Parent Directory | Revision Log
Revision 769 - (view) (download)
1 : | Isibaar | 3 | /************************************************************************** |
2 : | * | ||
3 : | chl | 259 | * XVID MPEG-4 VIDEO CODEC |
4 : | syskin | 748 | * motion estimation |
5 : | chl | 259 | * |
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 : | chl | 769 | #include <math.h> // lrint |
36 : | Isibaar | 3 | |
37 : | #include "../encoder.h" | ||
38 : | #include "../utils/mbfunctions.h" | ||
39 : | #include "../prediction/mbprediction.h" | ||
40 : | #include "../global.h" | ||
41 : | #include "../utils/timer.h" | ||
42 : | Isibaar | 580 | #include "../image/interpolate8x8.h" |
43 : | chl | 530 | #include "motion_est.h" |
44 : | suxen_drol | 118 | #include "motion.h" |
45 : | Isibaar | 3 | #include "sad.h" |
46 : | Isibaar | 539 | #include "../utils/emms.h" |
47 : | Isibaar | 3 | |
48 : | chl | 530 | #define INITIAL_SKIP_THRESH (10) |
49 : | #define FINAL_SKIP_THRESH (50) | ||
50 : | #define MAX_SAD00_FOR_SKIP (20) | ||
51 : | Isibaar | 539 | #define MAX_CHROMA_SAD_FOR_SKIP (22) |
52 : | Isibaar | 3 | |
53 : | chl | 530 | #define CHECK_CANDIDATE(X,Y,D) { \ |
54 : | (*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } | ||
55 : | Isibaar | 3 | |
56 : | syskin | 768 | static __inline uint32_t |
57 : | syskin | 724 | d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv) |
58 : | Isibaar | 3 | { |
59 : | chl | 530 | int xb, yb; |
60 : | syskin | 768 | x += x * qpel; y += y * qpel; |
61 : | if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); } | ||
62 : | syskin | 748 | x -= pred.x; |
63 : | y -= pred.y; | ||
64 : | syskin | 768 | |
65 : | syskin | 748 | if (x) { |
66 : | syskin | 768 | x = ABS(x); |
67 : | chl | 530 | x += (1 << (iFcode - 1)) - 1; |
68 : | x >>= (iFcode - 1); | ||
69 : | if (x > 32) x = 32; | ||
70 : | xb = mvtab[x] + iFcode; | ||
71 : | syskin | 748 | } else xb = 1; |
72 : | Isibaar | 3 | |
73 : | syskin | 748 | if (y) { |
74 : | syskin | 768 | y = ABS(y); |
75 : | chl | 530 | y += (1 << (iFcode - 1)) - 1; |
76 : | y >>= (iFcode - 1); | ||
77 : | if (y > 32) y = 32; | ||
78 : | yb = mvtab[y] + iFcode; | ||
79 : | syskin | 748 | } else yb = 1; |
80 : | chl | 530 | return xb + yb; |
81 : | Isibaar | 3 | } |
82 : | |||
83 : | syskin | 748 | static int32_t |
84 : | syskin | 628 | ChromaSAD(int dx, int dy, const SearchData * const data) |
85 : | Isibaar | 3 | { |
86 : | syskin | 628 | int sad; |
87 : | syskin | 748 | const uint32_t stride = data->iEdgedWidth/2; |
88 : | Isibaar | 3 | |
89 : | syskin | 712 | if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently |
90 : | syskin | 768 | data->temp[5] = dx; data->temp[6] = dy; // backup |
91 : | syskin | 712 | |
92 : | switch (((dx & 1) << 1) | (dy & 1)) { | ||
93 : | syskin | 628 | case 0: |
94 : | syskin | 748 | dx = dx / 2; dy = dy / 2; |
95 : | sad = sad8(data->CurU, data->RefCU + dy * stride + dx, stride); | ||
96 : | sad += sad8(data->CurV, data->RefCV + dy * stride + dx, stride); | ||
97 : | syskin | 628 | break; |
98 : | case 1: | ||
99 : | dx = dx / 2; dy = (dy - 1) / 2; | ||
100 : | syskin | 748 | sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + (dy+1) * stride + dx, stride); |
101 : | sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + (dy+1) * stride + dx, stride); | ||
102 : | syskin | 628 | break; |
103 : | case 2: | ||
104 : | dx = (dx - 1) / 2; dy = dy / 2; | ||
105 : | syskin | 748 | sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + dy * stride + dx+1, stride); |
106 : | sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + dy * stride + dx+1, stride); | ||
107 : | syskin | 628 | break; |
108 : | default: | ||
109 : | dx = (dx - 1) / 2; dy = (dy - 1) / 2; | ||
110 : | syskin | 748 | interpolate8x8_halfpel_hv(data->RefQ, data->RefCU + dy * stride + dx, stride, data->rounding); |
111 : | sad = sad8(data->CurU, data->RefQ, stride); | ||
112 : | |||
113 : | interpolate8x8_halfpel_hv(data->RefQ, data->RefCV + dy * stride + dx, stride, data->rounding); | ||
114 : | sad += sad8(data->CurV, data->RefQ, stride); | ||
115 : | syskin | 628 | break; |
116 : | chl | 530 | } |
117 : | syskin | 748 | data->temp[7] = sad; //backup, part 2 |
118 : | syskin | 628 | return sad; |
119 : | } | ||
120 : | Isibaar | 3 | |
121 : | syskin | 663 | static __inline const uint8_t * |
122 : | syskin | 768 | GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data) |
123 : | Isibaar | 606 | { |
124 : | syskin | 663 | // dir : 0 = forward, 1 = backward |
125 : | switch ( (dir << 2) | ((x&1)<<1) | (y&1) ) { | ||
126 : | case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth); | ||
127 : | case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); | ||
128 : | case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); | ||
129 : | case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); | ||
130 : | case 4 : return data->bRef + x/2 + (y/2)*(data->iEdgedWidth); | ||
131 : | case 5 : return data->bRefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); | ||
132 : | case 6 : return data->bRefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); | ||
133 : | default : return data->bRefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); | ||
134 : | Isibaar | 606 | } |
135 : | } | ||
136 : | |||
137 : | syskin | 748 | // this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate |
138 : | static __inline const uint8_t * | ||
139 : | GetReference(const int x, const int y, const SearchData * const data) | ||
140 : | { | ||
141 : | switch ( ((x&1)<<1) | (y&1) ) { | ||
142 : | case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth); | ||
143 : | case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); | ||
144 : | case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); | ||
145 : | default : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); | ||
146 : | } | ||
147 : | } | ||
148 : | |||
149 : | static uint8_t * | ||
150 : | syskin | 768 | Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data) |
151 : | chl | 530 | { |
152 : | syskin | 663 | // create or find a qpel-precision reference picture; return pointer to it |
153 : | syskin | 768 | uint8_t * Reference = data->RefQ + 16*dir; |
154 : | const uint32_t iEdgedWidth = data->iEdgedWidth; | ||
155 : | syskin | 663 | const uint32_t rounding = data->rounding; |
156 : | const int halfpel_x = x/2; | ||
157 : | const int halfpel_y = y/2; | ||
158 : | const uint8_t *ref1, *ref2, *ref3, *ref4; | ||
159 : | chl | 326 | |
160 : | syskin | 748 | ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
161 : | syskin | 663 | ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
162 : | switch( ((x&1)<<1) + (y&1) ) { | ||
163 : | case 0: // pure halfpel position | ||
164 : | syskin | 740 | return (uint8_t *) ref1; |
165 : | syskin | 663 | break; |
166 : | Isibaar | 3 | |
167 : | syskin | 663 | case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
168 : | syskin | 748 | ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
169 : | syskin | 663 | ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
170 : | Isibaar | 665 | interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
171 : | syskin | 663 | break; |
172 : | Isibaar | 3 | |
173 : | syskin | 663 | case 2: // x qpel, y halfpel - left or right during qpel refinement |
174 : | syskin | 748 | ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
175 : | syskin | 663 | ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
176 : | Isibaar | 665 | interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
177 : | syskin | 663 | break; |
178 : | Isibaar | 3 | |
179 : | syskin | 663 | default: // x and y in qpel resolution - the "corners" (top left/right and |
180 : | // bottom left/right) during qpel refinement | ||
181 : | syskin | 748 | ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
182 : | ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); | ||
183 : | ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); | ||
184 : | syskin | 663 | ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
185 : | ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; | ||
186 : | ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; | ||
187 : | interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); | ||
188 : | break; | ||
189 : | } | ||
190 : | return Reference; | ||
191 : | Isibaar | 3 | } |
192 : | |||
193 : | syskin | 748 | static uint8_t * |
194 : | syskin | 768 | Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data) |
195 : | Isibaar | 580 | { |
196 : | syskin | 663 | // create or find a qpel-precision reference picture; return pointer to it |
197 : | syskin | 768 | uint8_t * Reference = data->RefQ + 16*dir; |
198 : | const uint32_t iEdgedWidth = data->iEdgedWidth; | ||
199 : | syskin | 663 | const uint32_t rounding = data->rounding; |
200 : | const int halfpel_x = x/2; | ||
201 : | const int halfpel_y = y/2; | ||
202 : | Isibaar | 580 | const uint8_t *ref1, *ref2, *ref3, *ref4; |
203 : | |||
204 : | syskin | 748 | ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
205 : | syskin | 663 | switch( ((x&1)<<1) + (y&1) ) { |
206 : | case 0: // pure halfpel position | ||
207 : | syskin | 740 | return (uint8_t *) ref1; |
208 : | Isibaar | 580 | case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
209 : | syskin | 748 | ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
210 : | Isibaar | 665 | interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
211 : | interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); | ||
212 : | interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); | ||
213 : | interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); | ||
214 : | Isibaar | 580 | break; |
215 : | |||
216 : | case 2: // x qpel, y halfpel - left or right during qpel refinement | ||
217 : | syskin | 768 | ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
218 : | Isibaar | 665 | interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
219 : | interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); | ||
220 : | interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); | ||
221 : | interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); | ||
222 : | Isibaar | 580 | break; |
223 : | |||
224 : | default: // x and y in qpel resolution - the "corners" (top left/right and | ||
225 : | // bottom left/right) during qpel refinement | ||
226 : | syskin | 748 | ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
227 : | ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); | ||
228 : | ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); | ||
229 : | Isibaar | 580 | interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
230 : | interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); | ||
231 : | interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); | ||
232 : | interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); | ||
233 : | break; | ||
234 : | } | ||
235 : | syskin | 663 | return Reference; |
236 : | } | ||
237 : | |||
238 : | /* CHECK_CANDIATE FUNCTIONS START */ | ||
239 : | |||
240 : | syskin | 748 | static void |
241 : | syskin | 663 | CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
242 : | { | ||
243 : | int t, xc, yc; | ||
244 : | const uint8_t * Reference; | ||
245 : | VECTOR * current; | ||
246 : | |||
247 : | syskin | 768 | if ( (x > data->max_dx) | (x < data->min_dx) |
248 : | | (y > data->max_dy) | (y < data->min_dy) ) return; | ||
249 : | syskin | 663 | |
250 : | if (data->qpel_precision) { // x and y are in 1/4 precision | ||
251 : | Reference = Interpolate16x16qpel(x, y, 0, data); | ||
252 : | xc = x/2; yc = y/2; //for chroma sad | ||
253 : | current = data->currentQMV; | ||
254 : | } else { | ||
255 : | syskin | 748 | Reference = GetReference(x, y, data); |
256 : | syskin | 663 | current = data->currentMV; |
257 : | xc = x; yc = y; | ||
258 : | } | ||
259 : | syskin | 768 | t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
260 : | |||
261 : | syskin | 663 | data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
262 : | Isibaar | 580 | |
263 : | syskin | 768 | data->temp[0] += (data->lambda16 * t * data->temp[0])>>10; |
264 : | data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; | ||
265 : | Isibaar | 580 | |
266 : | syskin | 753 | if (data->chroma) data->temp[0] += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
267 : | (yc >> 1) + roundtab_79[yc & 0x3], data); | ||
268 : | syskin | 628 | |
269 : | Isibaar | 580 | if (data->temp[0] < data->iMinSAD[0]) { |
270 : | data->iMinSAD[0] = data->temp[0]; | ||
271 : | syskin | 663 | current[0].x = x; current[0].y = y; |
272 : | *dir = Direction; } | ||
273 : | Isibaar | 580 | |
274 : | if (data->temp[1] < data->iMinSAD[1]) { | ||
275 : | syskin | 663 | data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y= y; } |
276 : | Isibaar | 580 | if (data->temp[2] < data->iMinSAD[2]) { |
277 : | syskin | 663 | data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; } |
278 : | Isibaar | 580 | if (data->temp[3] < data->iMinSAD[3]) { |
279 : | syskin | 663 | data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; } |
280 : | Isibaar | 580 | if (data->temp[4] < data->iMinSAD[4]) { |
281 : | syskin | 663 | data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; } |
282 : | |||
283 : | Isibaar | 580 | } |
284 : | |||
285 : | syskin | 748 | static void |
286 : | syskin | 704 | CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
287 : | { | ||
288 : | syskin | 768 | uint32_t t; |
289 : | syskin | 704 | const uint8_t * Reference; |
290 : | |||
291 : | syskin | 768 | if ( (!(x&1) && x !=0) | (!(y&1) && y !=0) || //non-zero integer value |
292 : | (x > data->max_dx) | (x < data->min_dx) | ||
293 : | | (y > data->max_dy) | (y < data->min_dy) ) return; | ||
294 : | syskin | 704 | |
295 : | syskin | 748 | Reference = GetReference(x, y, data); |
296 : | syskin | 724 | t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1); |
297 : | syskin | 704 | |
298 : | data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); | ||
299 : | |||
300 : | syskin | 768 | data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10; |
301 : | data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; | ||
302 : | syskin | 704 | |
303 : | if (data->temp[0] < data->iMinSAD[0]) { | ||
304 : | data->iMinSAD[0] = data->temp[0]; | ||
305 : | data->currentMV[0].x = x; data->currentMV[0].y = y; | ||
306 : | *dir = Direction; } | ||
307 : | |||
308 : | if (data->temp[1] < data->iMinSAD[1]) { | ||
309 : | data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } | ||
310 : | if (data->temp[2] < data->iMinSAD[2]) { | ||
311 : | data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } | ||
312 : | if (data->temp[3] < data->iMinSAD[3]) { | ||
313 : | data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } | ||
314 : | if (data->temp[4] < data->iMinSAD[4]) { | ||
315 : | data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } | ||
316 : | } | ||
317 : | |||
318 : | syskin | 748 | static void |
319 : | syskin | 663 | CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
320 : | syskin | 658 | { |
321 : | int32_t sad; | ||
322 : | syskin | 663 | const uint8_t * Reference; |
323 : | syskin | 768 | uint32_t t; |
324 : | syskin | 663 | VECTOR * current; |
325 : | syskin | 658 | |
326 : | syskin | 768 | if ( (x > data->max_dx) | ( x < data->min_dx) |
327 : | | (y > data->max_dy) | (y < data->min_dy) ) return; | ||
328 : | syskin | 658 | |
329 : | syskin | 768 | if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; //non-zero even value |
330 : | syskin | 704 | |
331 : | syskin | 663 | if (data->qpel_precision) { // x and y are in 1/4 precision |
332 : | Reference = Interpolate16x16qpel(x, y, 0, data); | ||
333 : | current = data->currentQMV; | ||
334 : | } else { | ||
335 : | syskin | 748 | Reference = GetReference(x, y, data); |
336 : | syskin | 663 | current = data->currentMV; |
337 : | syskin | 658 | } |
338 : | syskin | 724 | t = d_mv_bits(x, y, data->predMV, data->iFcode, |
339 : | syskin | 768 | data->qpel^data->qpel_precision, data->rrv); |
340 : | |||
341 : | syskin | 658 | sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
342 : | syskin | 768 | sad += (data->lambda16 * t * sad)>>10; |
343 : | syskin | 658 | |
344 : | syskin | 663 | if (sad < *(data->iMinSAD)) { |
345 : | *(data->iMinSAD) = sad; | ||
346 : | current->x = x; current->y = y; | ||
347 : | *dir = Direction; } | ||
348 : | syskin | 658 | } |
349 : | |||
350 : | syskin | 748 | static void |
351 : | syskin | 739 | CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
352 : | Isibaar | 539 | { |
353 : | syskin | 658 | // maximum speed - for P/B/I decision |
354 : | Isibaar | 539 | |
355 : | syskin | 768 | if ( (x > data->max_dx) | (x < data->min_dx) |
356 : | | (y > data->max_dy) | (y < data->min_dy) ) return; | ||
357 : | Isibaar | 539 | |
358 : | syskin | 739 | data->temp[0] = sad32v_c(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), |
359 : | data->iEdgedWidth, data->temp+1); | ||
360 : | syskin | 768 | |
361 : | syskin | 739 | if (data->temp[0] < *(data->iMinSAD)) { |
362 : | *(data->iMinSAD) = data->temp[0]; | ||
363 : | Isibaar | 539 | data->currentMV[0].x = x; data->currentMV[0].y = y; |
364 : | *dir = Direction; } | ||
365 : | syskin | 739 | if (data->temp[1] < data->iMinSAD[1]) { |
366 : | data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } | ||
367 : | if (data->temp[2] < data->iMinSAD[2]) { | ||
368 : | data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } | ||
369 : | if (data->temp[3] < data->iMinSAD[3]) { | ||
370 : | data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } | ||
371 : | if (data->temp[4] < data->iMinSAD[4]) { | ||
372 : | data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } | ||
373 : | |||
374 : | Isibaar | 539 | } |
375 : | |||
376 : | syskin | 748 | static void |
377 : | chl | 530 | CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) |
378 : | { | ||
379 : | syskin | 768 | int32_t sad, xb, yb; |
380 : | uint32_t t; | ||
381 : | chl | 530 | const uint8_t *ReferenceF, *ReferenceB; |
382 : | syskin | 663 | VECTOR *current; |
383 : | Isibaar | 3 | |
384 : | syskin | 768 | if ( (xf > data->max_dx) | (xf < data->min_dx) |
385 : | | (yf > data->max_dy) | (yf < data->min_dy) ) return; | ||
386 : | Isibaar | 3 | |
387 : | syskin | 768 | if (!data->qpel_precision) { |
388 : | syskin | 748 | ReferenceF = GetReference(xf, yf, data); |
389 : | syskin | 663 | xb = data->currentMV[1].x; yb = data->currentMV[1].y; |
390 : | syskin | 748 | ReferenceB = GetReferenceB(xb, yb, 1, data); |
391 : | syskin | 663 | current = data->currentMV; |
392 : | syskin | 768 | } else { |
393 : | ReferenceF = Interpolate16x16qpel(xf, yf, 0, data); | ||
394 : | xb = data->currentQMV[1].x; yb = data->currentQMV[1].y; | ||
395 : | current = data->currentQMV; | ||
396 : | ReferenceB = Interpolate16x16qpel(xb, yb, 1, data); | ||
397 : | chl | 530 | } |
398 : | edgomez | 195 | |
399 : | syskin | 768 | t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0) |
400 : | + d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0); | ||
401 : | |||
402 : | syskin | 628 | sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
403 : | syskin | 768 | sad += (data->lambda16 * t * sad)>>10; |
404 : | syskin | 628 | |
405 : | chl | 530 | if (sad < *(data->iMinSAD)) { |
406 : | *(data->iMinSAD) = sad; | ||
407 : | syskin | 663 | current->x = xf; current->y = yf; |
408 : | chl | 530 | *dir = Direction; } |
409 : | Isibaar | 3 | } |
410 : | |||
411 : | syskin | 748 | static void |
412 : | chl | 530 | CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
413 : | Isibaar | 3 | { |
414 : | syskin | 628 | int32_t sad = 0; |
415 : | syskin | 768 | uint32_t k; |
416 : | chl | 530 | const uint8_t *ReferenceF; |
417 : | const uint8_t *ReferenceB; | ||
418 : | VECTOR mvs, b_mvs; | ||
419 : | Isibaar | 3 | |
420 : | syskin | 768 | if (( x > 31) | ( x < -32) | ( y > 31) | (y < -32)) return; |
421 : | Isibaar | 3 | |
422 : | chl | 530 | for (k = 0; k < 4; k++) { |
423 : | mvs.x = data->directmvF[k].x + x; | ||
424 : | b_mvs.x = ((x == 0) ? | ||
425 : | data->directmvB[k].x | ||
426 : | : mvs.x - data->referencemv[k].x); | ||
427 : | Isibaar | 3 | |
428 : | chl | 530 | mvs.y = data->directmvF[k].y + y; |
429 : | b_mvs.y = ((y == 0) ? | ||
430 : | data->directmvB[k].y | ||
431 : | : mvs.y - data->referencemv[k].y); | ||
432 : | syskin | 768 | |
433 : | chl | 530 | if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
434 : | || ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) | ||
435 : | || ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) | ||
436 : | || ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; | ||
437 : | Isibaar | 3 | |
438 : | syskin | 768 | |
439 : | mvs.x *= 2 - data->qpel; mvs.y *= 2 - data->qpel; | ||
440 : | b_mvs.x *= 2 - data->qpel; b_mvs.y *= 2 - data->qpel; //we move to qpel precision anyway | ||
441 : | |||
442 : | syskin | 663 | ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); |
443 : | ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); | ||
444 : | syskin | 768 | |
445 : | chl | 530 | sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
446 : | syskin | 768 | ReferenceF, ReferenceB, data->iEdgedWidth); |
447 : | chl | 530 | if (sad > *(data->iMinSAD)) return; |
448 : | edgomez | 78 | } |
449 : | chl | 530 | |
450 : | syskin | 768 | sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
451 : | syskin | 628 | |
452 : | chl | 530 | if (sad < *(data->iMinSAD)) { |
453 : | *(data->iMinSAD) = sad; | ||
454 : | data->currentMV->x = x; data->currentMV->y = y; | ||
455 : | *dir = Direction; } | ||
456 : | Isibaar | 3 | } |
457 : | |||
458 : | syskin | 748 | static void |
459 : | chl | 530 | CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
460 : | chl | 96 | { |
461 : | chl | 530 | int32_t sad; |
462 : | const uint8_t *ReferenceF; | ||
463 : | const uint8_t *ReferenceB; | ||
464 : | VECTOR mvs, b_mvs; | ||
465 : | chl | 96 | |
466 : | syskin | 768 | if (( x > 31) | ( x < -32) | ( y > 31) | (y < -32)) return; |
467 : | edgomez | 195 | |
468 : | chl | 530 | mvs.x = data->directmvF[0].x + x; |
469 : | b_mvs.x = ((x == 0) ? | ||
470 : | data->directmvB[0].x | ||
471 : | : mvs.x - data->referencemv[0].x); | ||
472 : | edgomez | 195 | |
473 : | chl | 530 | mvs.y = data->directmvF[0].y + y; |
474 : | b_mvs.y = ((y == 0) ? | ||
475 : | data->directmvB[0].y | ||
476 : | : mvs.y - data->referencemv[0].y); | ||
477 : | syskin | 768 | |
478 : | chl | 530 | if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
479 : | || ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) | ||
480 : | || ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) | ||
481 : | || ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; | ||
482 : | chl | 96 | |
483 : | syskin | 768 | mvs.x *= 2 - data->qpel; mvs.y *= 2 - data->qpel; |
484 : | b_mvs.x *= 2 - data->qpel; b_mvs.y *= 2 - data->qpel; //we move to qpel precision anyway | ||
485 : | |||
486 : | syskin | 663 | ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data); |
487 : | ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); | ||
488 : | chl | 96 | |
489 : | syskin | 628 | sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
490 : | syskin | 768 | sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
491 : | chl | 96 | |
492 : | chl | 530 | if (sad < *(data->iMinSAD)) { |
493 : | *(data->iMinSAD) = sad; | ||
494 : | data->currentMV->x = x; data->currentMV->y = y; | ||
495 : | *dir = Direction; } | ||
496 : | } | ||
497 : | chl | 96 | |
498 : | chl | 530 | static void |
499 : | CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) | ||
500 : | { | ||
501 : | syskin | 768 | int32_t sad; uint32_t t; |
502 : | chl | 530 | const uint8_t * Reference; |
503 : | edgomez | 195 | |
504 : | syskin | 768 | if ( (x > data->max_dx) | (x < data->min_dx) |
505 : | | (y > data->max_dy) | (y < data->min_dy) ) return; | ||
506 : | chl | 96 | |
507 : | syskin | 663 | if (data->qpel) Reference = Interpolate16x16qpel(x, y, 0, data); |
508 : | syskin | 768 | else Reference = GetReference(x, y, data); |
509 : | edgomez | 195 | |
510 : | chl | 530 | sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
511 : | syskin | 724 | t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0); |
512 : | edgomez | 195 | |
513 : | syskin | 768 | sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10; |
514 : | chl | 96 | |
515 : | Isibaar | 606 | if (sad < *(data->iMinSAD)) { |
516 : | *(data->iMinSAD) = sad; | ||
517 : | data->currentMV->x = x; data->currentMV->y = y; | ||
518 : | *dir = Direction; } | ||
519 : | } | ||
520 : | |||
521 : | Isibaar | 539 | /* CHECK_CANDIATE FUNCTIONS END */ |
522 : | chl | 96 | |
523 : | chl | 530 | /* MAINSEARCH FUNCTIONS START */ |
524 : | edgomez | 195 | |
525 : | chl | 530 | static void |
526 : | AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection) | ||
527 : | chl | 181 | { |
528 : | |||
529 : | /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ | ||
530 : | |||
531 : | syskin | 712 | int iDirection; |
532 : | edgomez | 195 | |
533 : | syskin | 712 | for(;;) { //forever |
534 : | iDirection = 0; | ||
535 : | if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); | ||
536 : | if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); | ||
537 : | if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); | ||
538 : | if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); | ||
539 : | chl | 181 | |
540 : | syskin | 712 | /* now we're doing diagonal checks near our candidate */ |
541 : | chl | 181 | |
542 : | syskin | 768 | if (iDirection) { //if anything found |
543 : | syskin | 712 | bDirection = iDirection; |
544 : | iDirection = 0; | ||
545 : | x = data->currentMV->x; y = data->currentMV->y; | ||
546 : | if (bDirection & 3) { //our candidate is left or right | ||
547 : | CHECK_CANDIDATE(x, y + iDiamondSize, 8); | ||
548 : | CHECK_CANDIDATE(x, y - iDiamondSize, 4); | ||
549 : | } else { // what remains here is up or down | ||
550 : | CHECK_CANDIDATE(x + iDiamondSize, y, 2); | ||
551 : | syskin | 748 | CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
552 : | syskin | 712 | } |
553 : | chl | 181 | |
554 : | syskin | 712 | if (iDirection) { |
555 : | bDirection += iDirection; | ||
556 : | syskin | 748 | x = data->currentMV->x; y = data->currentMV->y; |
557 : | chl | 181 | } |
558 : | syskin | 712 | } else { //about to quit, eh? not so fast.... |
559 : | switch (bDirection) { | ||
560 : | case 2: | ||
561 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
562 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
563 : | break; | ||
564 : | case 1: | ||
565 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
566 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
567 : | break; | ||
568 : | case 2 + 4: | ||
569 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
570 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
571 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
572 : | break; | ||
573 : | case 4: | ||
574 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
575 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
576 : | break; | ||
577 : | case 8: | ||
578 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
579 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
580 : | break; | ||
581 : | case 1 + 4: | ||
582 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
583 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
584 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
585 : | break; | ||
586 : | case 2 + 8: | ||
587 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
588 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
589 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
590 : | break; | ||
591 : | case 1 + 8: | ||
592 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
593 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
594 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
595 : | break; | ||
596 : | default: //1+2+4+8 == we didn't find anything at all | ||
597 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
598 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
599 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
600 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
601 : | break; | ||
602 : | } | ||
603 : | if (!iDirection) break; //ok, the end. really | ||
604 : | bDirection = iDirection; | ||
605 : | x = data->currentMV->x; y = data->currentMV->y; | ||
606 : | chl | 181 | } |
607 : | syskin | 712 | } |
608 : | chl | 181 | } |
609 : | |||
610 : | chl | 530 | static void |
611 : | SquareSearch(int x, int y, const SearchData * const data, int bDirection) | ||
612 : | chl | 326 | { |
613 : | chl | 530 | int iDirection; |
614 : | chl | 326 | |
615 : | chl | 530 | do { |
616 : | iDirection = 0; | ||
617 : | if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64); | ||
618 : | if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128); | ||
619 : | if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32); | ||
620 : | if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128); | ||
621 : | if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64); | ||
622 : | if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128); | ||
623 : | if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128); | ||
624 : | if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128); | ||
625 : | chl | 326 | |
626 : | chl | 530 | bDirection = iDirection; |
627 : | x = data->currentMV->x; y = data->currentMV->y; | ||
628 : | } while (iDirection); | ||
629 : | chl | 346 | } |
630 : | |||
631 : | chl | 530 | static void |
632 : | DiamondSearch(int x, int y, const SearchData * const data, int bDirection) | ||
633 : | chl | 346 | { |
634 : | |||
635 : | chl | 181 | /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
636 : | |||
637 : | syskin | 712 | int iDirection; |
638 : | edgomez | 195 | |
639 : | syskin | 712 | do { |
640 : | iDirection = 0; | ||
641 : | if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); | ||
642 : | if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); | ||
643 : | if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); | ||
644 : | if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); | ||
645 : | chl | 181 | |
646 : | syskin | 712 | /* now we're doing diagonal checks near our candidate */ |
647 : | chl | 181 | |
648 : | syskin | 712 | if (iDirection) { //checking if anything found |
649 : | bDirection = iDirection; | ||
650 : | iDirection = 0; | ||
651 : | x = data->currentMV->x; y = data->currentMV->y; | ||
652 : | if (bDirection & 3) { //our candidate is left or right | ||
653 : | CHECK_CANDIDATE(x, y + iDiamondSize, 8); | ||
654 : | CHECK_CANDIDATE(x, y - iDiamondSize, 4); | ||
655 : | } else { // what remains here is up or down | ||
656 : | CHECK_CANDIDATE(x + iDiamondSize, y, 2); | ||
657 : | syskin | 748 | CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
658 : | chl | 181 | } |
659 : | syskin | 712 | bDirection += iDirection; |
660 : | x = data->currentMV->x; y = data->currentMV->y; | ||
661 : | chl | 181 | } |
662 : | syskin | 712 | } |
663 : | while (iDirection); | ||
664 : | chl | 181 | } |
665 : | |||
666 : | chl | 530 | /* MAINSEARCH FUNCTIONS END */ |
667 : | chl | 181 | |
668 : | chl | 530 | static void |
669 : | syskin | 663 | SubpelRefine(const SearchData * const data) |
670 : | chl | 96 | { |
671 : | syskin | 663 | /* Do a half-pel or q-pel refinement */ |
672 : | syskin | 768 | const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV; |
673 : | int iDirection; //only needed because macro expects it | ||
674 : | chl | 96 | |
675 : | syskin | 768 | CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0); |
676 : | CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0); | ||
677 : | CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0); | ||
678 : | CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0); | ||
679 : | CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0); | ||
680 : | CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0); | ||
681 : | CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0); | ||
682 : | CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0); | ||
683 : | Isibaar | 580 | } |
684 : | |||
685 : | chl | 530 | static __inline int |
686 : | SkipDecisionP(const IMAGE * current, const IMAGE * reference, | ||
687 : | const int x, const int y, | ||
688 : | syskin | 768 | const uint32_t stride, const uint32_t iQuant, int rrv) |
689 : | chl | 96 | |
690 : | Isibaar | 3 | { |
691 : | syskin | 768 | if(!rrv) { |
692 : | uint32_t sadC = sad8(current->u + x*8 + y*stride*8, | ||
693 : | reference->u + x*8 + y*stride*8, stride); | ||
694 : | syskin | 704 | if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
695 : | syskin | 768 | sadC += sad8(current->v + (x + y*stride)*8, |
696 : | reference->v + (x + y*stride)*8, stride); | ||
697 : | syskin | 704 | if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
698 : | return 1; | ||
699 : | syskin | 768 | |
700 : | } else { | ||
701 : | uint32_t sadC = sad16(current->u + x*16 + y*stride*16, | ||
702 : | reference->u + x*16 + y*stride*16, stride, 256*4096); | ||
703 : | if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; | ||
704 : | sadC += sad16(current->v + (x + y*stride)*16, | ||
705 : | reference->v + (x + y*stride)*16, stride, 256*4096); | ||
706 : | if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; | ||
707 : | return 1; | ||
708 : | syskin | 704 | } |
709 : | chl | 530 | } |
710 : | edgomez | 195 | |
711 : | chl | 530 | static __inline void |
712 : | SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) | ||
713 : | { | ||
714 : | pMB->mode = MODE_NOT_CODED; | ||
715 : | syskin | 768 | pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV; |
716 : | pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV; | ||
717 : | chl | 530 | pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
718 : | Isibaar | 3 | } |
719 : | |||
720 : | syskin | 628 | bool |
721 : | chl | 530 | MotionEstimation(MBParam * const pParam, |
722 : | FRAMEINFO * const current, | ||
723 : | FRAMEINFO * const reference, | ||
724 : | const IMAGE * const pRefH, | ||
725 : | const IMAGE * const pRefV, | ||
726 : | const IMAGE * const pRefHV, | ||
727 : | const uint32_t iLimit) | ||
728 : | Isibaar | 3 | { |
729 : | chl | 530 | MACROBLOCK *const pMBs = current->mbs; |
730 : | const IMAGE *const pCurrent = ¤t->image; | ||
731 : | const IMAGE *const pRef = &reference->image; | ||
732 : | Isibaar | 3 | |
733 : | syskin | 704 | uint32_t mb_width = pParam->mb_width; |
734 : | uint32_t mb_height = pParam->mb_height; | ||
735 : | syskin | 768 | const uint32_t iEdgedWidth = pParam->edged_width; |
736 : | suxen_drol | 702 | |
737 : | chl | 530 | uint32_t x, y; |
738 : | uint32_t iIntra = 0; | ||
739 : | syskin | 628 | int32_t InterBias, quant = current->quant, sad00; |
740 : | edgomez | 195 | |
741 : | Isibaar | 539 | // some pre-initialized thingies for SearchP |
742 : | syskin | 712 | int32_t temp[8]; |
743 : | Isibaar | 539 | VECTOR currentMV[5]; |
744 : | Isibaar | 580 | VECTOR currentQMV[5]; |
745 : | Isibaar | 539 | int32_t iMinSAD[5]; |
746 : | SearchData Data; | ||
747 : | syskin | 704 | memset(&Data, 0, sizeof(SearchData)); |
748 : | syskin | 768 | Data.iEdgedWidth = iEdgedWidth; |
749 : | Isibaar | 539 | Data.currentMV = currentMV; |
750 : | Isibaar | 580 | Data.currentQMV = currentQMV; |
751 : | Isibaar | 539 | Data.iMinSAD = iMinSAD; |
752 : | Data.temp = temp; | ||
753 : | Data.iFcode = current->fcode; | ||
754 : | Isibaar | 580 | Data.rounding = pParam->m_rounding_type; |
755 : | syskin | 628 | Data.qpel = pParam->m_quarterpel; |
756 : | chl | 769 | Data.chroma = current->global_flags & ( PMV_CHROMA16 | PMV_CHROMA8 ); |
757 : | syskin | 704 | Data.rrv = current->global_flags & XVID_REDUCED; |
758 : | Isibaar | 539 | |
759 : | syskin | 704 | if ((current->global_flags & XVID_REDUCED)) { |
760 : | suxen_drol | 702 | mb_width = (pParam->width + 31) / 32; |
761 : | mb_height = (pParam->height + 31) / 32; | ||
762 : | syskin | 704 | Data.qpel = Data.chroma = 0; |
763 : | suxen_drol | 702 | } |
764 : | |||
765 : | syskin | 754 | Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose) |
766 : | chl | 530 | if (sadInit) (*sadInit) (); |
767 : | edgomez | 195 | |
768 : | suxen_drol | 702 | for (y = 0; y < mb_height; y++) { |
769 : | for (x = 0; x < mb_width; x++) { | ||
770 : | chl | 530 | MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
771 : | syskin | 768 | |
772 : | if (!Data.rrv) pMB->sad16 = | ||
773 : | sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16, | ||
774 : | pRef->y + (x + y * iEdgedWidth) * 16, | ||
775 : | syskin | 704 | pParam->edged_width, pMB->sad8 ); |
776 : | syskin | 768 | |
777 : | syskin | 748 | else pMB->sad16 = |
778 : | syskin | 768 | sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32, |
779 : | pRef->y + (x + y * iEdgedWidth) * 32, | ||
780 : | chl | 530 | pParam->edged_width, pMB->sad8 ); |
781 : | edgomez | 195 | |
782 : | syskin | 628 | if (Data.chroma) { |
783 : | syskin | 768 | Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8, |
784 : | pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2) | ||
785 : | + sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8, | ||
786 : | pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); | ||
787 : | pMB->sad16 += Data.temp[7]; | ||
788 : | syskin | 628 | } |
789 : | |||
790 : | syskin | 768 | sad00 = pMB->sad16; |
791 : | syskin | 628 | |
792 : | chl | 530 | if (!(current->global_flags & XVID_LUMIMASKING)) { |
793 : | pMB->dquant = NO_CHANGE; | ||
794 : | syskin | 601 | } else { |
795 : | syskin | 574 | if (pMB->dquant != NO_CHANGE) { |
796 : | quant += DQtab[pMB->dquant]; | ||
797 : | if (quant > 31) quant = 31; | ||
798 : | else if (quant < 1) quant = 1; | ||
799 : | } | ||
800 : | syskin | 601 | } |
801 : | syskin | 768 | pMB->quant = current->quant; |
802 : | chl | 181 | |
803 : | syskin | 628 | //initial skip decision |
804 : | chl | 769 | /* no early skip for GMC (global vector = skip vector is unknown!) */ |
805 : | if (!(current->global_flags & XVID_GMC)) { /* no fast SKIP for S(GMC)-VOPs */ | ||
806 : | if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH * (Data.rrv ? 4:1) ) | ||
807 : | syskin | 768 | if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) { |
808 : | chl | 619 | SkipMacroblockP(pMB, sad00); |
809 : | continue; | ||
810 : | } | ||
811 : | } | ||
812 : | syskin | 768 | |
813 : | syskin | 628 | SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
814 : | chl | 530 | y, current->motion_flags, pMB->quant, |
815 : | Isibaar | 539 | &Data, pParam, pMBs, reference->mbs, |
816 : | chl | 530 | current->global_flags & XVID_INTER4V, pMB); |
817 : | edgomez | 195 | |
818 : | chl | 530 | /* final skip decision, a.k.a. "the vector you found, really that good?" */ |
819 : | chl | 769 | if (!(current->global_flags & XVID_GMC)) { |
820 : | chl | 619 | if ( (pMB->dquant == NO_CHANGE) && (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
821 : | syskin | 704 | && ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1)) ) |
822 : | syskin | 768 | if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) { |
823 : | chl | 619 | SkipMacroblockP(pMB, sad00); |
824 : | continue; | ||
825 : | } | ||
826 : | } | ||
827 : | syskin | 768 | |
828 : | chl | 530 | /* finally, intra decision */ |
829 : | Isibaar | 3 | |
830 : | chl | 530 | InterBias = MV16_INTER_BIAS; |
831 : | syskin | 768 | if (pMB->quant > 8) InterBias += 100 * (pMB->quant - 8); // to make high quants work |
832 : | chl | 530 | if (y != 0) |
833 : | syskin | 639 | if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; |
834 : | chl | 530 | if (x != 0) |
835 : | syskin | 639 | if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; |
836 : | syskin | 768 | |
837 : | syskin | 628 | if (Data.chroma) InterBias += 50; // to compensate bigger SAD |
838 : | syskin | 768 | if (Data.rrv) InterBias *= 4; |
839 : | Isibaar | 3 | |
840 : | syskin | 768 | if (InterBias < pMB->sad16) { |
841 : | syskin | 704 | int32_t deviation; |
842 : | syskin | 768 | if (!Data.rrv) |
843 : | deviation = dev16(pCurrent->y + (x + y * iEdgedWidth) * 16, iEdgedWidth); | ||
844 : | else { | ||
845 : | deviation = dev16(pCurrent->y + (x + y * iEdgedWidth) * 32, iEdgedWidth) | ||
846 : | + dev16(pCurrent->y + (x + y * iEdgedWidth) * 32 + 16, iEdgedWidth) | ||
847 : | + dev16(pCurrent->y + (x + y * iEdgedWidth) * 32 + 16 * iEdgedWidth, iEdgedWidth) | ||
848 : | + dev16(pCurrent->y + (x + y * iEdgedWidth) * 32 + 16 * (iEdgedWidth+1), iEdgedWidth); | ||
849 : | } | ||
850 : | if (deviation < (pMB->sad16 - InterBias)) { | ||
851 : | syskin | 754 | if (++iIntra >= iLimit) return 1; |
852 : | syskin | 768 | SkipMacroblockP(pMB, 0); //same thing |
853 : | chl | 530 | pMB->mode = MODE_INTRA; |
854 : | } | ||
855 : | } | ||
856 : | } | ||
857 : | Isibaar | 3 | } |
858 : | chl | 619 | |
859 : | chl | 769 | if (current->global_flags & XVID_GMC ) /* first GMC step only for S(GMC)-VOPs */ |
860 : | { | ||
861 : | current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV); | ||
862 : | } | ||
863 : | |||
864 : | syskin | 628 | return 0; |
865 : | chl | 530 | } |
866 : | Isibaar | 3 | |
867 : | edgomez | 195 | |
868 : | chl | 530 | static __inline int |
869 : | make_mask(const VECTOR * const pmv, const int i) | ||
870 : | { | ||
871 : | Isibaar | 539 | int mask = 255, j; |
872 : | chl | 530 | for (j = 0; j < i; j++) { |
873 : | if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already | ||
874 : | if (pmv[i].x == pmv[j].x) { | ||
875 : | syskin | 768 | if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4; |
876 : | else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8; | ||
877 : | syskin | 748 | } else |
878 : | chl | 530 | if (pmv[i].y == pmv[j].y) { |
879 : | syskin | 768 | if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1; |
880 : | else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2; | ||
881 : | chl | 530 | } |
882 : | Isibaar | 3 | } |
883 : | chl | 530 | return mask; |
884 : | } | ||
885 : | edgomez | 195 | |
886 : | syskin | 748 | static __inline void |
887 : | syskin | 704 | PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, |
888 : | int iHcount, const MACROBLOCK * const prevMB, int rrv) | ||
889 : | chl | 530 | { |
890 : | edgomez | 195 | |
891 : | chl | 530 | //this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself |
892 : | syskin | 704 | if (rrv) { iWcount /= 2; iHcount /= 2; } |
893 : | edgomez | 195 | |
894 : | syskin | 704 | if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour |
895 : | chl | 530 | pmv[5].x = EVEN(pmv[3].x); |
896 : | syskin | 748 | pmv[5].y = EVEN(pmv[3].y); |
897 : | Isibaar | 539 | } else pmv[5].x = pmv[5].y = 0; |
898 : | chl | 169 | |
899 : | chl | 530 | if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour |
900 : | else pmv[3].x = pmv[3].y = 0; | ||
901 : | Isibaar | 3 | |
902 : | chl | 530 | if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour |
903 : | syskin | 768 | else pmv[4].x = pmv[4].y = 0; |
904 : | chl | 169 | |
905 : | chl | 530 | // [1] median prediction |
906 : | syskin | 767 | pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); |
907 : | chl | 169 | |
908 : | 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 |
909 : | chl | 169 | |
910 : | chl | 530 | pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
911 : | pmv[2].y = EVEN(prevMB->mvs[0].y); | ||
912 : | chl | 169 | |
913 : | syskin | 704 | if ((x < iWcount-1) && (y < iHcount-1)) { |
914 : | chl | 530 | pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame |
915 : | syskin | 748 | pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
916 : | Isibaar | 539 | } else pmv[6].x = pmv[6].y = 0; |
917 : | syskin | 704 | |
918 : | if (rrv) { | ||
919 : | int i; | ||
920 : | for (i = 0; i < 7; i++) { | ||
921 : | syskin | 768 | pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); |
922 : | syskin | 767 | pmv[i].y = RRV_MV_SCALEUP(pmv[i].y); |
923 : | syskin | 704 | } |
924 : | } | ||
925 : | chl | 530 | } |
926 : | chl | 169 | |
927 : | chl | 530 | static void |
928 : | syskin | 628 | SearchP(const IMAGE * const pRef, |
929 : | chl | 530 | const uint8_t * const pRefH, |
930 : | const uint8_t * const pRefV, | ||
931 : | const uint8_t * const pRefHV, | ||
932 : | const IMAGE * const pCur, | ||
933 : | const int x, | ||
934 : | const int y, | ||
935 : | const uint32_t MotionFlags, | ||
936 : | const uint32_t iQuant, | ||
937 : | Isibaar | 539 | SearchData * const Data, |
938 : | chl | 530 | const MBParam * const pParam, |
939 : | const MACROBLOCK * const pMBs, | ||
940 : | const MACROBLOCK * const prevMBs, | ||
941 : | int inter4v, | ||
942 : | MACROBLOCK * const pMB) | ||
943 : | { | ||
944 : | chl | 169 | |
945 : | chl | 530 | int i, iDirection = 255, mask, threshA; |
946 : | Isibaar | 539 | VECTOR pmv[7]; |
947 : | Isibaar | 3 | |
948 : | syskin | 712 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
949 : | pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); | ||
950 : | Isibaar | 3 | |
951 : | syskin | 768 | get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); |
952 : | syskin | 628 | |
953 : | syskin | 768 | Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache |
954 : | i = Data->rrv ? 2 : 1; | ||
955 : | syskin | 704 | Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i; |
956 : | Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; | ||
957 : | Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; | ||
958 : | Isibaar | 3 | |
959 : | syskin | 704 | Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16*i; |
960 : | Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16*i; | ||
961 : | Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16*i; | ||
962 : | Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16*i; | ||
963 : | Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; | ||
964 : | Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; | ||
965 : | |||
966 : | syskin | 601 | Data->lambda16 = lambda_vec16[iQuant]; |
967 : | Isibaar | 606 | Data->lambda8 = lambda_vec8[iQuant]; |
968 : | syskin | 663 | Data->qpel_precision = 0; |
969 : | chl | 169 | |
970 : | edgomez | 568 | if (pMB->dquant != NO_CHANGE) inter4v = 0; |
971 : | edgomez | 195 | |
972 : | syskin | 768 | for(i = 0; i < 5; i++) |
973 : | Isibaar | 580 | Data->currentMV[i].x = Data->currentMV[i].y = 0; |
974 : | Isibaar | 539 | |
975 : | syskin | 768 | if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
976 : | syskin | 663 | else Data->predMV = pmv[0]; |
977 : | Isibaar | 606 | |
978 : | syskin | 724 | i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0); |
979 : | syskin | 768 | Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10); |
980 : | Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10); | ||
981 : | Isibaar | 539 | Data->iMinSAD[2] = pMB->sad8[1]; |
982 : | Data->iMinSAD[3] = pMB->sad8[2]; | ||
983 : | Data->iMinSAD[4] = pMB->sad8[3]; | ||
984 : | |||
985 : | syskin | 768 | if (x | y) { |
986 : | edgomez | 568 | threshA = Data->temp[0]; // that's when we keep this SAD atm |
987 : | chl | 530 | if (threshA < 512) threshA = 512; |
988 : | syskin | 768 | else if (threshA > 1024) threshA = 1024; |
989 : | } else threshA = 512; | ||
990 : | Isibaar | 3 | |
991 : | chl | 530 | PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
992 : | syskin | 704 | prevMBs + x + y * pParam->mb_width, Data->rrv); |
993 : | edgomez | 195 | |
994 : | syskin | 768 | if (!Data->rrv) { |
995 : | if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16; | ||
996 : | else CheckCandidate = CheckCandidate16no4v; //for extra speed | ||
997 : | } else CheckCandidate = CheckCandidate32; | ||
998 : | edgomez | 568 | |
999 : | syskin | 768 | /* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/ |
1000 : | chl | 530 | |
1001 : | for (i = 1; i < 7; i++) { | ||
1002 : | if (!(mask = make_mask(pmv, i)) ) continue; | ||
1003 : | edgomez | 568 | (*CheckCandidate)(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1004 : | Isibaar | 539 | if (Data->iMinSAD[0] <= threshA) break; |
1005 : | chl | 530 | } |
1006 : | |||
1007 : | Isibaar | 539 | if ((Data->iMinSAD[0] <= threshA) || |
1008 : | (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && | ||
1009 : | syskin | 768 | (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) |
1010 : | chl | 530 | inter4v = 0; |
1011 : | syskin | 768 | else { |
1012 : | Isibaar | 3 | |
1013 : | syskin | 748 | MainSearchFunc * MainSearchPtr; |
1014 : | Isibaar | 539 | if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1015 : | else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; | ||
1016 : | else MainSearchPtr = DiamondSearch; | ||
1017 : | chl | 181 | |
1018 : | Isibaar | 539 | (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
1019 : | Isibaar | 3 | |
1020 : | chl | 530 | /* extended search, diamond starting in 0,0 and in prediction. |
1021 : | note that this search is/might be done in halfpel positions, | ||
1022 : | which makes it more different than the diamond above */ | ||
1023 : | chl | 259 | |
1024 : | Isibaar | 539 | if (MotionFlags & PMV_EXTSEARCH16) { |
1025 : | int32_t bSAD; | ||
1026 : | VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; | ||
1027 : | syskin | 704 | if (Data->rrv) { |
1028 : | startMV.x = RRV_MV_SCALEUP(startMV.x); | ||
1029 : | startMV.y = RRV_MV_SCALEUP(startMV.y); | ||
1030 : | syskin | 748 | } else |
1031 : | syskin | 704 | if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? |
1032 : | startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); | ||
1033 : | Isibaar | 539 | if (!(MVequal(startMV, backupMV))) { |
1034 : | bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; | ||
1035 : | Isibaar | 3 | |
1036 : | Isibaar | 606 | (*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); |
1037 : | Isibaar | 539 | (*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
1038 : | if (bSAD < Data->iMinSAD[0]) { | ||
1039 : | Data->currentMV[0] = backupMV; | ||
1040 : | Data->iMinSAD[0] = bSAD; } | ||
1041 : | } | ||
1042 : | Isibaar | 3 | |
1043 : | Isibaar | 539 | backupMV = Data->currentMV[0]; |
1044 : | syskin | 768 | if (MotionFlags & PMV_HALFPELREFINE16 && !Data->rrv) startMV.x = startMV.y = 1; |
1045 : | else startMV.x = startMV.y = 0; | ||
1046 : | Isibaar | 539 | if (!(MVequal(startMV, backupMV))) { |
1047 : | bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; | ||
1048 : | edgomez | 195 | |
1049 : | Isibaar | 606 | (*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); |
1050 : | Isibaar | 539 | (*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
1051 : | if (bSAD < Data->iMinSAD[0]) { | ||
1052 : | Data->currentMV[0] = backupMV; | ||
1053 : | Data->iMinSAD[0] = bSAD; } | ||
1054 : | } | ||
1055 : | Isibaar | 3 | } |
1056 : | } | ||
1057 : | |||
1058 : | syskin | 663 | if (MotionFlags & PMV_HALFPELREFINE16) SubpelRefine(Data); |
1059 : | Isibaar | 3 | |
1060 : | Isibaar | 580 | for(i = 0; i < 5; i++) { |
1061 : | Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors | ||
1062 : | Data->currentQMV[i].y = 2 * Data->currentMV[i].y; | ||
1063 : | } | ||
1064 : | |||
1065 : | syskin | 768 | if (Data->qpel && MotionFlags & PMV_QUARTERPELREFINE16) { |
1066 : | syskin | 663 | Data->qpel_precision = 1; |
1067 : | syskin | 712 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1068 : | pParam->width, pParam->height, Data->iFcode, 1, 0); | ||
1069 : | syskin | 601 | |
1070 : | syskin | 663 | SubpelRefine(Data); |
1071 : | Isibaar | 580 | } |
1072 : | |||
1073 : | syskin | 768 | if (Data->iMinSAD[0] < (int32_t)iQuant * 30) inter4v = 0; |
1074 : | Isibaar | 539 | if (inter4v) { |
1075 : | SearchData Data8; | ||
1076 : | syskin | 768 | memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data |
1077 : | |||
1078 : | Isibaar | 539 | Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
1079 : | Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); | ||
1080 : | Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); | ||
1081 : | Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); | ||
1082 : | syskin | 768 | |
1083 : | syskin | 628 | if (Data->chroma) { |
1084 : | syskin | 768 | int sumx = 0, sumy = 0; |
1085 : | const int div = 1 + Data->qpel; | ||
1086 : | const VECTOR * const mv = Data->qpel ? pMB->qmvs : pMB->mvs; | ||
1087 : | syskin | 628 | |
1088 : | syskin | 768 | for (i = 0; i < 4; i++) { |
1089 : | sumx += mv[i].x / div; | ||
1090 : | sumy += mv[i].y / div; | ||
1091 : | syskin | 663 | } |
1092 : | syskin | 768 | |
1093 : | syskin | 753 | Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf], |
1094 : | (sumy >> 3) + roundtab_76[sumy & 0xf], Data); | ||
1095 : | syskin | 628 | } |
1096 : | Isibaar | 539 | } |
1097 : | Isibaar | 3 | |
1098 : | syskin | 704 | if (Data->rrv) { |
1099 : | Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); | ||
1100 : | Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); | ||
1101 : | } | ||
1102 : | syskin | 740 | |
1103 : | chl | 530 | if (!(inter4v) || |
1104 : | syskin | 748 | (Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
1105 : | Isibaar | 539 | Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
1106 : | chl | 530 | // INTER MODE |
1107 : | pMB->mode = MODE_INTER; | ||
1108 : | syskin | 768 | pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
1109 : | pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; | ||
1110 : | Isibaar | 3 | |
1111 : | syskin | 768 | if(Data->qpel) { |
1112 : | syskin | 663 | pMB->qmvs[0] = pMB->qmvs[1] |
1113 : | = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; | ||
1114 : | pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; | ||
1115 : | pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; | ||
1116 : | syskin | 628 | } else { |
1117 : | Isibaar | 580 | pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
1118 : | pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; | ||
1119 : | } | ||
1120 : | edgomez | 195 | } else { |
1121 : | syskin | 748 | // INTER4V MODE; all other things are already set in Search8 |
1122 : | chl | 530 | pMB->mode = MODE_INTER4V; |
1123 : | syskin | 748 | pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + |
1124 : | Isibaar | 539 | Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant; |
1125 : | edgomez | 78 | } |
1126 : | Isibaar | 3 | } |
1127 : | |||
1128 : | chl | 530 | static void |
1129 : | Search8(const SearchData * const OldData, | ||
1130 : | const int x, const int y, | ||
1131 : | const uint32_t MotionFlags, | ||
1132 : | const MBParam * const pParam, | ||
1133 : | MACROBLOCK * const pMB, | ||
1134 : | const MACROBLOCK * const pMBs, | ||
1135 : | Isibaar | 539 | const int block, |
1136 : | SearchData * const Data) | ||
1137 : | chl | 345 | { |
1138 : | syskin | 704 | int i = 0; |
1139 : | Isibaar | 539 | Data->iMinSAD = OldData->iMinSAD + 1 + block; |
1140 : | Data->currentMV = OldData->currentMV + 1 + block; | ||
1141 : | Isibaar | 580 | Data->currentQMV = OldData->currentQMV + 1 + block; |
1142 : | chl | 345 | |
1143 : | syskin | 768 | if(Data->qpel) { |
1144 : | syskin | 704 | Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
1145 : | syskin | 724 | if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y, |
1146 : | Data->predMV, Data->iFcode, 0, 0); | ||
1147 : | Isibaar | 606 | } else { |
1148 : | syskin | 704 | Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
1149 : | syskin | 768 | if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y, |
1150 : | Data->predMV, Data->iFcode, 0, Data->rrv); | ||
1151 : | Isibaar | 606 | } |
1152 : | chl | 345 | |
1153 : | syskin | 768 | *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10; |
1154 : | syskin | 704 | |
1155 : | syskin | 768 | if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8|PMV_QUARTERPELREFINE8)) { |
1156 : | syskin | 704 | if (Data->rrv) i = 2; else i = 1; |
1157 : | chl | 345 | |
1158 : | syskin | 768 | Data->Ref = OldData->Ref + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1159 : | Data->RefH = OldData->RefH + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); | ||
1160 : | Data->RefV = OldData->RefV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); | ||
1161 : | Data->RefHV = OldData->RefHV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); | ||
1162 : | chl | 345 | |
1163 : | syskin | 768 | Data->Cur = OldData->Cur + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1164 : | syskin | 663 | Data->qpel_precision = 0; |
1165 : | chl | 345 | |
1166 : | syskin | 712 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
1167 : | syskin | 768 | pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
1168 : | syskin | 712 | |
1169 : | syskin | 768 | if (!Data->rrv) CheckCandidate = CheckCandidate8; |
1170 : | else CheckCandidate = CheckCandidate16no4v; | ||
1171 : | syskin | 712 | |
1172 : | chl | 530 | if (MotionFlags & PMV_EXTSEARCH8) { |
1173 : | Isibaar | 580 | int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
1174 : | syskin | 768 | |
1175 : | chl | 530 | MainSearchFunc *MainSearchPtr; |
1176 : | if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; | ||
1177 : | else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; | ||
1178 : | else MainSearchPtr = DiamondSearch; | ||
1179 : | chl | 345 | |
1180 : | Isibaar | 580 | (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1181 : | chl | 345 | |
1182 : | syskin | 601 | if(*(Data->iMinSAD) < temp_sad) { |
1183 : | Isibaar | 580 | Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
1184 : | Data->currentQMV->y = 2 * Data->currentMV->y; | ||
1185 : | } | ||
1186 : | } | ||
1187 : | |||
1188 : | if (MotionFlags & PMV_HALFPELREFINE8) { | ||
1189 : | int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD | ||
1190 : | |||
1191 : | syskin | 663 | SubpelRefine(Data); // perform halfpel refine of current best vector |
1192 : | Isibaar | 580 | |
1193 : | if(*(Data->iMinSAD) < temp_sad) { // we have found a better match | ||
1194 : | Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector | ||
1195 : | Data->currentQMV->y = 2 * Data->currentMV->y; | ||
1196 : | } | ||
1197 : | } | ||
1198 : | |||
1199 : | syskin | 768 | if (Data->qpel && MotionFlags & PMV_QUARTERPELREFINE8) { |
1200 : | Data->qpel_precision = 1; | ||
1201 : | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, | ||
1202 : | pParam->width, pParam->height, Data->iFcode, 1, 0); | ||
1203 : | SubpelRefine(Data); | ||
1204 : | Isibaar | 580 | } |
1205 : | chl | 345 | } |
1206 : | chl | 530 | |
1207 : | syskin | 704 | if (Data->rrv) { |
1208 : | Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x); | ||
1209 : | Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y); | ||
1210 : | } | ||
1211 : | syskin | 768 | |
1212 : | syskin | 712 | if(Data->qpel) { |
1213 : | syskin | 663 | pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x; |
1214 : | pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y; | ||
1215 : | syskin | 768 | pMB->qmvs[block] = *Data->currentQMV; |
1216 : | syskin | 704 | } else { |
1217 : | Isibaar | 580 | pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
1218 : | pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; | ||
1219 : | } | ||
1220 : | |||
1221 : | syskin | 768 | pMB->mvs[block] = *Data->currentMV; |
1222 : | pMB->sad8[block] = 4 * *Data->iMinSAD; | ||
1223 : | chl | 345 | } |
1224 : | |||
1225 : | syskin | 768 | /* motion estimation for B-frames */ |
1226 : | chl | 345 | |
1227 : | chl | 530 | static __inline VECTOR |
1228 : | ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) | ||
1229 : | Isibaar | 3 | { |
1230 : | chl | 530 | /* the stupidiest function ever */ |
1231 : | syskin | 768 | return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]); |
1232 : | Isibaar | 3 | } |
1233 : | |||
1234 : | chl | 530 | static void __inline |
1235 : | PreparePredictionsBF(VECTOR * const pmv, const int x, const int y, | ||
1236 : | const uint32_t iWcount, | ||
1237 : | const MACROBLOCK * const pMB, | ||
1238 : | const uint32_t mode_curr) | ||
1239 : | Isibaar | 3 | { |
1240 : | |||
1241 : | chl | 530 | // [0] is prediction |
1242 : | pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); | ||
1243 : | Isibaar | 3 | |
1244 : | chl | 530 | pmv[1].x = pmv[1].y = 0; // [1] is zero |
1245 : | Isibaar | 3 | |
1246 : | chl | 530 | pmv[2] = ChoosePred(pMB, mode_curr); |
1247 : | pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); | ||
1248 : | edgomez | 195 | |
1249 : | chl | 530 | if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour |
1250 : | pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); | ||
1251 : | syskin | 748 | pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); |
1252 : | Isibaar | 539 | } else pmv[3].x = pmv[3].y = 0; |
1253 : | Isibaar | 3 | |
1254 : | chl | 530 | if (y != 0) { |
1255 : | pmv[4] = ChoosePred(pMB-iWcount, mode_curr); | ||
1256 : | pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y); | ||
1257 : | } else pmv[4].x = pmv[4].y = 0; | ||
1258 : | Isibaar | 3 | |
1259 : | chl | 530 | if (x != 0) { |
1260 : | pmv[5] = ChoosePred(pMB-1, mode_curr); | ||
1261 : | pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); | ||
1262 : | } else pmv[5].x = pmv[5].y = 0; | ||
1263 : | Isibaar | 3 | |
1264 : | syskin | 768 | if (x != 0 && y != 0) { |
1265 : | chl | 530 | pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); |
1266 : | syskin | 767 | pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y); |
1267 : | chl | 530 | } else pmv[6].x = pmv[6].y = 0; |
1268 : | } | ||
1269 : | chl | 181 | |
1270 : | edgomez | 170 | |
1271 : | syskin | 768 | /* search backward or forward */ |
1272 : | chl | 530 | static void |
1273 : | SearchBF( const uint8_t * const pRef, | ||
1274 : | const uint8_t * const pRefH, | ||
1275 : | const uint8_t * const pRefV, | ||
1276 : | const uint8_t * const pRefHV, | ||
1277 : | const IMAGE * const pCur, | ||
1278 : | const int x, const int y, | ||
1279 : | const uint32_t MotionFlags, | ||
1280 : | const uint32_t iFcode, | ||
1281 : | const MBParam * const pParam, | ||
1282 : | MACROBLOCK * const pMB, | ||
1283 : | const VECTOR * const predMV, | ||
1284 : | int32_t * const best_sad, | ||
1285 : | h | 545 | const int32_t mode_current, |
1286 : | SearchData * const Data) | ||
1287 : | chl | 530 | { |
1288 : | Isibaar | 3 | |
1289 : | syskin | 740 | int i, iDirection = 255, mask; |
1290 : | h | 545 | VECTOR pmv[7]; |
1291 : | chl | 530 | MainSearchFunc *MainSearchPtr; |
1292 : | h | 545 | *Data->iMinSAD = MV_MAX_ERROR; |
1293 : | Data->iFcode = iFcode; | ||
1294 : | syskin | 663 | Data->qpel_precision = 0; |
1295 : | Isibaar | 3 | |
1296 : | syskin | 768 | Data->Ref = pRef + (x + y * Data->iEdgedWidth) * 16; |
1297 : | Data->RefH = pRefH + (x + y * Data->iEdgedWidth) * 16; | ||
1298 : | Data->RefV = pRefV + (x + y * Data->iEdgedWidth) * 16; | ||
1299 : | Data->RefHV = pRefHV + (x + y * Data->iEdgedWidth) * 16; | ||
1300 : | Isibaar | 3 | |
1301 : | h | 545 | Data->predMV = *predMV; |
1302 : | edgomez | 195 | |
1303 : | h | 545 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1304 : | syskin | 712 | pParam->width, pParam->height, iFcode - Data->qpel, 0, 0); |
1305 : | Isibaar | 3 | |
1306 : | h | 545 | pmv[0] = Data->predMV; |
1307 : | syskin | 658 | if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } |
1308 : | syskin | 768 | |
1309 : | h | 545 | PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
1310 : | edgomez | 195 | |
1311 : | h | 545 | Data->currentMV->x = Data->currentMV->y = 0; |
1312 : | chl | 530 | CheckCandidate = CheckCandidate16no4v; |
1313 : | chl | 181 | |
1314 : | chl | 530 | // main loop. checking all predictions |
1315 : | syskin | 740 | for (i = 0; i < 7; i++) { |
1316 : | chl | 530 | if (!(mask = make_mask(pmv, i)) ) continue; |
1317 : | h | 545 | CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1318 : | edgomez | 78 | } |
1319 : | Isibaar | 3 | |
1320 : | syskin | 768 | if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1321 : | else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; | ||
1322 : | chl | 530 | else MainSearchPtr = DiamondSearch; |
1323 : | chl | 96 | |
1324 : | syskin | 740 | (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
1325 : | chl | 169 | |
1326 : | syskin | 663 | SubpelRefine(Data); |
1327 : | syskin | 768 | |
1328 : | syskin | 740 | if (Data->qpel && *Data->iMinSAD < *best_sad + 300) { |
1329 : | syskin | 658 | Data->currentQMV->x = 2*Data->currentMV->x; |
1330 : | Data->currentQMV->y = 2*Data->currentMV->y; | ||
1331 : | syskin | 663 | Data->qpel_precision = 1; |
1332 : | syskin | 712 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1333 : | pParam->width, pParam->height, iFcode, 1, 0); | ||
1334 : | syskin | 663 | SubpelRefine(Data); |
1335 : | syskin | 658 | } |
1336 : | chl | 169 | |
1337 : | chl | 530 | // three bits are needed to code backward mode. four for forward |
1338 : | syskin | 740 | |
1339 : | syskin | 768 | if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
1340 : | else *Data->iMinSAD += 3 * Data->lambda16; | ||
1341 : | chl | 169 | |
1342 : | h | 545 | if (*Data->iMinSAD < *best_sad) { |
1343 : | *best_sad = *Data->iMinSAD; | ||
1344 : | chl | 530 | pMB->mode = mode_current; |
1345 : | syskin | 658 | if (Data->qpel) { |
1346 : | pMB->pmvs[0].x = Data->currentQMV->x - predMV->x; | ||
1347 : | pMB->pmvs[0].y = Data->currentQMV->y - predMV->y; | ||
1348 : | syskin | 748 | if (mode_current == MODE_FORWARD) |
1349 : | syskin | 658 | pMB->qmvs[0] = *Data->currentQMV; |
1350 : | syskin | 748 | else |
1351 : | syskin | 658 | pMB->b_qmvs[0] = *Data->currentQMV; |
1352 : | } else { | ||
1353 : | pMB->pmvs[0].x = Data->currentMV->x - predMV->x; | ||
1354 : | pMB->pmvs[0].y = Data->currentMV->y - predMV->y; | ||
1355 : | } | ||
1356 : | syskin | 739 | if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; |
1357 : | else pMB->b_mvs[0] = *Data->currentMV; | ||
1358 : | } | ||
1359 : | syskin | 658 | |
1360 : | syskin | 768 | if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; |
1361 : | else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search | ||
1362 : | chl | 530 | } |
1363 : | chl | 169 | |
1364 : | syskin | 748 | static void |
1365 : | syskin | 712 | SkipDecisionB(const IMAGE * const pCur, |
1366 : | syskin | 767 | const IMAGE * const f_Ref, |
1367 : | const IMAGE * const b_Ref, | ||
1368 : | MACROBLOCK * const pMB, | ||
1369 : | const uint32_t x, const uint32_t y, | ||
1370 : | const SearchData * const Data) | ||
1371 : | syskin | 712 | { |
1372 : | syskin | 767 | int dx = 0, dy = 0, b_dx = 0, b_dy = 0; |
1373 : | syskin | 768 | int32_t sum; |
1374 : | syskin | 767 | const int div = 1 + Data->qpel; |
1375 : | int k; | ||
1376 : | syskin | 768 | const uint32_t stride = Data->iEdgedWidth/2; |
1377 : | syskin | 712 | //this is not full chroma compensation, only it's fullpel approximation. should work though |
1378 : | |||
1379 : | syskin | 767 | for (k = 0; k < 4; k++) { |
1380 : | dy += Data->directmvF[k].y / div; | ||
1381 : | dx += Data->directmvF[0].x / div; | ||
1382 : | b_dy += Data->directmvB[0].y / div; | ||
1383 : | b_dx += Data->directmvB[0].x / div; | ||
1384 : | syskin | 712 | } |
1385 : | |||
1386 : | dy = (dy >> 3) + roundtab_76[dy & 0xf]; | ||
1387 : | dx = (dx >> 3) + roundtab_76[dx & 0xf]; | ||
1388 : | b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf]; | ||
1389 : | b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf]; | ||
1390 : | |||
1391 : | syskin | 768 | sum = sad8bi(pCur->u + 8 * x + 8 * y * stride, |
1392 : | f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2, | ||
1393 : | b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, | ||
1394 : | stride); | ||
1395 : | syskin | 767 | |
1396 : | syskin | 768 | if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; //no skip |
1397 : | syskin | 767 | |
1398 : | syskin | 768 | sum += sad8bi(pCur->v + 8*x + 8 * y * stride, |
1399 : | f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2, | ||
1400 : | b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, | ||
1401 : | stride); | ||
1402 : | syskin | 712 | |
1403 : | syskin | 768 | if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) pMB->mode = MODE_DIRECT_NONE_MV; //skipped |
1404 : | syskin | 712 | } |
1405 : | |||
1406 : | static __inline uint32_t | ||
1407 : | Isibaar | 539 | SearchDirect(const IMAGE * const f_Ref, |
1408 : | chl | 530 | const uint8_t * const f_RefH, |
1409 : | const uint8_t * const f_RefV, | ||
1410 : | const uint8_t * const f_RefHV, | ||
1411 : | Isibaar | 539 | const IMAGE * const b_Ref, |
1412 : | chl | 530 | const uint8_t * const b_RefH, |
1413 : | const uint8_t * const b_RefV, | ||
1414 : | const uint8_t * const b_RefHV, | ||
1415 : | const IMAGE * const pCur, | ||
1416 : | const int x, const int y, | ||
1417 : | const uint32_t MotionFlags, | ||
1418 : | const int32_t TRB, const int32_t TRD, | ||
1419 : | const MBParam * const pParam, | ||
1420 : | MACROBLOCK * const pMB, | ||
1421 : | const MACROBLOCK * const b_mb, | ||
1422 : | h | 545 | int32_t * const best_sad, |
1423 : | SearchData * const Data) | ||
1424 : | edgomez | 195 | |
1425 : | chl | 530 | { |
1426 : | h | 545 | int32_t skip_sad; |
1427 : | syskin | 768 | int k = (x + Data->iEdgedWidth*y) * 16; |
1428 : | chl | 530 | MainSearchFunc *MainSearchPtr; |
1429 : | Isibaar | 3 | |
1430 : | h | 545 | *Data->iMinSAD = 256*4096; |
1431 : | syskin | 768 | Data->Ref = f_Ref->y + k; |
1432 : | Data->RefH = f_RefH + k; | ||
1433 : | Data->RefV = f_RefV + k; | ||
1434 : | Data->RefHV = f_RefHV + k; | ||
1435 : | Data->bRef = b_Ref->y + k; | ||
1436 : | Data->bRefH = b_RefH + k; | ||
1437 : | Data->bRefV = b_RefV + k; | ||
1438 : | Data->bRefHV = b_RefHV + k; | ||
1439 : | chl | 140 | |
1440 : | syskin | 768 | k = Data->qpel ? 4 : 2; |
1441 : | Data->max_dx = k * (pParam->width - x * 16); | ||
1442 : | Data->max_dy = k * (pParam->height - y * 16); | ||
1443 : | Data->min_dx = -k * (16 + x * 16); | ||
1444 : | Data->min_dy = -k * (16 + y * 16); | ||
1445 : | chl | 140 | |
1446 : | syskin | 768 | Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs; |
1447 : | Data->qpel_precision = 0; | ||
1448 : | Isibaar | 3 | |
1449 : | chl | 530 | for (k = 0; k < 4; k++) { |
1450 : | h | 545 | pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); |
1451 : | pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD; | ||
1452 : | pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); | ||
1453 : | pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; | ||
1454 : | Isibaar | 3 | |
1455 : | syskin | 768 | if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx) |
1456 : | | (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) { | ||
1457 : | Isibaar | 3 | |
1458 : | Isibaar | 539 | *best_sad = 256*4096; // in that case, we won't use direct mode |
1459 : | pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV" | ||
1460 : | pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; | ||
1461 : | syskin | 712 | return 256*4096; |
1462 : | Isibaar | 539 | } |
1463 : | if (b_mb->mode != MODE_INTER4V) { | ||
1464 : | pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; | ||
1465 : | pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0]; | ||
1466 : | h | 545 | Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0]; |
1467 : | Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0]; | ||
1468 : | Isibaar | 539 | break; |
1469 : | } | ||
1470 : | } | ||
1471 : | Isibaar | 3 | |
1472 : | syskin | 768 | CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v; |
1473 : | |||
1474 : | h | 545 | (*CheckCandidate)(0, 0, 255, &k, Data); |
1475 : | Isibaar | 3 | |
1476 : | syskin | 712 | // initial (fast) skip decision |
1477 : | syskin | 768 | if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * 2) { |
1478 : | //possible skip - checking chroma | ||
1479 : | SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data); | ||
1480 : | syskin | 712 | if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip. |
1481 : | Isibaar | 539 | } |
1482 : | |||
1483 : | h | 545 | skip_sad = *Data->iMinSAD; |
1484 : | edgomez | 195 | |
1485 : | syskin | 768 | // DIRECT MODE DELTA VECTOR SEARCH. |
1486 : | chl | 530 | // This has to be made more effective, but at the moment I'm happy it's running at all |
1487 : | Isibaar | 3 | |
1488 : | chl | 530 | if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1489 : | else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; | ||
1490 : | else MainSearchPtr = DiamondSearch; | ||
1491 : | edgomez | 195 | |
1492 : | h | 545 | (*MainSearchPtr)(0, 0, Data, 255); |
1493 : | Isibaar | 3 | |
1494 : | syskin | 663 | SubpelRefine(Data); |
1495 : | Isibaar | 3 | |
1496 : | h | 545 | *best_sad = *Data->iMinSAD; |
1497 : | Isibaar | 3 | |
1498 : | syskin | 768 | if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT; |
1499 : | syskin | 672 | else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
1500 : | edgomez | 195 | |
1501 : | h | 545 | pMB->pmvs[3] = *Data->currentMV; |
1502 : | edgomez | 195 | |
1503 : | chl | 530 | for (k = 0; k < 4; k++) { |
1504 : | h | 545 | pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x; |
1505 : | syskin | 658 | pMB->b_mvs[k].x = ( (Data->currentMV->x == 0) |
1506 : | h | 545 | ? Data->directmvB[k].x |
1507 : | syskin | 658 | :pMB->mvs[k].x - Data->referencemv[k].x); |
1508 : | h | 545 | pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y); |
1509 : | pMB->b_mvs[k].y = ((Data->currentMV->y == 0) | ||
1510 : | ? Data->directmvB[k].y | ||
1511 : | : pMB->mvs[k].y - Data->referencemv[k].y); | ||
1512 : | syskin | 658 | if (Data->qpel) { |
1513 : | pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2; | ||
1514 : | pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2; | ||
1515 : | pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2; | ||
1516 : | pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2; | ||
1517 : | } | ||
1518 : | |||
1519 : | chl | 530 | if (b_mb->mode != MODE_INTER4V) { |
1520 : | pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; | ||
1521 : | pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; | ||
1522 : | syskin | 658 | pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0]; |
1523 : | pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0]; | ||
1524 : | chl | 530 | break; |
1525 : | } | ||
1526 : | } | ||
1527 : | h | 545 | return skip_sad; |
1528 : | Isibaar | 3 | } |
1529 : | chl | 96 | |
1530 : | syskin | 712 | static void |
1531 : | chl | 530 | SearchInterpolate(const uint8_t * const f_Ref, |
1532 : | const uint8_t * const f_RefH, | ||
1533 : | const uint8_t * const f_RefV, | ||
1534 : | const uint8_t * const f_RefHV, | ||
1535 : | const uint8_t * const b_Ref, | ||
1536 : | const uint8_t * const b_RefH, | ||
1537 : | const uint8_t * const b_RefV, | ||
1538 : | const uint8_t * const b_RefHV, | ||
1539 : | const IMAGE * const pCur, | ||
1540 : | const int x, const int y, | ||
1541 : | const uint32_t fcode, | ||
1542 : | const uint32_t bcode, | ||
1543 : | const uint32_t MotionFlags, | ||
1544 : | const MBParam * const pParam, | ||
1545 : | const VECTOR * const f_predMV, | ||
1546 : | const VECTOR * const b_predMV, | ||
1547 : | MACROBLOCK * const pMB, | ||
1548 : | h | 545 | int32_t * const best_sad, |
1549 : | SearchData * const fData) | ||
1550 : | chl | 530 | |
1551 : | chl | 96 | { |
1552 : | |||
1553 : | chl | 530 | int iDirection, i, j; |
1554 : | h | 545 | SearchData bData; |
1555 : | chl | 96 | |
1556 : | syskin | 712 | fData->qpel_precision = 0; |
1557 : | memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data | ||
1558 : | *fData->iMinSAD = 4096*256; | ||
1559 : | syskin | 739 | bData.currentMV++; bData.currentQMV++; |
1560 : | h | 545 | fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
1561 : | edgomez | 195 | |
1562 : | syskin | 712 | i = (x + y * fData->iEdgedWidth) * 16; |
1563 : | bData.bRef = fData->Ref = f_Ref + i; | ||
1564 : | bData.bRefH = fData->RefH = f_RefH + i; | ||
1565 : | bData.bRefV = fData->RefV = f_RefV + i; | ||
1566 : | bData.bRefHV = fData->RefHV = f_RefHV + i; | ||
1567 : | bData.Ref = fData->bRef = b_Ref + i; | ||
1568 : | bData.RefH = fData->bRefH = b_RefH + i; | ||
1569 : | bData.RefV = fData->bRefV = b_RefV + i; | ||
1570 : | bData.RefHV = fData->bRefHV = b_RefHV + i; | ||
1571 : | chl | 96 | |
1572 : | h | 545 | bData.bpredMV = fData->predMV = *f_predMV; |
1573 : | fData->bpredMV = bData.predMV = *b_predMV; | ||
1574 : | syskin | 661 | fData->currentMV[0] = fData->currentMV[2]; |
1575 : | chl | 96 | |
1576 : | 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); |
1577 : | 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); | ||
1578 : | |||
1579 : | h | 545 | if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; |
1580 : | syskin | 675 | if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx; |
1581 : | if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy; | ||
1582 : | if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy; | ||
1583 : | Isibaar | 539 | |
1584 : | h | 545 | if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx; |
1585 : | syskin | 675 | if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx; |
1586 : | if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy; | ||
1587 : | if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy; | ||
1588 : | Isibaar | 539 | |
1589 : | h | 545 | CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
1590 : | chl | 96 | |
1591 : | syskin | 768 | //diamond |
1592 : | chl | 530 | do { |
1593 : | iDirection = 255; | ||
1594 : | // forward MV moves | ||
1595 : | h | 545 | i = fData->currentMV[0].x; j = fData->currentMV[0].y; |
1596 : | chl | 96 | |
1597 : | h | 545 | CheckCandidateInt(i + 1, j, 0, &iDirection, fData); |
1598 : | CheckCandidateInt(i, j + 1, 0, &iDirection, fData); | ||
1599 : | CheckCandidateInt(i - 1, j, 0, &iDirection, fData); | ||
1600 : | CheckCandidateInt(i, j - 1, 0, &iDirection, fData); | ||
1601 : | chl | 96 | |
1602 : | chl | 530 | // backward MV moves |
1603 : | h | 545 | i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
1604 : | fData->currentMV[2] = fData->currentMV[0]; | ||
1605 : | Isibaar | 539 | CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); |
1606 : | CheckCandidateInt(i, j + 1, 0, &iDirection, &bData); | ||
1607 : | CheckCandidateInt(i - 1, j, 0, &iDirection, &bData); | ||
1608 : | CheckCandidateInt(i, j - 1, 0, &iDirection, &bData); | ||
1609 : | chl | 96 | |
1610 : | chl | 530 | } while (!(iDirection)); |
1611 : | edgomez | 195 | |
1612 : | syskin | 768 | //qpel refinement |
1613 : | syskin | 658 | if (fData->qpel) { |
1614 : | syskin | 740 | if (*fData->iMinSAD > *best_sad + 500) return; |
1615 : | syskin | 675 | CheckCandidate = CheckCandidateInt; |
1616 : | syskin | 663 | fData->qpel_precision = bData.qpel_precision = 1; |
1617 : | 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); |
1618 : | get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0); | ||
1619 : | syskin | 658 | fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x; |
1620 : | fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y; | ||
1621 : | fData->currentQMV[1].x = 2 * fData->currentMV[1].x; | ||
1622 : | fData->currentQMV[1].y = 2 * fData->currentMV[1].y; | ||
1623 : | syskin | 663 | SubpelRefine(fData); |
1624 : | syskin | 740 | if (*fData->iMinSAD > *best_sad + 300) return; |
1625 : | syskin | 658 | fData->currentQMV[2] = fData->currentQMV[0]; |
1626 : | syskin | 663 | SubpelRefine(&bData); |
1627 : | syskin | 658 | } |
1628 : | |||
1629 : | syskin | 768 | *fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode. |
1630 : | |||
1631 : | h | 545 | if (*fData->iMinSAD < *best_sad) { |
1632 : | *best_sad = *fData->iMinSAD; | ||
1633 : | pMB->mvs[0] = fData->currentMV[0]; | ||
1634 : | pMB->b_mvs[0] = fData->currentMV[1]; | ||
1635 : | chl | 530 | pMB->mode = MODE_INTERPOLATE; |
1636 : | syskin | 658 | if (fData->qpel) { |
1637 : | pMB->qmvs[0] = fData->currentQMV[0]; | ||
1638 : | pMB->b_qmvs[0] = fData->currentQMV[1]; | ||
1639 : | pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x; | ||
1640 : | pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y; | ||
1641 : | pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x; | ||
1642 : | pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y; | ||
1643 : | } else { | ||
1644 : | pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; | ||
1645 : | pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; | ||
1646 : | pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; | ||
1647 : | pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; | ||
1648 : | } | ||
1649 : | chl | 530 | } |
1650 : | } | ||
1651 : | chl | 96 | |
1652 : | chl | 530 | void |
1653 : | MotionEstimationBVOP(MBParam * const pParam, | ||
1654 : | FRAMEINFO * const frame, | ||
1655 : | const int32_t time_bp, | ||
1656 : | const int32_t time_pp, | ||
1657 : | // forward (past) reference | ||
1658 : | const MACROBLOCK * const f_mbs, | ||
1659 : | const IMAGE * const f_ref, | ||
1660 : | const IMAGE * const f_refH, | ||
1661 : | const IMAGE * const f_refV, | ||
1662 : | const IMAGE * const f_refHV, | ||
1663 : | // backward (future) reference | ||
1664 : | syskin | 644 | const FRAMEINFO * const b_reference, |
1665 : | chl | 530 | const IMAGE * const b_ref, |
1666 : | const IMAGE * const b_refH, | ||
1667 : | const IMAGE * const b_refV, | ||
1668 : | const IMAGE * const b_refHV) | ||
1669 : | { | ||
1670 : | uint32_t i, j; | ||
1671 : | syskin | 712 | int32_t best_sad; |
1672 : | uint32_t skip_sad; | ||
1673 : | chl | 530 | int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; |
1674 : | syskin | 644 | const MACROBLOCK * const b_mbs = b_reference->mbs; |
1675 : | chl | 96 | |
1676 : | chl | 530 | VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
1677 : | chl | 96 | |
1678 : | chl | 530 | const int32_t TRB = time_pp - time_bp; |
1679 : | const int32_t TRD = time_pp; | ||
1680 : | chl | 96 | |
1681 : | h | 545 | // some pre-inintialized data for the rest of the search |
1682 : | |||
1683 : | SearchData Data; | ||
1684 : | int32_t iMinSAD; | ||
1685 : | VECTOR currentMV[3]; | ||
1686 : | syskin | 658 | VECTOR currentQMV[3]; |
1687 : | syskin | 704 | memset(&Data, 0, sizeof(SearchData)); |
1688 : | h | 545 | Data.iEdgedWidth = pParam->edged_width; |
1689 : | syskin | 658 | Data.currentMV = currentMV; Data.currentQMV = currentQMV; |
1690 : | h | 545 | Data.iMinSAD = &iMinSAD; |
1691 : | syskin | 712 | Data.lambda16 = lambda_vec16[frame->quant]; |
1692 : | Data.chroma = frame->quant; | ||
1693 : | syskin | 658 | Data.qpel = pParam->m_quarterpel; |
1694 : | syskin | 663 | Data.rounding = 0; |
1695 : | h | 545 | |
1696 : | syskin | 754 | Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose) |
1697 : | chl | 530 | // note: i==horizontal, j==vertical |
1698 : | for (j = 0; j < pParam->mb_height; j++) { | ||
1699 : | chl | 96 | |
1700 : | chl | 530 | f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
1701 : | edgomez | 195 | |
1702 : | chl | 530 | for (i = 0; i < pParam->mb_width; i++) { |
1703 : | MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; | ||
1704 : | const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; | ||
1705 : | edgomez | 195 | |
1706 : | chl | 619 | /* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */ |
1707 : | syskin | 644 | if (b_reference->coding_type != S_VOP) |
1708 : | if (b_mb->mode == MODE_NOT_CODED) { | ||
1709 : | pMB->mode = MODE_NOT_CODED; | ||
1710 : | continue; | ||
1711 : | } | ||
1712 : | chl | 96 | |
1713 : | h | 545 | Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
1714 : | syskin | 601 | pMB->quant = frame->quant; |
1715 : | syskin | 628 | |
1716 : | chl | 530 | /* direct search comes first, because it (1) checks for SKIP-mode |
1717 : | and (2) sets very good predictions for forward and backward search */ | ||
1718 : | Isibaar | 539 | skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
1719 : | b_ref, b_refH->y, b_refV->y, b_refHV->y, | ||
1720 : | chl | 530 | &frame->image, |
1721 : | i, j, | ||
1722 : | frame->motion_flags, | ||
1723 : | TRB, TRD, | ||
1724 : | pParam, | ||
1725 : | pMB, b_mb, | ||
1726 : | h | 545 | &best_sad, |
1727 : | &Data); | ||
1728 : | chl | 96 | |
1729 : | Isibaar | 539 | if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
1730 : | syskin | 739 | |
1731 : | chl | 530 | // forward search |
1732 : | SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, | ||
1733 : | &frame->image, i, j, | ||
1734 : | frame->motion_flags, | ||
1735 : | h | 545 | frame->fcode, pParam, |
1736 : | chl | 530 | pMB, &f_predMV, &best_sad, |
1737 : | h | 545 | MODE_FORWARD, &Data); |
1738 : | chl | 96 | |
1739 : | chl | 530 | // backward search |
1740 : | SearchBF(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, | ||
1741 : | &frame->image, i, j, | ||
1742 : | frame->motion_flags, | ||
1743 : | h | 545 | frame->bcode, pParam, |
1744 : | chl | 530 | pMB, &b_predMV, &best_sad, |
1745 : | h | 545 | MODE_BACKWARD, &Data); |
1746 : | chl | 96 | |
1747 : | chl | 530 | // interpolate search comes last, because it uses data from forward and backward as prediction |
1748 : | SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, | ||
1749 : | b_ref->y, b_refH->y, b_refV->y, b_refHV->y, | ||
1750 : | &frame->image, | ||
1751 : | i, j, | ||
1752 : | frame->fcode, frame->bcode, | ||
1753 : | frame->motion_flags, | ||
1754 : | h | 545 | pParam, |
1755 : | chl | 530 | &f_predMV, &b_predMV, |
1756 : | h | 545 | pMB, &best_sad, |
1757 : | &Data); | ||
1758 : | syskin | 662 | |
1759 : | syskin | 712 | // final skip decision |
1760 : | if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP*2) | ||
1761 : | && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) | ||
1762 : | syskin | 767 | SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data); |
1763 : | syskin | 712 | |
1764 : | chl | 530 | switch (pMB->mode) { |
1765 : | case MODE_FORWARD: | ||
1766 : | f_count++; | ||
1767 : | syskin | 768 | f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; |
1768 : | chl | 530 | break; |
1769 : | case MODE_BACKWARD: | ||
1770 : | b_count++; | ||
1771 : | syskin | 768 | b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; |
1772 : | chl | 530 | break; |
1773 : | case MODE_INTERPOLATE: | ||
1774 : | i_count++; | ||
1775 : | syskin | 768 | f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; |
1776 : | b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; | ||
1777 : | chl | 530 | break; |
1778 : | case MODE_DIRECT: | ||
1779 : | case MODE_DIRECT_NO4V: | ||
1780 : | d_count++; | ||
1781 : | default: | ||
1782 : | break; | ||
1783 : | chl | 96 | } |
1784 : | } | ||
1785 : | } | ||
1786 : | } | ||
1787 : | |||
1788 : | syskin | 739 | static __inline void |
1789 : | Isibaar | 539 | MEanalyzeMB ( const uint8_t * const pRef, |
1790 : | const uint8_t * const pCur, | ||
1791 : | const int x, | ||
1792 : | const int y, | ||
1793 : | const MBParam * const pParam, | ||
1794 : | syskin | 739 | MACROBLOCK * const pMBs, |
1795 : | h | 545 | SearchData * const Data) |
1796 : | Isibaar | 539 | { |
1797 : | chl | 317 | |
1798 : | syskin | 739 | int i, mask; |
1799 : | h | 545 | VECTOR pmv[3]; |
1800 : | syskin | 739 | MACROBLOCK * pMB = &pMBs[x + y * pParam->mb_width]; |
1801 : | syskin | 768 | |
1802 : | syskin | 739 | for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
1803 : | suxen_drol | 118 | |
1804 : | syskin | 640 | //median is only used as prediction. it doesn't have to be real |
1805 : | if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; | ||
1806 : | else | ||
1807 : | if (x == 1) //left macroblock does not have any vector now | ||
1808 : | Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median | ||
1809 : | syskin | 748 | else if (y == 1) // top macroblock doesn't have it's vector |
1810 : | syskin | 640 | Data->predMV = (pMB - 1)->mvs[0]; // left instead of median |
1811 : | else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median | ||
1812 : | |||
1813 : | h | 545 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1814 : | syskin | 739 | pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, Data->rrv); |
1815 : | chl | 326 | |
1816 : | h | 545 | Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
1817 : | Data->Ref = pRef + (x + y * pParam->edged_width) * 16; | ||
1818 : | syskin | 739 | |
1819 : | Isibaar | 539 | pmv[1].x = EVEN(pMB->mvs[0].x); |
1820 : | pmv[1].y = EVEN(pMB->mvs[0].y); | ||
1821 : | syskin | 640 | pmv[2].x = EVEN(Data->predMV.x); |
1822 : | pmv[2].y = EVEN(Data->predMV.y); | ||
1823 : | pmv[0].x = pmv[0].y = 0; | ||
1824 : | chl | 326 | |
1825 : | syskin | 739 | CheckCandidate32I(0, 0, 255, &i, Data); |
1826 : | syskin | 640 | |
1827 : | syskin | 739 | if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) { |
1828 : | syskin | 640 | |
1829 : | syskin | 739 | if (!(mask = make_mask(pmv, 1))) |
1830 : | CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data); | ||
1831 : | if (!(mask = make_mask(pmv, 2))) | ||
1832 : | CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data); | ||
1833 : | Isibaar | 539 | |
1834 : | syskin | 739 | if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) // diamond only if needed |
1835 : | DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); | ||
1836 : | Isibaar | 539 | |
1837 : | syskin | 739 | for (i = 0; i < 4; i++) { |
1838 : | syskin | 768 | MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width]; |
1839 : | syskin | 739 | MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i]; |
1840 : | MB->mode = MODE_INTER; | ||
1841 : | MB->sad16 = Data->iMinSAD[i+1]; | ||
1842 : | } | ||
1843 : | } | ||
1844 : | Isibaar | 539 | } |
1845 : | |||
1846 : | syskin | 739 | #define INTRA_BIAS 2500 |
1847 : | #define INTRA_THRESH 1500 | ||
1848 : | #define INTER_THRESH 1400 | ||
1849 : | Isibaar | 539 | |
1850 : | int | ||
1851 : | syskin | 768 | MEanalysis( const IMAGE * const pRef, |
1852 : | syskin | 644 | FRAMEINFO * const Current, |
1853 : | Isibaar | 539 | MBParam * const pParam, |
1854 : | syskin | 644 | int maxIntra, //maximum number if non-I frames |
1855 : | int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame | ||
1856 : | syskin | 768 | int bCount) // number of B frames in a row |
1857 : | Isibaar | 539 | { |
1858 : | uint32_t x, y, intra = 0; | ||
1859 : | int sSAD = 0; | ||
1860 : | syskin | 644 | MACROBLOCK * const pMBs = Current->mbs; |
1861 : | const IMAGE * const pCurrent = &Current->image; | ||
1862 : | int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH; | ||
1863 : | h | 545 | |
1864 : | syskin | 739 | int32_t iMinSAD[5], temp[5]; |
1865 : | VECTOR currentMV[5]; | ||
1866 : | h | 545 | SearchData Data; |
1867 : | Data.iEdgedWidth = pParam->edged_width; | ||
1868 : | syskin | 739 | Data.currentMV = currentMV; |
1869 : | Data.iMinSAD = iMinSAD; | ||
1870 : | syskin | 644 | Data.iFcode = Current->fcode; |
1871 : | syskin | 739 | Data.rrv = Current->global_flags & XVID_REDUCED; |
1872 : | Data.temp = temp; | ||
1873 : | CheckCandidate = CheckCandidate32I; | ||
1874 : | h | 545 | |
1875 : | syskin | 768 | if (intraCount != 0 && intraCount < 10) // we're right after an I frame |
1876 : | syskin | 658 | IntraThresh += 4 * (intraCount - 10) * (intraCount - 10); |
1877 : | syskin | 644 | else |
1878 : | if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec | ||
1879 : | IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; | ||
1880 : | |||
1881 : | syskin | 672 | InterThresh += 400 * (1 - bCount); |
1882 : | syskin | 739 | if (InterThresh < 300) InterThresh = 300; |
1883 : | syskin | 644 | |
1884 : | Isibaar | 539 | if (sadInit) (*sadInit) (); |
1885 : | |||
1886 : | syskin | 768 | for (y = 1; y < pParam->mb_height-1; y += 2) { |
1887 : | for (x = 1; x < pParam->mb_width-1; x += 2) { | ||
1888 : | syskin | 739 | int i; |
1889 : | Isibaar | 539 | |
1890 : | syskin | 739 | if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; |
1891 : | |||
1892 : | MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data); | ||
1893 : | syskin | 768 | |
1894 : | syskin | 739 | for (i = 0; i < 4; i++) { |
1895 : | int dev; | ||
1896 : | syskin | 768 | MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width]; |
1897 : | syskin | 739 | if (pMB->sad16 > IntraThresh) { |
1898 : | syskin | 768 | dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16, |
1899 : | pParam->edged_width); | ||
1900 : | syskin | 739 | if (dev + IntraThresh < pMB->sad16) { |
1901 : | pMB->mode = MODE_INTRA; | ||
1902 : | if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return I_VOP; | ||
1903 : | } | ||
1904 : | Isibaar | 539 | } |
1905 : | syskin | 739 | sSAD += pMB->sad16; |
1906 : | syskin | 644 | } |
1907 : | suxen_drol | 118 | } |
1908 : | } | ||
1909 : | syskin | 704 | sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); |
1910 : | syskin | 768 | if (sSAD > IntraThresh + INTRA_BIAS) return I_VOP; |
1911 : | suxen_drol | 709 | if (sSAD > InterThresh ) return P_VOP; |
1912 : | Isibaar | 539 | emms(); |
1913 : | suxen_drol | 709 | return B_VOP; |
1914 : | syskin | 712 | |
1915 : | suxen_drol | 118 | } |
1916 : | syskin | 576 | |
1917 : | chl | 769 | |
1918 : | static WARPPOINTS | ||
1919 : | GlobalMotionEst(const MACROBLOCK * const pMBs, | ||
1920 : | const MBParam * const pParam, | ||
1921 : | const FRAMEINFO * const current, | ||
1922 : | const FRAMEINFO * const reference, | ||
1923 : | const IMAGE * const pRefH, | ||
1924 : | const IMAGE * const pRefV, | ||
1925 : | const IMAGE * const pRefHV ) | ||
1926 : | syskin | 628 | { |
1927 : | chl | 619 | |
1928 : | chl | 769 | const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs |
1929 : | const int deltay=8; | ||
1930 : | const int grad=512; // lower bound for deviation in MB | ||
1931 : | |||
1932 : | WARPPOINTS gmc; | ||
1933 : | |||
1934 : | uint32_t mx, my; | ||
1935 : | chl | 619 | |
1936 : | chl | 769 | int MBh = pParam->mb_height; |
1937 : | int MBw = pParam->mb_width; | ||
1938 : | |||
1939 : | int *MBmask= calloc(MBh*MBw,sizeof(int)); | ||
1940 : | double DtimesF[4] = { 0.,0., 0., 0. }; | ||
1941 : | double sol[4] = { 0., 0., 0., 0. }; | ||
1942 : | double a,b,c,n,denom; | ||
1943 : | double meanx,meany; | ||
1944 : | int num,oldnum; | ||
1945 : | |||
1946 : | if (!MBmask) { fprintf(stderr,"Mem error\n"); return gmc;} | ||
1947 : | |||
1948 : | // filter mask of all blocks | ||
1949 : | |||
1950 : | for (my = 1; my < MBh-1; my++) | ||
1951 : | for (mx = 1; mx < MBw-1; mx++) | ||
1952 : | { | ||
1953 : | const int mbnum = mx + my * MBw; | ||
1954 : | const MACROBLOCK *pMB = &pMBs[mbnum]; | ||
1955 : | const VECTOR mv = pMB->mvs[0]; | ||
1956 : | |||
1957 : | if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) | ||
1958 : | continue; | ||
1959 : | |||
1960 : | if ( ( (ABS(mv.x - (pMB-1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-1)->mvs[0].y) < deltay) ) | ||
1961 : | && ( (ABS(mv.x - (pMB+1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+1)->mvs[0].y) < deltay) ) | ||
1962 : | && ( (ABS(mv.x - (pMB-MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-MBw)->mvs[0].y) < deltay) ) | ||
1963 : | && ( (ABS(mv.x - (pMB+MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) ) | ||
1964 : | MBmask[mbnum]=1; | ||
1965 : | syskin | 628 | } |
1966 : | chl | 769 | |
1967 : | for (my = 1; my < MBh-1; my++) | ||
1968 : | for (mx = 1; mx < MBw-1; mx++) | ||
1969 : | { | ||
1970 : | const uint8_t *const pCur = current->image.y + 16*my*pParam->edged_width + 16*mx; | ||
1971 : | |||
1972 : | const int mbnum = mx + my * MBw; | ||
1973 : | if (!MBmask[mbnum]) | ||
1974 : | continue; | ||
1975 : | syskin | 628 | |
1976 : | chl | 769 | if (sad16 ( pCur, pCur+1 , pParam->edged_width, 65536) <= grad ) |
1977 : | MBmask[mbnum] = 0; | ||
1978 : | if (sad16 ( pCur, pCur+pParam->edged_width, pParam->edged_width, 65536) <= grad ) | ||
1979 : | MBmask[mbnum] = 0; | ||
1980 : | |||
1981 : | } | ||
1982 : | syskin | 628 | |
1983 : | chl | 769 | emms(); |
1984 : | chl | 619 | |
1985 : | chl | 769 | do { /* until convergence */ |
1986 : | chl | 619 | |
1987 : | chl | 769 | a = b = c = n = 0; |
1988 : | DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.; | ||
1989 : | for (my = 0; my < MBh; my++) | ||
1990 : | for (mx = 0; mx < MBw; mx++) | ||
1991 : | { | ||
1992 : | const int mbnum = mx + my * MBw; | ||
1993 : | const MACROBLOCK *pMB = &pMBs[mbnum]; | ||
1994 : | const VECTOR mv = pMB->mvs[0]; | ||
1995 : | chl | 619 | |
1996 : | chl | 769 | if (!MBmask[mbnum]) |
1997 : | continue; | ||
1998 : | |||
1999 : | n++; | ||
2000 : | a += 16*mx+8; | ||
2001 : | b += 16*my+8; | ||
2002 : | c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8); | ||
2003 : | syskin | 768 | |
2004 : | chl | 769 | DtimesF[0] += (double)mv.x; |
2005 : | DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8); | ||
2006 : | DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8); | ||
2007 : | DtimesF[3] += (double)mv.y; | ||
2008 : | } | ||
2009 : | syskin | 768 | |
2010 : | chl | 769 | denom = a*a+b*b-c*n; |
2011 : | |||
2012 : | /* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */ | ||
2013 : | /* D'*E*F has been calculated in the same loop as matrix */ | ||
2014 : | syskin | 768 | |
2015 : | chl | 769 | sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2]; |
2016 : | sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3]; | ||
2017 : | sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3]; | ||
2018 : | sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3]; | ||
2019 : | syskin | 628 | |
2020 : | chl | 769 | sol[0] /= denom; |
2021 : | sol[1] /= denom; | ||
2022 : | sol[2] /= denom; | ||
2023 : | sol[3] /= denom; | ||
2024 : | |||
2025 : | meanx = meany = 0.; | ||
2026 : | oldnum = 0; | ||
2027 : | for (my = 0; my < MBh; my++) | ||
2028 : | for (mx = 0; mx < MBw; mx++) | ||
2029 : | { | ||
2030 : | const int mbnum = mx + my * MBw; | ||
2031 : | const MACROBLOCK *pMB = &pMBs[mbnum]; | ||
2032 : | const VECTOR mv = pMB->mvs[0]; | ||
2033 : | syskin | 628 | |
2034 : | chl | 769 | if (!MBmask[mbnum]) |
2035 : | continue; | ||
2036 : | syskin | 628 | |
2037 : | chl | 769 | oldnum++; |
2038 : | meanx += ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ); | ||
2039 : | meany += ABS(( sol[3] + (16*my+8)*sol[1] - (16*mx+8)*sol[2] ) - mv.y ); | ||
2040 : | |||
2041 : | } | ||
2042 : | |||
2043 : | if (2*meanx > oldnum) /* mittlere Abweichung von Ebene */ | ||
2044 : | meanx /= oldnum; | ||
2045 : | else | ||
2046 : | meanx = 0.5; | ||
2047 : | if (2*meany > oldnum) | ||
2048 : | meany /= oldnum; | ||
2049 : | else | ||
2050 : | meanx = 0.5; | ||
2051 : | |||
2052 : | // fprintf(stderr,"meanx = %8.5f meany = %8.5f %d\n",meanx,meany, oldnum); | ||
2053 : | |||
2054 : | num = 0; | ||
2055 : | for (my = 0; my < MBh; my++) | ||
2056 : | for (mx = 0; mx < MBw; mx++) | ||
2057 : | { | ||
2058 : | const int mbnum = mx + my * MBw; | ||
2059 : | const MACROBLOCK *pMB = &pMBs[mbnum]; | ||
2060 : | const VECTOR mv = pMB->mvs[0]; | ||
2061 : | |||
2062 : | if (!MBmask[mbnum]) | ||
2063 : | continue; | ||
2064 : | |||
2065 : | if ( ( ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ) > meanx ) | ||
2066 : | || ( ABS(( sol[3] + (16*my+8)*sol[1] - (16*mx+8)*sol[2] ) - mv.y ) > meany ) ) | ||
2067 : | MBmask[mbnum]=0; | ||
2068 : | else | ||
2069 : | num++; | ||
2070 : | } | ||
2071 : | |||
2072 : | } while ( (oldnum != num) && (num>=4) ); | ||
2073 : | |||
2074 : | if (num < 4) | ||
2075 : | { | ||
2076 : | gmc.duv[0].x= gmc.duv[0].y= gmc.duv[1].x= gmc.duv[1].y= gmc.duv[2].x= gmc.duv[2].y=0; | ||
2077 : | } else { | ||
2078 : | |||
2079 : | gmc.duv[0].x=(int)(sol[0]+0.5); | ||
2080 : | gmc.duv[0].y=(int)(sol[3]+0.5); | ||
2081 : | |||
2082 : | gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5); | ||
2083 : | gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5); | ||
2084 : | |||
2085 : | gmc.duv[2].x=0; | ||
2086 : | gmc.duv[2].y=0; | ||
2087 : | chl | 619 | } |
2088 : | chl | 769 | // fprintf(stderr,"wp1 = ( %4d, %4d) wp2 = ( %4d, %4d) \n", gmc.duv[0].x, gmc.duv[0].y, gmc.duv[1].x, gmc.duv[1].y); |
2089 : | chl | 619 | |
2090 : | chl | 769 | free(MBmask); |
2091 : | |||
2092 : | chl | 619 | return gmc; |
2093 : | } |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |