmv[0].x, currMV->y - pmv[0].y, (uint8_t)iFcode) * iQuant; |
1081 : | |||
1082 : | if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,pMB->mvs[iSubBlock])) && (iMinSAD < pMB->sad8[iSubBlock]) ) ) | ||
1083 : | edgomez | 78 | { |
1084 : | if (MotionFlags & PMV_QUICKSTOP8) | ||
1085 : | goto step10_8b; | ||
1086 : | if (MotionFlags & PMV_EARLYSTOP8) | ||
1087 : | goto step10_8; | ||
1088 : | } | ||
1089 : | Isibaar | 3 | |
1090 : | /* | ||
1091 : | edgomez | 78 | Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
1092 : | Also calculate (0,0) but do not subtract offset. | ||
1093 : | Let MinSAD be the smallest SAD up to this point. | ||
1094 : | If MV is (0,0) subtract offset. ******** WHAT'S THIS 'OFFSET' ??? *********** | ||
1095 : | Isibaar | 3 | */ |
1096 : | |||
1097 : | // the prediction might be even better than mv16 | ||
1098 : | CHECK_MV8_CANDIDATE(pmv[0].x,pmv[0].y); | ||
1099 : | |||
1100 : | // (0,0) is always possible | ||
1101 : | CHECK_MV8_ZERO; | ||
1102 : | |||
1103 : | // previous frame MV is always possible | ||
1104 : | CHECK_MV8_CANDIDATE(pMB->mvs[iSubBlock].x,pMB->mvs[iSubBlock].y); | ||
1105 : | |||
1106 : | // left neighbour, if allowed | ||
1107 : | if (psad[1] != MV_MAX_ERROR) | ||
1108 : | { | ||
1109 : | if (!(MotionFlags & PMV_HALFPEL8 )) | ||
1110 : | { pmv[1].x = EVEN(pmv[1].x); | ||
1111 : | edgomez | 78 | pmv[1].y = EVEN(pmv[1].y); |
1112 : | Isibaar | 3 | } |
1113 : | CHECK_MV8_CANDIDATE(pmv[1].x,pmv[1].y); | ||
1114 : | } | ||
1115 : | |||
1116 : | // top neighbour, if allowed | ||
1117 : | if (psad[2] != MV_MAX_ERROR) | ||
1118 : | { | ||
1119 : | if (!(MotionFlags & PMV_HALFPEL8 )) | ||
1120 : | { pmv[2].x = EVEN(pmv[2].x); | ||
1121 : | edgomez | 78 | pmv[2].y = EVEN(pmv[2].y); |
1122 : | Isibaar | 3 | } |
1123 : | CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); | ||
1124 : | |||
1125 : | // top right neighbour, if allowed | ||
1126 : | if (psad[3] != MV_MAX_ERROR) | ||
1127 : | { | ||
1128 : | edgomez | 78 | if (!(MotionFlags & PMV_HALFPEL8 )) |
1129 : | { pmv[3].x = EVEN(pmv[3].x); | ||
1130 : | Isibaar | 3 | pmv[3].y = EVEN(pmv[3].y); |
1131 : | edgomez | 78 | } |
1132 : | Isibaar | 3 | CHECK_MV8_CANDIDATE(pmv[3].x,pmv[3].y); |
1133 : | } | ||
1134 : | } | ||
1135 : | |||
1136 : | /* Step 6: If MinSAD <= thresa goto Step 10. | ||
1137 : | If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. | ||
1138 : | */ | ||
1139 : | |||
1140 : | if ( (iMinSAD <= threshA) || ( MVequal(*currMV,pMB->mvs[iSubBlock]) && (iMinSAD < pMB->sad8[iSubBlock]) ) ) | ||
1141 : | edgomez | 78 | { |
1142 : | if (MotionFlags & PMV_QUICKSTOP8) | ||
1143 : | goto step10_8b; | ||
1144 : | if (MotionFlags & PMV_EARLYSTOP8) | ||
1145 : | goto step10_8; | ||
1146 : | } | ||
1147 : | Isibaar | 3 | |
1148 : | /************ (Diamond Search) **************/ | ||
1149 : | /* | ||
1150 : | edgomez | 78 | Step 7: Perform Diamond search, with either the small or large diamond. |
1151 : | If Found=2 only examine one Diamond pattern, and afterwards goto step 10 | ||
1152 : | Step 8: If small diamond, iterate small diamond search pattern until motion vector lies in the center of the diamond. | ||
1153 : | If center then goto step 10. | ||
1154 : | Step 9: If large diamond, iterate large diamond search pattern until motion vector lies in the center. | ||
1155 : | Refine by using small diamond and goto step 10. | ||
1156 : | Isibaar | 3 | */ |
1157 : | |||
1158 : | backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ | ||
1159 : | |||
1160 : | /* default: use best prediction as starting point for one call of PMVfast_MainSearch */ | ||
1161 : | iSAD = PMVfastSearch8_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, | ||
1162 : | edgomez | 78 | x, y, |
1163 : | currMV->x, currMV->y, iMinSAD, &newMV, | ||
1164 : | pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); | ||
1165 : | Isibaar | 3 | |
1166 : | if (iSAD < iMinSAD) | ||
1167 : | { | ||
1168 : | *currMV = newMV; | ||
1169 : | iMinSAD = iSAD; | ||
1170 : | } | ||
1171 : | |||
1172 : | if (MotionFlags & PMV_EXTSEARCH8) | ||
1173 : | { | ||
1174 : | /* extended: search (up to) two more times: orignal prediction and (0,0) */ | ||
1175 : | |||
1176 : | if (!(MVequal(pmv[0],backupMV)) ) | ||
1177 : | { iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, | ||
1178 : | edgomez | 78 | x, y, |
1179 : | pmv[0].x, pmv[0].y, iMinSAD, &newMV, | ||
1180 : | pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); | ||
1181 : | Isibaar | 3 | |
1182 : | edgomez | 78 | if (iSAD < iMinSAD) |
1183 : | { | ||
1184 : | *currMV = newMV; | ||
1185 : | iMinSAD = iSAD; | ||
1186 : | Isibaar | 3 | } |
1187 : | edgomez | 78 | } |
1188 : | Isibaar | 3 | |
1189 : | if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) | ||
1190 : | { iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, | ||
1191 : | edgomez | 78 | x, y, |
1192 : | 0, 0, iMinSAD, &newMV, | ||
1193 : | pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); | ||
1194 : | Isibaar | 3 | |
1195 : | edgomez | 78 | if (iSAD < iMinSAD) |
1196 : | { | ||
1197 : | *currMV = newMV; | ||
1198 : | iMinSAD = iSAD; | ||
1199 : | Isibaar | 3 | } |
1200 : | edgomez | 78 | } |
1201 : | Isibaar | 3 | } |
1202 : | |||
1203 : | /* Step 10: The motion vector is chosen according to the block corresponding to MinSAD. | ||
1204 : | edgomez | 78 | By performing an optional local half-pixel search, we can refine this result even further. |
1205 : | Isibaar | 3 | */ |
1206 : | |||
1207 : | edgomez | 78 | step10_8: |
1208 : | Isibaar | 3 | if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
1209 : | iMinSAD = PMVfastSearch8_Refine( pRef, pRefH, pRefV, pRefHV, cur, | ||
1210 : | edgomez | 78 | x, y, |
1211 : | currMV, iMinSAD, | ||
1212 : | pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); | ||
1213 : | Isibaar | 3 | |
1214 : | edgomez | 78 | step10_8b: |
1215 : | Isibaar | 3 | |
1216 : | currPMV->x = currMV->x - pmv[0].x; | ||
1217 : | currPMV->y = currMV->y - pmv[0].y; | ||
1218 : | |||
1219 : | return iMinSAD; | ||
1220 : | } |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |