Parent Directory | Revision Log
Revision 743 - (view) (download)
1 : | Isibaar | 3 | /************************************************************************** |
2 : | * | ||
3 : | chl | 259 | * XVID MPEG-4 VIDEO CODEC |
4 : | * motion estimation | ||
5 : | * | ||
6 : | * This program is an implementation of a part of one or more MPEG-4 | ||
7 : | * Video tools as specified in ISO/IEC 14496-2 standard. Those intending | ||
8 : | * to use this software module in hardware or software products are | ||
9 : | * advised that its use may infringe existing patents or copyrights, and | ||
10 : | * any such use would be at such party's own risk. The original | ||
11 : | * developer of this software module and his/her company, and subsequent | ||
12 : | * editors and their companies, will have no liability for use of this | ||
13 : | * software or modifications or derivatives thereof. | ||
14 : | * | ||
15 : | * This program is free software; you can redistribute it and/or modify | ||
16 : | * it under the terms of the GNU General Public License as published by | ||
17 : | * the Free Software Foundation; either version 2 of the License, or | ||
18 : | * (at your option) any later version. | ||
19 : | * | ||
20 : | * This program is distributed in the hope that it will be useful, | ||
21 : | * but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
22 : | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
23 : | * GNU General Public License for more details. | ||
24 : | * | ||
25 : | * You should have received a copy of the GNU General Public License | ||
26 : | * along with this program; if not, write to the Free Software | ||
27 : | * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. | ||
28 : | * | ||
29 : | *************************************************************************/ | ||
30 : | |||
31 : | Isibaar | 3 | #include <assert.h> |
32 : | #include <stdio.h> | ||
33 : | chl | 96 | #include <stdlib.h> |
34 : | suxen_drol | 721 | #include <string.h> // memcpy |
35 : | Isibaar | 3 | |
36 : | #include "../encoder.h" | ||
37 : | #include "../utils/mbfunctions.h" | ||
38 : | #include "../prediction/mbprediction.h" | ||
39 : | #include "../global.h" | ||
40 : | #include "../utils/timer.h" | ||
41 : | Isibaar | 580 | #include "../image/interpolate8x8.h" |
42 : | chl | 530 | #include "motion_est.h" |
43 : | suxen_drol | 118 | #include "motion.h" |
44 : | Isibaar | 3 | #include "sad.h" |
45 : | Isibaar | 539 | #include "../utils/emms.h" |
46 : | Isibaar | 3 | |
47 : | chl | 530 | #define INITIAL_SKIP_THRESH (10) |
48 : | #define FINAL_SKIP_THRESH (50) | ||
49 : | #define MAX_SAD00_FOR_SKIP (20) | ||
50 : | Isibaar | 539 | #define MAX_CHROMA_SAD_FOR_SKIP (22) |
51 : | Isibaar | 3 | |
52 : | chl | 530 | #define CHECK_CANDIDATE(X,Y,D) { \ |
53 : | (*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } | ||
54 : | Isibaar | 3 | |
55 : | chl | 530 | static __inline int |
56 : | syskin | 724 | d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv) |
57 : | Isibaar | 3 | { |
58 : | chl | 530 | int xb, yb; |
59 : | syskin | 712 | if (qpel) { x *= 2; y *= 2;} |
60 : | else if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); } | ||
61 : | syskin | 724 | x = pred.x - x; |
62 : | y = pred.y - y; | ||
63 : | chl | 530 | |
64 : | if (x == 0) xb = 1; | ||
65 : | else { | ||
66 : | if (x < 0) x = -x; | ||
67 : | x += (1 << (iFcode - 1)) - 1; | ||
68 : | x >>= (iFcode - 1); | ||
69 : | if (x > 32) x = 32; | ||
70 : | xb = mvtab[x] + iFcode; | ||
71 : | edgomez | 195 | } |
72 : | Isibaar | 3 | |
73 : | chl | 530 | if (y == 0) yb = 1; |
74 : | else { | ||
75 : | if (y < 0) y = -y; | ||
76 : | y += (1 << (iFcode - 1)) - 1; | ||
77 : | y >>= (iFcode - 1); | ||
78 : | if (y > 32) y = 32; | ||
79 : | yb = mvtab[y] + iFcode; | ||
80 : | } | ||
81 : | return xb + yb; | ||
82 : | Isibaar | 3 | } |
83 : | |||
84 : | syskin | 628 | static int32_t |
85 : | ChromaSAD(int dx, int dy, const SearchData * const data) | ||
86 : | Isibaar | 3 | { |
87 : | syskin | 628 | int sad; |
88 : | dx = (dx >> 1) + roundtab_79[dx & 0x3]; | ||
89 : | dy = (dy >> 1) + roundtab_79[dy & 0x3]; | ||
90 : | Isibaar | 3 | |
91 : | syskin | 712 | if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently |
92 : | |||
93 : | switch (((dx & 1) << 1) | (dy & 1)) { | ||
94 : | syskin | 628 | case 0: |
95 : | sad = sad8(data->CurU, data->RefCU + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); | ||
96 : | sad += sad8(data->CurV, data->RefCV + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); | ||
97 : | break; | ||
98 : | case 1: | ||
99 : | dx = dx / 2; dy = (dy - 1) / 2; | ||
100 : | sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); | ||
101 : | sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); | ||
102 : | break; | ||
103 : | case 2: | ||
104 : | dx = (dx - 1) / 2; dy = dy / 2; | ||
105 : | sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); | ||
106 : | sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); | ||
107 : | break; | ||
108 : | default: | ||
109 : | dx = (dx - 1) / 2; dy = (dy - 1) / 2; | ||
110 : | interpolate8x8_halfpel_hv(data->RefQ, | ||
111 : | data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, | ||
112 : | data->rounding); | ||
113 : | sad = sad8(data->CurU, data->RefQ, data->iEdgedWidth/2); | ||
114 : | interpolate8x8_halfpel_hv(data->RefQ, | ||
115 : | data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, | ||
116 : | data->rounding); | ||
117 : | sad += sad8(data->CurV, data->RefQ, data->iEdgedWidth/2); | ||
118 : | break; | ||
119 : | chl | 530 | } |
120 : | syskin | 712 | data->temp[5] = dx; data->temp[6] = dy; data->temp[7] = sad; //backup |
121 : | syskin | 628 | return sad; |
122 : | } | ||
123 : | Isibaar | 3 | |
124 : | syskin | 663 | static __inline const uint8_t * |
125 : | GetReference(const int x, const int y, const int dir, const SearchData * const data) | ||
126 : | Isibaar | 606 | { |
127 : | syskin | 663 | // dir : 0 = forward, 1 = backward |
128 : | switch ( (dir << 2) | ((x&1)<<1) | (y&1) ) { | ||
129 : | case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth); | ||
130 : | case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); | ||
131 : | case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); | ||
132 : | case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); | ||
133 : | case 4 : return data->bRef + x/2 + (y/2)*(data->iEdgedWidth); | ||
134 : | case 5 : return data->bRefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); | ||
135 : | case 6 : return data->bRefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); | ||
136 : | default : return data->bRefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); | ||
137 : | Isibaar | 606 | } |
138 : | } | ||
139 : | |||
140 : | syskin | 663 | static uint8_t * |
141 : | Interpolate8x8qpel(const int x, const int y, const int block, const int dir, const SearchData * const data) | ||
142 : | chl | 530 | { |
143 : | syskin | 663 | // create or find a qpel-precision reference picture; return pointer to it |
144 : | syskin | 668 | uint8_t * Reference = (uint8_t *)data->RefQ + 16*dir; |
145 : | syskin | 663 | const int32_t iEdgedWidth = data->iEdgedWidth; |
146 : | const uint32_t rounding = data->rounding; | ||
147 : | const int halfpel_x = x/2; | ||
148 : | const int halfpel_y = y/2; | ||
149 : | const uint8_t *ref1, *ref2, *ref3, *ref4; | ||
150 : | chl | 326 | |
151 : | syskin | 740 | ref1 = GetReference(halfpel_x, halfpel_y, dir, data); |
152 : | syskin | 663 | ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
153 : | switch( ((x&1)<<1) + (y&1) ) { | ||
154 : | case 0: // pure halfpel position | ||
155 : | syskin | 740 | return (uint8_t *) ref1; |
156 : | syskin | 663 | break; |
157 : | Isibaar | 3 | |
158 : | syskin | 663 | case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
159 : | ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); | ||
160 : | ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; | ||
161 : | Isibaar | 665 | interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
162 : | syskin | 663 | break; |
163 : | Isibaar | 3 | |
164 : | syskin | 663 | case 2: // x qpel, y halfpel - left or right during qpel refinement |
165 : | ref2 = GetReference(x - halfpel_x, halfpel_y, dir, data); | ||
166 : | ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; | ||
167 : | Isibaar | 665 | interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
168 : | syskin | 663 | break; |
169 : | Isibaar | 3 | |
170 : | syskin | 663 | default: // x and y in qpel resolution - the "corners" (top left/right and |
171 : | // bottom left/right) during qpel refinement | ||
172 : | ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); | ||
173 : | ref3 = GetReference(x - halfpel_x, halfpel_y, dir, data); | ||
174 : | ref4 = GetReference(x - halfpel_x, y - halfpel_y, dir, data); | ||
175 : | ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; | ||
176 : | ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; | ||
177 : | ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; | ||
178 : | interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); | ||
179 : | break; | ||
180 : | } | ||
181 : | return Reference; | ||
182 : | Isibaar | 3 | } |
183 : | |||
184 : | syskin | 663 | static uint8_t * |
185 : | Interpolate16x16qpel(const int x, const int y, const int dir, const SearchData * const data) | ||
186 : | Isibaar | 580 | { |
187 : | syskin | 663 | // create or find a qpel-precision reference picture; return pointer to it |
188 : | syskin | 668 | uint8_t * Reference = (uint8_t *)data->RefQ + 16*dir; |
189 : | syskin | 663 | const int32_t iEdgedWidth = data->iEdgedWidth; |
190 : | const uint32_t rounding = data->rounding; | ||
191 : | const int halfpel_x = x/2; | ||
192 : | const int halfpel_y = y/2; | ||
193 : | Isibaar | 580 | const uint8_t *ref1, *ref2, *ref3, *ref4; |
194 : | |||
195 : | syskin | 740 | ref1 = GetReference(halfpel_x, halfpel_y, dir, data); |
196 : | syskin | 663 | switch( ((x&1)<<1) + (y&1) ) { |
197 : | case 0: // pure halfpel position | ||
198 : | syskin | 740 | return (uint8_t *) ref1; |
199 : | Isibaar | 580 | case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
200 : | syskin | 663 | ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
201 : | Isibaar | 665 | interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
202 : | interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); | ||
203 : | interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); | ||
204 : | interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); | ||
205 : | Isibaar | 580 | break; |
206 : | |||
207 : | case 2: // x qpel, y halfpel - left or right during qpel refinement | ||
208 : | syskin | 663 | ref2 = GetReference(x - halfpel_x, halfpel_y, dir, data); |
209 : | Isibaar | 665 | interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
210 : | interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); | ||
211 : | interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); | ||
212 : | interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); | ||
213 : | Isibaar | 580 | break; |
214 : | |||
215 : | default: // x and y in qpel resolution - the "corners" (top left/right and | ||
216 : | // bottom left/right) during qpel refinement | ||
217 : | syskin | 663 | ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
218 : | ref3 = GetReference(x - halfpel_x, halfpel_y, dir, data); | ||
219 : | ref4 = GetReference(x - halfpel_x, y - halfpel_y, dir, data); | ||
220 : | Isibaar | 580 | interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
221 : | interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); | ||
222 : | interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); | ||
223 : | interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); | ||
224 : | break; | ||
225 : | } | ||
226 : | syskin | 663 | return Reference; |
227 : | } | ||
228 : | |||
229 : | /* CHECK_CANDIATE FUNCTIONS START */ | ||
230 : | |||
231 : | static void | ||
232 : | CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) | ||
233 : | { | ||
234 : | int t, xc, yc; | ||
235 : | const uint8_t * Reference; | ||
236 : | VECTOR * current; | ||
237 : | |||
238 : | if (( x > data->max_dx) || ( x < data->min_dx) | ||
239 : | || ( y > data->max_dy) || (y < data->min_dy)) return; | ||
240 : | |||
241 : | if (data->qpel_precision) { // x and y are in 1/4 precision | ||
242 : | Reference = Interpolate16x16qpel(x, y, 0, data); | ||
243 : | xc = x/2; yc = y/2; //for chroma sad | ||
244 : | current = data->currentQMV; | ||
245 : | } else { | ||
246 : | syskin | 712 | Reference = GetReference(x, y, 0, data); |
247 : | syskin | 663 | current = data->currentMV; |
248 : | xc = x; yc = y; | ||
249 : | } | ||
250 : | syskin | 724 | t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0); |
251 : | Isibaar | 580 | |
252 : | syskin | 663 | data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
253 : | Isibaar | 580 | |
254 : | syskin | 628 | data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; |
255 : | data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; | ||
256 : | Isibaar | 580 | |
257 : | syskin | 663 | if (data->chroma) data->temp[0] += ChromaSAD(xc, yc, data); |
258 : | syskin | 628 | |
259 : | Isibaar | 580 | if (data->temp[0] < data->iMinSAD[0]) { |
260 : | data->iMinSAD[0] = data->temp[0]; | ||
261 : | syskin | 663 | current[0].x = x; current[0].y = y; |
262 : | *dir = Direction; } | ||
263 : | Isibaar | 580 | |
264 : | if (data->temp[1] < data->iMinSAD[1]) { | ||
265 : | syskin | 663 | data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y= y; } |
266 : | Isibaar | 580 | if (data->temp[2] < data->iMinSAD[2]) { |
267 : | syskin | 663 | data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; } |
268 : | Isibaar | 580 | if (data->temp[3] < data->iMinSAD[3]) { |
269 : | syskin | 663 | data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; } |
270 : | Isibaar | 580 | if (data->temp[4] < data->iMinSAD[4]) { |
271 : | syskin | 663 | data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; } |
272 : | |||
273 : | Isibaar | 580 | } |
274 : | |||
275 : | static void | ||
276 : | syskin | 704 | CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
277 : | { | ||
278 : | int t; | ||
279 : | const uint8_t * Reference; | ||
280 : | |||
281 : | if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero integer value | ||
282 : | ( x > data->max_dx) || ( x < data->min_dx) | ||
283 : | || ( y > data->max_dy) || (y < data->min_dy)) return; | ||
284 : | |||
285 : | syskin | 712 | Reference = GetReference(x, y, 0, data); |
286 : | syskin | 724 | t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1); |
287 : | syskin | 704 | |
288 : | data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); | ||
289 : | |||
290 : | data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; | ||
291 : | data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; | ||
292 : | |||
293 : | if (data->temp[0] < data->iMinSAD[0]) { | ||
294 : | data->iMinSAD[0] = data->temp[0]; | ||
295 : | data->currentMV[0].x = x; data->currentMV[0].y = y; | ||
296 : | *dir = Direction; } | ||
297 : | |||
298 : | if (data->temp[1] < data->iMinSAD[1]) { | ||
299 : | data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } | ||
300 : | if (data->temp[2] < data->iMinSAD[2]) { | ||
301 : | data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } | ||
302 : | if (data->temp[3] < data->iMinSAD[3]) { | ||
303 : | data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } | ||
304 : | if (data->temp[4] < data->iMinSAD[4]) { | ||
305 : | data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } | ||
306 : | } | ||
307 : | |||
308 : | static void | ||
309 : | syskin | 663 | CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
310 : | syskin | 658 | { |
311 : | int32_t sad; | ||
312 : | syskin | 663 | const uint8_t * Reference; |
313 : | int t; | ||
314 : | VECTOR * current; | ||
315 : | syskin | 658 | |
316 : | if (( x > data->max_dx) || ( x < data->min_dx) | ||
317 : | || ( y > data->max_dy) || (y < data->min_dy)) return; | ||
318 : | |||
319 : | syskin | 712 | if (data->rrv) |
320 : | syskin | 704 | if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) ) return; //non-zero integer value |
321 : | |||
322 : | syskin | 663 | if (data->qpel_precision) { // x and y are in 1/4 precision |
323 : | Reference = Interpolate16x16qpel(x, y, 0, data); | ||
324 : | current = data->currentQMV; | ||
325 : | } else { | ||
326 : | syskin | 712 | Reference = GetReference(x, y, 0, data); |
327 : | syskin | 663 | current = data->currentMV; |
328 : | syskin | 658 | } |
329 : | syskin | 724 | t = d_mv_bits(x, y, data->predMV, data->iFcode, |
330 : | syskin | 712 | data->qpel && !data->qpel_precision && !data->rrv, data->rrv); |
331 : | syskin | 658 | |
332 : | sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); | ||
333 : | syskin | 663 | sad += (data->lambda16 * t * sad)/1000; |
334 : | syskin | 658 | |
335 : | syskin | 663 | if (sad < *(data->iMinSAD)) { |
336 : | *(data->iMinSAD) = sad; | ||
337 : | current->x = x; current->y = y; | ||
338 : | *dir = Direction; } | ||
339 : | syskin | 658 | } |
340 : | |||
341 : | static void | ||
342 : | syskin | 739 | CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
343 : | Isibaar | 539 | { |
344 : | syskin | 658 | // maximum speed - for P/B/I decision |
345 : | Isibaar | 539 | |
346 : | if (( x > data->max_dx) || ( x < data->min_dx) | ||
347 : | || ( y > data->max_dy) || (y < data->min_dy)) return; | ||
348 : | |||
349 : | syskin | 739 | data->temp[0] = sad32v_c(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), |
350 : | data->iEdgedWidth, data->temp+1); | ||
351 : | if (data->temp[0] < *(data->iMinSAD)) { | ||
352 : | *(data->iMinSAD) = data->temp[0]; | ||
353 : | Isibaar | 539 | data->currentMV[0].x = x; data->currentMV[0].y = y; |
354 : | *dir = Direction; } | ||
355 : | syskin | 739 | if (data->temp[1] < data->iMinSAD[1]) { |
356 : | data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } | ||
357 : | if (data->temp[2] < data->iMinSAD[2]) { | ||
358 : | data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } | ||
359 : | if (data->temp[3] < data->iMinSAD[3]) { | ||
360 : | data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } | ||
361 : | if (data->temp[4] < data->iMinSAD[4]) { | ||
362 : | data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } | ||
363 : | |||
364 : | Isibaar | 539 | } |
365 : | |||
366 : | |||
367 : | static void | ||
368 : | chl | 530 | CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) |
369 : | { | ||
370 : | int32_t sad; | ||
371 : | syskin | 663 | int xb, yb, t; |
372 : | chl | 530 | const uint8_t *ReferenceF, *ReferenceB; |
373 : | syskin | 663 | VECTOR *current; |
374 : | Isibaar | 3 | |
375 : | chl | 530 | if (( xf > data->max_dx) || ( xf < data->min_dx) |
376 : | || ( yf > data->max_dy) || (yf < data->min_dy)) return; | ||
377 : | Isibaar | 3 | |
378 : | syskin | 663 | if (data->qpel_precision) { |
379 : | ReferenceF = Interpolate16x16qpel(xf, yf, 0, data); | ||
380 : | xb = data->currentQMV[1].x; yb = data->currentQMV[1].y; | ||
381 : | current = data->currentQMV; | ||
382 : | ReferenceB = Interpolate16x16qpel(xb, yb, 1, data); | ||
383 : | } else { | ||
384 : | syskin | 712 | ReferenceF = GetReference(xf, yf, 0, data); |
385 : | syskin | 663 | xb = data->currentMV[1].x; yb = data->currentMV[1].y; |
386 : | syskin | 712 | ReferenceB = GetReference(xb, yb, 1, data); |
387 : | syskin | 663 | current = data->currentMV; |
388 : | chl | 530 | } |
389 : | edgomez | 195 | |
390 : | syskin | 724 | t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0) |
391 : | + d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel && !data->qpel_precision, 0); | ||
392 : | syskin | 712 | |
393 : | syskin | 628 | sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
394 : | syskin | 663 | sad += (data->lambda16 * t * sad)/1000; |
395 : | syskin | 628 | |
396 : | chl | 530 | if (sad < *(data->iMinSAD)) { |
397 : | *(data->iMinSAD) = sad; | ||
398 : | syskin | 663 | current->x = xf; current->y = yf; |
399 : | chl | 530 | *dir = Direction; } |
400 : | Isibaar | 3 | } |
401 : | |||
402 : | chl | 530 | static void |
403 : | CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) | ||
404 : | Isibaar | 3 | { |
405 : | syskin | 628 | int32_t sad = 0; |
406 : | chl | 530 | int k; |
407 : | const uint8_t *ReferenceF; | ||
408 : | const uint8_t *ReferenceB; | ||
409 : | VECTOR mvs, b_mvs; | ||
410 : | syskin | 724 | const VECTOR zeroMV={0,0}; |
411 : | Isibaar | 3 | |
412 : | chl | 530 | if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
413 : | Isibaar | 3 | |
414 : | chl | 530 | for (k = 0; k < 4; k++) { |
415 : | mvs.x = data->directmvF[k].x + x; | ||
416 : | b_mvs.x = ((x == 0) ? | ||
417 : | data->directmvB[k].x | ||
418 : | : mvs.x - data->referencemv[k].x); | ||
419 : | Isibaar | 3 | |
420 : | chl | 530 | mvs.y = data->directmvF[k].y + y; |
421 : | b_mvs.y = ((y == 0) ? | ||
422 : | data->directmvB[k].y | ||
423 : | : mvs.y - data->referencemv[k].y); | ||
424 : | |||
425 : | if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) | ||
426 : | || ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) | ||
427 : | || ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) | ||
428 : | || ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; | ||
429 : | Isibaar | 3 | |
430 : | syskin | 663 | if (!data->qpel) { |
431 : | mvs.x *= 2; mvs.y *= 2; | ||
432 : | b_mvs.x *= 2; b_mvs.y *= 2; //we move to qpel precision anyway | ||
433 : | chl | 530 | } |
434 : | syskin | 663 | ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); |
435 : | ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); | ||
436 : | chl | 530 | |
437 : | sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), | ||
438 : | syskin | 663 | ReferenceF, ReferenceB, |
439 : | chl | 530 | data->iEdgedWidth); |
440 : | if (sad > *(data->iMinSAD)) return; | ||
441 : | edgomez | 78 | } |
442 : | chl | 530 | |
443 : | syskin | 724 | sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)/1000; |
444 : | syskin | 628 | |
445 : | chl | 530 | if (sad < *(data->iMinSAD)) { |
446 : | *(data->iMinSAD) = sad; | ||
447 : | data->currentMV->x = x; data->currentMV->y = y; | ||
448 : | *dir = Direction; } | ||
449 : | Isibaar | 3 | } |
450 : | |||
451 : | chl | 530 | static void |
452 : | CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) | ||
453 : | chl | 96 | { |
454 : | chl | 530 | int32_t sad; |
455 : | const uint8_t *ReferenceF; | ||
456 : | const uint8_t *ReferenceB; | ||
457 : | VECTOR mvs, b_mvs; | ||
458 : | syskin | 724 | const VECTOR zeroMV = {0,0}; |
459 : | chl | 96 | |
460 : | syskin | 576 | if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
461 : | edgomez | 195 | |
462 : | chl | 530 | mvs.x = data->directmvF[0].x + x; |
463 : | b_mvs.x = ((x == 0) ? | ||
464 : | data->directmvB[0].x | ||
465 : | : mvs.x - data->referencemv[0].x); | ||
466 : | edgomez | 195 | |
467 : | chl | 530 | mvs.y = data->directmvF[0].y + y; |
468 : | b_mvs.y = ((y == 0) ? | ||
469 : | data->directmvB[0].y | ||
470 : | : mvs.y - data->referencemv[0].y); | ||
471 : | |||
472 : | if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) | ||
473 : | || ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) | ||
474 : | || ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) | ||
475 : | || ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; | ||
476 : | chl | 96 | |
477 : | syskin | 663 | if (!data->qpel) { |
478 : | mvs.x *= 2; mvs.y *= 2; | ||
479 : | b_mvs.x *= 2; b_mvs.y *= 2; //we move to qpel precision anyway | ||
480 : | syskin | 712 | } |
481 : | syskin | 663 | ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data); |
482 : | ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); | ||
483 : | chl | 96 | |
484 : | syskin | 628 | sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
485 : | syskin | 724 | sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)/1000; |
486 : | chl | 96 | |
487 : | chl | 530 | if (sad < *(data->iMinSAD)) { |
488 : | *(data->iMinSAD) = sad; | ||
489 : | data->currentMV->x = x; data->currentMV->y = y; | ||
490 : | *dir = Direction; } | ||
491 : | } | ||
492 : | chl | 96 | |
493 : | chl | 530 | static void |
494 : | CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) | ||
495 : | { | ||
496 : | syskin | 628 | int32_t sad; int t; |
497 : | chl | 530 | const uint8_t * Reference; |
498 : | edgomez | 195 | |
499 : | chl | 530 | if (( x > data->max_dx) || ( x < data->min_dx) |
500 : | || ( y > data->max_dy) || (y < data->min_dy)) return; | ||
501 : | chl | 96 | |
502 : | syskin | 663 | if (data->qpel) Reference = Interpolate16x16qpel(x, y, 0, data); |
503 : | syskin | 712 | else Reference = GetReference(x, y, 0, data); |
504 : | edgomez | 195 | |
505 : | chl | 530 | sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
506 : | syskin | 724 | t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0); |
507 : | edgomez | 195 | |
508 : | syskin | 628 | sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))/100; |
509 : | chl | 96 | |
510 : | Isibaar | 606 | if (sad < *(data->iMinSAD)) { |
511 : | *(data->iMinSAD) = sad; | ||
512 : | data->currentMV->x = x; data->currentMV->y = y; | ||
513 : | *dir = Direction; } | ||
514 : | } | ||
515 : | |||
516 : | Isibaar | 539 | /* CHECK_CANDIATE FUNCTIONS END */ |
517 : | chl | 96 | |
518 : | chl | 530 | /* MAINSEARCH FUNCTIONS START */ |
519 : | edgomez | 195 | |
520 : | chl | 530 | static void |
521 : | AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection) | ||
522 : | chl | 181 | { |
523 : | |||
524 : | /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ | ||
525 : | |||
526 : | syskin | 712 | int iDirection; |
527 : | edgomez | 195 | |
528 : | syskin | 712 | for(;;) { //forever |
529 : | iDirection = 0; | ||
530 : | if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); | ||
531 : | if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); | ||
532 : | if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); | ||
533 : | if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); | ||
534 : | chl | 181 | |
535 : | syskin | 712 | /* now we're doing diagonal checks near our candidate */ |
536 : | chl | 181 | |
537 : | syskin | 712 | if (iDirection) { //checking if anything found |
538 : | bDirection = iDirection; | ||
539 : | iDirection = 0; | ||
540 : | x = data->currentMV->x; y = data->currentMV->y; | ||
541 : | if (bDirection & 3) { //our candidate is left or right | ||
542 : | CHECK_CANDIDATE(x, y + iDiamondSize, 8); | ||
543 : | CHECK_CANDIDATE(x, y - iDiamondSize, 4); | ||
544 : | } else { // what remains here is up or down | ||
545 : | CHECK_CANDIDATE(x + iDiamondSize, y, 2); | ||
546 : | CHECK_CANDIDATE(x - iDiamondSize, y, 1); | ||
547 : | } | ||
548 : | chl | 181 | |
549 : | syskin | 712 | if (iDirection) { |
550 : | bDirection += iDirection; | ||
551 : | x = data->currentMV->x; y = data->currentMV->y; | ||
552 : | chl | 181 | } |
553 : | syskin | 712 | } else { //about to quit, eh? not so fast.... |
554 : | switch (bDirection) { | ||
555 : | case 2: | ||
556 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
557 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
558 : | break; | ||
559 : | case 1: | ||
560 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
561 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
562 : | break; | ||
563 : | case 2 + 4: | ||
564 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
565 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
566 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
567 : | break; | ||
568 : | case 4: | ||
569 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
570 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
571 : | break; | ||
572 : | case 8: | ||
573 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
574 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
575 : | break; | ||
576 : | case 1 + 4: | ||
577 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
578 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
579 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
580 : | break; | ||
581 : | case 2 + 8: | ||
582 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
583 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
584 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
585 : | break; | ||
586 : | case 1 + 8: | ||
587 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
588 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
589 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
590 : | break; | ||
591 : | default: //1+2+4+8 == we didn't find anything at all | ||
592 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
593 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
594 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
595 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
596 : | break; | ||
597 : | } | ||
598 : | if (!iDirection) break; //ok, the end. really | ||
599 : | bDirection = iDirection; | ||
600 : | x = data->currentMV->x; y = data->currentMV->y; | ||
601 : | chl | 181 | } |
602 : | syskin | 712 | } |
603 : | chl | 181 | } |
604 : | |||
605 : | chl | 530 | static void |
606 : | SquareSearch(int x, int y, const SearchData * const data, int bDirection) | ||
607 : | chl | 326 | { |
608 : | chl | 530 | int iDirection; |
609 : | chl | 326 | |
610 : | chl | 530 | do { |
611 : | iDirection = 0; | ||
612 : | if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64); | ||
613 : | if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128); | ||
614 : | if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32); | ||
615 : | if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128); | ||
616 : | if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64); | ||
617 : | if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128); | ||
618 : | if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128); | ||
619 : | if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128); | ||
620 : | chl | 326 | |
621 : | chl | 530 | bDirection = iDirection; |
622 : | x = data->currentMV->x; y = data->currentMV->y; | ||
623 : | } while (iDirection); | ||
624 : | chl | 346 | } |
625 : | |||
626 : | chl | 530 | static void |
627 : | DiamondSearch(int x, int y, const SearchData * const data, int bDirection) | ||
628 : | chl | 346 | { |
629 : | |||
630 : | chl | 181 | /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
631 : | |||
632 : | syskin | 712 | int iDirection; |
633 : | edgomez | 195 | |
634 : | syskin | 712 | do { |
635 : | iDirection = 0; | ||
636 : | if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); | ||
637 : | if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); | ||
638 : | if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); | ||
639 : | if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); | ||
640 : | chl | 181 | |
641 : | syskin | 712 | /* now we're doing diagonal checks near our candidate */ |
642 : | chl | 181 | |
643 : | syskin | 712 | if (iDirection) { //checking if anything found |
644 : | bDirection = iDirection; | ||
645 : | iDirection = 0; | ||
646 : | x = data->currentMV->x; y = data->currentMV->y; | ||
647 : | if (bDirection & 3) { //our candidate is left or right | ||
648 : | CHECK_CANDIDATE(x, y + iDiamondSize, 8); | ||
649 : | CHECK_CANDIDATE(x, y - iDiamondSize, 4); | ||
650 : | } else { // what remains here is up or down | ||
651 : | CHECK_CANDIDATE(x + iDiamondSize, y, 2); | ||
652 : | CHECK_CANDIDATE(x - iDiamondSize, y, 1); | ||
653 : | chl | 181 | } |
654 : | syskin | 712 | bDirection += iDirection; |
655 : | x = data->currentMV->x; y = data->currentMV->y; | ||
656 : | chl | 181 | } |
657 : | syskin | 712 | } |
658 : | while (iDirection); | ||
659 : | chl | 181 | } |
660 : | |||
661 : | chl | 530 | /* MAINSEARCH FUNCTIONS END */ |
662 : | chl | 181 | |
663 : | chl | 530 | /* HALFPELREFINE COULD BE A MAINSEARCH FUNCTION, BUT THERE IS NO NEED FOR IT */ |
664 : | |||
665 : | static void | ||
666 : | syskin | 663 | SubpelRefine(const SearchData * const data) |
667 : | chl | 96 | { |
668 : | syskin | 663 | /* Do a half-pel or q-pel refinement */ |
669 : | VECTOR backupMV; | ||
670 : | chl | 530 | int iDirection; //not needed |
671 : | chl | 96 | |
672 : | syskin | 663 | if (data->qpel_precision) |
673 : | backupMV = *(data->currentQMV); | ||
674 : | else backupMV = *(data->currentMV); | ||
675 : | edgomez | 195 | |
676 : | Isibaar | 580 | CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); |
677 : | CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); | ||
678 : | CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); | ||
679 : | CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0); | ||
680 : | |||
681 : | CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); | ||
682 : | CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); | ||
683 : | |||
684 : | CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); | ||
685 : | CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); | ||
686 : | } | ||
687 : | |||
688 : | chl | 530 | static __inline int |
689 : | SkipDecisionP(const IMAGE * current, const IMAGE * reference, | ||
690 : | const int x, const int y, | ||
691 : | syskin | 704 | const uint32_t iEdgedWidth, const uint32_t iQuant, int rrv) |
692 : | chl | 96 | |
693 : | Isibaar | 3 | { |
694 : | chl | 530 | /* keep repeating checks for all b-frames before this P frame, |
695 : | to make sure that SKIP is possible (todo) | ||
696 : | how: if skip is not possible set sad00 to a very high value */ | ||
697 : | syskin | 704 | if(rrv) { |
698 : | uint32_t sadC = sad16(current->u + x*16 + y*(iEdgedWidth/2)*16, | ||
699 : | reference->u + x*16 + y*(iEdgedWidth/2)*16, iEdgedWidth/2, 256*4096); | ||
700 : | if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; | ||
701 : | sadC += sad16(current->v + (x + y*(iEdgedWidth/2))*16, | ||
702 : | reference->v + (x + y*(iEdgedWidth/2))*16, iEdgedWidth/2, 256*4096); | ||
703 : | if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; | ||
704 : | return 1; | ||
705 : | } else { | ||
706 : | uint32_t sadC = sad8(current->u + x*8 + y*(iEdgedWidth/2)*8, | ||
707 : | reference->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); | ||
708 : | if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; | ||
709 : | sadC += sad8(current->v + (x + y*(iEdgedWidth/2))*8, | ||
710 : | reference->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); | ||
711 : | if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; | ||
712 : | return 1; | ||
713 : | } | ||
714 : | chl | 530 | } |
715 : | edgomez | 195 | |
716 : | chl | 530 | static __inline void |
717 : | SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) | ||
718 : | { | ||
719 : | pMB->mode = MODE_NOT_CODED; | ||
720 : | Isibaar | 539 | pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
721 : | pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; | ||
722 : | syskin | 628 | |
723 : | Isibaar | 580 | pMB->qmvs[0].x = pMB->qmvs[1].x = pMB->qmvs[2].x = pMB->qmvs[3].x = 0; |
724 : | pMB->qmvs[0].y = pMB->qmvs[1].y = pMB->qmvs[2].y = pMB->qmvs[3].y = 0; | ||
725 : | |||
726 : | chl | 530 | pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
727 : | Isibaar | 3 | } |
728 : | |||
729 : | syskin | 628 | bool |
730 : | chl | 530 | MotionEstimation(MBParam * const pParam, |
731 : | FRAMEINFO * const current, | ||
732 : | FRAMEINFO * const reference, | ||
733 : | const IMAGE * const pRefH, | ||
734 : | const IMAGE * const pRefV, | ||
735 : | const IMAGE * const pRefHV, | ||
736 : | const uint32_t iLimit) | ||
737 : | Isibaar | 3 | { |
738 : | chl | 530 | MACROBLOCK *const pMBs = current->mbs; |
739 : | const IMAGE *const pCurrent = ¤t->image; | ||
740 : | const IMAGE *const pRef = &reference->image; | ||
741 : | Isibaar | 3 | |
742 : | chl | 530 | const VECTOR zeroMV = { 0, 0 }; |
743 : | Isibaar | 3 | |
744 : | syskin | 704 | uint32_t mb_width = pParam->mb_width; |
745 : | uint32_t mb_height = pParam->mb_height; | ||
746 : | suxen_drol | 702 | |
747 : | chl | 530 | uint32_t x, y; |
748 : | uint32_t iIntra = 0; | ||
749 : | syskin | 628 | int32_t InterBias, quant = current->quant, sad00; |
750 : | Isibaar | 580 | uint8_t *qimage; |
751 : | edgomez | 195 | |
752 : | Isibaar | 539 | // some pre-initialized thingies for SearchP |
753 : | syskin | 712 | int32_t temp[8]; |
754 : | Isibaar | 539 | VECTOR currentMV[5]; |
755 : | Isibaar | 580 | VECTOR currentQMV[5]; |
756 : | Isibaar | 539 | int32_t iMinSAD[5]; |
757 : | SearchData Data; | ||
758 : | syskin | 704 | memset(&Data, 0, sizeof(SearchData)); |
759 : | Isibaar | 539 | Data.iEdgedWidth = pParam->edged_width; |
760 : | Data.currentMV = currentMV; | ||
761 : | Isibaar | 580 | Data.currentQMV = currentQMV; |
762 : | Isibaar | 539 | Data.iMinSAD = iMinSAD; |
763 : | Data.temp = temp; | ||
764 : | Data.iFcode = current->fcode; | ||
765 : | Isibaar | 580 | Data.rounding = pParam->m_rounding_type; |
766 : | syskin | 628 | Data.qpel = pParam->m_quarterpel; |
767 : | Data.chroma = current->global_flags & XVID_ME_COLOUR; | ||
768 : | syskin | 704 | Data.rrv = current->global_flags & XVID_REDUCED; |
769 : | Isibaar | 539 | |
770 : | syskin | 704 | if ((current->global_flags & XVID_REDUCED)) { |
771 : | suxen_drol | 702 | mb_width = (pParam->width + 31) / 32; |
772 : | mb_height = (pParam->height + 31) / 32; | ||
773 : | syskin | 704 | Data.qpel = Data.chroma = 0; |
774 : | suxen_drol | 702 | } |
775 : | |||
776 : | Isibaar | 580 | if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
777 : | syskin | 594 | return 1; // allocate some mem for qpel interpolated blocks |
778 : | Isibaar | 580 | // somehow this is dirty since I think we shouldn't use malloc outside |
779 : | // encoder_create() - so please fix me! | ||
780 : | syskin | 628 | Data.RefQ = qimage; |
781 : | chl | 530 | if (sadInit) (*sadInit) (); |
782 : | edgomez | 195 | |
783 : | suxen_drol | 702 | for (y = 0; y < mb_height; y++) { |
784 : | for (x = 0; x < mb_width; x++) { | ||
785 : | chl | 530 | MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
786 : | syskin | 628 | |
787 : | syskin | 704 | if (Data.rrv) pMB->sad16 = |
788 : | sad32v_c(pCurrent->y + (x + y * pParam->edged_width) * 32, | ||
789 : | pRef->y + (x + y * pParam->edged_width) * 32, | ||
790 : | pParam->edged_width, pMB->sad8 ); | ||
791 : | |||
792 : | else pMB->sad16 = | ||
793 : | sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, | ||
794 : | chl | 530 | pRef->y + (x + y * pParam->edged_width) * 16, |
795 : | pParam->edged_width, pMB->sad8 ); | ||
796 : | edgomez | 195 | |
797 : | syskin | 628 | if (Data.chroma) { |
798 : | pMB->sad16 += sad8(pCurrent->u + x*8 + y*(pParam->edged_width/2)*8, | ||
799 : | pRef->u + x*8 + y*(pParam->edged_width/2)*8, pParam->edged_width/2); | ||
800 : | |||
801 : | pMB->sad16 += sad8(pCurrent->v + (x + y*(pParam->edged_width/2))*8, | ||
802 : | pRef->v + (x + y*(pParam->edged_width/2))*8, pParam->edged_width/2); | ||
803 : | } | ||
804 : | |||
805 : | sad00 = pMB->sad16; //if no gmc; else sad00 = (..) | ||
806 : | |||
807 : | chl | 530 | if (!(current->global_flags & XVID_LUMIMASKING)) { |
808 : | pMB->dquant = NO_CHANGE; | ||
809 : | syskin | 601 | pMB->quant = current->quant; |
810 : | } else { | ||
811 : | syskin | 574 | if (pMB->dquant != NO_CHANGE) { |
812 : | quant += DQtab[pMB->dquant]; | ||
813 : | if (quant > 31) quant = 31; | ||
814 : | else if (quant < 1) quant = 1; | ||
815 : | } | ||
816 : | syskin | 601 | pMB->quant = quant; |
817 : | } | ||
818 : | chl | 181 | |
819 : | syskin | 628 | //initial skip decision |
820 : | chl | 619 | /* no early skip for GMC (global vector = skip vector is unknown!) */ |
821 : | if (current->coding_type == P_VOP) { /* no fast SKIP for S(GMC)-VOPs */ | ||
822 : | syskin | 704 | if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH * (Data.rrv ? 4:1) ) |
823 : | if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant, Data.rrv)) { | ||
824 : | chl | 619 | SkipMacroblockP(pMB, sad00); |
825 : | continue; | ||
826 : | } | ||
827 : | } | ||
828 : | syskin | 628 | |
829 : | SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, | ||
830 : | chl | 530 | y, current->motion_flags, pMB->quant, |
831 : | Isibaar | 539 | &Data, pParam, pMBs, reference->mbs, |
832 : | chl | 530 | current->global_flags & XVID_INTER4V, pMB); |
833 : | edgomez | 195 | |
834 : | chl | 530 | /* final skip decision, a.k.a. "the vector you found, really that good?" */ |
835 : | chl | 619 | if (current->coding_type == P_VOP) { |
836 : | if ( (pMB->dquant == NO_CHANGE) && (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) | ||
837 : | syskin | 704 | && ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1)) ) |
838 : | if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant, Data.rrv)) { | ||
839 : | chl | 619 | SkipMacroblockP(pMB, sad00); |
840 : | continue; | ||
841 : | } | ||
842 : | } | ||
843 : | syskin | 628 | |
844 : | chl | 530 | /* finally, intra decision */ |
845 : | Isibaar | 3 | |
846 : | chl | 530 | InterBias = MV16_INTER_BIAS; |
847 : | syskin | 639 | if (pMB->quant > 8) InterBias += 100 * (pMB->quant - 8); // to make high quants work |
848 : | chl | 530 | if (y != 0) |
849 : | syskin | 639 | if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; |
850 : | chl | 530 | if (x != 0) |
851 : | syskin | 639 | if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; |
852 : | syskin | 628 | |
853 : | if (Data.chroma) InterBias += 50; // to compensate bigger SAD | ||
854 : | syskin | 704 | if (Data.rrv) InterBias *= 4; //?? |
855 : | Isibaar | 3 | |
856 : | chl | 530 | if (InterBias < pMB->sad16) { |
857 : | syskin | 704 | int32_t deviation; |
858 : | if (Data.rrv) { | ||
859 : | deviation = dev16(pCurrent->y + (x + y * pParam->edged_width) * 32, | ||
860 : | pParam->edged_width) | ||
861 : | + dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16, | ||
862 : | pParam->edged_width) | ||
863 : | + dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16 * pParam->edged_width, | ||
864 : | pParam->edged_width) | ||
865 : | + dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16 * (pParam->edged_width+1), | ||
866 : | pParam->edged_width); | ||
867 : | } else | ||
868 : | deviation = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, | ||
869 : | chl | 530 | pParam->edged_width); |
870 : | Isibaar | 3 | |
871 : | syskin | 704 | if (deviation < (pMB->sad16 - InterBias)) { |
872 : | Isibaar | 580 | if (++iIntra >= iLimit) { free(qimage); return 1; } |
873 : | chl | 530 | pMB->mode = MODE_INTRA; |
874 : | Isibaar | 539 | pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = |
875 : | pMB->mvs[3] = zeroMV; | ||
876 : | Isibaar | 580 | pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = |
877 : | pMB->qmvs[3] = zeroMV; | ||
878 : | chl | 530 | pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = |
879 : | pMB->sad8[3] = 0; | ||
880 : | } | ||
881 : | } | ||
882 : | } | ||
883 : | Isibaar | 3 | } |
884 : | Isibaar | 580 | free(qimage); |
885 : | chl | 619 | |
886 : | if (current->coding_type == S_VOP) /* first GMC step only for S(GMC)-VOPs */ | ||
887 : | syskin | 628 | current->GMC_MV = GlobalMotionEst( pMBs, pParam, current->fcode ); |
888 : | chl | 619 | else |
889 : | current->GMC_MV = zeroMV; | ||
890 : | |||
891 : | syskin | 628 | return 0; |
892 : | chl | 530 | } |
893 : | Isibaar | 3 | |
894 : | edgomez | 195 | |
895 : | chl | 530 | #define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
896 : | Isibaar | 3 | |
897 : | chl | 530 | static __inline int |
898 : | make_mask(const VECTOR * const pmv, const int i) | ||
899 : | { | ||
900 : | Isibaar | 539 | int mask = 255, j; |
901 : | chl | 530 | for (j = 0; j < i; j++) { |
902 : | if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already | ||
903 : | if (pmv[i].x == pmv[j].x) { | ||
904 : | if (pmv[i].y == pmv[j].y + iDiamondSize) { mask &= ~4; continue; } | ||
905 : | if (pmv[i].y == pmv[j].y - iDiamondSize) { mask &= ~8; continue; } | ||
906 : | } else | ||
907 : | if (pmv[i].y == pmv[j].y) { | ||
908 : | if (pmv[i].x == pmv[j].x + iDiamondSize) { mask &= ~1; continue; } | ||
909 : | if (pmv[i].x == pmv[j].x - iDiamondSize) { mask &= ~2; continue; } | ||
910 : | } | ||
911 : | Isibaar | 3 | } |
912 : | chl | 530 | return mask; |
913 : | } | ||
914 : | edgomez | 195 | |
915 : | chl | 530 | static __inline void |
916 : | syskin | 704 | PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, |
917 : | int iHcount, const MACROBLOCK * const prevMB, int rrv) | ||
918 : | chl | 530 | { |
919 : | edgomez | 195 | |
920 : | chl | 530 | //this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself |
921 : | syskin | 704 | if (rrv) { iWcount /= 2; iHcount /= 2; } |
922 : | edgomez | 195 | |
923 : | syskin | 704 | if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour |
924 : | chl | 530 | pmv[5].x = EVEN(pmv[3].x); |
925 : | Isibaar | 539 | pmv[5].y = EVEN(pmv[3].y); |
926 : | } else pmv[5].x = pmv[5].y = 0; | ||
927 : | chl | 169 | |
928 : | chl | 530 | if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour |
929 : | else pmv[3].x = pmv[3].y = 0; | ||
930 : | Isibaar | 3 | |
931 : | chl | 530 | if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour |
932 : | else pmv[4].x = pmv[4].y = 0; | ||
933 : | chl | 169 | |
934 : | chl | 530 | // [1] median prediction |
935 : | syskin | 704 | if (rrv) { //median is in halfzero-precision |
936 : | pmv[1].x = RRV_MV_SCALEUP(pmv[0].x); | ||
937 : | pmv[1].y = RRV_MV_SCALEUP(pmv[0].y); | ||
938 : | } else { pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); } | ||
939 : | chl | 169 | |
940 : | chl | 530 | pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask |
941 : | chl | 169 | |
942 : | chl | 530 | pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
943 : | pmv[2].y = EVEN(prevMB->mvs[0].y); | ||
944 : | chl | 169 | |
945 : | syskin | 704 | if ((x < iWcount-1) && (y < iHcount-1)) { |
946 : | chl | 530 | pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame |
947 : | Isibaar | 539 | pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
948 : | } else pmv[6].x = pmv[6].y = 0; | ||
949 : | syskin | 704 | |
950 : | if (rrv) { | ||
951 : | int i; | ||
952 : | for (i = 0; i < 7; i++) { | ||
953 : | pmv[i].x = RRV_MV_SCALEDOWN(pmv[i].x); | ||
954 : | pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); // a trick | ||
955 : | } | ||
956 : | } | ||
957 : | chl | 530 | } |
958 : | chl | 169 | |
959 : | chl | 530 | static void |
960 : | syskin | 628 | SearchP(const IMAGE * const pRef, |
961 : | chl | 530 | const uint8_t * const pRefH, |
962 : | const uint8_t * const pRefV, | ||
963 : | const uint8_t * const pRefHV, | ||
964 : | const IMAGE * const pCur, | ||
965 : | const int x, | ||
966 : | const int y, | ||
967 : | const uint32_t MotionFlags, | ||
968 : | const uint32_t iQuant, | ||
969 : | Isibaar | 539 | SearchData * const Data, |
970 : | chl | 530 | const MBParam * const pParam, |
971 : | const MACROBLOCK * const pMBs, | ||
972 : | const MACROBLOCK * const prevMBs, | ||
973 : | int inter4v, | ||
974 : | MACROBLOCK * const pMB) | ||
975 : | { | ||
976 : | chl | 169 | |
977 : | chl | 530 | int i, iDirection = 255, mask, threshA; |
978 : | Isibaar | 539 | VECTOR pmv[7]; |
979 : | Isibaar | 3 | |
980 : | syskin | 712 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
981 : | pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); | ||
982 : | Isibaar | 3 | |
983 : | syskin | 712 | get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); //has to be changed to get_pmv(2)() |
984 : | syskin | 628 | |
985 : | syskin | 712 | Data->temp[5] = Data->temp[7] = 256*4096; // to reset chroma-sad cache |
986 : | syskin | 704 | if (Data->rrv) i = 2; else i = 1; |
987 : | Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i; | ||
988 : | Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; | ||
989 : | Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; | ||
990 : | Isibaar | 3 | |
991 : | syskin | 704 | Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16*i; |
992 : | Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16*i; | ||
993 : | Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16*i; | ||
994 : | Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16*i; | ||
995 : | Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; | ||
996 : | Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; | ||
997 : | |||
998 : | syskin | 601 | Data->lambda16 = lambda_vec16[iQuant]; |
999 : | Isibaar | 606 | Data->lambda8 = lambda_vec8[iQuant]; |
1000 : | syskin | 663 | Data->qpel_precision = 0; |
1001 : | chl | 169 | |
1002 : | edgomez | 568 | if (pMB->dquant != NO_CHANGE) inter4v = 0; |
1003 : | edgomez | 195 | |
1004 : | Isibaar | 580 | for(i = 0; i < 5; i++) |
1005 : | Data->currentMV[i].x = Data->currentMV[i].y = 0; | ||
1006 : | Isibaar | 539 | |
1007 : | syskin | 663 | if (pParam->m_quarterpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
1008 : | else Data->predMV = pmv[0]; | ||
1009 : | Isibaar | 606 | |
1010 : | syskin | 724 | i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0); |
1011 : | syskin | 628 | Data->iMinSAD[0] = pMB->sad16 + (Data->lambda16 * i * pMB->sad16)/1000; |
1012 : | Data->iMinSAD[1] = pMB->sad8[0] + (Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS))/100; | ||
1013 : | Isibaar | 539 | Data->iMinSAD[2] = pMB->sad8[1]; |
1014 : | Data->iMinSAD[3] = pMB->sad8[2]; | ||
1015 : | Data->iMinSAD[4] = pMB->sad8[3]; | ||
1016 : | |||
1017 : | chl | 530 | if ((x == 0) && (y == 0)) threshA = 512; |
1018 : | else { | ||
1019 : | edgomez | 568 | threshA = Data->temp[0]; // that's when we keep this SAD atm |
1020 : | chl | 530 | if (threshA < 512) threshA = 512; |
1021 : | if (threshA > 1024) threshA = 1024; } | ||
1022 : | Isibaar | 3 | |
1023 : | chl | 530 | PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
1024 : | syskin | 704 | prevMBs + x + y * pParam->mb_width, Data->rrv); |
1025 : | edgomez | 195 | |
1026 : | syskin | 704 | if (Data->rrv) CheckCandidate = CheckCandidate32; |
1027 : | else if (inter4v || Data->chroma) CheckCandidate = CheckCandidate16; | ||
1028 : | else CheckCandidate = CheckCandidate16no4v; //for extra speed | ||
1029 : | edgomez | 568 | |
1030 : | chl | 530 | /* main loop. checking all predictions */ |
1031 : | |||
1032 : | for (i = 1; i < 7; i++) { | ||
1033 : | if (!(mask = make_mask(pmv, i)) ) continue; | ||
1034 : | edgomez | 568 | (*CheckCandidate)(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1035 : | Isibaar | 539 | if (Data->iMinSAD[0] <= threshA) break; |
1036 : | chl | 530 | } |
1037 : | |||
1038 : | Isibaar | 539 | if ((Data->iMinSAD[0] <= threshA) || |
1039 : | (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && | ||
1040 : | (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { | ||
1041 : | chl | 530 | inter4v = 0; |
1042 : | Isibaar | 539 | } else { |
1043 : | Isibaar | 3 | |
1044 : | syskin | 712 | MainSearchFunc * MainSearchPtr; |
1045 : | Isibaar | 539 | if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1046 : | else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; | ||
1047 : | else MainSearchPtr = DiamondSearch; | ||
1048 : | chl | 181 | |
1049 : | Isibaar | 539 | (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
1050 : | Isibaar | 3 | |
1051 : | chl | 530 | /* extended search, diamond starting in 0,0 and in prediction. |
1052 : | note that this search is/might be done in halfpel positions, | ||
1053 : | which makes it more different than the diamond above */ | ||
1054 : | chl | 259 | |
1055 : | Isibaar | 539 | if (MotionFlags & PMV_EXTSEARCH16) { |
1056 : | int32_t bSAD; | ||
1057 : | VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; | ||
1058 : | syskin | 704 | if (Data->rrv) { |
1059 : | startMV.x = RRV_MV_SCALEUP(startMV.x); | ||
1060 : | startMV.y = RRV_MV_SCALEUP(startMV.y); | ||
1061 : | } else | ||
1062 : | if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? | ||
1063 : | startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); | ||
1064 : | Isibaar | 539 | if (!(MVequal(startMV, backupMV))) { |
1065 : | bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; | ||
1066 : | Isibaar | 3 | |
1067 : | Isibaar | 606 | (*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); |
1068 : | Isibaar | 539 | (*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
1069 : | if (bSAD < Data->iMinSAD[0]) { | ||
1070 : | Data->currentMV[0] = backupMV; | ||
1071 : | Data->iMinSAD[0] = bSAD; } | ||
1072 : | } | ||
1073 : | Isibaar | 3 | |
1074 : | Isibaar | 539 | backupMV = Data->currentMV[0]; |
1075 : | syskin | 704 | if (!MotionFlags & PMV_HALFPELREFINE16 || Data->rrv) startMV.x = startMV.y = 0; |
1076 : | else startMV.x = startMV.y = 1; | ||
1077 : | Isibaar | 539 | if (!(MVequal(startMV, backupMV))) { |
1078 : | bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; | ||
1079 : | edgomez | 195 | |
1080 : | Isibaar | 606 | (*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); |
1081 : | Isibaar | 539 | (*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
1082 : | if (bSAD < Data->iMinSAD[0]) { | ||
1083 : | Data->currentMV[0] = backupMV; | ||
1084 : | Data->iMinSAD[0] = bSAD; } | ||
1085 : | } | ||
1086 : | Isibaar | 3 | } |
1087 : | } | ||
1088 : | |||
1089 : | syskin | 663 | if (MotionFlags & PMV_HALFPELREFINE16) SubpelRefine(Data); |
1090 : | Isibaar | 3 | |
1091 : | Isibaar | 580 | for(i = 0; i < 5; i++) { |
1092 : | Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors | ||
1093 : | Data->currentQMV[i].y = 2 * Data->currentMV[i].y; | ||
1094 : | } | ||
1095 : | |||
1096 : | syskin | 704 | if((!Data->rrv) && (pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { |
1097 : | Isibaar | 580 | |
1098 : | syskin | 663 | Data->qpel_precision = 1; |
1099 : | syskin | 712 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1100 : | pParam->width, pParam->height, Data->iFcode, 1, 0); | ||
1101 : | syskin | 601 | |
1102 : | syskin | 663 | SubpelRefine(Data); |
1103 : | Isibaar | 580 | } |
1104 : | |||
1105 : | syskin | 628 | if (Data->iMinSAD[0] < (int32_t)iQuant * 30 ) inter4v = 0; |
1106 : | Isibaar | 539 | if (inter4v) { |
1107 : | SearchData Data8; | ||
1108 : | Data8.iFcode = Data->iFcode; | ||
1109 : | Isibaar | 606 | Data8.lambda8 = Data->lambda8; |
1110 : | Isibaar | 539 | Data8.iEdgedWidth = Data->iEdgedWidth; |
1111 : | Isibaar | 606 | Data8.RefQ = Data->RefQ; |
1112 : | syskin | 628 | Data8.qpel = Data->qpel; |
1113 : | syskin | 704 | Data8.rrv = Data->rrv; |
1114 : | Isibaar | 539 | Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
1115 : | Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); | ||
1116 : | Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); | ||
1117 : | Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); | ||
1118 : | syskin | 628 | |
1119 : | if (Data->chroma) { | ||
1120 : | syskin | 663 | int sumx, sumy, dx, dy; |
1121 : | syskin | 628 | |
1122 : | if(pParam->m_quarterpel) { | ||
1123 : | syskin | 663 | sumx= pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2; |
1124 : | sumy = pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2; | ||
1125 : | } else { | ||
1126 : | sumx = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; | ||
1127 : | sumy = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; | ||
1128 : | } | ||
1129 : | dx = (sumx >> 3) + roundtab_76[sumx & 0xf]; | ||
1130 : | dy = (sumy >> 3) + roundtab_76[sumy & 0xf]; | ||
1131 : | syskin | 628 | |
1132 : | Data->iMinSAD[1] += ChromaSAD(dx, dy, Data); | ||
1133 : | } | ||
1134 : | Isibaar | 539 | } |
1135 : | Isibaar | 3 | |
1136 : | syskin | 704 | if (Data->rrv) { |
1137 : | Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); | ||
1138 : | Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); | ||
1139 : | } | ||
1140 : | syskin | 740 | |
1141 : | chl | 530 | if (!(inter4v) || |
1142 : | Isibaar | 539 | (Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
1143 : | Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { | ||
1144 : | chl | 530 | // INTER MODE |
1145 : | pMB->mode = MODE_INTER; | ||
1146 : | Isibaar | 539 | pMB->mvs[0] = pMB->mvs[1] |
1147 : | = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; | ||
1148 : | Isibaar | 3 | |
1149 : | chl | 530 | pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
1150 : | Isibaar | 539 | pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
1151 : | Isibaar | 3 | |
1152 : | Isibaar | 580 | if(pParam->m_quarterpel) { |
1153 : | syskin | 663 | pMB->qmvs[0] = pMB->qmvs[1] |
1154 : | = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; | ||
1155 : | pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; | ||
1156 : | pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; | ||
1157 : | syskin | 628 | } else { |
1158 : | Isibaar | 580 | pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
1159 : | pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; | ||
1160 : | } | ||
1161 : | edgomez | 195 | } else { |
1162 : | chl | 530 | // INTER4V MODE; all other things are already set in Search8 |
1163 : | pMB->mode = MODE_INTER4V; | ||
1164 : | Isibaar | 539 | pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + |
1165 : | Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant; | ||
1166 : | edgomez | 78 | } |
1167 : | Isibaar | 3 | } |
1168 : | |||
1169 : | chl | 530 | static void |
1170 : | Search8(const SearchData * const OldData, | ||
1171 : | const int x, const int y, | ||
1172 : | const uint32_t MotionFlags, | ||
1173 : | const MBParam * const pParam, | ||
1174 : | MACROBLOCK * const pMB, | ||
1175 : | const MACROBLOCK * const pMBs, | ||
1176 : | Isibaar | 539 | const int block, |
1177 : | SearchData * const Data) | ||
1178 : | chl | 345 | { |
1179 : | syskin | 704 | int i = 0; |
1180 : | Isibaar | 539 | Data->iMinSAD = OldData->iMinSAD + 1 + block; |
1181 : | Data->currentMV = OldData->currentMV + 1 + block; | ||
1182 : | Isibaar | 580 | Data->currentQMV = OldData->currentQMV + 1 + block; |
1183 : | chl | 345 | |
1184 : | syskin | 601 | if(pParam->m_quarterpel) { |
1185 : | syskin | 704 | Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
1186 : | syskin | 724 | if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y, |
1187 : | Data->predMV, Data->iFcode, 0, 0); | ||
1188 : | syskin | 712 | |
1189 : | Isibaar | 606 | } else { |
1190 : | syskin | 704 | Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
1191 : | if (block != 0) { | ||
1192 : | syskin | 724 | if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y, |
1193 : | Data->predMV, Data->iFcode, 0, Data->rrv); | ||
1194 : | syskin | 704 | } |
1195 : | Isibaar | 606 | } |
1196 : | chl | 345 | |
1197 : | syskin | 704 | *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100; |
1198 : | |||
1199 : | chl | 530 | if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { |
1200 : | syskin | 704 | if (Data->rrv) i = 2; else i = 1; |
1201 : | chl | 345 | |
1202 : | syskin | 704 | Data->Ref = OldData->Ref + i*8 * ((block&1) + pParam->edged_width*(block>>1)); |
1203 : | Data->RefH = OldData->RefH + i*8 * ((block&1) + pParam->edged_width*(block>>1)); | ||
1204 : | Data->RefV = OldData->RefV + i*8 * ((block&1) + pParam->edged_width*(block>>1)); | ||
1205 : | Data->RefHV = OldData->RefHV + i*8 * ((block&1) + pParam->edged_width*(block>>1)); | ||
1206 : | chl | 345 | |
1207 : | syskin | 704 | Data->Cur = OldData->Cur + i*8 * ((block&1) + pParam->edged_width*(block>>1)); |
1208 : | syskin | 663 | Data->qpel_precision = 0; |
1209 : | chl | 345 | |
1210 : | syskin | 712 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
1211 : | pParam->width, pParam->height, OldData->iFcode - Data->qpel, 0, Data->rrv); | ||
1212 : | |||
1213 : | if (Data->rrv) CheckCandidate = CheckCandidate16no4v; | ||
1214 : | else CheckCandidate = CheckCandidate8; | ||
1215 : | |||
1216 : | chl | 530 | if (MotionFlags & PMV_EXTSEARCH8) { |
1217 : | Isibaar | 580 | int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
1218 : | |||
1219 : | chl | 530 | MainSearchFunc *MainSearchPtr; |
1220 : | if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; | ||
1221 : | else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; | ||
1222 : | else MainSearchPtr = DiamondSearch; | ||
1223 : | chl | 345 | |
1224 : | Isibaar | 580 | (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1225 : | chl | 345 | |
1226 : | syskin | 601 | if(*(Data->iMinSAD) < temp_sad) { |
1227 : | Isibaar | 580 | Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
1228 : | Data->currentQMV->y = 2 * Data->currentMV->y; | ||
1229 : | } | ||
1230 : | } | ||
1231 : | |||
1232 : | if (MotionFlags & PMV_HALFPELREFINE8) { | ||
1233 : | int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD | ||
1234 : | |||
1235 : | syskin | 663 | SubpelRefine(Data); // perform halfpel refine of current best vector |
1236 : | Isibaar | 580 | |
1237 : | if(*(Data->iMinSAD) < temp_sad) { // we have found a better match | ||
1238 : | Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector | ||
1239 : | Data->currentQMV->y = 2 * Data->currentMV->y; | ||
1240 : | } | ||
1241 : | } | ||
1242 : | |||
1243 : | syskin | 712 | if(!Data->rrv && Data->qpel) { |
1244 : | syskin | 601 | if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && |
1245 : | (MotionFlags & PMV_QUARTERPELREFINE8)) { | ||
1246 : | syskin | 663 | Data->qpel_precision = 1; |
1247 : | syskin | 712 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
1248 : | pParam->width, pParam->height, OldData->iFcode, 1, 0); | ||
1249 : | syskin | 663 | SubpelRefine(Data); |
1250 : | syskin | 601 | } |
1251 : | Isibaar | 580 | } |
1252 : | chl | 345 | } |
1253 : | chl | 530 | |
1254 : | syskin | 704 | if (Data->rrv) { |
1255 : | Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x); | ||
1256 : | Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y); | ||
1257 : | } | ||
1258 : | |||
1259 : | syskin | 712 | if(Data->qpel) { |
1260 : | syskin | 663 | pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x; |
1261 : | pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y; | ||
1262 : | pMB->qmvs[block] = *(Data->currentQMV); | ||
1263 : | syskin | 704 | } else { |
1264 : | Isibaar | 580 | pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
1265 : | pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; | ||
1266 : | } | ||
1267 : | |||
1268 : | Isibaar | 539 | pMB->mvs[block] = *(Data->currentMV); |
1269 : | syskin | 601 | pMB->sad8[block] = 4 * (*Data->iMinSAD); |
1270 : | chl | 345 | } |
1271 : | |||
1272 : | chl | 530 | /* B-frames code starts here */ |
1273 : | chl | 345 | |
1274 : | chl | 530 | static __inline VECTOR |
1275 : | ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) | ||
1276 : | Isibaar | 3 | { |
1277 : | chl | 530 | /* the stupidiest function ever */ |
1278 : | if (mode == MODE_FORWARD) return pMB->mvs[0]; | ||
1279 : | else return pMB->b_mvs[0]; | ||
1280 : | Isibaar | 3 | } |
1281 : | |||
1282 : | chl | 530 | static void __inline |
1283 : | PreparePredictionsBF(VECTOR * const pmv, const int x, const int y, | ||
1284 : | const uint32_t iWcount, | ||
1285 : | const MACROBLOCK * const pMB, | ||
1286 : | const uint32_t mode_curr) | ||
1287 : | Isibaar | 3 | { |
1288 : | |||
1289 : | chl | 530 | // [0] is prediction |
1290 : | pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); | ||
1291 : | Isibaar | 3 | |
1292 : | chl | 530 | pmv[1].x = pmv[1].y = 0; // [1] is zero |
1293 : | Isibaar | 3 | |
1294 : | chl | 530 | pmv[2] = ChoosePred(pMB, mode_curr); |
1295 : | pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); | ||
1296 : | edgomez | 195 | |
1297 : | chl | 530 | if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour |
1298 : | pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); | ||
1299 : | Isibaar | 539 | pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); |
1300 : | } else pmv[3].x = pmv[3].y = 0; | ||
1301 : | Isibaar | 3 | |
1302 : | chl | 530 | if (y != 0) { |
1303 : | pmv[4] = ChoosePred(pMB-iWcount, mode_curr); | ||
1304 : | pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y); | ||
1305 : | } else pmv[4].x = pmv[4].y = 0; | ||
1306 : | Isibaar | 3 | |
1307 : | chl | 530 | if (x != 0) { |
1308 : | pmv[5] = ChoosePred(pMB-1, mode_curr); | ||
1309 : | pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); | ||
1310 : | } else pmv[5].x = pmv[5].y = 0; | ||
1311 : | Isibaar | 3 | |
1312 : | chl | 530 | if ((x != 0)&&(y != 0)) { |
1313 : | pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); | ||
1314 : | pmv[6].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); | ||
1315 : | } else pmv[6].x = pmv[6].y = 0; | ||
1316 : | edgomez | 195 | |
1317 : | chl | 530 | // more? |
1318 : | } | ||
1319 : | chl | 181 | |
1320 : | edgomez | 170 | |
1321 : | chl | 530 | /* search backward or forward, for b-frames */ |
1322 : | static void | ||
1323 : | SearchBF( const uint8_t * const pRef, | ||
1324 : | const uint8_t * const pRefH, | ||
1325 : | const uint8_t * const pRefV, | ||
1326 : | const uint8_t * const pRefHV, | ||
1327 : | const IMAGE * const pCur, | ||
1328 : | const int x, const int y, | ||
1329 : | const uint32_t MotionFlags, | ||
1330 : | const uint32_t iFcode, | ||
1331 : | const MBParam * const pParam, | ||
1332 : | MACROBLOCK * const pMB, | ||
1333 : | const VECTOR * const predMV, | ||
1334 : | int32_t * const best_sad, | ||
1335 : | h | 545 | const int32_t mode_current, |
1336 : | SearchData * const Data) | ||
1337 : | chl | 530 | { |
1338 : | Isibaar | 3 | |
1339 : | chl | 530 | const int32_t iEdgedWidth = pParam->edged_width; |
1340 : | |||
1341 : | syskin | 740 | int i, iDirection = 255, mask; |
1342 : | h | 545 | VECTOR pmv[7]; |
1343 : | chl | 530 | MainSearchFunc *MainSearchPtr; |
1344 : | h | 545 | *Data->iMinSAD = MV_MAX_ERROR; |
1345 : | Data->iFcode = iFcode; | ||
1346 : | syskin | 663 | Data->qpel_precision = 0; |
1347 : | Isibaar | 3 | |
1348 : | h | 545 | Data->Ref = pRef + (x + y * iEdgedWidth) * 16; |
1349 : | Data->RefH = pRefH + (x + y * iEdgedWidth) * 16; | ||
1350 : | Data->RefV = pRefV + (x + y * iEdgedWidth) * 16; | ||
1351 : | Data->RefHV = pRefHV + (x + y * iEdgedWidth) * 16; | ||
1352 : | Isibaar | 3 | |
1353 : | h | 545 | Data->predMV = *predMV; |
1354 : | edgomez | 195 | |
1355 : | h | 545 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1356 : | syskin | 712 | pParam->width, pParam->height, iFcode - Data->qpel, 0, 0); |
1357 : | Isibaar | 3 | |
1358 : | h | 545 | pmv[0] = Data->predMV; |
1359 : | syskin | 658 | if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } |
1360 : | h | 545 | PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
1361 : | edgomez | 195 | |
1362 : | h | 545 | Data->currentMV->x = Data->currentMV->y = 0; |
1363 : | chl | 530 | CheckCandidate = CheckCandidate16no4v; |
1364 : | chl | 181 | |
1365 : | chl | 530 | // main loop. checking all predictions |
1366 : | syskin | 740 | for (i = 0; i < 7; i++) { |
1367 : | chl | 530 | if (!(mask = make_mask(pmv, i)) ) continue; |
1368 : | h | 545 | CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1369 : | edgomez | 78 | } |
1370 : | Isibaar | 3 | |
1371 : | chl | 530 | if (MotionFlags & PMV_USESQUARES16) |
1372 : | MainSearchPtr = SquareSearch; | ||
1373 : | else if (MotionFlags & PMV_ADVANCEDDIAMOND16) | ||
1374 : | MainSearchPtr = AdvDiamondSearch; | ||
1375 : | else MainSearchPtr = DiamondSearch; | ||
1376 : | chl | 96 | |
1377 : | syskin | 740 | (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
1378 : | chl | 169 | |
1379 : | syskin | 663 | SubpelRefine(Data); |
1380 : | syskin | 658 | |
1381 : | syskin | 740 | if (Data->qpel && *Data->iMinSAD < *best_sad + 300) { |
1382 : | syskin | 658 | Data->currentQMV->x = 2*Data->currentMV->x; |
1383 : | Data->currentQMV->y = 2*Data->currentMV->y; | ||
1384 : | syskin | 663 | Data->qpel_precision = 1; |
1385 : | syskin | 712 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1386 : | pParam->width, pParam->height, iFcode, 1, 0); | ||
1387 : | syskin | 663 | SubpelRefine(Data); |
1388 : | syskin | 658 | } |
1389 : | chl | 169 | |
1390 : | chl | 530 | // three bits are needed to code backward mode. four for forward |
1391 : | syskin | 740 | |
1392 : | syskin | 628 | if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
1393 : | else *Data->iMinSAD += 3 * Data->lambda16; | ||
1394 : | chl | 169 | |
1395 : | h | 545 | if (*Data->iMinSAD < *best_sad) { |
1396 : | *best_sad = *Data->iMinSAD; | ||
1397 : | chl | 530 | pMB->mode = mode_current; |
1398 : | syskin | 658 | if (Data->qpel) { |
1399 : | pMB->pmvs[0].x = Data->currentQMV->x - predMV->x; | ||
1400 : | pMB->pmvs[0].y = Data->currentQMV->y - predMV->y; | ||
1401 : | if (mode_current == MODE_FORWARD) | ||
1402 : | pMB->qmvs[0] = *Data->currentQMV; | ||
1403 : | else | ||
1404 : | pMB->b_qmvs[0] = *Data->currentQMV; | ||
1405 : | } else { | ||
1406 : | pMB->pmvs[0].x = Data->currentMV->x - predMV->x; | ||
1407 : | pMB->pmvs[0].y = Data->currentMV->y - predMV->y; | ||
1408 : | } | ||
1409 : | syskin | 739 | if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; |
1410 : | else pMB->b_mvs[0] = *Data->currentMV; | ||
1411 : | } | ||
1412 : | if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; | ||
1413 : | else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search | ||
1414 : | syskin | 658 | |
1415 : | chl | 530 | } |
1416 : | chl | 169 | |
1417 : | syskin | 712 | static void |
1418 : | SkipDecisionB(const IMAGE * const pCur, | ||
1419 : | const IMAGE * const f_Ref, | ||
1420 : | const IMAGE * const b_Ref, | ||
1421 : | MACROBLOCK * const pMB, | ||
1422 : | const uint32_t quant, | ||
1423 : | const uint32_t x, const uint32_t y, | ||
1424 : | const SearchData * const Data) | ||
1425 : | { | ||
1426 : | int dx, dy, b_dx, b_dy; | ||
1427 : | uint32_t sum; | ||
1428 : | //this is not full chroma compensation, only it's fullpel approximation. should work though | ||
1429 : | if (Data->qpel) { | ||
1430 : | dy = Data->directmvF[0].y/2 + Data->directmvF[1].y/2 + | ||
1431 : | Data->directmvF[2].y/2 + Data->directmvF[3].y/2; | ||
1432 : | |||
1433 : | dx = Data->directmvF[0].x/2 + Data->directmvF[1].x/2 + | ||
1434 : | Data->directmvF[2].x/2 + Data->directmvF[3].x/2; | ||
1435 : | |||
1436 : | b_dy = Data->directmvB[0].y/2 + Data->directmvB[1].y/2 + | ||
1437 : | Data->directmvB[2].y/2 + Data->directmvB[3].y/2; | ||
1438 : | |||
1439 : | b_dx = Data->directmvB[0].x/2 + Data->directmvB[1].x/2 + | ||
1440 : | Data->directmvB[2].x/2 + Data->directmvB[3].x/2; | ||
1441 : | |||
1442 : | } else { | ||
1443 : | dy = Data->directmvF[0].y + Data->directmvF[1].y + | ||
1444 : | Data->directmvF[2].y + Data->directmvF[3].y; | ||
1445 : | |||
1446 : | dx = Data->directmvF[0].x + Data->directmvF[1].x + | ||
1447 : | Data->directmvF[2].x + Data->directmvF[3].x; | ||
1448 : | |||
1449 : | b_dy = Data->directmvB[0].y + Data->directmvB[1].y + | ||
1450 : | Data->directmvB[2].y + Data->directmvB[3].y; | ||
1451 : | |||
1452 : | b_dx = Data->directmvB[0].x + Data->directmvB[1].x + | ||
1453 : | Data->directmvB[2].x + Data->directmvB[3].x; | ||
1454 : | } | ||
1455 : | |||
1456 : | |||
1457 : | dy = (dy >> 3) + roundtab_76[dy & 0xf]; | ||
1458 : | dx = (dx >> 3) + roundtab_76[dx & 0xf]; | ||
1459 : | b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf]; | ||
1460 : | b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf]; | ||
1461 : | |||
1462 : | sum = sad8bi(pCur->u + 8*x + 8*y*(Data->iEdgedWidth/2), | ||
1463 : | f_Ref->u + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, | ||
1464 : | b_Ref->u + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, | ||
1465 : | Data->iEdgedWidth/2); | ||
1466 : | sum += sad8bi(pCur->v + 8*x + 8*y*(Data->iEdgedWidth/2), | ||
1467 : | f_Ref->v + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, | ||
1468 : | b_Ref->v + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, | ||
1469 : | Data->iEdgedWidth/2); | ||
1470 : | |||
1471 : | if (sum < 2*MAX_CHROMA_SAD_FOR_SKIP * quant) pMB->mode = MODE_DIRECT_NONE_MV; //skipped | ||
1472 : | } | ||
1473 : | |||
1474 : | |||
1475 : | |||
1476 : | static __inline uint32_t | ||
1477 : | Isibaar | 539 | SearchDirect(const IMAGE * const f_Ref, |
1478 : | chl | 530 | const uint8_t * const f_RefH, |
1479 : | const uint8_t * const f_RefV, | ||
1480 : | const uint8_t * const f_RefHV, | ||
1481 : | Isibaar | 539 | const IMAGE * const b_Ref, |
1482 : | chl | 530 | const uint8_t * const b_RefH, |
1483 : | const uint8_t * const b_RefV, | ||
1484 : | const uint8_t * const b_RefHV, | ||
1485 : | const IMAGE * const pCur, | ||
1486 : | const int x, const int y, | ||
1487 : | const uint32_t MotionFlags, | ||
1488 : | const int32_t TRB, const int32_t TRD, | ||
1489 : | const MBParam * const pParam, | ||
1490 : | MACROBLOCK * const pMB, | ||
1491 : | const MACROBLOCK * const b_mb, | ||
1492 : | h | 545 | int32_t * const best_sad, |
1493 : | SearchData * const Data) | ||
1494 : | edgomez | 195 | |
1495 : | chl | 530 | { |
1496 : | h | 545 | int32_t skip_sad; |
1497 : | chl | 530 | int k; |
1498 : | h | 545 | |
1499 : | chl | 530 | MainSearchFunc *MainSearchPtr; |
1500 : | Isibaar | 3 | |
1501 : | h | 545 | *Data->iMinSAD = 256*4096; |
1502 : | chl | 140 | |
1503 : | h | 545 | Data->Ref = f_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
1504 : | Data->RefH = f_RefH + (x + Data->iEdgedWidth*y) * 16; | ||
1505 : | Data->RefV = f_RefV + (x + Data->iEdgedWidth*y) * 16; | ||
1506 : | Data->RefHV = f_RefHV + (x + Data->iEdgedWidth*y) * 16; | ||
1507 : | Data->bRef = b_Ref->y + (x + Data->iEdgedWidth*y) * 16; | ||
1508 : | Data->bRefH = b_RefH + (x + Data->iEdgedWidth*y) * 16; | ||
1509 : | Data->bRefV = b_RefV + (x + Data->iEdgedWidth*y) * 16; | ||
1510 : | Data->bRefHV = b_RefHV + (x + Data->iEdgedWidth*y) * 16; | ||
1511 : | chl | 140 | |
1512 : | h | 545 | Data->max_dx = 2 * pParam->width - 2 * (x) * 16; |
1513 : | Data->max_dy = 2 * pParam->height - 2 * (y) * 16; | ||
1514 : | Data->min_dx = -(2 * 16 + 2 * (x) * 16); | ||
1515 : | Data->min_dy = -(2 * 16 + 2 * (y) * 16); | ||
1516 : | syskin | 658 | if (Data->qpel) { //we measure in qpixels |
1517 : | Data->max_dx *= 2; | ||
1518 : | Data->max_dy *= 2; | ||
1519 : | Data->min_dx *= 2; | ||
1520 : | Data->min_dy *= 2; | ||
1521 : | Data->referencemv = b_mb->qmvs; | ||
1522 : | } else Data->referencemv = b_mb->mvs; | ||
1523 : | syskin | 672 | Data->qpel_precision = 0; // it's a trick. it's 1 not 0, but we need 0 here |
1524 : | Isibaar | 3 | |
1525 : | chl | 530 | for (k = 0; k < 4; k++) { |
1526 : | h | 545 | pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); |
1527 : | pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD; | ||
1528 : | pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); | ||
1529 : | pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; | ||
1530 : | Isibaar | 3 | |
1531 : | h | 545 | if ( ( pMB->b_mvs[k].x > Data->max_dx ) || ( pMB->b_mvs[k].x < Data->min_dx ) |
1532 : | || ( pMB->b_mvs[k].y > Data->max_dy ) || ( pMB->b_mvs[k].y < Data->min_dy )) { | ||
1533 : | Isibaar | 3 | |
1534 : | Isibaar | 539 | *best_sad = 256*4096; // in that case, we won't use direct mode |
1535 : | pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV" | ||
1536 : | pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; | ||
1537 : | syskin | 712 | return 256*4096; |
1538 : | Isibaar | 539 | } |
1539 : | if (b_mb->mode != MODE_INTER4V) { | ||
1540 : | pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; | ||
1541 : | pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0]; | ||
1542 : | h | 545 | Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0]; |
1543 : | Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0]; | ||
1544 : | Isibaar | 539 | break; |
1545 : | } | ||
1546 : | } | ||
1547 : | Isibaar | 3 | |
1548 : | syskin | 663 | if (b_mb->mode == MODE_INTER4V) CheckCandidate = CheckCandidateDirect; |
1549 : | else CheckCandidate = CheckCandidateDirectno4v; | ||
1550 : | syskin | 658 | |
1551 : | h | 545 | (*CheckCandidate)(0, 0, 255, &k, Data); |
1552 : | Isibaar | 3 | |
1553 : | syskin | 712 | // initial (fast) skip decision |
1554 : | if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH*2) { | ||
1555 : | SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data->chroma, Data); //possible skip - checking chroma | ||
1556 : | if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip. | ||
1557 : | Isibaar | 539 | } |
1558 : | |||
1559 : | h | 545 | skip_sad = *Data->iMinSAD; |
1560 : | edgomez | 195 | |
1561 : | chl | 530 | // DIRECT MODE DELTA VECTOR SEARCH. |
1562 : | // This has to be made more effective, but at the moment I'm happy it's running at all | ||
1563 : | Isibaar | 3 | |
1564 : | chl | 530 | if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1565 : | else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; | ||
1566 : | else MainSearchPtr = DiamondSearch; | ||
1567 : | edgomez | 195 | |
1568 : | h | 545 | (*MainSearchPtr)(0, 0, Data, 255); |
1569 : | Isibaar | 3 | |
1570 : | syskin | 663 | SubpelRefine(Data); |
1571 : | Isibaar | 3 | |
1572 : | h | 545 | *best_sad = *Data->iMinSAD; |
1573 : | Isibaar | 3 | |
1574 : | syskin | 743 | if (b_mb->mode == MODE_INTER4V || Data->qpel) pMB->mode = MODE_DIRECT; |
1575 : | syskin | 672 | else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
1576 : | edgomez | 195 | |
1577 : | h | 545 | pMB->pmvs[3] = *Data->currentMV; |
1578 : | edgomez | 195 | |
1579 : | chl | 530 | for (k = 0; k < 4; k++) { |
1580 : | h | 545 | pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x; |
1581 : | syskin | 658 | pMB->b_mvs[k].x = ( (Data->currentMV->x == 0) |
1582 : | h | 545 | ? Data->directmvB[k].x |
1583 : | syskin | 658 | :pMB->mvs[k].x - Data->referencemv[k].x); |
1584 : | h | 545 | pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y); |
1585 : | pMB->b_mvs[k].y = ((Data->currentMV->y == 0) | ||
1586 : | ? Data->directmvB[k].y | ||
1587 : | : pMB->mvs[k].y - Data->referencemv[k].y); | ||
1588 : | syskin | 658 | if (Data->qpel) { |
1589 : | pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2; | ||
1590 : | pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2; | ||
1591 : | pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2; | ||
1592 : | pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2; | ||
1593 : | } | ||
1594 : | |||
1595 : | chl | 530 | if (b_mb->mode != MODE_INTER4V) { |
1596 : | pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; | ||
1597 : | pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; | ||
1598 : | syskin | 658 | pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0]; |
1599 : | pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0]; | ||
1600 : | chl | 530 | break; |
1601 : | } | ||
1602 : | } | ||
1603 : | h | 545 | return skip_sad; |
1604 : | Isibaar | 3 | } |
1605 : | chl | 96 | |
1606 : | h | 545 | |
1607 : | syskin | 712 | static void |
1608 : | chl | 530 | SearchInterpolate(const uint8_t * const f_Ref, |
1609 : | const uint8_t * const f_RefH, | ||
1610 : | const uint8_t * const f_RefV, | ||
1611 : | const uint8_t * const f_RefHV, | ||
1612 : | const uint8_t * const b_Ref, | ||
1613 : | const uint8_t * const b_RefH, | ||
1614 : | const uint8_t * const b_RefV, | ||
1615 : | const uint8_t * const b_RefHV, | ||
1616 : | const IMAGE * const pCur, | ||
1617 : | const int x, const int y, | ||
1618 : | const uint32_t fcode, | ||
1619 : | const uint32_t bcode, | ||
1620 : | const uint32_t MotionFlags, | ||
1621 : | const MBParam * const pParam, | ||
1622 : | const VECTOR * const f_predMV, | ||
1623 : | const VECTOR * const b_predMV, | ||
1624 : | MACROBLOCK * const pMB, | ||
1625 : | h | 545 | int32_t * const best_sad, |
1626 : | SearchData * const fData) | ||
1627 : | chl | 530 | |
1628 : | chl | 96 | { |
1629 : | |||
1630 : | chl | 530 | int iDirection, i, j; |
1631 : | h | 545 | SearchData bData; |
1632 : | chl | 96 | |
1633 : | syskin | 712 | fData->qpel_precision = 0; |
1634 : | memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data | ||
1635 : | *fData->iMinSAD = 4096*256; | ||
1636 : | syskin | 739 | bData.currentMV++; bData.currentQMV++; |
1637 : | h | 545 | fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
1638 : | edgomez | 195 | |
1639 : | syskin | 712 | i = (x + y * fData->iEdgedWidth) * 16; |
1640 : | bData.bRef = fData->Ref = f_Ref + i; | ||
1641 : | bData.bRefH = fData->RefH = f_RefH + i; | ||
1642 : | bData.bRefV = fData->RefV = f_RefV + i; | ||
1643 : | bData.bRefHV = fData->RefHV = f_RefHV + i; | ||
1644 : | bData.Ref = fData->bRef = b_Ref + i; | ||
1645 : | bData.RefH = fData->bRefH = b_RefH + i; | ||
1646 : | bData.RefV = fData->bRefV = b_RefV + i; | ||
1647 : | bData.RefHV = fData->bRefHV = b_RefHV + i; | ||
1648 : | chl | 96 | |
1649 : | h | 545 | bData.bpredMV = fData->predMV = *f_predMV; |
1650 : | fData->bpredMV = bData.predMV = *b_predMV; | ||
1651 : | syskin | 661 | fData->currentMV[0] = fData->currentMV[2]; |
1652 : | chl | 96 | |
1653 : | syskin | 712 | get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode - fData->qpel, 0, 0); |
1654 : | get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode - fData->qpel, 0, 0); | ||
1655 : | |||
1656 : | h | 545 | if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; |
1657 : | syskin | 675 | if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx; |
1658 : | if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy; | ||
1659 : | if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy; | ||
1660 : | Isibaar | 539 | |
1661 : | h | 545 | if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx; |
1662 : | syskin | 675 | if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx; |
1663 : | if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy; | ||
1664 : | if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy; | ||
1665 : | Isibaar | 539 | |
1666 : | h | 545 | CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
1667 : | chl | 96 | |
1668 : | chl | 530 | //diamond. I wish we could use normal mainsearch functions (square, advdiamond) |
1669 : | chl | 96 | |
1670 : | chl | 530 | do { |
1671 : | iDirection = 255; | ||
1672 : | // forward MV moves | ||
1673 : | h | 545 | i = fData->currentMV[0].x; j = fData->currentMV[0].y; |
1674 : | chl | 96 | |
1675 : | h | 545 | CheckCandidateInt(i + 1, j, 0, &iDirection, fData); |
1676 : | CheckCandidateInt(i, j + 1, 0, &iDirection, fData); | ||
1677 : | CheckCandidateInt(i - 1, j, 0, &iDirection, fData); | ||
1678 : | CheckCandidateInt(i, j - 1, 0, &iDirection, fData); | ||
1679 : | chl | 96 | |
1680 : | chl | 530 | // backward MV moves |
1681 : | h | 545 | i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
1682 : | fData->currentMV[2] = fData->currentMV[0]; | ||
1683 : | Isibaar | 539 | CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); |
1684 : | CheckCandidateInt(i, j + 1, 0, &iDirection, &bData); | ||
1685 : | CheckCandidateInt(i - 1, j, 0, &iDirection, &bData); | ||
1686 : | CheckCandidateInt(i, j - 1, 0, &iDirection, &bData); | ||
1687 : | chl | 96 | |
1688 : | chl | 530 | } while (!(iDirection)); |
1689 : | edgomez | 195 | |
1690 : | syskin | 658 | if (fData->qpel) { |
1691 : | syskin | 740 | if (*fData->iMinSAD > *best_sad + 500) return; |
1692 : | syskin | 675 | CheckCandidate = CheckCandidateInt; |
1693 : | syskin | 663 | fData->qpel_precision = bData.qpel_precision = 1; |
1694 : | syskin | 712 | get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0); |
1695 : | get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0); | ||
1696 : | syskin | 658 | fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x; |
1697 : | fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y; | ||
1698 : | fData->currentQMV[1].x = 2 * fData->currentMV[1].x; | ||
1699 : | fData->currentQMV[1].y = 2 * fData->currentMV[1].y; | ||
1700 : | syskin | 663 | SubpelRefine(fData); |
1701 : | syskin | 740 | if (*fData->iMinSAD > *best_sad + 300) return; |
1702 : | syskin | 658 | fData->currentQMV[2] = fData->currentQMV[0]; |
1703 : | syskin | 663 | SubpelRefine(&bData); |
1704 : | syskin | 658 | } |
1705 : | syskin | 676 | |
1706 : | syskin | 740 | *fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode. |
1707 : | syskin | 658 | |
1708 : | h | 545 | if (*fData->iMinSAD < *best_sad) { |
1709 : | *best_sad = *fData->iMinSAD; | ||
1710 : | pMB->mvs[0] = fData->currentMV[0]; | ||
1711 : | pMB->b_mvs[0] = fData->currentMV[1]; | ||
1712 : | chl | 530 | pMB->mode = MODE_INTERPOLATE; |
1713 : | syskin | 658 | if (fData->qpel) { |
1714 : | pMB->qmvs[0] = fData->currentQMV[0]; | ||
1715 : | pMB->b_qmvs[0] = fData->currentQMV[1]; | ||
1716 : | pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x; | ||
1717 : | pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y; | ||
1718 : | pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x; | ||
1719 : | pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y; | ||
1720 : | } else { | ||
1721 : | pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; | ||
1722 : | pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; | ||
1723 : | pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; | ||
1724 : | pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; | ||
1725 : | } | ||
1726 : | chl | 530 | } |
1727 : | } | ||
1728 : | chl | 96 | |
1729 : | chl | 530 | void |
1730 : | MotionEstimationBVOP(MBParam * const pParam, | ||
1731 : | FRAMEINFO * const frame, | ||
1732 : | const int32_t time_bp, | ||
1733 : | const int32_t time_pp, | ||
1734 : | // forward (past) reference | ||
1735 : | const MACROBLOCK * const f_mbs, | ||
1736 : | const IMAGE * const f_ref, | ||
1737 : | const IMAGE * const f_refH, | ||
1738 : | const IMAGE * const f_refV, | ||
1739 : | const IMAGE * const f_refHV, | ||
1740 : | // backward (future) reference | ||
1741 : | syskin | 644 | const FRAMEINFO * const b_reference, |
1742 : | chl | 530 | const IMAGE * const b_ref, |
1743 : | const IMAGE * const b_refH, | ||
1744 : | const IMAGE * const b_refV, | ||
1745 : | const IMAGE * const b_refHV) | ||
1746 : | { | ||
1747 : | uint32_t i, j; | ||
1748 : | syskin | 712 | int32_t best_sad; |
1749 : | uint32_t skip_sad; | ||
1750 : | chl | 530 | int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; |
1751 : | static const VECTOR zeroMV={0,0}; | ||
1752 : | syskin | 644 | const MACROBLOCK * const b_mbs = b_reference->mbs; |
1753 : | chl | 96 | |
1754 : | chl | 530 | VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
1755 : | chl | 96 | |
1756 : | chl | 530 | const int32_t TRB = time_pp - time_bp; |
1757 : | const int32_t TRD = time_pp; | ||
1758 : | syskin | 658 | uint8_t * qimage; |
1759 : | chl | 96 | |
1760 : | h | 545 | // some pre-inintialized data for the rest of the search |
1761 : | |||
1762 : | SearchData Data; | ||
1763 : | int32_t iMinSAD; | ||
1764 : | VECTOR currentMV[3]; | ||
1765 : | syskin | 658 | VECTOR currentQMV[3]; |
1766 : | syskin | 704 | memset(&Data, 0, sizeof(SearchData)); |
1767 : | h | 545 | Data.iEdgedWidth = pParam->edged_width; |
1768 : | syskin | 658 | Data.currentMV = currentMV; Data.currentQMV = currentQMV; |
1769 : | h | 545 | Data.iMinSAD = &iMinSAD; |
1770 : | syskin | 712 | Data.lambda16 = lambda_vec16[frame->quant]; |
1771 : | Data.chroma = frame->quant; | ||
1772 : | syskin | 658 | Data.qpel = pParam->m_quarterpel; |
1773 : | syskin | 663 | Data.rounding = 0; |
1774 : | h | 545 | |
1775 : | syskin | 658 | if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
1776 : | return; // allocate some mem for qpel interpolated blocks | ||
1777 : | // somehow this is dirty since I think we shouldn't use malloc outside | ||
1778 : | // encoder_create() - so please fix me! | ||
1779 : | Data.RefQ = qimage; | ||
1780 : | |||
1781 : | chl | 530 | // note: i==horizontal, j==vertical |
1782 : | for (j = 0; j < pParam->mb_height; j++) { | ||
1783 : | chl | 96 | |
1784 : | chl | 530 | f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
1785 : | edgomez | 195 | |
1786 : | chl | 530 | for (i = 0; i < pParam->mb_width; i++) { |
1787 : | MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; | ||
1788 : | const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; | ||
1789 : | edgomez | 195 | |
1790 : | chl | 619 | /* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */ |
1791 : | syskin | 644 | if (b_reference->coding_type != S_VOP) |
1792 : | if (b_mb->mode == MODE_NOT_CODED) { | ||
1793 : | pMB->mode = MODE_NOT_CODED; | ||
1794 : | continue; | ||
1795 : | } | ||
1796 : | chl | 96 | |
1797 : | h | 545 | Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
1798 : | syskin | 601 | pMB->quant = frame->quant; |
1799 : | syskin | 628 | |
1800 : | chl | 530 | /* direct search comes first, because it (1) checks for SKIP-mode |
1801 : | and (2) sets very good predictions for forward and backward search */ | ||
1802 : | Isibaar | 539 | skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
1803 : | b_ref, b_refH->y, b_refV->y, b_refHV->y, | ||
1804 : | chl | 530 | &frame->image, |
1805 : | i, j, | ||
1806 : | frame->motion_flags, | ||
1807 : | TRB, TRD, | ||
1808 : | pParam, | ||
1809 : | pMB, b_mb, | ||
1810 : | h | 545 | &best_sad, |
1811 : | &Data); | ||
1812 : | chl | 96 | |
1813 : | Isibaar | 539 | if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
1814 : | syskin | 739 | |
1815 : | chl | 530 | // forward search |
1816 : | SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, | ||
1817 : | &frame->image, i, j, | ||
1818 : | frame->motion_flags, | ||
1819 : | h | 545 | frame->fcode, pParam, |
1820 : | chl | 530 | pMB, &f_predMV, &best_sad, |
1821 : | h | 545 | MODE_FORWARD, &Data); |
1822 : | chl | 96 | |
1823 : | chl | 530 | // backward search |
1824 : | SearchBF(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, | ||
1825 : | &frame->image, i, j, | ||
1826 : | frame->motion_flags, | ||
1827 : | h | 545 | frame->bcode, pParam, |
1828 : | chl | 530 | pMB, &b_predMV, &best_sad, |
1829 : | h | 545 | MODE_BACKWARD, &Data); |
1830 : | chl | 96 | |
1831 : | chl | 530 | // interpolate search comes last, because it uses data from forward and backward as prediction |
1832 : | SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, | ||
1833 : | b_ref->y, b_refH->y, b_refV->y, b_refHV->y, | ||
1834 : | &frame->image, | ||
1835 : | i, j, | ||
1836 : | frame->fcode, frame->bcode, | ||
1837 : | frame->motion_flags, | ||
1838 : | h | 545 | pParam, |
1839 : | chl | 530 | &f_predMV, &b_predMV, |
1840 : | h | 545 | pMB, &best_sad, |
1841 : | &Data); | ||
1842 : | syskin | 662 | |
1843 : | syskin | 712 | // final skip decision |
1844 : | if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP*2) | ||
1845 : | && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) | ||
1846 : | SkipDecisionB(&frame->image, f_ref, b_ref, pMB,frame->quant, i, j, &Data); | ||
1847 : | |||
1848 : | chl | 530 | switch (pMB->mode) { |
1849 : | case MODE_FORWARD: | ||
1850 : | f_count++; | ||
1851 : | syskin | 712 | if (Data.qpel) f_predMV = pMB->qmvs[0]; |
1852 : | syskin | 658 | else f_predMV = pMB->mvs[0]; |
1853 : | chl | 530 | break; |
1854 : | case MODE_BACKWARD: | ||
1855 : | b_count++; | ||
1856 : | syskin | 712 | if (Data.qpel) b_predMV = pMB->b_qmvs[0]; |
1857 : | syskin | 658 | else b_predMV = pMB->b_mvs[0]; |
1858 : | chl | 530 | break; |
1859 : | case MODE_INTERPOLATE: | ||
1860 : | i_count++; | ||
1861 : | syskin | 712 | if (Data.qpel) { |
1862 : | syskin | 658 | f_predMV = pMB->qmvs[0]; |
1863 : | b_predMV = pMB->b_qmvs[0]; | ||
1864 : | } else { | ||
1865 : | f_predMV = pMB->mvs[0]; | ||
1866 : | b_predMV = pMB->b_mvs[0]; | ||
1867 : | } | ||
1868 : | chl | 530 | break; |
1869 : | case MODE_DIRECT: | ||
1870 : | case MODE_DIRECT_NO4V: | ||
1871 : | d_count++; | ||
1872 : | default: | ||
1873 : | break; | ||
1874 : | chl | 96 | } |
1875 : | } | ||
1876 : | } | ||
1877 : | syskin | 658 | free(qimage); |
1878 : | chl | 96 | } |
1879 : | |||
1880 : | syskin | 739 | static __inline void |
1881 : | Isibaar | 539 | MEanalyzeMB ( const uint8_t * const pRef, |
1882 : | const uint8_t * const pCur, | ||
1883 : | const int x, | ||
1884 : | const int y, | ||
1885 : | const MBParam * const pParam, | ||
1886 : | syskin | 739 | MACROBLOCK * const pMBs, |
1887 : | h | 545 | SearchData * const Data) |
1888 : | Isibaar | 539 | { |
1889 : | chl | 317 | |
1890 : | syskin | 739 | int i, mask; |
1891 : | h | 545 | VECTOR pmv[3]; |
1892 : | syskin | 739 | MACROBLOCK * pMB = &pMBs[x + y * pParam->mb_width]; |
1893 : | |||
1894 : | for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; | ||
1895 : | suxen_drol | 118 | |
1896 : | syskin | 640 | //median is only used as prediction. it doesn't have to be real |
1897 : | if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; | ||
1898 : | else | ||
1899 : | if (x == 1) //left macroblock does not have any vector now | ||
1900 : | Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median | ||
1901 : | syskin | 739 | else if (y == 1) // top macroblock doesn't have it's vector |
1902 : | syskin | 640 | Data->predMV = (pMB - 1)->mvs[0]; // left instead of median |
1903 : | else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median | ||
1904 : | |||
1905 : | h | 545 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1906 : | syskin | 739 | pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, Data->rrv); |
1907 : | chl | 326 | |
1908 : | h | 545 | Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
1909 : | Data->Ref = pRef + (x + y * pParam->edged_width) * 16; | ||
1910 : | syskin | 739 | |
1911 : | Isibaar | 539 | pmv[1].x = EVEN(pMB->mvs[0].x); |
1912 : | pmv[1].y = EVEN(pMB->mvs[0].y); | ||
1913 : | syskin | 640 | pmv[2].x = EVEN(Data->predMV.x); |
1914 : | pmv[2].y = EVEN(Data->predMV.y); | ||
1915 : | pmv[0].x = pmv[0].y = 0; | ||
1916 : | chl | 326 | |
1917 : | syskin | 739 | CheckCandidate32I(0, 0, 255, &i, Data); |
1918 : | syskin | 640 | |
1919 : | syskin | 739 | if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) { |
1920 : | syskin | 640 | |
1921 : | syskin | 739 | if (!(mask = make_mask(pmv, 1))) |
1922 : | CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data); | ||
1923 : | if (!(mask = make_mask(pmv, 2))) | ||
1924 : | CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data); | ||
1925 : | Isibaar | 539 | |
1926 : | syskin | 739 | if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) // diamond only if needed |
1927 : | DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); | ||
1928 : | Isibaar | 539 | |
1929 : | syskin | 739 | for (i = 0; i < 4; i++) { |
1930 : | MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1) * pParam->mb_width)]; | ||
1931 : | MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i]; | ||
1932 : | MB->mode = MODE_INTER; | ||
1933 : | MB->sad16 = Data->iMinSAD[i+1]; | ||
1934 : | } | ||
1935 : | } | ||
1936 : | Isibaar | 539 | } |
1937 : | |||
1938 : | syskin | 739 | #define INTRA_BIAS 2500 |
1939 : | #define INTRA_THRESH 1500 | ||
1940 : | #define INTER_THRESH 1400 | ||
1941 : | Isibaar | 539 | |
1942 : | syskin | 644 | |
1943 : | Isibaar | 539 | int |
1944 : | MEanalysis( const IMAGE * const pRef, | ||
1945 : | syskin | 644 | FRAMEINFO * const Current, |
1946 : | Isibaar | 539 | MBParam * const pParam, |
1947 : | syskin | 644 | int maxIntra, //maximum number if non-I frames |
1948 : | int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame | ||
1949 : | int bCount) // number if B frames in a row | ||
1950 : | Isibaar | 539 | { |
1951 : | uint32_t x, y, intra = 0; | ||
1952 : | int sSAD = 0; | ||
1953 : | syskin | 644 | MACROBLOCK * const pMBs = Current->mbs; |
1954 : | const IMAGE * const pCurrent = &Current->image; | ||
1955 : | int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH; | ||
1956 : | syskin | 739 | const VECTOR zeroMV = {0,0}; |
1957 : | h | 545 | |
1958 : | syskin | 739 | int32_t iMinSAD[5], temp[5]; |
1959 : | VECTOR currentMV[5]; | ||
1960 : | h | 545 | SearchData Data; |
1961 : | Data.iEdgedWidth = pParam->edged_width; | ||
1962 : | syskin | 739 | Data.currentMV = currentMV; |
1963 : | Data.iMinSAD = iMinSAD; | ||
1964 : | syskin | 644 | Data.iFcode = Current->fcode; |
1965 : | syskin | 739 | Data.rrv = Current->global_flags & XVID_REDUCED; |
1966 : | Data.temp = temp; | ||
1967 : | CheckCandidate = CheckCandidate32I; | ||
1968 : | h | 545 | |
1969 : | syskin | 658 | if (intraCount < 10) // we're right after an I frame |
1970 : | IntraThresh += 4 * (intraCount - 10) * (intraCount - 10); | ||
1971 : | syskin | 644 | else |
1972 : | if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec | ||
1973 : | IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; | ||
1974 : | |||
1975 : | syskin | 672 | InterThresh += 400 * (1 - bCount); |
1976 : | syskin | 739 | if (InterThresh < 300) InterThresh = 300; |
1977 : | syskin | 644 | |
1978 : | Isibaar | 539 | if (sadInit) (*sadInit) (); |
1979 : | |||
1980 : | syskin | 739 | for (y = 1; y < pParam->mb_height-1; y+=2) { |
1981 : | for (x = 1; x < pParam->mb_width-1; x+=2) { | ||
1982 : | int i; | ||
1983 : | Isibaar | 539 | |
1984 : | syskin | 739 | if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; |
1985 : | |||
1986 : | MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data); | ||
1987 : | Isibaar | 539 | |
1988 : | syskin | 739 | for (i = 0; i < 4; i++) { |
1989 : | int dev; | ||
1990 : | MACROBLOCK *pMB = &pMBs[x+(i&1) + y+(i>>1) * pParam->mb_width]; | ||
1991 : | if (pMB->sad16 > IntraThresh) { | ||
1992 : | dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1))* pParam->edged_width) * 16, | ||
1993 : | pParam->edged_width); | ||
1994 : | if (dev + IntraThresh < pMB->sad16) { | ||
1995 : | pMB->mode = MODE_INTRA; | ||
1996 : | if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return I_VOP; | ||
1997 : | } | ||
1998 : | Isibaar | 539 | } |
1999 : | syskin | 739 | sSAD += pMB->sad16; |
2000 : | syskin | 644 | } |
2001 : | suxen_drol | 118 | } |
2002 : | } | ||
2003 : | syskin | 704 | sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); |
2004 : | syskin | 739 | if (sSAD > IntraThresh + INTRA_BIAS ) return I_VOP; |
2005 : | suxen_drol | 709 | if (sSAD > InterThresh ) return P_VOP; |
2006 : | Isibaar | 539 | emms(); |
2007 : | suxen_drol | 709 | return B_VOP; |
2008 : | syskin | 712 | |
2009 : | suxen_drol | 118 | } |
2010 : | syskin | 576 | |
2011 : | syskin | 628 | static void |
2012 : | CheckGMC(int x, int y, const int dir, int * iDirection, | ||
2013 : | const MACROBLOCK * const pMBs, uint32_t * bestcount, VECTOR * GMC, | ||
2014 : | const MBParam * const pParam) | ||
2015 : | { | ||
2016 : | uint32_t mx, my, a, count = 0; | ||
2017 : | chl | 619 | |
2018 : | syskin | 628 | for (my = 1; my < pParam->mb_height-1; my++) |
2019 : | for (mx = 1; mx < pParam->mb_width-1; mx++) { | ||
2020 : | VECTOR mv; | ||
2021 : | const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; | ||
2022 : | if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) continue; | ||
2023 : | mv = pMB->mvs[0]; | ||
2024 : | a = ABS(mv.x - x) + ABS(mv.y - y); | ||
2025 : | if (a < 6) count += 6 - a; | ||
2026 : | } | ||
2027 : | chl | 619 | |
2028 : | syskin | 628 | if (count > *bestcount) { |
2029 : | *bestcount = count; | ||
2030 : | *iDirection = dir; | ||
2031 : | GMC->x = x; GMC->y = y; | ||
2032 : | } | ||
2033 : | } | ||
2034 : | |||
2035 : | |||
2036 : | suxen_drol | 721 | static VECTOR |
2037 : | chl | 619 | GlobalMotionEst(const MACROBLOCK * const pMBs, const MBParam * const pParam, const uint32_t iFcode) |
2038 : | { | ||
2039 : | |||
2040 : | syskin | 628 | uint32_t count, bestcount = 0; |
2041 : | chl | 619 | int x, y; |
2042 : | syskin | 628 | VECTOR gmc = {0,0}; |
2043 : | chl | 619 | int step, min_x, max_x, min_y, max_y; |
2044 : | uint32_t mx, my; | ||
2045 : | syskin | 628 | int iDirection, bDirection; |
2046 : | chl | 619 | |
2047 : | min_x = min_y = -32<<iFcode; | ||
2048 : | max_x = max_y = 32<<iFcode; | ||
2049 : | |||
2050 : | syskin | 628 | //step1: let's find a rough camera panning |
2051 : | chl | 619 | for (step = 32; step >= 2; step /= 2) { |
2052 : | bestcount = 0; | ||
2053 : | for (y = min_y; y <= max_y; y += step) | ||
2054 : | for (x = min_x ; x <= max_x; x += step) { | ||
2055 : | count = 0; | ||
2056 : | //for all macroblocks | ||
2057 : | for (my = 1; my < pParam->mb_height-1; my++) | ||
2058 : | for (mx = 1; mx < pParam->mb_width-1; mx++) { | ||
2059 : | const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; | ||
2060 : | VECTOR mv; | ||
2061 : | |||
2062 : | if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) | ||
2063 : | continue; | ||
2064 : | |||
2065 : | mv = pMB->mvs[0]; | ||
2066 : | if ( ABS(mv.x - x) <= step && ABS(mv.y - y) <= step ) /* GMC translation is always halfpel-res */ | ||
2067 : | count++; | ||
2068 : | } | ||
2069 : | if (count >= bestcount) { bestcount = count; gmc.x = x; gmc.y = y; } | ||
2070 : | } | ||
2071 : | min_x = gmc.x - step; | ||
2072 : | max_x = gmc.x + step; | ||
2073 : | min_y = gmc.y - step; | ||
2074 : | max_y = gmc.y + step; | ||
2075 : | |||
2076 : | } | ||
2077 : | syskin | 628 | |
2078 : | if (bestcount < (pParam->mb_height-2)*(pParam->mb_width-2)/10) | ||
2079 : | gmc.x = gmc.y = 0; //no camara pan, no GMC | ||
2080 : | |||
2081 : | // step2: let's refine camera panning using gradiend-descent approach. | ||
2082 : | // TODO: more warping points may be evaluated here (like in interpolate mode search - two vectors in one diamond) | ||
2083 : | bestcount = 0; | ||
2084 : | CheckGMC(gmc.x, gmc.y, 255, &iDirection, pMBs, &bestcount, &gmc, pParam); | ||
2085 : | do { | ||
2086 : | x = gmc.x; y = gmc.y; | ||
2087 : | bDirection = iDirection; iDirection = 0; | ||
2088 : | if (bDirection & 1) CheckGMC(x - 1, y, 1+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); | ||
2089 : | if (bDirection & 2) CheckGMC(x + 1, y, 2+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); | ||
2090 : | if (bDirection & 4) CheckGMC(x, y - 1, 1+2+4, &iDirection, pMBs, &bestcount, &gmc, pParam); | ||
2091 : | if (bDirection & 8) CheckGMC(x, y + 1, 1+2+8, &iDirection, pMBs, &bestcount, &gmc, pParam); | ||
2092 : | |||
2093 : | } while (iDirection); | ||
2094 : | |||
2095 : | chl | 619 | if (pParam->m_quarterpel) { |
2096 : | gmc.x *= 2; | ||
2097 : | gmc.y *= 2; /* we store the halfpel value as pseudo-qpel to make comparison easier */ | ||
2098 : | } | ||
2099 : | |||
2100 : | return gmc; | ||
2101 : | } |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |