Parent Directory | Revision Log
Revision 1.44.2.43 - (view) (download)
1 : | Isibaar | 1.1 | /************************************************************************** |
2 : | * | ||
3 : | chl | 1.26 | * XVID MPEG-4 VIDEO CODEC |
4 : | syskin | 1.44.2.40 | * motion estimation |
5 : | chl | 1.26 | * |
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 | 1.1 | #include <assert.h> |
32 : | #include <stdio.h> | ||
33 : | chl | 1.3 | #include <stdlib.h> |
34 : | suxen_drol | 1.44.2.35 | #include <string.h> // memcpy |
35 : | Isibaar | 1.1 | |
36 : | #include "../encoder.h" | ||
37 : | #include "../utils/mbfunctions.h" | ||
38 : | #include "../prediction/mbprediction.h" | ||
39 : | #include "../global.h" | ||
40 : | #include "../utils/timer.h" | ||
41 : | Isibaar | 1.44.2.7 | #include "../image/interpolate8x8.h" |
42 : | chl | 1.44.2.1 | #include "motion_est.h" |
43 : | suxen_drol | 1.6 | #include "motion.h" |
44 : | Isibaar | 1.1 | #include "sad.h" |
45 : | Isibaar | 1.44.2.2 | #include "../utils/emms.h" |
46 : | Isibaar | 1.1 | |
47 : | chl | 1.44.2.1 | #define INITIAL_SKIP_THRESH (10) |
48 : | #define FINAL_SKIP_THRESH (50) | ||
49 : | #define MAX_SAD00_FOR_SKIP (20) | ||
50 : | Isibaar | 1.44.2.2 | #define MAX_CHROMA_SAD_FOR_SKIP (22) |
51 : | chl | 1.44.2.1 | |
52 : | #define CHECK_CANDIDATE(X,Y,D) { \ | ||
53 : | (*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } | ||
54 : | edgomez | 1.20 | |
55 : | chl | 1.44.2.1 | static __inline int |
56 : | syskin | 1.44.2.36 | d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv) |
57 : | chl | 1.44.2.1 | { |
58 : | int xb, yb; | ||
59 : | syskin | 1.44.2.34 | if (qpel) { x *= 2; y *= 2;} |
60 : | else if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); } | ||
61 : | syskin | 1.44.2.40 | x -= pred.x; |
62 : | y -= pred.y; | ||
63 : | chl | 1.44.2.1 | |
64 : | syskin | 1.44.2.40 | if (x) { |
65 : | chl | 1.44.2.1 | if (x < 0) x = -x; |
66 : | x += (1 << (iFcode - 1)) - 1; | ||
67 : | x >>= (iFcode - 1); | ||
68 : | if (x > 32) x = 32; | ||
69 : | xb = mvtab[x] + iFcode; | ||
70 : | syskin | 1.44.2.40 | } else xb = 1; |
71 : | chl | 1.44.2.1 | |
72 : | syskin | 1.44.2.40 | if (y) { |
73 : | chl | 1.44.2.1 | if (y < 0) y = -y; |
74 : | y += (1 << (iFcode - 1)) - 1; | ||
75 : | y >>= (iFcode - 1); | ||
76 : | if (y > 32) y = 32; | ||
77 : | yb = mvtab[y] + iFcode; | ||
78 : | syskin | 1.44.2.40 | } else yb = 1; |
79 : | chl | 1.44.2.1 | return xb + yb; |
80 : | } | ||
81 : | |||
82 : | syskin | 1.44.2.40 | static int32_t |
83 : | syskin | 1.44.2.14 | ChromaSAD(int dx, int dy, const SearchData * const data) |
84 : | chl | 1.44.2.1 | { |
85 : | syskin | 1.44.2.14 | int sad; |
86 : | syskin | 1.44.2.40 | const uint32_t stride = data->iEdgedWidth/2; |
87 : | syskin | 1.44.2.14 | |
88 : | syskin | 1.44.2.34 | if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently |
89 : | syskin | 1.44.2.40 | data->temp[5] = dx; data->temp[6] = dy; // backup |
90 : | syskin | 1.44.2.34 | |
91 : | switch (((dx & 1) << 1) | (dy & 1)) { | ||
92 : | syskin | 1.44.2.14 | case 0: |
93 : | syskin | 1.44.2.40 | dx = dx / 2; dy = dy / 2; |
94 : | sad = sad8(data->CurU, data->RefCU + dy * stride + dx, stride); | ||
95 : | sad += sad8(data->CurV, data->RefCV + dy * stride + dx, stride); | ||
96 : | syskin | 1.44.2.14 | break; |
97 : | case 1: | ||
98 : | dx = dx / 2; dy = (dy - 1) / 2; | ||
99 : | syskin | 1.44.2.40 | sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + (dy+1) * stride + dx, stride); |
100 : | sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + (dy+1) * stride + dx, stride); | ||
101 : | syskin | 1.44.2.14 | break; |
102 : | case 2: | ||
103 : | dx = (dx - 1) / 2; dy = dy / 2; | ||
104 : | syskin | 1.44.2.40 | sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + dy * stride + dx+1, stride); |
105 : | sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + dy * stride + dx+1, stride); | ||
106 : | syskin | 1.44.2.14 | break; |
107 : | default: | ||
108 : | dx = (dx - 1) / 2; dy = (dy - 1) / 2; | ||
109 : | syskin | 1.44.2.40 | interpolate8x8_halfpel_hv(data->RefQ, data->RefCU + dy * stride + dx, stride, data->rounding); |
110 : | sad = sad8(data->CurU, data->RefQ, stride); | ||
111 : | |||
112 : | interpolate8x8_halfpel_hv(data->RefQ, data->RefCV + dy * stride + dx, stride, data->rounding); | ||
113 : | sad += sad8(data->CurV, data->RefQ, stride); | ||
114 : | syskin | 1.44.2.14 | break; |
115 : | chl | 1.44.2.1 | } |
116 : | syskin | 1.44.2.40 | data->temp[7] = sad; //backup, part 2 |
117 : | syskin | 1.44.2.14 | return sad; |
118 : | } | ||
119 : | chl | 1.10 | |
120 : | syskin | 1.44.2.22 | static __inline const uint8_t * |
121 : | syskin | 1.44.2.40 | GetReferenceB(const int x, const int y, const int dir, const SearchData * const data) |
122 : | Isibaar | 1.44.2.12 | { |
123 : | syskin | 1.44.2.22 | // dir : 0 = forward, 1 = backward |
124 : | switch ( (dir << 2) | ((x&1)<<1) | (y&1) ) { | ||
125 : | case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth); | ||
126 : | case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); | ||
127 : | case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); | ||
128 : | case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); | ||
129 : | case 4 : return data->bRef + x/2 + (y/2)*(data->iEdgedWidth); | ||
130 : | case 5 : return data->bRefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); | ||
131 : | case 6 : return data->bRefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); | ||
132 : | default : return data->bRefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); | ||
133 : | Isibaar | 1.44.2.12 | } |
134 : | } | ||
135 : | |||
136 : | syskin | 1.44.2.40 | // this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate |
137 : | static __inline const uint8_t * | ||
138 : | GetReference(const int x, const int y, const SearchData * const data) | ||
139 : | { | ||
140 : | switch ( ((x&1)<<1) | (y&1) ) { | ||
141 : | case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth); | ||
142 : | case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); | ||
143 : | case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); | ||
144 : | default : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); | ||
145 : | } | ||
146 : | } | ||
147 : | |||
148 : | static uint8_t * | ||
149 : | syskin | 1.44.2.22 | Interpolate8x8qpel(const int x, const int y, const int block, const int dir, const SearchData * const data) |
150 : | Isibaar | 1.1 | { |
151 : | syskin | 1.44.2.22 | // create or find a qpel-precision reference picture; return pointer to it |
152 : | syskin | 1.44.2.24 | uint8_t * Reference = (uint8_t *)data->RefQ + 16*dir; |
153 : | syskin | 1.44.2.22 | const int32_t iEdgedWidth = data->iEdgedWidth; |
154 : | const uint32_t rounding = data->rounding; | ||
155 : | const int halfpel_x = x/2; | ||
156 : | const int halfpel_y = y/2; | ||
157 : | const uint8_t *ref1, *ref2, *ref3, *ref4; | ||
158 : | Isibaar | 1.1 | |
159 : | syskin | 1.44.2.40 | ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
160 : | syskin | 1.44.2.22 | ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
161 : | switch( ((x&1)<<1) + (y&1) ) { | ||
162 : | case 0: // pure halfpel position | ||
163 : | syskin | 1.44.2.38 | return (uint8_t *) ref1; |
164 : | syskin | 1.44.2.22 | break; |
165 : | Isibaar | 1.1 | |
166 : | syskin | 1.44.2.22 | case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
167 : | syskin | 1.44.2.40 | ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
168 : | syskin | 1.44.2.22 | ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
169 : | Isibaar | 1.44.2.23 | interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
170 : | syskin | 1.44.2.22 | break; |
171 : | Isibaar | 1.1 | |
172 : | syskin | 1.44.2.22 | case 2: // x qpel, y halfpel - left or right during qpel refinement |
173 : | syskin | 1.44.2.40 | ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
174 : | syskin | 1.44.2.22 | ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
175 : | Isibaar | 1.44.2.23 | interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
176 : | syskin | 1.44.2.22 | break; |
177 : | Isibaar | 1.1 | |
178 : | syskin | 1.44.2.22 | default: // x and y in qpel resolution - the "corners" (top left/right and |
179 : | // bottom left/right) during qpel refinement | ||
180 : | syskin | 1.44.2.40 | ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
181 : | ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); | ||
182 : | ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); | ||
183 : | syskin | 1.44.2.22 | ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
184 : | ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; | ||
185 : | ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; | ||
186 : | interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); | ||
187 : | break; | ||
188 : | } | ||
189 : | return Reference; | ||
190 : | Isibaar | 1.1 | } |
191 : | |||
192 : | syskin | 1.44.2.40 | static uint8_t * |
193 : | syskin | 1.44.2.22 | Interpolate16x16qpel(const int x, const int y, const int dir, const SearchData * const data) |
194 : | Isibaar | 1.44.2.7 | { |
195 : | syskin | 1.44.2.22 | // create or find a qpel-precision reference picture; return pointer to it |
196 : | syskin | 1.44.2.24 | uint8_t * Reference = (uint8_t *)data->RefQ + 16*dir; |
197 : | syskin | 1.44.2.22 | const int32_t iEdgedWidth = data->iEdgedWidth; |
198 : | const uint32_t rounding = data->rounding; | ||
199 : | const int halfpel_x = x/2; | ||
200 : | const int halfpel_y = y/2; | ||
201 : | Isibaar | 1.44.2.7 | const uint8_t *ref1, *ref2, *ref3, *ref4; |
202 : | |||
203 : | syskin | 1.44.2.40 | ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
204 : | syskin | 1.44.2.22 | switch( ((x&1)<<1) + (y&1) ) { |
205 : | case 0: // pure halfpel position | ||
206 : | syskin | 1.44.2.38 | return (uint8_t *) ref1; |
207 : | Isibaar | 1.44.2.7 | case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
208 : | syskin | 1.44.2.40 | ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
209 : | Isibaar | 1.44.2.23 | interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
210 : | interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); | ||
211 : | interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); | ||
212 : | interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); | ||
213 : | Isibaar | 1.44.2.7 | break; |
214 : | |||
215 : | case 2: // x qpel, y halfpel - left or right during qpel refinement | ||
216 : | syskin | 1.44.2.40 | ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
217 : | Isibaar | 1.44.2.23 | interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
218 : | interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); | ||
219 : | interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); | ||
220 : | interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); | ||
221 : | Isibaar | 1.44.2.7 | break; |
222 : | |||
223 : | default: // x and y in qpel resolution - the "corners" (top left/right and | ||
224 : | // bottom left/right) during qpel refinement | ||
225 : | syskin | 1.44.2.40 | ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
226 : | ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); | ||
227 : | ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); | ||
228 : | Isibaar | 1.44.2.7 | interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
229 : | interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); | ||
230 : | interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); | ||
231 : | interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); | ||
232 : | break; | ||
233 : | } | ||
234 : | syskin | 1.44.2.22 | return Reference; |
235 : | } | ||
236 : | |||
237 : | /* CHECK_CANDIATE FUNCTIONS START */ | ||
238 : | |||
239 : | syskin | 1.44.2.40 | static void |
240 : | syskin | 1.44.2.22 | CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
241 : | { | ||
242 : | int t, xc, yc; | ||
243 : | const uint8_t * Reference; | ||
244 : | VECTOR * current; | ||
245 : | |||
246 : | if (( x > data->max_dx) || ( x < data->min_dx) | ||
247 : | || ( y > data->max_dy) || (y < data->min_dy)) return; | ||
248 : | |||
249 : | if (data->qpel_precision) { // x and y are in 1/4 precision | ||
250 : | Reference = Interpolate16x16qpel(x, y, 0, data); | ||
251 : | xc = x/2; yc = y/2; //for chroma sad | ||
252 : | current = data->currentQMV; | ||
253 : | } else { | ||
254 : | syskin | 1.44.2.40 | Reference = GetReference(x, y, data); |
255 : | syskin | 1.44.2.22 | current = data->currentMV; |
256 : | xc = x; yc = y; | ||
257 : | } | ||
258 : | syskin | 1.44.2.36 | t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0); |
259 : | Isibaar | 1.44.2.7 | |
260 : | syskin | 1.44.2.22 | data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
261 : | Isibaar | 1.44.2.7 | |
262 : | syskin | 1.44.2.14 | data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; |
263 : | data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; | ||
264 : | |||
265 : | syskin | 1.44.2.41 | if (data->chroma) data->temp[0] += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
266 : | (yc >> 1) + roundtab_79[yc & 0x3], data); | ||
267 : | Isibaar | 1.44.2.7 | |
268 : | if (data->temp[0] < data->iMinSAD[0]) { | ||
269 : | data->iMinSAD[0] = data->temp[0]; | ||
270 : | syskin | 1.44.2.22 | current[0].x = x; current[0].y = y; |
271 : | *dir = Direction; } | ||
272 : | Isibaar | 1.44.2.7 | |
273 : | if (data->temp[1] < data->iMinSAD[1]) { | ||
274 : | syskin | 1.44.2.22 | data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y= y; } |
275 : | Isibaar | 1.44.2.7 | if (data->temp[2] < data->iMinSAD[2]) { |
276 : | syskin | 1.44.2.22 | data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; } |
277 : | Isibaar | 1.44.2.7 | if (data->temp[3] < data->iMinSAD[3]) { |
278 : | syskin | 1.44.2.22 | data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; } |
279 : | Isibaar | 1.44.2.7 | if (data->temp[4] < data->iMinSAD[4]) { |
280 : | syskin | 1.44.2.22 | data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; } |
281 : | |||
282 : | Isibaar | 1.44.2.7 | } |
283 : | |||
284 : | syskin | 1.44.2.40 | static void |
285 : | syskin | 1.44.2.32 | CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
286 : | { | ||
287 : | int t; | ||
288 : | const uint8_t * Reference; | ||
289 : | |||
290 : | if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero integer value | ||
291 : | ( x > data->max_dx) || ( x < data->min_dx) | ||
292 : | || ( y > data->max_dy) || (y < data->min_dy)) return; | ||
293 : | |||
294 : | syskin | 1.44.2.40 | Reference = GetReference(x, y, data); |
295 : | syskin | 1.44.2.36 | t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1); |
296 : | syskin | 1.44.2.32 | |
297 : | data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); | ||
298 : | |||
299 : | data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; | ||
300 : | data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; | ||
301 : | |||
302 : | if (data->temp[0] < data->iMinSAD[0]) { | ||
303 : | data->iMinSAD[0] = data->temp[0]; | ||
304 : | data->currentMV[0].x = x; data->currentMV[0].y = y; | ||
305 : | *dir = Direction; } | ||
306 : | |||
307 : | if (data->temp[1] < data->iMinSAD[1]) { | ||
308 : | data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } | ||
309 : | if (data->temp[2] < data->iMinSAD[2]) { | ||
310 : | data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } | ||
311 : | if (data->temp[3] < data->iMinSAD[3]) { | ||
312 : | data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } | ||
313 : | if (data->temp[4] < data->iMinSAD[4]) { | ||
314 : | data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } | ||
315 : | } | ||
316 : | |||
317 : | syskin | 1.44.2.40 | static void |
318 : | syskin | 1.44.2.22 | CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
319 : | syskin | 1.44.2.19 | { |
320 : | int32_t sad; | ||
321 : | syskin | 1.44.2.22 | const uint8_t * Reference; |
322 : | int t; | ||
323 : | VECTOR * current; | ||
324 : | syskin | 1.44.2.19 | |
325 : | if (( x > data->max_dx) || ( x < data->min_dx) | ||
326 : | || ( y > data->max_dy) || (y < data->min_dy)) return; | ||
327 : | |||
328 : | syskin | 1.44.2.34 | if (data->rrv) |
329 : | syskin | 1.44.2.32 | if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) ) return; //non-zero integer value |
330 : | |||
331 : | syskin | 1.44.2.22 | 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 | 1.44.2.40 | Reference = GetReference(x, y, data); |
336 : | syskin | 1.44.2.22 | current = data->currentMV; |
337 : | syskin | 1.44.2.19 | } |
338 : | syskin | 1.44.2.36 | t = d_mv_bits(x, y, data->predMV, data->iFcode, |
339 : | syskin | 1.44.2.40 | data->qpel && !data->qpel_precision, data->rrv); |
340 : | syskin | 1.44.2.19 | |
341 : | sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); | ||
342 : | syskin | 1.44.2.22 | sad += (data->lambda16 * t * sad)/1000; |
343 : | syskin | 1.44.2.19 | |
344 : | syskin | 1.44.2.22 | if (sad < *(data->iMinSAD)) { |
345 : | *(data->iMinSAD) = sad; | ||
346 : | current->x = x; current->y = y; | ||
347 : | *dir = Direction; } | ||
348 : | syskin | 1.44.2.19 | } |
349 : | |||
350 : | syskin | 1.44.2.40 | static void |
351 : | syskin | 1.44.2.37 | CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
352 : | Isibaar | 1.44.2.2 | { |
353 : | syskin | 1.44.2.19 | // maximum speed - for P/B/I decision |
354 : | Isibaar | 1.44.2.2 | |
355 : | if (( x > data->max_dx) || ( x < data->min_dx) | ||
356 : | || ( y > data->max_dy) || (y < data->min_dy)) return; | ||
357 : | |||
358 : | syskin | 1.44.2.37 | data->temp[0] = sad32v_c(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), |
359 : | data->iEdgedWidth, data->temp+1); | ||
360 : | if (data->temp[0] < *(data->iMinSAD)) { | ||
361 : | *(data->iMinSAD) = data->temp[0]; | ||
362 : | Isibaar | 1.44.2.2 | data->currentMV[0].x = x; data->currentMV[0].y = y; |
363 : | *dir = Direction; } | ||
364 : | syskin | 1.44.2.37 | if (data->temp[1] < data->iMinSAD[1]) { |
365 : | data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } | ||
366 : | if (data->temp[2] < data->iMinSAD[2]) { | ||
367 : | data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } | ||
368 : | if (data->temp[3] < data->iMinSAD[3]) { | ||
369 : | data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } | ||
370 : | if (data->temp[4] < data->iMinSAD[4]) { | ||
371 : | data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } | ||
372 : | |||
373 : | Isibaar | 1.44.2.2 | } |
374 : | |||
375 : | syskin | 1.44.2.40 | static void |
376 : | chl | 1.44.2.1 | CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) |
377 : | Isibaar | 1.1 | { |
378 : | chl | 1.44.2.1 | int32_t sad; |
379 : | syskin | 1.44.2.22 | int xb, yb, t; |
380 : | chl | 1.44.2.1 | const uint8_t *ReferenceF, *ReferenceB; |
381 : | syskin | 1.44.2.22 | VECTOR *current; |
382 : | suxen_drol | 1.8 | |
383 : | chl | 1.44.2.1 | if (( xf > data->max_dx) || ( xf < data->min_dx) |
384 : | || ( yf > data->max_dy) || (yf < data->min_dy)) return; | ||
385 : | suxen_drol | 1.8 | |
386 : | syskin | 1.44.2.22 | if (data->qpel_precision) { |
387 : | ReferenceF = Interpolate16x16qpel(xf, yf, 0, data); | ||
388 : | xb = data->currentQMV[1].x; yb = data->currentQMV[1].y; | ||
389 : | current = data->currentQMV; | ||
390 : | ReferenceB = Interpolate16x16qpel(xb, yb, 1, data); | ||
391 : | } else { | ||
392 : | syskin | 1.44.2.40 | ReferenceF = GetReference(xf, yf, data); |
393 : | syskin | 1.44.2.22 | xb = data->currentMV[1].x; yb = data->currentMV[1].y; |
394 : | syskin | 1.44.2.40 | ReferenceB = GetReferenceB(xb, yb, 1, data); |
395 : | syskin | 1.44.2.22 | current = data->currentMV; |
396 : | syskin | 1.44.2.19 | } |
397 : | |||
398 : | syskin | 1.44.2.36 | t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0) |
399 : | + d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel && !data->qpel_precision, 0); | ||
400 : | syskin | 1.44.2.34 | |
401 : | syskin | 1.44.2.19 | sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
402 : | syskin | 1.44.2.22 | sad += (data->lambda16 * t * sad)/1000; |
403 : | syskin | 1.44.2.19 | |
404 : | if (sad < *(data->iMinSAD)) { | ||
405 : | *(data->iMinSAD) = sad; | ||
406 : | syskin | 1.44.2.22 | current->x = xf; current->y = yf; |
407 : | syskin | 1.44.2.19 | *dir = Direction; } |
408 : | } | ||
409 : | |||
410 : | syskin | 1.44.2.40 | static void |
411 : | chl | 1.44.2.1 | CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
412 : | { | ||
413 : | syskin | 1.44.2.14 | int32_t sad = 0; |
414 : | chl | 1.44.2.1 | int k; |
415 : | const uint8_t *ReferenceF; | ||
416 : | const uint8_t *ReferenceB; | ||
417 : | VECTOR mvs, b_mvs; | ||
418 : | syskin | 1.44.2.36 | const VECTOR zeroMV={0,0}; |
419 : | edgomez | 1.20 | |
420 : | chl | 1.44.2.1 | if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
421 : | edgomez | 1.20 | |
422 : | chl | 1.44.2.1 | 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 : | edgomez | 1.20 | |
428 : | chl | 1.44.2.1 | 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 : | |||
433 : | 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 : | |||
438 : | syskin | 1.44.2.40 | if (!data->qpel) { |
439 : | mvs.x *= 2; mvs.y *= 2; | ||
440 : | syskin | 1.44.2.22 | b_mvs.x *= 2; b_mvs.y *= 2; //we move to qpel precision anyway |
441 : | chl | 1.44.2.1 | } |
442 : | syskin | 1.44.2.22 | ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); |
443 : | ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); | ||
444 : | chl | 1.44.2.1 | |
445 : | sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), | ||
446 : | syskin | 1.44.2.22 | ReferenceF, ReferenceB, |
447 : | chl | 1.44.2.1 | data->iEdgedWidth); |
448 : | if (sad > *(data->iMinSAD)) return; | ||
449 : | } | ||
450 : | chl | 1.17 | |
451 : | syskin | 1.44.2.36 | sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)/1000; |
452 : | syskin | 1.44.2.14 | |
453 : | chl | 1.44.2.1 | if (sad < *(data->iMinSAD)) { |
454 : | *(data->iMinSAD) = sad; | ||
455 : | data->currentMV->x = x; data->currentMV->y = y; | ||
456 : | *dir = Direction; } | ||
457 : | Isibaar | 1.1 | } |
458 : | |||
459 : | syskin | 1.44.2.40 | static void |
460 : | chl | 1.44.2.1 | CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
461 : | { | ||
462 : | int32_t sad; | ||
463 : | const uint8_t *ReferenceF; | ||
464 : | const uint8_t *ReferenceB; | ||
465 : | VECTOR mvs, b_mvs; | ||
466 : | syskin | 1.44.2.36 | const VECTOR zeroMV = {0,0}; |
467 : | chl | 1.35 | |
468 : | syskin | 1.44.2.6 | if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
469 : | Isibaar | 1.1 | |
470 : | chl | 1.44.2.1 | mvs.x = data->directmvF[0].x + x; |
471 : | b_mvs.x = ((x == 0) ? | ||
472 : | data->directmvB[0].x | ||
473 : | : mvs.x - data->referencemv[0].x); | ||
474 : | |||
475 : | mvs.y = data->directmvF[0].y + y; | ||
476 : | b_mvs.y = ((y == 0) ? | ||
477 : | data->directmvB[0].y | ||
478 : | : mvs.y - data->referencemv[0].y); | ||
479 : | |||
480 : | if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) | ||
481 : | || ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) | ||
482 : | || ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) | ||
483 : | || ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; | ||
484 : | |||
485 : | syskin | 1.44.2.40 | if (!data->qpel) { |
486 : | mvs.x *= 2; mvs.y *= 2; | ||
487 : | syskin | 1.44.2.22 | b_mvs.x *= 2; b_mvs.y *= 2; //we move to qpel precision anyway |
488 : | syskin | 1.44.2.34 | } |
489 : | syskin | 1.44.2.22 | ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data); |
490 : | ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); | ||
491 : | chl | 1.44.2.1 | |
492 : | syskin | 1.44.2.14 | sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
493 : | syskin | 1.44.2.36 | sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)/1000; |
494 : | Isibaar | 1.1 | |
495 : | chl | 1.44.2.1 | if (sad < *(data->iMinSAD)) { |
496 : | *(data->iMinSAD) = sad; | ||
497 : | data->currentMV->x = x; data->currentMV->y = y; | ||
498 : | *dir = Direction; } | ||
499 : | } | ||
500 : | Isibaar | 1.1 | |
501 : | chl | 1.44.2.1 | static void |
502 : | CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) | ||
503 : | { | ||
504 : | syskin | 1.44.2.14 | int32_t sad; int t; |
505 : | chl | 1.44.2.1 | const uint8_t * Reference; |
506 : | Isibaar | 1.1 | |
507 : | chl | 1.44.2.1 | if (( x > data->max_dx) || ( x < data->min_dx) |
508 : | || ( y > data->max_dy) || (y < data->min_dy)) return; | ||
509 : | Isibaar | 1.1 | |
510 : | syskin | 1.44.2.22 | if (data->qpel) Reference = Interpolate16x16qpel(x, y, 0, data); |
511 : | syskin | 1.44.2.40 | else Reference = GetReference(x, y, data); |
512 : | Isibaar | 1.1 | |
513 : | chl | 1.44.2.1 | sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
514 : | syskin | 1.44.2.36 | t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0); |
515 : | Isibaar | 1.44.2.12 | |
516 : | syskin | 1.44.2.14 | sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))/100; |
517 : | Isibaar | 1.44.2.12 | |
518 : | if (sad < *(data->iMinSAD)) { | ||
519 : | *(data->iMinSAD) = sad; | ||
520 : | data->currentMV->x = x; data->currentMV->y = y; | ||
521 : | *dir = Direction; } | ||
522 : | } | ||
523 : | |||
524 : | Isibaar | 1.44.2.2 | /* CHECK_CANDIATE FUNCTIONS END */ |
525 : | chl | 1.3 | |
526 : | chl | 1.44.2.1 | /* MAINSEARCH FUNCTIONS START */ |
527 : | chl | 1.18 | |
528 : | chl | 1.44.2.1 | static void |
529 : | AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection) | ||
530 : | { | ||
531 : | chl | 1.18 | |
532 : | /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ | ||
533 : | |||
534 : | syskin | 1.44.2.34 | int iDirection; |
535 : | |||
536 : | for(;;) { //forever | ||
537 : | iDirection = 0; | ||
538 : | if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); | ||
539 : | if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); | ||
540 : | if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); | ||
541 : | if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); | ||
542 : | edgomez | 1.20 | |
543 : | syskin | 1.44.2.34 | /* now we're doing diagonal checks near our candidate */ |
544 : | |||
545 : | if (iDirection) { //checking if anything found | ||
546 : | bDirection = iDirection; | ||
547 : | chl | 1.18 | iDirection = 0; |
548 : | syskin | 1.44.2.34 | x = data->currentMV->x; y = data->currentMV->y; |
549 : | if (bDirection & 3) { //our candidate is left or right | ||
550 : | CHECK_CANDIDATE(x, y + iDiamondSize, 8); | ||
551 : | CHECK_CANDIDATE(x, y - iDiamondSize, 4); | ||
552 : | } else { // what remains here is up or down | ||
553 : | CHECK_CANDIDATE(x + iDiamondSize, y, 2); | ||
554 : | syskin | 1.44.2.40 | CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
555 : | syskin | 1.44.2.34 | } |
556 : | |||
557 : | if (iDirection) { | ||
558 : | bDirection += iDirection; | ||
559 : | syskin | 1.44.2.40 | x = data->currentMV->x; y = data->currentMV->y; |
560 : | syskin | 1.44.2.34 | } |
561 : | } else { //about to quit, eh? not so fast.... | ||
562 : | switch (bDirection) { | ||
563 : | case 2: | ||
564 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
565 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
566 : | break; | ||
567 : | case 1: | ||
568 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
569 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
570 : | break; | ||
571 : | case 2 + 4: | ||
572 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
573 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
574 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
575 : | break; | ||
576 : | case 4: | ||
577 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
578 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
579 : | break; | ||
580 : | case 8: | ||
581 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
582 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
583 : | break; | ||
584 : | case 1 + 4: | ||
585 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
586 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
587 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
588 : | break; | ||
589 : | case 2 + 8: | ||
590 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
591 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
592 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
593 : | break; | ||
594 : | case 1 + 8: | ||
595 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
596 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
597 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
598 : | break; | ||
599 : | default: //1+2+4+8 == we didn't find anything at all | ||
600 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
601 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
602 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
603 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
604 : | break; | ||
605 : | chl | 1.18 | } |
606 : | syskin | 1.44.2.34 | if (!iDirection) break; //ok, the end. really |
607 : | bDirection = iDirection; | ||
608 : | x = data->currentMV->x; y = data->currentMV->y; | ||
609 : | chl | 1.18 | } |
610 : | syskin | 1.44.2.34 | } |
611 : | chl | 1.18 | } |
612 : | |||
613 : | chl | 1.44.2.1 | static void |
614 : | SquareSearch(int x, int y, const SearchData * const data, int bDirection) | ||
615 : | { | ||
616 : | int iDirection; | ||
617 : | chl | 1.35 | |
618 : | chl | 1.44.2.1 | do { |
619 : | iDirection = 0; | ||
620 : | if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64); | ||
621 : | if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128); | ||
622 : | if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32); | ||
623 : | if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128); | ||
624 : | if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64); | ||
625 : | if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128); | ||
626 : | if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128); | ||
627 : | if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128); | ||
628 : | |||
629 : | bDirection = iDirection; | ||
630 : | x = data->currentMV->x; y = data->currentMV->y; | ||
631 : | } while (iDirection); | ||
632 : | chl | 1.35 | } |
633 : | |||
634 : | chl | 1.44.2.1 | static void |
635 : | DiamondSearch(int x, int y, const SearchData * const data, int bDirection) | ||
636 : | chl | 1.18 | { |
637 : | |||
638 : | /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ | ||
639 : | |||
640 : | syskin | 1.44.2.34 | int iDirection; |
641 : | edgomez | 1.20 | |
642 : | syskin | 1.44.2.34 | do { |
643 : | iDirection = 0; | ||
644 : | if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); | ||
645 : | if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); | ||
646 : | if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); | ||
647 : | if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); | ||
648 : | chl | 1.18 | |
649 : | syskin | 1.44.2.34 | /* now we're doing diagonal checks near our candidate */ |
650 : | |||
651 : | if (iDirection) { //checking if anything found | ||
652 : | bDirection = iDirection; | ||
653 : | iDirection = 0; | ||
654 : | x = data->currentMV->x; y = data->currentMV->y; | ||
655 : | if (bDirection & 3) { //our candidate is left or right | ||
656 : | CHECK_CANDIDATE(x, y + iDiamondSize, 8); | ||
657 : | CHECK_CANDIDATE(x, y - iDiamondSize, 4); | ||
658 : | } else { // what remains here is up or down | ||
659 : | CHECK_CANDIDATE(x + iDiamondSize, y, 2); | ||
660 : | syskin | 1.44.2.40 | CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
661 : | chl | 1.18 | } |
662 : | syskin | 1.44.2.34 | bDirection += iDirection; |
663 : | x = data->currentMV->x; y = data->currentMV->y; | ||
664 : | chl | 1.18 | } |
665 : | syskin | 1.44.2.34 | } |
666 : | while (iDirection); | ||
667 : | chl | 1.18 | } |
668 : | |||
669 : | chl | 1.44.2.1 | /* MAINSEARCH FUNCTIONS END */ |
670 : | |||
671 : | /* HALFPELREFINE COULD BE A MAINSEARCH FUNCTION, BUT THERE IS NO NEED FOR IT */ | ||
672 : | chl | 1.18 | |
673 : | chl | 1.44.2.1 | static void |
674 : | syskin | 1.44.2.22 | SubpelRefine(const SearchData * const data) |
675 : | Isibaar | 1.1 | { |
676 : | syskin | 1.44.2.22 | /* Do a half-pel or q-pel refinement */ |
677 : | VECTOR backupMV; | ||
678 : | chl | 1.44.2.1 | int iDirection; //not needed |
679 : | edgomez | 1.20 | |
680 : | syskin | 1.44.2.22 | if (data->qpel_precision) |
681 : | backupMV = *(data->currentQMV); | ||
682 : | else backupMV = *(data->currentMV); | ||
683 : | Isibaar | 1.44.2.7 | |
684 : | syskin | 1.44.2.40 | CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); |
685 : | Isibaar | 1.44.2.7 | CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); |
686 : | CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); | ||
687 : | syskin | 1.44.2.40 | CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0); |
688 : | Isibaar | 1.44.2.7 | CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); |
689 : | syskin | 1.44.2.40 | CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); |
690 : | CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); | ||
691 : | CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); | ||
692 : | Isibaar | 1.44.2.7 | } |
693 : | |||
694 : | chl | 1.44.2.1 | static __inline int |
695 : | SkipDecisionP(const IMAGE * current, const IMAGE * reference, | ||
696 : | const int x, const int y, | ||
697 : | syskin | 1.44.2.32 | const uint32_t iEdgedWidth, const uint32_t iQuant, int rrv) |
698 : | Isibaar | 1.1 | |
699 : | chl | 1.44.2.1 | { |
700 : | /* keep repeating checks for all b-frames before this P frame, | ||
701 : | to make sure that SKIP is possible (todo) | ||
702 : | how: if skip is not possible set sad00 to a very high value */ | ||
703 : | syskin | 1.44.2.32 | if(rrv) { |
704 : | uint32_t sadC = sad16(current->u + x*16 + y*(iEdgedWidth/2)*16, | ||
705 : | reference->u + x*16 + y*(iEdgedWidth/2)*16, iEdgedWidth/2, 256*4096); | ||
706 : | if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; | ||
707 : | sadC += sad16(current->v + (x + y*(iEdgedWidth/2))*16, | ||
708 : | reference->v + (x + y*(iEdgedWidth/2))*16, iEdgedWidth/2, 256*4096); | ||
709 : | if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; | ||
710 : | return 1; | ||
711 : | } else { | ||
712 : | uint32_t sadC = sad8(current->u + x*8 + y*(iEdgedWidth/2)*8, | ||
713 : | reference->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); | ||
714 : | if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; | ||
715 : | sadC += sad8(current->v + (x + y*(iEdgedWidth/2))*8, | ||
716 : | reference->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); | ||
717 : | if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; | ||
718 : | return 1; | ||
719 : | } | ||
720 : | chl | 1.44.2.1 | } |
721 : | Isibaar | 1.1 | |
722 : | chl | 1.44.2.1 | static __inline void |
723 : | SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) | ||
724 : | { | ||
725 : | pMB->mode = MODE_NOT_CODED; | ||
726 : | Isibaar | 1.44.2.2 | pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
727 : | pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; | ||
728 : | syskin | 1.44.2.14 | |
729 : | Isibaar | 1.44.2.7 | pMB->qmvs[0].x = pMB->qmvs[1].x = pMB->qmvs[2].x = pMB->qmvs[3].x = 0; |
730 : | pMB->qmvs[0].y = pMB->qmvs[1].y = pMB->qmvs[2].y = pMB->qmvs[3].y = 0; | ||
731 : | |||
732 : | chl | 1.44.2.1 | pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
733 : | } | ||
734 : | |||
735 : | syskin | 1.44.2.14 | bool |
736 : | chl | 1.44.2.1 | MotionEstimation(MBParam * const pParam, |
737 : | FRAMEINFO * const current, | ||
738 : | FRAMEINFO * const reference, | ||
739 : | const IMAGE * const pRefH, | ||
740 : | const IMAGE * const pRefV, | ||
741 : | const IMAGE * const pRefHV, | ||
742 : | const uint32_t iLimit) | ||
743 : | { | ||
744 : | MACROBLOCK *const pMBs = current->mbs; | ||
745 : | const IMAGE *const pCurrent = ¤t->image; | ||
746 : | const IMAGE *const pRef = &reference->image; | ||
747 : | Isibaar | 1.1 | |
748 : | chl | 1.44.2.1 | const VECTOR zeroMV = { 0, 0 }; |
749 : | edgomez | 1.20 | |
750 : | syskin | 1.44.2.32 | uint32_t mb_width = pParam->mb_width; |
751 : | uint32_t mb_height = pParam->mb_height; | ||
752 : | suxen_drol | 1.44.2.31 | |
753 : | chl | 1.44.2.1 | uint32_t x, y; |
754 : | uint32_t iIntra = 0; | ||
755 : | syskin | 1.44.2.14 | int32_t InterBias, quant = current->quant, sad00; |
756 : | chl | 1.44.2.1 | |
757 : | Isibaar | 1.44.2.2 | // some pre-initialized thingies for SearchP |
758 : | syskin | 1.44.2.34 | int32_t temp[8]; |
759 : | Isibaar | 1.44.2.2 | VECTOR currentMV[5]; |
760 : | Isibaar | 1.44.2.7 | VECTOR currentQMV[5]; |
761 : | Isibaar | 1.44.2.2 | int32_t iMinSAD[5]; |
762 : | SearchData Data; | ||
763 : | syskin | 1.44.2.32 | memset(&Data, 0, sizeof(SearchData)); |
764 : | Isibaar | 1.44.2.2 | Data.iEdgedWidth = pParam->edged_width; |
765 : | Data.currentMV = currentMV; | ||
766 : | Isibaar | 1.44.2.7 | Data.currentQMV = currentQMV; |
767 : | Isibaar | 1.44.2.2 | Data.iMinSAD = iMinSAD; |
768 : | Data.temp = temp; | ||
769 : | Data.iFcode = current->fcode; | ||
770 : | Isibaar | 1.44.2.7 | Data.rounding = pParam->m_rounding_type; |
771 : | syskin | 1.44.2.14 | Data.qpel = pParam->m_quarterpel; |
772 : | Data.chroma = current->global_flags & XVID_ME_COLOUR; | ||
773 : | syskin | 1.44.2.32 | Data.rrv = current->global_flags & XVID_REDUCED; |
774 : | Isibaar | 1.44.2.7 | |
775 : | syskin | 1.44.2.32 | if ((current->global_flags & XVID_REDUCED)) { |
776 : | suxen_drol | 1.44.2.31 | mb_width = (pParam->width + 31) / 32; |
777 : | mb_height = (pParam->height + 31) / 32; | ||
778 : | syskin | 1.44.2.32 | Data.qpel = Data.chroma = 0; |
779 : | suxen_drol | 1.44.2.31 | } |
780 : | |||
781 : | syskin | 1.44.2.42 | Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose) |
782 : | chl | 1.44.2.1 | if (sadInit) (*sadInit) (); |
783 : | |||
784 : | suxen_drol | 1.44.2.31 | for (y = 0; y < mb_height; y++) { |
785 : | for (x = 0; x < mb_width; x++) { | ||
786 : | chl | 1.44.2.1 | MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
787 : | syskin | 1.44.2.14 | |
788 : | syskin | 1.44.2.40 | if (Data.rrv) pMB->sad16 = |
789 : | syskin | 1.44.2.32 | sad32v_c(pCurrent->y + (x + y * pParam->edged_width) * 32, |
790 : | pRef->y + (x + y * pParam->edged_width) * 32, | ||
791 : | pParam->edged_width, pMB->sad8 ); | ||
792 : | |||
793 : | syskin | 1.44.2.40 | else pMB->sad16 = |
794 : | syskin | 1.44.2.32 | sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, |
795 : | chl | 1.44.2.1 | pRef->y + (x + y * pParam->edged_width) * 16, |
796 : | pParam->edged_width, pMB->sad8 ); | ||
797 : | |||
798 : | syskin | 1.44.2.14 | if (Data.chroma) { |
799 : | pMB->sad16 += sad8(pCurrent->u + x*8 + y*(pParam->edged_width/2)*8, | ||
800 : | pRef->u + x*8 + y*(pParam->edged_width/2)*8, pParam->edged_width/2); | ||
801 : | |||
802 : | pMB->sad16 += sad8(pCurrent->v + (x + y*(pParam->edged_width/2))*8, | ||
803 : | pRef->v + (x + y*(pParam->edged_width/2))*8, pParam->edged_width/2); | ||
804 : | } | ||
805 : | |||
806 : | sad00 = pMB->sad16; //if no gmc; else sad00 = (..) | ||
807 : | |||
808 : | chl | 1.44.2.1 | if (!(current->global_flags & XVID_LUMIMASKING)) { |
809 : | pMB->dquant = NO_CHANGE; | ||
810 : | syskin | 1.44.2.40 | pMB->quant = current->quant; |
811 : | syskin | 1.44.2.11 | } else { |
812 : | syskin | 1.44.2.5 | if (pMB->dquant != NO_CHANGE) { |
813 : | quant += DQtab[pMB->dquant]; | ||
814 : | if (quant > 31) quant = 31; | ||
815 : | else if (quant < 1) quant = 1; | ||
816 : | } | ||
817 : | syskin | 1.44.2.11 | pMB->quant = quant; |
818 : | } | ||
819 : | chl | 1.44.2.1 | |
820 : | syskin | 1.44.2.14 | //initial skip decision |
821 : | chl | 1.44.2.13 | /* no early skip for GMC (global vector = skip vector is unknown!) */ |
822 : | if (current->coding_type == P_VOP) { /* no fast SKIP for S(GMC)-VOPs */ | ||
823 : | syskin | 1.44.2.32 | if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH * (Data.rrv ? 4:1) ) |
824 : | if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant, Data.rrv)) { | ||
825 : | chl | 1.44.2.13 | SkipMacroblockP(pMB, sad00); |
826 : | continue; | ||
827 : | } | ||
828 : | } | ||
829 : | syskin | 1.44.2.14 | |
830 : | SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, | ||
831 : | chl | 1.44.2.1 | y, current->motion_flags, pMB->quant, |
832 : | Isibaar | 1.44.2.2 | &Data, pParam, pMBs, reference->mbs, |
833 : | chl | 1.44.2.1 | current->global_flags & XVID_INTER4V, pMB); |
834 : | |||
835 : | /* final skip decision, a.k.a. "the vector you found, really that good?" */ | ||
836 : | syskin | 1.44.2.40 | if (current->coding_type == P_VOP) { |
837 : | chl | 1.44.2.13 | if ( (pMB->dquant == NO_CHANGE) && (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
838 : | syskin | 1.44.2.32 | && ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1)) ) |
839 : | if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant, Data.rrv)) { | ||
840 : | chl | 1.44.2.13 | SkipMacroblockP(pMB, sad00); |
841 : | continue; | ||
842 : | } | ||
843 : | } | ||
844 : | syskin | 1.44.2.14 | |
845 : | chl | 1.44.2.1 | /* finally, intra decision */ |
846 : | Isibaar | 1.1 | |
847 : | chl | 1.44.2.1 | InterBias = MV16_INTER_BIAS; |
848 : | syskin | 1.44.2.16 | if (pMB->quant > 8) InterBias += 100 * (pMB->quant - 8); // to make high quants work |
849 : | chl | 1.44.2.1 | if (y != 0) |
850 : | syskin | 1.44.2.16 | if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; |
851 : | chl | 1.44.2.1 | if (x != 0) |
852 : | syskin | 1.44.2.16 | if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; |
853 : | syskin | 1.44.2.14 | |
854 : | if (Data.chroma) InterBias += 50; // to compensate bigger SAD | ||
855 : | syskin | 1.44.2.32 | if (Data.rrv) InterBias *= 4; //?? |
856 : | chl | 1.44.2.1 | |
857 : | if (InterBias < pMB->sad16) { | ||
858 : | syskin | 1.44.2.32 | int32_t deviation; |
859 : | if (Data.rrv) { | ||
860 : | deviation = dev16(pCurrent->y + (x + y * pParam->edged_width) * 32, | ||
861 : | pParam->edged_width) | ||
862 : | + dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16, | ||
863 : | pParam->edged_width) | ||
864 : | + dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16 * pParam->edged_width, | ||
865 : | pParam->edged_width) | ||
866 : | + dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16 * (pParam->edged_width+1), | ||
867 : | pParam->edged_width); | ||
868 : | } else | ||
869 : | deviation = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, | ||
870 : | chl | 1.44.2.1 | pParam->edged_width); |
871 : | edgomez | 1.20 | |
872 : | syskin | 1.44.2.32 | if (deviation < (pMB->sad16 - InterBias)) { |
873 : | syskin | 1.44.2.42 | if (++iIntra >= iLimit) return 1; |
874 : | chl | 1.44.2.1 | pMB->mode = MODE_INTRA; |
875 : | Isibaar | 1.44.2.2 | pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = |
876 : | pMB->mvs[3] = zeroMV; | ||
877 : | Isibaar | 1.44.2.7 | pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = |
878 : | pMB->qmvs[3] = zeroMV; | ||
879 : | chl | 1.44.2.1 | pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = |
880 : | pMB->sad8[3] = 0; | ||
881 : | } | ||
882 : | } | ||
883 : | } | ||
884 : | } | ||
885 : | chl | 1.44.2.13 | |
886 : | if (current->coding_type == S_VOP) /* first GMC step only for S(GMC)-VOPs */ | ||
887 : | syskin | 1.44.2.14 | current->GMC_MV = GlobalMotionEst( pMBs, pParam, current->fcode ); |
888 : | chl | 1.44.2.13 | else |
889 : | current->GMC_MV = zeroMV; | ||
890 : | |||
891 : | syskin | 1.44.2.14 | return 0; |
892 : | chl | 1.44.2.1 | } |
893 : | chl | 1.18 | |
894 : | edgomez | 1.20 | |
895 : | chl | 1.44.2.1 | #define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
896 : | edgomez | 1.20 | |
897 : | chl | 1.44.2.1 | static __inline int |
898 : | make_mask(const VECTOR * const pmv, const int i) | ||
899 : | { | ||
900 : | Isibaar | 1.44.2.2 | int mask = 255, j; |
901 : | chl | 1.44.2.1 | for (j = 0; j < i; j++) { |
902 : | if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already | ||
903 : | if (pmv[i].x == pmv[j].x) { | ||
904 : | if (pmv[i].y == pmv[j].y + iDiamondSize) { mask &= ~4; continue; } | ||
905 : | if (pmv[i].y == pmv[j].y - iDiamondSize) { mask &= ~8; continue; } | ||
906 : | syskin | 1.44.2.40 | } else |
907 : | chl | 1.44.2.1 | if (pmv[i].y == pmv[j].y) { |
908 : | if (pmv[i].x == pmv[j].x + iDiamondSize) { mask &= ~1; continue; } | ||
909 : | if (pmv[i].x == pmv[j].x - iDiamondSize) { mask &= ~2; continue; } | ||
910 : | } | ||
911 : | } | ||
912 : | return mask; | ||
913 : | } | ||
914 : | |||
915 : | syskin | 1.44.2.40 | static __inline void |
916 : | syskin | 1.44.2.32 | PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, |
917 : | int iHcount, const MACROBLOCK * const prevMB, int rrv) | ||
918 : | chl | 1.44.2.1 | { |
919 : | |||
920 : | //this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself | ||
921 : | syskin | 1.44.2.32 | if (rrv) { iWcount /= 2; iHcount /= 2; } |
922 : | chl | 1.44.2.1 | |
923 : | syskin | 1.44.2.32 | if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour |
924 : | chl | 1.44.2.1 | pmv[5].x = EVEN(pmv[3].x); |
925 : | syskin | 1.44.2.40 | pmv[5].y = EVEN(pmv[3].y); |
926 : | Isibaar | 1.44.2.2 | } else pmv[5].x = pmv[5].y = 0; |
927 : | chl | 1.44.2.1 | |
928 : | if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour | ||
929 : | else pmv[3].x = pmv[3].y = 0; | ||
930 : | |||
931 : | if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour | ||
932 : | else pmv[4].x = pmv[4].y = 0; | ||
933 : | |||
934 : | // [1] median prediction | ||
935 : | syskin | 1.44.2.43 | pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); |
936 : | chl | 1.44.2.1 | |
937 : | pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask | ||
938 : | |||
939 : | pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame | ||
940 : | pmv[2].y = EVEN(prevMB->mvs[0].y); | ||
941 : | |||
942 : | syskin | 1.44.2.32 | if ((x < iWcount-1) && (y < iHcount-1)) { |
943 : | chl | 1.44.2.1 | pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame |
944 : | syskin | 1.44.2.40 | pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
945 : | Isibaar | 1.44.2.2 | } else pmv[6].x = pmv[6].y = 0; |
946 : | syskin | 1.44.2.32 | |
947 : | if (rrv) { | ||
948 : | int i; | ||
949 : | for (i = 0; i < 7; i++) { | ||
950 : | syskin | 1.44.2.43 | pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); // halfzero->halfpel |
951 : | pmv[i].y = RRV_MV_SCALEUP(pmv[i].y); | ||
952 : | syskin | 1.44.2.32 | } |
953 : | } | ||
954 : | chl | 1.44.2.1 | } |
955 : | |||
956 : | static void | ||
957 : | syskin | 1.44.2.14 | SearchP(const IMAGE * const pRef, |
958 : | chl | 1.44.2.1 | const uint8_t * const pRefH, |
959 : | const uint8_t * const pRefV, | ||
960 : | const uint8_t * const pRefHV, | ||
961 : | const IMAGE * const pCur, | ||
962 : | const int x, | ||
963 : | const int y, | ||
964 : | const uint32_t MotionFlags, | ||
965 : | const uint32_t iQuant, | ||
966 : | Isibaar | 1.44.2.2 | SearchData * const Data, |
967 : | chl | 1.44.2.1 | const MBParam * const pParam, |
968 : | const MACROBLOCK * const pMBs, | ||
969 : | const MACROBLOCK * const prevMBs, | ||
970 : | int inter4v, | ||
971 : | MACROBLOCK * const pMB) | ||
972 : | { | ||
973 : | Isibaar | 1.1 | |
974 : | chl | 1.44.2.1 | int i, iDirection = 255, mask, threshA; |
975 : | Isibaar | 1.44.2.2 | VECTOR pmv[7]; |
976 : | chl | 1.44.2.1 | |
977 : | syskin | 1.44.2.34 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
978 : | pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); | ||
979 : | syskin | 1.44.2.32 | |
980 : | syskin | 1.44.2.34 | get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); //has to be changed to get_pmv(2)() |
981 : | syskin | 1.44.2.32 | |
982 : | syskin | 1.44.2.34 | Data->temp[5] = Data->temp[7] = 256*4096; // to reset chroma-sad cache |
983 : | syskin | 1.44.2.32 | if (Data->rrv) i = 2; else i = 1; |
984 : | Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i; | ||
985 : | Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; | ||
986 : | Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; | ||
987 : | |||
988 : | Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16*i; | ||
989 : | Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16*i; | ||
990 : | Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16*i; | ||
991 : | Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16*i; | ||
992 : | Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; | ||
993 : | Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; | ||
994 : | Isibaar | 1.1 | |
995 : | syskin | 1.44.2.11 | Data->lambda16 = lambda_vec16[iQuant]; |
996 : | Isibaar | 1.44.2.12 | Data->lambda8 = lambda_vec8[iQuant]; |
997 : | syskin | 1.44.2.22 | Data->qpel_precision = 0; |
998 : | Isibaar | 1.1 | |
999 : | edgomez | 1.44.2.4 | if (pMB->dquant != NO_CHANGE) inter4v = 0; |
1000 : | Isibaar | 1.44.2.2 | |
1001 : | Isibaar | 1.44.2.7 | for(i = 0; i < 5; i++) |
1002 : | Data->currentMV[i].x = Data->currentMV[i].y = 0; | ||
1003 : | chl | 1.44.2.1 | |
1004 : | syskin | 1.44.2.22 | if (pParam->m_quarterpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
1005 : | else Data->predMV = pmv[0]; | ||
1006 : | Isibaar | 1.44.2.12 | |
1007 : | syskin | 1.44.2.36 | i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0); |
1008 : | syskin | 1.44.2.14 | Data->iMinSAD[0] = pMB->sad16 + (Data->lambda16 * i * pMB->sad16)/1000; |
1009 : | Data->iMinSAD[1] = pMB->sad8[0] + (Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS))/100; | ||
1010 : | Isibaar | 1.44.2.2 | Data->iMinSAD[2] = pMB->sad8[1]; |
1011 : | Data->iMinSAD[3] = pMB->sad8[2]; | ||
1012 : | Data->iMinSAD[4] = pMB->sad8[3]; | ||
1013 : | chl | 1.44.2.1 | |
1014 : | if ((x == 0) && (y == 0)) threshA = 512; | ||
1015 : | else { | ||
1016 : | edgomez | 1.44.2.4 | threshA = Data->temp[0]; // that's when we keep this SAD atm |
1017 : | chl | 1.44.2.1 | if (threshA < 512) threshA = 512; |
1018 : | if (threshA > 1024) threshA = 1024; } | ||
1019 : | |||
1020 : | PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, | ||
1021 : | syskin | 1.44.2.32 | prevMBs + x + y * pParam->mb_width, Data->rrv); |
1022 : | chl | 1.44.2.1 | |
1023 : | syskin | 1.44.2.32 | if (Data->rrv) CheckCandidate = CheckCandidate32; |
1024 : | else if (inter4v || Data->chroma) CheckCandidate = CheckCandidate16; | ||
1025 : | else CheckCandidate = CheckCandidate16no4v; //for extra speed | ||
1026 : | edgomez | 1.44.2.4 | |
1027 : | chl | 1.44.2.1 | /* main loop. checking all predictions */ |
1028 : | |||
1029 : | for (i = 1; i < 7; i++) { | ||
1030 : | if (!(mask = make_mask(pmv, i)) ) continue; | ||
1031 : | edgomez | 1.44.2.4 | (*CheckCandidate)(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1032 : | Isibaar | 1.44.2.2 | if (Data->iMinSAD[0] <= threshA) break; |
1033 : | chl | 1.44.2.1 | } |
1034 : | |||
1035 : | Isibaar | 1.44.2.2 | if ((Data->iMinSAD[0] <= threshA) || |
1036 : | (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && | ||
1037 : | (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { | ||
1038 : | chl | 1.44.2.1 | inter4v = 0; |
1039 : | Isibaar | 1.44.2.2 | } else { |
1040 : | edgomez | 1.20 | |
1041 : | syskin | 1.44.2.40 | MainSearchFunc * MainSearchPtr; |
1042 : | Isibaar | 1.44.2.2 | if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1043 : | else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; | ||
1044 : | else MainSearchPtr = DiamondSearch; | ||
1045 : | chl | 1.35 | |
1046 : | Isibaar | 1.44.2.2 | (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
1047 : | chl | 1.13 | |
1048 : | chl | 1.44.2.1 | /* extended search, diamond starting in 0,0 and in prediction. |
1049 : | note that this search is/might be done in halfpel positions, | ||
1050 : | which makes it more different than the diamond above */ | ||
1051 : | Isibaar | 1.1 | |
1052 : | Isibaar | 1.44.2.2 | if (MotionFlags & PMV_EXTSEARCH16) { |
1053 : | int32_t bSAD; | ||
1054 : | VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; | ||
1055 : | syskin | 1.44.2.32 | if (Data->rrv) { |
1056 : | startMV.x = RRV_MV_SCALEUP(startMV.x); | ||
1057 : | startMV.y = RRV_MV_SCALEUP(startMV.y); | ||
1058 : | syskin | 1.44.2.40 | } else |
1059 : | syskin | 1.44.2.32 | if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? |
1060 : | startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); | ||
1061 : | Isibaar | 1.44.2.2 | if (!(MVequal(startMV, backupMV))) { |
1062 : | bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; | ||
1063 : | |||
1064 : | Isibaar | 1.44.2.12 | (*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); |
1065 : | Isibaar | 1.44.2.2 | (*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
1066 : | if (bSAD < Data->iMinSAD[0]) { | ||
1067 : | Data->currentMV[0] = backupMV; | ||
1068 : | Data->iMinSAD[0] = bSAD; } | ||
1069 : | } | ||
1070 : | chl | 1.13 | |
1071 : | Isibaar | 1.44.2.2 | backupMV = Data->currentMV[0]; |
1072 : | syskin | 1.44.2.32 | if (!MotionFlags & PMV_HALFPELREFINE16 || Data->rrv) startMV.x = startMV.y = 0; |
1073 : | else startMV.x = startMV.y = 1; | ||
1074 : | Isibaar | 1.44.2.2 | if (!(MVequal(startMV, backupMV))) { |
1075 : | bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; | ||
1076 : | |||
1077 : | Isibaar | 1.44.2.12 | (*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); |
1078 : | Isibaar | 1.44.2.2 | (*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
1079 : | if (bSAD < Data->iMinSAD[0]) { | ||
1080 : | Data->currentMV[0] = backupMV; | ||
1081 : | Data->iMinSAD[0] = bSAD; } | ||
1082 : | } | ||
1083 : | chl | 1.44.2.1 | } |
1084 : | } | ||
1085 : | chl | 1.13 | |
1086 : | syskin | 1.44.2.22 | if (MotionFlags & PMV_HALFPELREFINE16) SubpelRefine(Data); |
1087 : | Isibaar | 1.1 | |
1088 : | Isibaar | 1.44.2.7 | for(i = 0; i < 5; i++) { |
1089 : | Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors | ||
1090 : | Data->currentQMV[i].y = 2 * Data->currentMV[i].y; | ||
1091 : | } | ||
1092 : | |||
1093 : | syskin | 1.44.2.40 | if((Data->qpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { |
1094 : | Isibaar | 1.44.2.7 | |
1095 : | syskin | 1.44.2.22 | Data->qpel_precision = 1; |
1096 : | syskin | 1.44.2.34 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1097 : | pParam->width, pParam->height, Data->iFcode, 1, 0); | ||
1098 : | syskin | 1.44.2.8 | |
1099 : | syskin | 1.44.2.22 | SubpelRefine(Data); |
1100 : | Isibaar | 1.44.2.7 | } |
1101 : | |||
1102 : | syskin | 1.44.2.14 | if (Data->iMinSAD[0] < (int32_t)iQuant * 30 ) inter4v = 0; |
1103 : | Isibaar | 1.44.2.2 | if (inter4v) { |
1104 : | SearchData Data8; | ||
1105 : | Data8.iFcode = Data->iFcode; | ||
1106 : | Isibaar | 1.44.2.12 | Data8.lambda8 = Data->lambda8; |
1107 : | Isibaar | 1.44.2.2 | Data8.iEdgedWidth = Data->iEdgedWidth; |
1108 : | Isibaar | 1.44.2.12 | Data8.RefQ = Data->RefQ; |
1109 : | syskin | 1.44.2.14 | Data8.qpel = Data->qpel; |
1110 : | syskin | 1.44.2.32 | Data8.rrv = Data->rrv; |
1111 : | Isibaar | 1.44.2.2 | Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
1112 : | Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); | ||
1113 : | Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); | ||
1114 : | Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); | ||
1115 : | syskin | 1.44.2.14 | |
1116 : | if (Data->chroma) { | ||
1117 : | syskin | 1.44.2.41 | int sumx, sumy; |
1118 : | syskin | 1.44.2.14 | |
1119 : | if(pParam->m_quarterpel) { | ||
1120 : | syskin | 1.44.2.41 | sumx = pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2; |
1121 : | syskin | 1.44.2.22 | sumy = pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2; |
1122 : | } else { | ||
1123 : | sumx = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; | ||
1124 : | sumy = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; | ||
1125 : | } | ||
1126 : | syskin | 1.44.2.14 | |
1127 : | syskin | 1.44.2.41 | Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf], |
1128 : | (sumy >> 3) + roundtab_76[sumy & 0xf], Data); | ||
1129 : | syskin | 1.44.2.14 | } |
1130 : | Isibaar | 1.44.2.2 | } |
1131 : | chl | 1.13 | |
1132 : | syskin | 1.44.2.32 | if (Data->rrv) { |
1133 : | Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); | ||
1134 : | Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); | ||
1135 : | } | ||
1136 : | syskin | 1.44.2.38 | |
1137 : | chl | 1.44.2.1 | if (!(inter4v) || |
1138 : | syskin | 1.44.2.40 | (Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
1139 : | Isibaar | 1.44.2.2 | Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
1140 : | chl | 1.44.2.1 | // INTER MODE |
1141 : | pMB->mode = MODE_INTER; | ||
1142 : | Isibaar | 1.44.2.2 | pMB->mvs[0] = pMB->mvs[1] |
1143 : | = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; | ||
1144 : | edgomez | 1.20 | |
1145 : | chl | 1.44.2.1 | pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
1146 : | Isibaar | 1.44.2.2 | pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
1147 : | chl | 1.13 | |
1148 : | Isibaar | 1.44.2.7 | if(pParam->m_quarterpel) { |
1149 : | syskin | 1.44.2.22 | pMB->qmvs[0] = pMB->qmvs[1] |
1150 : | = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; | ||
1151 : | pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; | ||
1152 : | pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; | ||
1153 : | syskin | 1.44.2.14 | } else { |
1154 : | Isibaar | 1.44.2.7 | pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
1155 : | pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; | ||
1156 : | } | ||
1157 : | chl | 1.44.2.1 | } else { |
1158 : | syskin | 1.44.2.40 | // INTER4V MODE; all other things are already set in Search8 |
1159 : | chl | 1.44.2.1 | pMB->mode = MODE_INTER4V; |
1160 : | syskin | 1.44.2.40 | pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + |
1161 : | Isibaar | 1.44.2.2 | Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant; |
1162 : | chl | 1.44.2.1 | } |
1163 : | } | ||
1164 : | edgomez | 1.20 | |
1165 : | chl | 1.44.2.1 | static void |
1166 : | Search8(const SearchData * const OldData, | ||
1167 : | const int x, const int y, | ||
1168 : | const uint32_t MotionFlags, | ||
1169 : | const MBParam * const pParam, | ||
1170 : | MACROBLOCK * const pMB, | ||
1171 : | const MACROBLOCK * const pMBs, | ||
1172 : | Isibaar | 1.44.2.2 | const int block, |
1173 : | SearchData * const Data) | ||
1174 : | chl | 1.44.2.1 | { |
1175 : | syskin | 1.44.2.32 | int i = 0; |
1176 : | Isibaar | 1.44.2.2 | Data->iMinSAD = OldData->iMinSAD + 1 + block; |
1177 : | Data->currentMV = OldData->currentMV + 1 + block; | ||
1178 : | Isibaar | 1.44.2.7 | Data->currentQMV = OldData->currentQMV + 1 + block; |
1179 : | |||
1180 : | syskin | 1.44.2.11 | if(pParam->m_quarterpel) { |
1181 : | syskin | 1.44.2.32 | Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
1182 : | syskin | 1.44.2.36 | if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y, |
1183 : | Data->predMV, Data->iFcode, 0, 0); | ||
1184 : | syskin | 1.44.2.34 | |
1185 : | Isibaar | 1.44.2.12 | } else { |
1186 : | syskin | 1.44.2.32 | Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
1187 : | if (block != 0) { | ||
1188 : | syskin | 1.44.2.36 | if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y, |
1189 : | Data->predMV, Data->iFcode, 0, Data->rrv); | ||
1190 : | syskin | 1.44.2.32 | } |
1191 : | Isibaar | 1.44.2.12 | } |
1192 : | Isibaar | 1.1 | |
1193 : | syskin | 1.44.2.32 | *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100; |
1194 : | |||
1195 : | chl | 1.44.2.1 | if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { |
1196 : | syskin | 1.44.2.32 | if (Data->rrv) i = 2; else i = 1; |
1197 : | edgomez | 1.20 | |
1198 : | syskin | 1.44.2.32 | Data->Ref = OldData->Ref + i*8 * ((block&1) + pParam->edged_width*(block>>1)); |
1199 : | Data->RefH = OldData->RefH + i*8 * ((block&1) + pParam->edged_width*(block>>1)); | ||
1200 : | Data->RefV = OldData->RefV + i*8 * ((block&1) + pParam->edged_width*(block>>1)); | ||
1201 : | Data->RefHV = OldData->RefHV + i*8 * ((block&1) + pParam->edged_width*(block>>1)); | ||
1202 : | Isibaar | 1.1 | |
1203 : | syskin | 1.44.2.32 | Data->Cur = OldData->Cur + i*8 * ((block&1) + pParam->edged_width*(block>>1)); |
1204 : | syskin | 1.44.2.22 | Data->qpel_precision = 0; |
1205 : | syskin | 1.44.2.34 | |
1206 : | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, | ||
1207 : | pParam->width, pParam->height, OldData->iFcode - Data->qpel, 0, Data->rrv); | ||
1208 : | |||
1209 : | if (Data->rrv) CheckCandidate = CheckCandidate16no4v; | ||
1210 : | else CheckCandidate = CheckCandidate8; | ||
1211 : | Isibaar | 1.1 | |
1212 : | chl | 1.44.2.1 | if (MotionFlags & PMV_EXTSEARCH8) { |
1213 : | Isibaar | 1.44.2.7 | int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
1214 : | |||
1215 : | chl | 1.44.2.1 | MainSearchFunc *MainSearchPtr; |
1216 : | if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; | ||
1217 : | else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; | ||
1218 : | else MainSearchPtr = DiamondSearch; | ||
1219 : | Isibaar | 1.1 | |
1220 : | Isibaar | 1.44.2.7 | (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1221 : | Isibaar | 1.1 | |
1222 : | syskin | 1.44.2.11 | if(*(Data->iMinSAD) < temp_sad) { |
1223 : | Isibaar | 1.44.2.7 | Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
1224 : | Data->currentQMV->y = 2 * Data->currentMV->y; | ||
1225 : | } | ||
1226 : | } | ||
1227 : | |||
1228 : | if (MotionFlags & PMV_HALFPELREFINE8) { | ||
1229 : | int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD | ||
1230 : | |||
1231 : | syskin | 1.44.2.22 | SubpelRefine(Data); // perform halfpel refine of current best vector |
1232 : | Isibaar | 1.44.2.7 | |
1233 : | if(*(Data->iMinSAD) < temp_sad) { // we have found a better match | ||
1234 : | Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector | ||
1235 : | Data->currentQMV->y = 2 * Data->currentMV->y; | ||
1236 : | } | ||
1237 : | } | ||
1238 : | |||
1239 : | syskin | 1.44.2.40 | if(Data->qpel) { |
1240 : | syskin | 1.44.2.11 | if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && |
1241 : | (MotionFlags & PMV_QUARTERPELREFINE8)) { | ||
1242 : | syskin | 1.44.2.22 | Data->qpel_precision = 1; |
1243 : | syskin | 1.44.2.34 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
1244 : | pParam->width, pParam->height, OldData->iFcode, 1, 0); | ||
1245 : | syskin | 1.44.2.22 | SubpelRefine(Data); |
1246 : | syskin | 1.44.2.11 | } |
1247 : | Isibaar | 1.44.2.7 | } |
1248 : | } | ||
1249 : | |||
1250 : | syskin | 1.44.2.32 | if (Data->rrv) { |
1251 : | Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x); | ||
1252 : | Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y); | ||
1253 : | } | ||
1254 : | |||
1255 : | syskin | 1.44.2.34 | if(Data->qpel) { |
1256 : | syskin | 1.44.2.22 | pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x; |
1257 : | pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y; | ||
1258 : | pMB->qmvs[block] = *(Data->currentQMV); | ||
1259 : | syskin | 1.44.2.32 | } else { |
1260 : | Isibaar | 1.44.2.7 | pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
1261 : | pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; | ||
1262 : | Isibaar | 1.1 | } |
1263 : | |||
1264 : | Isibaar | 1.44.2.2 | pMB->mvs[block] = *(Data->currentMV); |
1265 : | syskin | 1.44.2.11 | pMB->sad8[block] = 4 * (*Data->iMinSAD); |
1266 : | Isibaar | 1.1 | } |
1267 : | |||
1268 : | chl | 1.44.2.1 | /* B-frames code starts here */ |
1269 : | Isibaar | 1.1 | |
1270 : | chl | 1.44.2.1 | static __inline VECTOR |
1271 : | ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) | ||
1272 : | { | ||
1273 : | /* the stupidiest function ever */ | ||
1274 : | if (mode == MODE_FORWARD) return pMB->mvs[0]; | ||
1275 : | else return pMB->b_mvs[0]; | ||
1276 : | } | ||
1277 : | Isibaar | 1.1 | |
1278 : | chl | 1.44.2.1 | static void __inline |
1279 : | PreparePredictionsBF(VECTOR * const pmv, const int x, const int y, | ||
1280 : | const uint32_t iWcount, | ||
1281 : | const MACROBLOCK * const pMB, | ||
1282 : | const uint32_t mode_curr) | ||
1283 : | { | ||
1284 : | Isibaar | 1.1 | |
1285 : | chl | 1.44.2.1 | // [0] is prediction |
1286 : | pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); | ||
1287 : | Isibaar | 1.1 | |
1288 : | chl | 1.44.2.1 | pmv[1].x = pmv[1].y = 0; // [1] is zero |
1289 : | Isibaar | 1.1 | |
1290 : | chl | 1.44.2.1 | pmv[2] = ChoosePred(pMB, mode_curr); |
1291 : | pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); | ||
1292 : | Isibaar | 1.1 | |
1293 : | chl | 1.44.2.1 | if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour |
1294 : | pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); | ||
1295 : | syskin | 1.44.2.40 | pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); |
1296 : | Isibaar | 1.44.2.2 | } else pmv[3].x = pmv[3].y = 0; |
1297 : | chl | 1.39 | |
1298 : | chl | 1.44.2.1 | if (y != 0) { |
1299 : | pmv[4] = ChoosePred(pMB-iWcount, mode_curr); | ||
1300 : | pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y); | ||
1301 : | } else pmv[4].x = pmv[4].y = 0; | ||
1302 : | chl | 1.39 | |
1303 : | chl | 1.44.2.1 | if (x != 0) { |
1304 : | pmv[5] = ChoosePred(pMB-1, mode_curr); | ||
1305 : | pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); | ||
1306 : | } else pmv[5].x = pmv[5].y = 0; | ||
1307 : | chl | 1.39 | |
1308 : | chl | 1.44.2.1 | if ((x != 0)&&(y != 0)) { |
1309 : | pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); | ||
1310 : | syskin | 1.44.2.43 | pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y); |
1311 : | chl | 1.44.2.1 | } else pmv[6].x = pmv[6].y = 0; |
1312 : | chl | 1.39 | |
1313 : | chl | 1.44.2.1 | // more? |
1314 : | chl | 1.39 | } |
1315 : | |||
1316 : | |||
1317 : | chl | 1.44.2.1 | /* search backward or forward, for b-frames */ |
1318 : | static void | ||
1319 : | SearchBF( const uint8_t * const pRef, | ||
1320 : | const uint8_t * const pRefH, | ||
1321 : | const uint8_t * const pRefV, | ||
1322 : | const uint8_t * const pRefHV, | ||
1323 : | const IMAGE * const pCur, | ||
1324 : | const int x, const int y, | ||
1325 : | const uint32_t MotionFlags, | ||
1326 : | const uint32_t iFcode, | ||
1327 : | const MBParam * const pParam, | ||
1328 : | MACROBLOCK * const pMB, | ||
1329 : | const VECTOR * const predMV, | ||
1330 : | int32_t * const best_sad, | ||
1331 : | h | 1.44.2.3 | const int32_t mode_current, |
1332 : | SearchData * const Data) | ||
1333 : | Isibaar | 1.1 | { |
1334 : | |||
1335 : | chl | 1.44.2.1 | const int32_t iEdgedWidth = pParam->edged_width; |
1336 : | |||
1337 : | syskin | 1.44.2.38 | int i, iDirection = 255, mask; |
1338 : | h | 1.44.2.3 | VECTOR pmv[7]; |
1339 : | chl | 1.44.2.1 | MainSearchFunc *MainSearchPtr; |
1340 : | h | 1.44.2.3 | *Data->iMinSAD = MV_MAX_ERROR; |
1341 : | Data->iFcode = iFcode; | ||
1342 : | syskin | 1.44.2.22 | Data->qpel_precision = 0; |
1343 : | chl | 1.44.2.1 | |
1344 : | h | 1.44.2.3 | Data->Ref = pRef + (x + y * iEdgedWidth) * 16; |
1345 : | Data->RefH = pRefH + (x + y * iEdgedWidth) * 16; | ||
1346 : | Data->RefV = pRefV + (x + y * iEdgedWidth) * 16; | ||
1347 : | Data->RefHV = pRefHV + (x + y * iEdgedWidth) * 16; | ||
1348 : | chl | 1.44.2.1 | |
1349 : | h | 1.44.2.3 | Data->predMV = *predMV; |
1350 : | edgomez | 1.20 | |
1351 : | h | 1.44.2.3 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1352 : | syskin | 1.44.2.34 | pParam->width, pParam->height, iFcode - Data->qpel, 0, 0); |
1353 : | Isibaar | 1.1 | |
1354 : | h | 1.44.2.3 | pmv[0] = Data->predMV; |
1355 : | syskin | 1.44.2.19 | if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } |
1356 : | h | 1.44.2.3 | PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
1357 : | Isibaar | 1.1 | |
1358 : | h | 1.44.2.3 | Data->currentMV->x = Data->currentMV->y = 0; |
1359 : | chl | 1.44.2.1 | CheckCandidate = CheckCandidate16no4v; |
1360 : | edgomez | 1.20 | |
1361 : | chl | 1.44.2.1 | // main loop. checking all predictions |
1362 : | syskin | 1.44.2.38 | for (i = 0; i < 7; i++) { |
1363 : | chl | 1.44.2.1 | if (!(mask = make_mask(pmv, i)) ) continue; |
1364 : | h | 1.44.2.3 | CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1365 : | edgomez | 1.20 | } |
1366 : | |||
1367 : | chl | 1.44.2.1 | if (MotionFlags & PMV_USESQUARES16) |
1368 : | MainSearchPtr = SquareSearch; | ||
1369 : | else if (MotionFlags & PMV_ADVANCEDDIAMOND16) | ||
1370 : | MainSearchPtr = AdvDiamondSearch; | ||
1371 : | else MainSearchPtr = DiamondSearch; | ||
1372 : | Isibaar | 1.1 | |
1373 : | syskin | 1.44.2.38 | (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
1374 : | Isibaar | 1.1 | |
1375 : | syskin | 1.44.2.22 | SubpelRefine(Data); |
1376 : | syskin | 1.44.2.19 | |
1377 : | syskin | 1.44.2.38 | if (Data->qpel && *Data->iMinSAD < *best_sad + 300) { |
1378 : | syskin | 1.44.2.19 | Data->currentQMV->x = 2*Data->currentMV->x; |
1379 : | Data->currentQMV->y = 2*Data->currentMV->y; | ||
1380 : | syskin | 1.44.2.22 | Data->qpel_precision = 1; |
1381 : | syskin | 1.44.2.34 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1382 : | pParam->width, pParam->height, iFcode, 1, 0); | ||
1383 : | syskin | 1.44.2.22 | SubpelRefine(Data); |
1384 : | syskin | 1.44.2.19 | } |
1385 : | Isibaar | 1.1 | |
1386 : | chl | 1.44.2.1 | // three bits are needed to code backward mode. four for forward |
1387 : | syskin | 1.44.2.38 | |
1388 : | syskin | 1.44.2.14 | if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
1389 : | else *Data->iMinSAD += 3 * Data->lambda16; | ||
1390 : | chl | 1.44.2.1 | |
1391 : | h | 1.44.2.3 | if (*Data->iMinSAD < *best_sad) { |
1392 : | *best_sad = *Data->iMinSAD; | ||
1393 : | chl | 1.44.2.1 | pMB->mode = mode_current; |
1394 : | syskin | 1.44.2.19 | if (Data->qpel) { |
1395 : | pMB->pmvs[0].x = Data->currentQMV->x - predMV->x; | ||
1396 : | pMB->pmvs[0].y = Data->currentQMV->y - predMV->y; | ||
1397 : | syskin | 1.44.2.40 | if (mode_current == MODE_FORWARD) |
1398 : | syskin | 1.44.2.19 | pMB->qmvs[0] = *Data->currentQMV; |
1399 : | syskin | 1.44.2.40 | else |
1400 : | syskin | 1.44.2.19 | pMB->b_qmvs[0] = *Data->currentQMV; |
1401 : | } else { | ||
1402 : | pMB->pmvs[0].x = Data->currentMV->x - predMV->x; | ||
1403 : | pMB->pmvs[0].y = Data->currentMV->y - predMV->y; | ||
1404 : | } | ||
1405 : | syskin | 1.44.2.37 | if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; |
1406 : | else pMB->b_mvs[0] = *Data->currentMV; | ||
1407 : | Isibaar | 1.1 | } |
1408 : | syskin | 1.44.2.37 | if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; |
1409 : | else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search | ||
1410 : | |||
1411 : | chl | 1.44.2.1 | } |
1412 : | |||
1413 : | syskin | 1.44.2.40 | static void |
1414 : | syskin | 1.44.2.34 | SkipDecisionB(const IMAGE * const pCur, |
1415 : | syskin | 1.44.2.43 | const IMAGE * const f_Ref, |
1416 : | const IMAGE * const b_Ref, | ||
1417 : | MACROBLOCK * const pMB, | ||
1418 : | const uint32_t x, const uint32_t y, | ||
1419 : | const SearchData * const Data) | ||
1420 : | syskin | 1.44.2.34 | { |
1421 : | syskin | 1.44.2.43 | int dx = 0, dy = 0, b_dx = 0, b_dy = 0; |
1422 : | syskin | 1.44.2.34 | uint32_t sum; |
1423 : | syskin | 1.44.2.43 | const int div = 1 + Data->qpel; |
1424 : | int k; | ||
1425 : | const uint32_t quant = pMB->quant; | ||
1426 : | syskin | 1.44.2.34 | //this is not full chroma compensation, only it's fullpel approximation. should work though |
1427 : | |||
1428 : | syskin | 1.44.2.43 | for (k = 0; k < 4; k++) { |
1429 : | dy += Data->directmvF[k].y / div; | ||
1430 : | dx += Data->directmvF[0].x / div; | ||
1431 : | b_dy += Data->directmvB[0].y / div; | ||
1432 : | b_dx += Data->directmvB[0].x / div; | ||
1433 : | syskin | 1.44.2.34 | } |
1434 : | |||
1435 : | dy = (dy >> 3) + roundtab_76[dy & 0xf]; | ||
1436 : | dx = (dx >> 3) + roundtab_76[dx & 0xf]; | ||
1437 : | b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf]; | ||
1438 : | b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf]; | ||
1439 : | |||
1440 : | sum = sad8bi(pCur->u + 8*x + 8*y*(Data->iEdgedWidth/2), | ||
1441 : | f_Ref->u + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, | ||
1442 : | b_Ref->u + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, | ||
1443 : | Data->iEdgedWidth/2); | ||
1444 : | syskin | 1.44.2.43 | |
1445 : | if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * quant) return; //no skip | ||
1446 : | |||
1447 : | syskin | 1.44.2.34 | sum += sad8bi(pCur->v + 8*x + 8*y*(Data->iEdgedWidth/2), |
1448 : | f_Ref->v + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, | ||
1449 : | b_Ref->v + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, | ||
1450 : | Data->iEdgedWidth/2); | ||
1451 : | |||
1452 : | syskin | 1.44.2.43 | if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * quant) pMB->mode = MODE_DIRECT_NONE_MV; //skipped |
1453 : | syskin | 1.44.2.34 | } |
1454 : | |||
1455 : | |||
1456 : | |||
1457 : | static __inline uint32_t | ||
1458 : | Isibaar | 1.44.2.2 | SearchDirect(const IMAGE * const f_Ref, |
1459 : | chl | 1.44.2.1 | const uint8_t * const f_RefH, |
1460 : | const uint8_t * const f_RefV, | ||
1461 : | const uint8_t * const f_RefHV, | ||
1462 : | Isibaar | 1.44.2.2 | const IMAGE * const b_Ref, |
1463 : | chl | 1.44.2.1 | const uint8_t * const b_RefH, |
1464 : | const uint8_t * const b_RefV, | ||
1465 : | const uint8_t * const b_RefHV, | ||
1466 : | const IMAGE * const pCur, | ||
1467 : | const int x, const int y, | ||
1468 : | const uint32_t MotionFlags, | ||
1469 : | const int32_t TRB, const int32_t TRD, | ||
1470 : | const MBParam * const pParam, | ||
1471 : | MACROBLOCK * const pMB, | ||
1472 : | const MACROBLOCK * const b_mb, | ||
1473 : | h | 1.44.2.3 | int32_t * const best_sad, |
1474 : | SearchData * const Data) | ||
1475 : | chl | 1.44.2.1 | |
1476 : | { | ||
1477 : | h | 1.44.2.3 | int32_t skip_sad; |
1478 : | chl | 1.44.2.1 | int k; |
1479 : | h | 1.44.2.3 | |
1480 : | chl | 1.44.2.1 | MainSearchFunc *MainSearchPtr; |
1481 : | |||
1482 : | h | 1.44.2.3 | *Data->iMinSAD = 256*4096; |
1483 : | chl | 1.44.2.1 | |
1484 : | h | 1.44.2.3 | Data->Ref = f_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
1485 : | Data->RefH = f_RefH + (x + Data->iEdgedWidth*y) * 16; | ||
1486 : | Data->RefV = f_RefV + (x + Data->iEdgedWidth*y) * 16; | ||
1487 : | Data->RefHV = f_RefHV + (x + Data->iEdgedWidth*y) * 16; | ||
1488 : | Data->bRef = b_Ref->y + (x + Data->iEdgedWidth*y) * 16; | ||
1489 : | Data->bRefH = b_RefH + (x + Data->iEdgedWidth*y) * 16; | ||
1490 : | Data->bRefV = b_RefV + (x + Data->iEdgedWidth*y) * 16; | ||
1491 : | Data->bRefHV = b_RefHV + (x + Data->iEdgedWidth*y) * 16; | ||
1492 : | |||
1493 : | Data->max_dx = 2 * pParam->width - 2 * (x) * 16; | ||
1494 : | Data->max_dy = 2 * pParam->height - 2 * (y) * 16; | ||
1495 : | Data->min_dx = -(2 * 16 + 2 * (x) * 16); | ||
1496 : | Data->min_dy = -(2 * 16 + 2 * (y) * 16); | ||
1497 : | syskin | 1.44.2.19 | if (Data->qpel) { //we measure in qpixels |
1498 : | Data->max_dx *= 2; | ||
1499 : | Data->max_dy *= 2; | ||
1500 : | Data->min_dx *= 2; | ||
1501 : | Data->min_dy *= 2; | ||
1502 : | Data->referencemv = b_mb->qmvs; | ||
1503 : | } else Data->referencemv = b_mb->mvs; | ||
1504 : | syskin | 1.44.2.26 | Data->qpel_precision = 0; // it's a trick. it's 1 not 0, but we need 0 here |
1505 : | chl | 1.44.2.1 | |
1506 : | for (k = 0; k < 4; k++) { | ||
1507 : | h | 1.44.2.3 | pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); |
1508 : | pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD; | ||
1509 : | pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); | ||
1510 : | pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; | ||
1511 : | |||
1512 : | if ( ( pMB->b_mvs[k].x > Data->max_dx ) || ( pMB->b_mvs[k].x < Data->min_dx ) | ||
1513 : | || ( pMB->b_mvs[k].y > Data->max_dy ) || ( pMB->b_mvs[k].y < Data->min_dy )) { | ||
1514 : | chl | 1.44.2.1 | |
1515 : | Isibaar | 1.44.2.2 | *best_sad = 256*4096; // in that case, we won't use direct mode |
1516 : | pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV" | ||
1517 : | pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; | ||
1518 : | syskin | 1.44.2.34 | return 256*4096; |
1519 : | Isibaar | 1.44.2.2 | } |
1520 : | if (b_mb->mode != MODE_INTER4V) { | ||
1521 : | pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; | ||
1522 : | pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0]; | ||
1523 : | h | 1.44.2.3 | Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0]; |
1524 : | Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0]; | ||
1525 : | Isibaar | 1.44.2.2 | break; |
1526 : | } | ||
1527 : | } | ||
1528 : | chl | 1.44.2.1 | |
1529 : | syskin | 1.44.2.22 | if (b_mb->mode == MODE_INTER4V) CheckCandidate = CheckCandidateDirect; |
1530 : | else CheckCandidate = CheckCandidateDirectno4v; | ||
1531 : | syskin | 1.44.2.19 | |
1532 : | h | 1.44.2.3 | (*CheckCandidate)(0, 0, 255, &k, Data); |
1533 : | Isibaar | 1.44.2.2 | |
1534 : | syskin | 1.44.2.34 | // initial (fast) skip decision |
1535 : | if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH*2) { | ||
1536 : | syskin | 1.44.2.43 | SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data); //possible skip - checking chroma |
1537 : | syskin | 1.44.2.34 | if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip. |
1538 : | Isibaar | 1.44.2.2 | } |
1539 : | |||
1540 : | h | 1.44.2.3 | skip_sad = *Data->iMinSAD; |
1541 : | Isibaar | 1.44.2.2 | |
1542 : | chl | 1.44.2.1 | // DIRECT MODE DELTA VECTOR SEARCH. |
1543 : | // This has to be made more effective, but at the moment I'm happy it's running at all | ||
1544 : | |||
1545 : | if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; | ||
1546 : | else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; | ||
1547 : | else MainSearchPtr = DiamondSearch; | ||
1548 : | |||
1549 : | h | 1.44.2.3 | (*MainSearchPtr)(0, 0, Data, 255); |
1550 : | chl | 1.44.2.1 | |
1551 : | syskin | 1.44.2.22 | SubpelRefine(Data); |
1552 : | chl | 1.44.2.1 | |
1553 : | h | 1.44.2.3 | *best_sad = *Data->iMinSAD; |
1554 : | chl | 1.44.2.1 | |
1555 : | syskin | 1.44.2.39 | if (b_mb->mode == MODE_INTER4V || Data->qpel) pMB->mode = MODE_DIRECT; |
1556 : | syskin | 1.44.2.26 | else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
1557 : | chl | 1.44.2.1 | |
1558 : | h | 1.44.2.3 | pMB->pmvs[3] = *Data->currentMV; |
1559 : | chl | 1.44.2.1 | |
1560 : | for (k = 0; k < 4; k++) { | ||
1561 : | h | 1.44.2.3 | pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x; |
1562 : | syskin | 1.44.2.19 | pMB->b_mvs[k].x = ( (Data->currentMV->x == 0) |
1563 : | h | 1.44.2.3 | ? Data->directmvB[k].x |
1564 : | syskin | 1.44.2.19 | :pMB->mvs[k].x - Data->referencemv[k].x); |
1565 : | h | 1.44.2.3 | pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y); |
1566 : | pMB->b_mvs[k].y = ((Data->currentMV->y == 0) | ||
1567 : | ? Data->directmvB[k].y | ||
1568 : | : pMB->mvs[k].y - Data->referencemv[k].y); | ||
1569 : | syskin | 1.44.2.19 | if (Data->qpel) { |
1570 : | pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2; | ||
1571 : | pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2; | ||
1572 : | pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2; | ||
1573 : | pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2; | ||
1574 : | } | ||
1575 : | |||
1576 : | chl | 1.44.2.1 | if (b_mb->mode != MODE_INTER4V) { |
1577 : | pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; | ||
1578 : | pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; | ||
1579 : | syskin | 1.44.2.19 | pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0]; |
1580 : | pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0]; | ||
1581 : | chl | 1.44.2.1 | break; |
1582 : | } | ||
1583 : | } | ||
1584 : | h | 1.44.2.3 | return skip_sad; |
1585 : | chl | 1.44.2.1 | } |
1586 : | |||
1587 : | h | 1.44.2.3 | |
1588 : | syskin | 1.44.2.34 | static void |
1589 : | chl | 1.44.2.1 | SearchInterpolate(const uint8_t * const f_Ref, |
1590 : | const uint8_t * const f_RefH, | ||
1591 : | const uint8_t * const f_RefV, | ||
1592 : | const uint8_t * const f_RefHV, | ||
1593 : | const uint8_t * const b_Ref, | ||
1594 : | const uint8_t * const b_RefH, | ||
1595 : | const uint8_t * const b_RefV, | ||
1596 : | const uint8_t * const b_RefHV, | ||
1597 : | const IMAGE * const pCur, | ||
1598 : | const int x, const int y, | ||
1599 : | const uint32_t fcode, | ||
1600 : | const uint32_t bcode, | ||
1601 : | const uint32_t MotionFlags, | ||
1602 : | const MBParam * const pParam, | ||
1603 : | const VECTOR * const f_predMV, | ||
1604 : | const VECTOR * const b_predMV, | ||
1605 : | MACROBLOCK * const pMB, | ||
1606 : | h | 1.44.2.3 | int32_t * const best_sad, |
1607 : | SearchData * const fData) | ||
1608 : | edgomez | 1.20 | |
1609 : | chl | 1.3 | { |
1610 : | |||
1611 : | chl | 1.44.2.1 | int iDirection, i, j; |
1612 : | h | 1.44.2.3 | SearchData bData; |
1613 : | edgomez | 1.20 | |
1614 : | syskin | 1.44.2.34 | fData->qpel_precision = 0; |
1615 : | memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data | ||
1616 : | *fData->iMinSAD = 4096*256; | ||
1617 : | syskin | 1.44.2.37 | bData.currentMV++; bData.currentQMV++; |
1618 : | h | 1.44.2.3 | fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
1619 : | |||
1620 : | syskin | 1.44.2.34 | i = (x + y * fData->iEdgedWidth) * 16; |
1621 : | bData.bRef = fData->Ref = f_Ref + i; | ||
1622 : | bData.bRefH = fData->RefH = f_RefH + i; | ||
1623 : | bData.bRefV = fData->RefV = f_RefV + i; | ||
1624 : | bData.bRefHV = fData->RefHV = f_RefHV + i; | ||
1625 : | bData.Ref = fData->bRef = b_Ref + i; | ||
1626 : | bData.RefH = fData->bRefH = b_RefH + i; | ||
1627 : | bData.RefV = fData->bRefV = b_RefV + i; | ||
1628 : | bData.RefHV = fData->bRefHV = b_RefHV + i; | ||
1629 : | h | 1.44.2.3 | |
1630 : | bData.bpredMV = fData->predMV = *f_predMV; | ||
1631 : | fData->bpredMV = bData.predMV = *b_predMV; | ||
1632 : | syskin | 1.44.2.20 | fData->currentMV[0] = fData->currentMV[2]; |
1633 : | syskin | 1.44.2.34 | |
1634 : | 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); | ||
1635 : | 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); | ||
1636 : | chl | 1.3 | |
1637 : | h | 1.44.2.3 | if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; |
1638 : | syskin | 1.44.2.27 | if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx; |
1639 : | if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy; | ||
1640 : | if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy; | ||
1641 : | h | 1.44.2.3 | |
1642 : | if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx; | ||
1643 : | syskin | 1.44.2.27 | if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx; |
1644 : | if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy; | ||
1645 : | if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy; | ||
1646 : | Isibaar | 1.44.2.2 | |
1647 : | h | 1.44.2.3 | CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
1648 : | chl | 1.3 | |
1649 : | chl | 1.44.2.1 | //diamond. I wish we could use normal mainsearch functions (square, advdiamond) |
1650 : | chl | 1.3 | |
1651 : | chl | 1.44.2.1 | do { |
1652 : | iDirection = 255; | ||
1653 : | // forward MV moves | ||
1654 : | h | 1.44.2.3 | i = fData->currentMV[0].x; j = fData->currentMV[0].y; |
1655 : | chl | 1.35 | |
1656 : | h | 1.44.2.3 | CheckCandidateInt(i + 1, j, 0, &iDirection, fData); |
1657 : | CheckCandidateInt(i, j + 1, 0, &iDirection, fData); | ||
1658 : | CheckCandidateInt(i - 1, j, 0, &iDirection, fData); | ||
1659 : | CheckCandidateInt(i, j - 1, 0, &iDirection, fData); | ||
1660 : | edgomez | 1.20 | |
1661 : | chl | 1.44.2.1 | // backward MV moves |
1662 : | h | 1.44.2.3 | i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
1663 : | fData->currentMV[2] = fData->currentMV[0]; | ||
1664 : | Isibaar | 1.44.2.2 | CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); |
1665 : | CheckCandidateInt(i, j + 1, 0, &iDirection, &bData); | ||
1666 : | CheckCandidateInt(i - 1, j, 0, &iDirection, &bData); | ||
1667 : | CheckCandidateInt(i, j - 1, 0, &iDirection, &bData); | ||
1668 : | chl | 1.3 | |
1669 : | chl | 1.44.2.1 | } while (!(iDirection)); |
1670 : | chl | 1.3 | |
1671 : | syskin | 1.44.2.19 | if (fData->qpel) { |
1672 : | syskin | 1.44.2.38 | if (*fData->iMinSAD > *best_sad + 500) return; |
1673 : | syskin | 1.44.2.27 | CheckCandidate = CheckCandidateInt; |
1674 : | syskin | 1.44.2.22 | fData->qpel_precision = bData.qpel_precision = 1; |
1675 : | syskin | 1.44.2.34 | get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0); |
1676 : | get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0); | ||
1677 : | syskin | 1.44.2.19 | fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x; |
1678 : | fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y; | ||
1679 : | fData->currentQMV[1].x = 2 * fData->currentMV[1].x; | ||
1680 : | fData->currentQMV[1].y = 2 * fData->currentMV[1].y; | ||
1681 : | syskin | 1.44.2.22 | SubpelRefine(fData); |
1682 : | syskin | 1.44.2.38 | if (*fData->iMinSAD > *best_sad + 300) return; |
1683 : | syskin | 1.44.2.19 | fData->currentQMV[2] = fData->currentQMV[0]; |
1684 : | syskin | 1.44.2.22 | SubpelRefine(&bData); |
1685 : | syskin | 1.44.2.19 | } |
1686 : | syskin | 1.44.2.28 | |
1687 : | syskin | 1.44.2.38 | *fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode. |
1688 : | syskin | 1.44.2.19 | |
1689 : | h | 1.44.2.3 | if (*fData->iMinSAD < *best_sad) { |
1690 : | *best_sad = *fData->iMinSAD; | ||
1691 : | pMB->mvs[0] = fData->currentMV[0]; | ||
1692 : | pMB->b_mvs[0] = fData->currentMV[1]; | ||
1693 : | chl | 1.44.2.1 | pMB->mode = MODE_INTERPOLATE; |
1694 : | syskin | 1.44.2.19 | if (fData->qpel) { |
1695 : | pMB->qmvs[0] = fData->currentQMV[0]; | ||
1696 : | pMB->b_qmvs[0] = fData->currentQMV[1]; | ||
1697 : | pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x; | ||
1698 : | pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y; | ||
1699 : | pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x; | ||
1700 : | pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y; | ||
1701 : | } else { | ||
1702 : | pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; | ||
1703 : | pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; | ||
1704 : | pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; | ||
1705 : | pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; | ||
1706 : | } | ||
1707 : | edgomez | 1.20 | } |
1708 : | chl | 1.3 | } |
1709 : | |||
1710 : | chl | 1.44.2.1 | void |
1711 : | MotionEstimationBVOP(MBParam * const pParam, | ||
1712 : | FRAMEINFO * const frame, | ||
1713 : | const int32_t time_bp, | ||
1714 : | const int32_t time_pp, | ||
1715 : | // forward (past) reference | ||
1716 : | const MACROBLOCK * const f_mbs, | ||
1717 : | const IMAGE * const f_ref, | ||
1718 : | const IMAGE * const f_refH, | ||
1719 : | const IMAGE * const f_refV, | ||
1720 : | const IMAGE * const f_refHV, | ||
1721 : | // backward (future) reference | ||
1722 : | syskin | 1.44.2.18 | const FRAMEINFO * const b_reference, |
1723 : | chl | 1.44.2.1 | const IMAGE * const b_ref, |
1724 : | const IMAGE * const b_refH, | ||
1725 : | const IMAGE * const b_refV, | ||
1726 : | const IMAGE * const b_refHV) | ||
1727 : | { | ||
1728 : | uint32_t i, j; | ||
1729 : | syskin | 1.44.2.34 | int32_t best_sad; |
1730 : | uint32_t skip_sad; | ||
1731 : | chl | 1.44.2.1 | int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; |
1732 : | static const VECTOR zeroMV={0,0}; | ||
1733 : | syskin | 1.44.2.18 | const MACROBLOCK * const b_mbs = b_reference->mbs; |
1734 : | chl | 1.3 | |
1735 : | chl | 1.44.2.1 | VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
1736 : | chl | 1.3 | |
1737 : | chl | 1.44.2.1 | const int32_t TRB = time_pp - time_bp; |
1738 : | const int32_t TRD = time_pp; | ||
1739 : | chl | 1.3 | |
1740 : | h | 1.44.2.3 | // some pre-inintialized data for the rest of the search |
1741 : | |||
1742 : | SearchData Data; | ||
1743 : | int32_t iMinSAD; | ||
1744 : | VECTOR currentMV[3]; | ||
1745 : | syskin | 1.44.2.19 | VECTOR currentQMV[3]; |
1746 : | syskin | 1.44.2.32 | memset(&Data, 0, sizeof(SearchData)); |
1747 : | h | 1.44.2.3 | Data.iEdgedWidth = pParam->edged_width; |
1748 : | syskin | 1.44.2.19 | Data.currentMV = currentMV; Data.currentQMV = currentQMV; |
1749 : | h | 1.44.2.3 | Data.iMinSAD = &iMinSAD; |
1750 : | syskin | 1.44.2.34 | Data.lambda16 = lambda_vec16[frame->quant]; |
1751 : | Data.chroma = frame->quant; | ||
1752 : | syskin | 1.44.2.19 | Data.qpel = pParam->m_quarterpel; |
1753 : | syskin | 1.44.2.22 | Data.rounding = 0; |
1754 : | h | 1.44.2.3 | |
1755 : | syskin | 1.44.2.42 | Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose) |
1756 : | syskin | 1.44.2.19 | // note: i==horizontal, j==vertical |
1757 : | chl | 1.44.2.1 | for (j = 0; j < pParam->mb_height; j++) { |
1758 : | chl | 1.3 | |
1759 : | chl | 1.44.2.1 | f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
1760 : | edgomez | 1.20 | |
1761 : | chl | 1.44.2.1 | for (i = 0; i < pParam->mb_width; i++) { |
1762 : | MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; | ||
1763 : | const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; | ||
1764 : | edgomez | 1.20 | |
1765 : | chl | 1.44.2.13 | /* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */ |
1766 : | syskin | 1.44.2.18 | if (b_reference->coding_type != S_VOP) |
1767 : | if (b_mb->mode == MODE_NOT_CODED) { | ||
1768 : | pMB->mode = MODE_NOT_CODED; | ||
1769 : | continue; | ||
1770 : | } | ||
1771 : | edgomez | 1.20 | |
1772 : | h | 1.44.2.3 | Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
1773 : | syskin | 1.44.2.11 | pMB->quant = frame->quant; |
1774 : | syskin | 1.44.2.14 | |
1775 : | chl | 1.44.2.1 | /* direct search comes first, because it (1) checks for SKIP-mode |
1776 : | and (2) sets very good predictions for forward and backward search */ | ||
1777 : | Isibaar | 1.44.2.2 | skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
1778 : | b_ref, b_refH->y, b_refV->y, b_refHV->y, | ||
1779 : | chl | 1.44.2.1 | &frame->image, |
1780 : | i, j, | ||
1781 : | frame->motion_flags, | ||
1782 : | TRB, TRD, | ||
1783 : | pParam, | ||
1784 : | pMB, b_mb, | ||
1785 : | h | 1.44.2.3 | &best_sad, |
1786 : | &Data); | ||
1787 : | edgomez | 1.20 | |
1788 : | Isibaar | 1.44.2.2 | if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
1789 : | syskin | 1.44.2.37 | |
1790 : | chl | 1.44.2.1 | // forward search |
1791 : | SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, | ||
1792 : | &frame->image, i, j, | ||
1793 : | frame->motion_flags, | ||
1794 : | h | 1.44.2.3 | frame->fcode, pParam, |
1795 : | chl | 1.44.2.1 | pMB, &f_predMV, &best_sad, |
1796 : | h | 1.44.2.3 | MODE_FORWARD, &Data); |
1797 : | chl | 1.10 | |
1798 : | chl | 1.44.2.1 | // backward search |
1799 : | SearchBF(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, | ||
1800 : | &frame->image, i, j, | ||
1801 : | frame->motion_flags, | ||
1802 : | h | 1.44.2.3 | frame->bcode, pParam, |
1803 : | chl | 1.44.2.1 | pMB, &b_predMV, &best_sad, |
1804 : | h | 1.44.2.3 | MODE_BACKWARD, &Data); |
1805 : | chl | 1.10 | |
1806 : | chl | 1.44.2.1 | // interpolate search comes last, because it uses data from forward and backward as prediction |
1807 : | SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, | ||
1808 : | b_ref->y, b_refH->y, b_refV->y, b_refHV->y, | ||
1809 : | &frame->image, | ||
1810 : | i, j, | ||
1811 : | frame->fcode, frame->bcode, | ||
1812 : | frame->motion_flags, | ||
1813 : | h | 1.44.2.3 | pParam, |
1814 : | chl | 1.44.2.1 | &f_predMV, &b_predMV, |
1815 : | h | 1.44.2.3 | pMB, &best_sad, |
1816 : | &Data); | ||
1817 : | syskin | 1.44.2.21 | |
1818 : | syskin | 1.44.2.34 | // final skip decision |
1819 : | if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP*2) | ||
1820 : | && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) | ||
1821 : | syskin | 1.44.2.43 | SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data); |
1822 : | syskin | 1.44.2.34 | |
1823 : | chl | 1.44.2.1 | switch (pMB->mode) { |
1824 : | case MODE_FORWARD: | ||
1825 : | f_count++; | ||
1826 : | syskin | 1.44.2.34 | if (Data.qpel) f_predMV = pMB->qmvs[0]; |
1827 : | syskin | 1.44.2.19 | else f_predMV = pMB->mvs[0]; |
1828 : | chl | 1.44.2.1 | break; |
1829 : | case MODE_BACKWARD: | ||
1830 : | b_count++; | ||
1831 : | syskin | 1.44.2.34 | if (Data.qpel) b_predMV = pMB->b_qmvs[0]; |
1832 : | syskin | 1.44.2.19 | else b_predMV = pMB->b_mvs[0]; |
1833 : | chl | 1.44.2.1 | break; |
1834 : | case MODE_INTERPOLATE: | ||
1835 : | i_count++; | ||
1836 : | syskin | 1.44.2.34 | if (Data.qpel) { |
1837 : | syskin | 1.44.2.19 | f_predMV = pMB->qmvs[0]; |
1838 : | b_predMV = pMB->b_qmvs[0]; | ||
1839 : | syskin | 1.44.2.40 | } else { |
1840 : | syskin | 1.44.2.19 | f_predMV = pMB->mvs[0]; |
1841 : | b_predMV = pMB->b_mvs[0]; | ||
1842 : | } | ||
1843 : | chl | 1.44.2.1 | break; |
1844 : | case MODE_DIRECT: | ||
1845 : | case MODE_DIRECT_NO4V: | ||
1846 : | d_count++; | ||
1847 : | default: | ||
1848 : | break; | ||
1849 : | chl | 1.10 | } |
1850 : | } | ||
1851 : | } | ||
1852 : | chl | 1.44.2.1 | } |
1853 : | chl | 1.3 | |
1854 : | syskin | 1.44.2.37 | static __inline void |
1855 : | Isibaar | 1.44.2.2 | MEanalyzeMB ( const uint8_t * const pRef, |
1856 : | const uint8_t * const pCur, | ||
1857 : | const int x, | ||
1858 : | const int y, | ||
1859 : | const MBParam * const pParam, | ||
1860 : | syskin | 1.44.2.37 | MACROBLOCK * const pMBs, |
1861 : | h | 1.44.2.3 | SearchData * const Data) |
1862 : | Isibaar | 1.44.2.2 | { |
1863 : | chl | 1.32 | |
1864 : | syskin | 1.44.2.37 | int i, mask; |
1865 : | h | 1.44.2.3 | VECTOR pmv[3]; |
1866 : | syskin | 1.44.2.37 | MACROBLOCK * pMB = &pMBs[x + y * pParam->mb_width]; |
1867 : | |||
1868 : | for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; | ||
1869 : | syskin | 1.44.2.17 | |
1870 : | //median is only used as prediction. it doesn't have to be real | ||
1871 : | if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; | ||
1872 : | else | ||
1873 : | if (x == 1) //left macroblock does not have any vector now | ||
1874 : | Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median | ||
1875 : | syskin | 1.44.2.40 | else if (y == 1) // top macroblock doesn't have it's vector |
1876 : | syskin | 1.44.2.17 | Data->predMV = (pMB - 1)->mvs[0]; // left instead of median |
1877 : | else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median | ||
1878 : | |||
1879 : | h | 1.44.2.3 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1880 : | syskin | 1.44.2.37 | pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, Data->rrv); |
1881 : | chl | 1.35 | |
1882 : | h | 1.44.2.3 | Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
1883 : | Data->Ref = pRef + (x + y * pParam->edged_width) * 16; | ||
1884 : | syskin | 1.44.2.37 | |
1885 : | Isibaar | 1.44.2.2 | pmv[1].x = EVEN(pMB->mvs[0].x); |
1886 : | pmv[1].y = EVEN(pMB->mvs[0].y); | ||
1887 : | syskin | 1.44.2.17 | pmv[2].x = EVEN(Data->predMV.x); |
1888 : | pmv[2].y = EVEN(Data->predMV.y); | ||
1889 : | pmv[0].x = pmv[0].y = 0; | ||
1890 : | |||
1891 : | syskin | 1.44.2.37 | CheckCandidate32I(0, 0, 255, &i, Data); |
1892 : | syskin | 1.44.2.17 | |
1893 : | syskin | 1.44.2.37 | if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) { |
1894 : | Isibaar | 1.44.2.2 | |
1895 : | syskin | 1.44.2.37 | if (!(mask = make_mask(pmv, 1))) |
1896 : | CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data); | ||
1897 : | if (!(mask = make_mask(pmv, 2))) | ||
1898 : | CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data); | ||
1899 : | |||
1900 : | if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) // diamond only if needed | ||
1901 : | DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); | ||
1902 : | |||
1903 : | for (i = 0; i < 4; i++) { | ||
1904 : | MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1) * pParam->mb_width)]; | ||
1905 : | MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i]; | ||
1906 : | MB->mode = MODE_INTER; | ||
1907 : | MB->sad16 = Data->iMinSAD[i+1]; | ||
1908 : | } | ||
1909 : | } | ||
1910 : | Isibaar | 1.44.2.2 | } |
1911 : | |||
1912 : | syskin | 1.44.2.37 | #define INTRA_BIAS 2500 |
1913 : | #define INTRA_THRESH 1500 | ||
1914 : | #define INTER_THRESH 1400 | ||
1915 : | Isibaar | 1.44.2.2 | |
1916 : | syskin | 1.44.2.18 | |
1917 : | Isibaar | 1.44.2.2 | int |
1918 : | MEanalysis( const IMAGE * const pRef, | ||
1919 : | syskin | 1.44.2.18 | FRAMEINFO * const Current, |
1920 : | Isibaar | 1.44.2.2 | MBParam * const pParam, |
1921 : | syskin | 1.44.2.18 | int maxIntra, //maximum number if non-I frames |
1922 : | int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame | ||
1923 : | int bCount) // number if B frames in a row | ||
1924 : | Isibaar | 1.44.2.2 | { |
1925 : | uint32_t x, y, intra = 0; | ||
1926 : | int sSAD = 0; | ||
1927 : | syskin | 1.44.2.18 | MACROBLOCK * const pMBs = Current->mbs; |
1928 : | const IMAGE * const pCurrent = &Current->image; | ||
1929 : | int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH; | ||
1930 : | syskin | 1.44.2.37 | const VECTOR zeroMV = {0,0}; |
1931 : | h | 1.44.2.3 | |
1932 : | syskin | 1.44.2.37 | int32_t iMinSAD[5], temp[5]; |
1933 : | VECTOR currentMV[5]; | ||
1934 : | h | 1.44.2.3 | SearchData Data; |
1935 : | Data.iEdgedWidth = pParam->edged_width; | ||
1936 : | syskin | 1.44.2.37 | Data.currentMV = currentMV; |
1937 : | Data.iMinSAD = iMinSAD; | ||
1938 : | syskin | 1.44.2.18 | Data.iFcode = Current->fcode; |
1939 : | syskin | 1.44.2.37 | Data.rrv = Current->global_flags & XVID_REDUCED; |
1940 : | Data.temp = temp; | ||
1941 : | CheckCandidate = CheckCandidate32I; | ||
1942 : | h | 1.44.2.3 | |
1943 : | syskin | 1.44.2.19 | if (intraCount < 10) // we're right after an I frame |
1944 : | IntraThresh += 4 * (intraCount - 10) * (intraCount - 10); | ||
1945 : | syskin | 1.44.2.18 | else |
1946 : | if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec | ||
1947 : | IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; | ||
1948 : | |||
1949 : | syskin | 1.44.2.26 | InterThresh += 400 * (1 - bCount); |
1950 : | syskin | 1.44.2.37 | if (InterThresh < 300) InterThresh = 300; |
1951 : | syskin | 1.44.2.18 | |
1952 : | Isibaar | 1.44.2.2 | if (sadInit) (*sadInit) (); |
1953 : | |||
1954 : | syskin | 1.44.2.37 | for (y = 1; y < pParam->mb_height-1; y+=2) { |
1955 : | for (x = 1; x < pParam->mb_width-1; x+=2) { | ||
1956 : | int i; | ||
1957 : | |||
1958 : | if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; | ||
1959 : | Isibaar | 1.44.2.2 | |
1960 : | syskin | 1.44.2.37 | MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data); |
1961 : | Isibaar | 1.44.2.2 | |
1962 : | syskin | 1.44.2.37 | for (i = 0; i < 4; i++) { |
1963 : | int dev; | ||
1964 : | MACROBLOCK *pMB = &pMBs[x+(i&1) + y+(i>>1) * pParam->mb_width]; | ||
1965 : | if (pMB->sad16 > IntraThresh) { | ||
1966 : | dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1))* pParam->edged_width) * 16, | ||
1967 : | pParam->edged_width); | ||
1968 : | if (dev + IntraThresh < pMB->sad16) { | ||
1969 : | pMB->mode = MODE_INTRA; | ||
1970 : | if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return I_VOP; | ||
1971 : | } | ||
1972 : | Isibaar | 1.44.2.2 | } |
1973 : | syskin | 1.44.2.37 | sSAD += pMB->sad16; |
1974 : | syskin | 1.44.2.18 | } |
1975 : | suxen_drol | 1.6 | } |
1976 : | } | ||
1977 : | syskin | 1.44.2.32 | sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); |
1978 : | syskin | 1.44.2.37 | if (sSAD > IntraThresh + INTRA_BIAS ) return I_VOP; |
1979 : | suxen_drol | 1.44.2.33 | if (sSAD > InterThresh ) return P_VOP; |
1980 : | Isibaar | 1.44.2.2 | emms(); |
1981 : | suxen_drol | 1.44.2.33 | return B_VOP; |
1982 : | syskin | 1.44.2.34 | |
1983 : | chl | 1.44.2.13 | } |
1984 : | |||
1985 : | syskin | 1.44.2.14 | static void |
1986 : | syskin | 1.44.2.40 | CheckGMC(int x, int y, const int dir, int * iDirection, |
1987 : | syskin | 1.44.2.14 | const MACROBLOCK * const pMBs, uint32_t * bestcount, VECTOR * GMC, |
1988 : | const MBParam * const pParam) | ||
1989 : | { | ||
1990 : | uint32_t mx, my, a, count = 0; | ||
1991 : | |||
1992 : | for (my = 1; my < pParam->mb_height-1; my++) | ||
1993 : | for (mx = 1; mx < pParam->mb_width-1; mx++) { | ||
1994 : | VECTOR mv; | ||
1995 : | const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; | ||
1996 : | if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) continue; | ||
1997 : | mv = pMB->mvs[0]; | ||
1998 : | a = ABS(mv.x - x) + ABS(mv.y - y); | ||
1999 : | if (a < 6) count += 6 - a; | ||
2000 : | } | ||
2001 : | |||
2002 : | if (count > *bestcount) { | ||
2003 : | *bestcount = count; | ||
2004 : | *iDirection = dir; | ||
2005 : | syskin | 1.44.2.40 | GMC->x = x; GMC->y = y; |
2006 : | syskin | 1.44.2.14 | } |
2007 : | } | ||
2008 : | chl | 1.44.2.13 | |
2009 : | |||
2010 : | suxen_drol | 1.44.2.35 | static VECTOR |
2011 : | chl | 1.44.2.13 | GlobalMotionEst(const MACROBLOCK * const pMBs, const MBParam * const pParam, const uint32_t iFcode) |
2012 : | { | ||
2013 : | |||
2014 : | syskin | 1.44.2.14 | uint32_t count, bestcount = 0; |
2015 : | chl | 1.44.2.13 | int x, y; |
2016 : | syskin | 1.44.2.14 | VECTOR gmc = {0,0}; |
2017 : | chl | 1.44.2.13 | int step, min_x, max_x, min_y, max_y; |
2018 : | uint32_t mx, my; | ||
2019 : | syskin | 1.44.2.14 | int iDirection, bDirection; |
2020 : | chl | 1.44.2.13 | |
2021 : | min_x = min_y = -32<<iFcode; | ||
2022 : | max_x = max_y = 32<<iFcode; | ||
2023 : | |||
2024 : | syskin | 1.44.2.14 | //step1: let's find a rough camera panning |
2025 : | chl | 1.44.2.13 | for (step = 32; step >= 2; step /= 2) { |
2026 : | bestcount = 0; | ||
2027 : | for (y = min_y; y <= max_y; y += step) | ||
2028 : | for (x = min_x ; x <= max_x; x += step) { | ||
2029 : | count = 0; | ||
2030 : | //for all macroblocks | ||
2031 : | for (my = 1; my < pParam->mb_height-1; my++) | ||
2032 : | for (mx = 1; mx < pParam->mb_width-1; mx++) { | ||
2033 : | const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; | ||
2034 : | VECTOR mv; | ||
2035 : | |||
2036 : | syskin | 1.44.2.40 | if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) |
2037 : | chl | 1.44.2.13 | continue; |
2038 : | |||
2039 : | mv = pMB->mvs[0]; | ||
2040 : | if ( ABS(mv.x - x) <= step && ABS(mv.y - y) <= step ) /* GMC translation is always halfpel-res */ | ||
2041 : | count++; | ||
2042 : | } | ||
2043 : | if (count >= bestcount) { bestcount = count; gmc.x = x; gmc.y = y; } | ||
2044 : | } | ||
2045 : | min_x = gmc.x - step; | ||
2046 : | max_x = gmc.x + step; | ||
2047 : | min_y = gmc.y - step; | ||
2048 : | max_y = gmc.y + step; | ||
2049 : | } | ||
2050 : | syskin | 1.44.2.14 | |
2051 : | syskin | 1.44.2.40 | if (bestcount < (pParam->mb_height-2)*(pParam->mb_width-2)/10) |
2052 : | syskin | 1.44.2.14 | gmc.x = gmc.y = 0; //no camara pan, no GMC |
2053 : | |||
2054 : | syskin | 1.44.2.40 | // step2: let's refine camera panning using gradiend-descent approach |
2055 : | syskin | 1.44.2.14 | // TODO: more warping points may be evaluated here (like in interpolate mode search - two vectors in one diamond) |
2056 : | bestcount = 0; | ||
2057 : | CheckGMC(gmc.x, gmc.y, 255, &iDirection, pMBs, &bestcount, &gmc, pParam); | ||
2058 : | do { | ||
2059 : | x = gmc.x; y = gmc.y; | ||
2060 : | bDirection = iDirection; iDirection = 0; | ||
2061 : | if (bDirection & 1) CheckGMC(x - 1, y, 1+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); | ||
2062 : | if (bDirection & 2) CheckGMC(x + 1, y, 2+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); | ||
2063 : | if (bDirection & 4) CheckGMC(x, y - 1, 1+2+4, &iDirection, pMBs, &bestcount, &gmc, pParam); | ||
2064 : | if (bDirection & 8) CheckGMC(x, y + 1, 1+2+8, &iDirection, pMBs, &bestcount, &gmc, pParam); | ||
2065 : | |||
2066 : | } while (iDirection); | ||
2067 : | |||
2068 : | chl | 1.44.2.13 | if (pParam->m_quarterpel) { |
2069 : | gmc.x *= 2; | ||
2070 : | gmc.y *= 2; /* we store the halfpel value as pseudo-qpel to make comparison easier */ | ||
2071 : | } | ||
2072 : | |||
2073 : | return gmc; | ||
2074 : | Isibaar | 1.44.2.2 | } |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |