Parent Directory | Revision Log
Revision 1137 - (view) (download)
1 : | edgomez | 1054 | /***************************************************************************** |
2 : | Isibaar | 3 | * |
3 : | edgomez | 1054 | * XVID MPEG-4 VIDEO CODEC |
4 : | * - Motion Estimation related code - | ||
5 : | chl | 259 | * |
6 : | edgomez | 1054 | * Copyright(C) 2002 Christoph Lampert <gruel@web.de> |
7 : | * 2002 Michael Militzer <michael@xvid.org> | ||
8 : | * 2002-2003 Radoslaw Czyz <xvid@syskin.cjb.net> | ||
9 : | chl | 259 | * |
10 : | edgomez | 1054 | * This program is free software ; you can redistribute it and/or modify |
11 : | * it under the terms of the GNU General Public License as published by | ||
12 : | * the Free Software Foundation ; either version 2 of the License, or | ||
13 : | * (at your option) any later version. | ||
14 : | chl | 259 | * |
15 : | edgomez | 1054 | * This program is distributed in the hope that it will be useful, |
16 : | * but WITHOUT ANY WARRANTY ; without even the implied warranty of | ||
17 : | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
18 : | * GNU General Public License for more details. | ||
19 : | chl | 259 | * |
20 : | edgomez | 1054 | * You should have received a copy of the GNU General Public License |
21 : | * along with this program ; if not, write to the Free Software | ||
22 : | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
23 : | chl | 259 | * |
24 : | Isibaar | 1137 | * $Id: motion_est.c,v 1.58.2.34 2003-09-05 10:01:50 Isibaar Exp $ |
25 : | edgomez | 1054 | * |
26 : | ****************************************************************************/ | ||
27 : | chl | 259 | |
28 : | Isibaar | 3 | #include <assert.h> |
29 : | #include <stdio.h> | ||
30 : | chl | 96 | #include <stdlib.h> |
31 : | edgomez | 1053 | #include <string.h> /* memcpy */ |
32 : | #include <math.h> /* lrint */ | ||
33 : | Isibaar | 3 | |
34 : | #include "../encoder.h" | ||
35 : | #include "../utils/mbfunctions.h" | ||
36 : | #include "../prediction/mbprediction.h" | ||
37 : | syskin | 1084 | #include "../global.h" |
38 : | Isibaar | 3 | #include "../utils/timer.h" |
39 : | edgomez | 851 | #include "../image/interpolate8x8.h" |
40 : | #include "motion_est.h" | ||
41 : | suxen_drol | 118 | #include "motion.h" |
42 : | Isibaar | 3 | #include "sad.h" |
43 : | chl | 1077 | #include "gmc.h" |
44 : | edgomez | 851 | #include "../utils/emms.h" |
45 : | #include "../dct/fdct.h" | ||
46 : | Isibaar | 3 | |
47 : | edgomez | 872 | /***************************************************************************** |
48 : | * Modified rounding tables -- declared in motion.h | ||
49 : | * Original tables see ISO spec tables 7-6 -> 7-9 | ||
50 : | ****************************************************************************/ | ||
51 : | |||
52 : | const uint32_t roundtab[16] = | ||
53 : | {0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2 }; | ||
54 : | |||
55 : | /* K = 4 */ | ||
56 : | const uint32_t roundtab_76[16] = | ||
57 : | { 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1 }; | ||
58 : | |||
59 : | /* K = 2 */ | ||
60 : | const uint32_t roundtab_78[8] = | ||
61 : | { 0, 0, 1, 1, 0, 0, 0, 1 }; | ||
62 : | |||
63 : | /* K = 1 */ | ||
64 : | const uint32_t roundtab_79[4] = | ||
65 : | { 0, 1, 0, 0 }; | ||
66 : | |||
67 : | edgomez | 851 | #define INITIAL_SKIP_THRESH (10) |
68 : | #define FINAL_SKIP_THRESH (50) | ||
69 : | #define MAX_SAD00_FOR_SKIP (20) | ||
70 : | #define MAX_CHROMA_SAD_FOR_SKIP (22) | ||
71 : | Isibaar | 3 | |
72 : | edgomez | 851 | #define CHECK_CANDIDATE(X,Y,D) { \ |
73 : | syskin | 1132 | CheckCandidate((X),(Y), data, (D) ); } |
74 : | Isibaar | 3 | |
75 : | syskin | 1084 | |
76 : | edgomez | 872 | /***************************************************************************** |
77 : | * Code | ||
78 : | ****************************************************************************/ | ||
79 : | |||
80 : | edgomez | 851 | static __inline uint32_t |
81 : | d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv) | ||
82 : | { | ||
83 : | edgomez | 974 | int bits; |
84 : | const int q = (1 << (iFcode - 1)) - 1; | ||
85 : | |||
86 : | x <<= qpel; | ||
87 : | y <<= qpel; | ||
88 : | edgomez | 851 | if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); } |
89 : | chl | 141 | |
90 : | edgomez | 851 | x -= pred.x; |
91 : | edgomez | 974 | bits = (x != 0 ? iFcode:0); |
92 : | x = abs(x); | ||
93 : | x += q; | ||
94 : | x >>= (iFcode - 1); | ||
95 : | bits += mvtab[x]; | ||
96 : | |||
97 : | edgomez | 851 | y -= pred.y; |
98 : | edgomez | 974 | bits += (y != 0 ? iFcode:0); |
99 : | y = abs(y); | ||
100 : | y += q; | ||
101 : | y >>= (iFcode - 1); | ||
102 : | bits += mvtab[y]; | ||
103 : | chl | 141 | |
104 : | edgomez | 974 | return bits; |
105 : | edgomez | 851 | } |
106 : | chl | 141 | |
107 : | edgomez | 978 | static int32_t ChromaSAD2(const int fx, const int fy, const int bx, const int by, |
108 : | const SearchData * const data) | ||
109 : | edgomez | 851 | { |
110 : | int sad; | ||
111 : | const uint32_t stride = data->iEdgedWidth/2; | ||
112 : | syskin | 1129 | uint8_t *f_refu, *f_refv, *b_refu, *b_refv; |
113 : | |||
114 : | const INTERPOLATE8X8_PTR interpolate8x8_halfpel[] = { | ||
115 : | NULL, | ||
116 : | interpolate8x8_halfpel_v, | ||
117 : | interpolate8x8_halfpel_h, | ||
118 : | interpolate8x8_halfpel_hv | ||
119 : | }; | ||
120 : | |||
121 : | edgomez | 978 | int offset = (fx>>1) + (fy>>1)*stride; |
122 : | syskin | 1129 | int filter = ((fx & 1) << 1) | (fy & 1); |
123 : | Isibaar | 3 | |
124 : | syskin | 1129 | if (filter != 0) { |
125 : | f_refu = data->RefQ; | ||
126 : | f_refv = data->RefQ + 8; | ||
127 : | interpolate8x8_halfpel[filter](f_refu, data->RefP[4] + offset, stride, data->rounding); | ||
128 : | interpolate8x8_halfpel[filter](f_refv, data->RefP[5] + offset, stride, data->rounding); | ||
129 : | } else { | ||
130 : | f_refu = (uint8_t*)data->RefP[4] + offset; | ||
131 : | f_refv = (uint8_t*)data->RefP[5] + offset; | ||
132 : | edgomez | 851 | } |
133 : | Isibaar | 3 | |
134 : | edgomez | 978 | offset = (bx>>1) + (by>>1)*stride; |
135 : | syskin | 1129 | filter = ((bx & 1) << 1) | (by & 1); |
136 : | |||
137 : | if (filter != 0) { | ||
138 : | b_refu = data->RefQ + 16; | ||
139 : | b_refv = data->RefQ + 24; | ||
140 : | interpolate8x8_halfpel[filter](b_refu, data->b_RefP[4] + offset, stride, data->rounding); | ||
141 : | interpolate8x8_halfpel[filter](b_refv, data->b_RefP[5] + offset, stride, data->rounding); | ||
142 : | } else { | ||
143 : | b_refu = (uint8_t*)data->b_RefP[4] + offset; | ||
144 : | b_refv = (uint8_t*)data->b_RefP[5] + offset; | ||
145 : | edgomez | 851 | } |
146 : | Isibaar | 3 | |
147 : | edgomez | 851 | sad = sad8bi(data->CurU, b_refu, f_refu, stride); |
148 : | sad += sad8bi(data->CurV, b_refv, f_refv, stride); | ||
149 : | Isibaar | 3 | |
150 : | edgomez | 851 | return sad; |
151 : | } | ||
152 : | Isibaar | 3 | |
153 : | edgomez | 851 | static int32_t |
154 : | edgomez | 978 | ChromaSAD(const int dx, const int dy, const SearchData * const data) |
155 : | edgomez | 851 | { |
156 : | int sad; | ||
157 : | const uint32_t stride = data->iEdgedWidth/2; | ||
158 : | edgomez | 978 | int offset = (dx>>1) + (dy>>1)*stride; |
159 : | syskin | 1132 | int next = 1; |
160 : | Isibaar | 3 | |
161 : | edgomez | 1053 | if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; /* it has been checked recently */ |
162 : | data->temp[5] = dx; data->temp[6] = dy; /* backup */ | ||
163 : | Isibaar | 3 | |
164 : | edgomez | 851 | switch (((dx & 1) << 1) | (dy & 1)) { |
165 : | case 0: | ||
166 : | edgomez | 978 | sad = sad8(data->CurU, data->RefP[4] + offset, stride); |
167 : | sad += sad8(data->CurV, data->RefP[5] + offset, stride); | ||
168 : | edgomez | 851 | break; |
169 : | case 1: | ||
170 : | syskin | 1132 | next = stride; |
171 : | edgomez | 851 | case 2: |
172 : | syskin | 1132 | sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + next, stride); |
173 : | sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + next, stride); | ||
174 : | edgomez | 851 | break; |
175 : | default: | ||
176 : | edgomez | 978 | interpolate8x8_halfpel_hv(data->RefQ, data->RefP[4] + offset, stride, data->rounding); |
177 : | edgomez | 851 | sad = sad8(data->CurU, data->RefQ, stride); |
178 : | |||
179 : | edgomez | 978 | interpolate8x8_halfpel_hv(data->RefQ, data->RefP[5] + offset, stride, data->rounding); |
180 : | edgomez | 851 | sad += sad8(data->CurV, data->RefQ, stride); |
181 : | break; | ||
182 : | } | ||
183 : | edgomez | 1053 | data->temp[7] = sad; /* backup, part 2 */ |
184 : | edgomez | 851 | return sad; |
185 : | Isibaar | 3 | } |
186 : | |||
187 : | edgomez | 851 | static __inline const uint8_t * |
188 : | GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data) | ||
189 : | Isibaar | 3 | { |
190 : | edgomez | 1053 | /* dir : 0 = forward, 1 = backward */ |
191 : | edgomez | 987 | const uint8_t *const *const direction = ( dir == 0 ? data->RefP : data->b_RefP ); |
192 : | edgomez | 978 | const int picture = ((x&1)<<1) | (y&1); |
193 : | const int offset = (x>>1) + (y>>1)*data->iEdgedWidth; | ||
194 : | return direction[picture] + offset; | ||
195 : | Isibaar | 3 | } |
196 : | |||
197 : | edgomez | 1053 | /* this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate */ |
198 : | edgomez | 851 | static __inline const uint8_t * |
199 : | GetReference(const int x, const int y, const SearchData * const data) | ||
200 : | Isibaar | 3 | { |
201 : | edgomez | 978 | const int picture = ((x&1)<<1) | (y&1); |
202 : | const int offset = (x>>1) + (y>>1)*data->iEdgedWidth; | ||
203 : | return data->RefP[picture] + offset; | ||
204 : | Isibaar | 3 | } |
205 : | edgomez | 851 | |
206 : | static uint8_t * | ||
207 : | Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data) | ||
208 : | Isibaar | 3 | { |
209 : | edgomez | 1053 | /* create or find a qpel-precision reference picture; return pointer to it */ |
210 : | edgomez | 851 | uint8_t * Reference = data->RefQ + 16*dir; |
211 : | const uint32_t iEdgedWidth = data->iEdgedWidth; | ||
212 : | const uint32_t rounding = data->rounding; | ||
213 : | const int halfpel_x = x/2; | ||
214 : | const int halfpel_y = y/2; | ||
215 : | const uint8_t *ref1, *ref2, *ref3, *ref4; | ||
216 : | suxen_drol | 136 | |
217 : | edgomez | 851 | ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
218 : | ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; | ||
219 : | switch( ((x&1)<<1) + (y&1) ) { | ||
220 : | edgomez | 1053 | case 3: /* x and y in qpel resolution - the "corners" (top left/right and */ |
221 : | /* bottom left/right) during qpel refinement */ | ||
222 : | edgomez | 963 | ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
223 : | ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); | ||
224 : | ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); | ||
225 : | ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; | ||
226 : | ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; | ||
227 : | ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; | ||
228 : | interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); | ||
229 : | edgomez | 851 | break; |
230 : | chl | 184 | |
231 : | edgomez | 1053 | case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */ |
232 : | edgomez | 851 | ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
233 : | ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; | ||
234 : | interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); | ||
235 : | break; | ||
236 : | suxen_drol | 136 | |
237 : | edgomez | 1053 | case 2: /* x qpel, y halfpel - left or right during qpel refinement */ |
238 : | edgomez | 851 | ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
239 : | ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; | ||
240 : | interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); | ||
241 : | break; | ||
242 : | edgomez | 195 | |
243 : | edgomez | 1053 | default: /* pure halfpel position */ |
244 : | edgomez | 963 | return (uint8_t *) ref1; |
245 : | edgomez | 978 | |
246 : | edgomez | 851 | } |
247 : | return Reference; | ||
248 : | } | ||
249 : | Isibaar | 3 | |
250 : | edgomez | 851 | static uint8_t * |
251 : | Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data) | ||
252 : | { | ||
253 : | edgomez | 1053 | /* create or find a qpel-precision reference picture; return pointer to it */ |
254 : | edgomez | 851 | uint8_t * Reference = data->RefQ + 16*dir; |
255 : | const uint32_t iEdgedWidth = data->iEdgedWidth; | ||
256 : | const uint32_t rounding = data->rounding; | ||
257 : | const int halfpel_x = x/2; | ||
258 : | const int halfpel_y = y/2; | ||
259 : | const uint8_t *ref1, *ref2, *ref3, *ref4; | ||
260 : | chl | 326 | |
261 : | edgomez | 851 | ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
262 : | switch( ((x&1)<<1) + (y&1) ) { | ||
263 : | edgomez | 1053 | case 3: |
264 : | /* | ||
265 : | * x and y in qpel resolution - the "corners" (top left/right and | ||
266 : | * bottom left/right) during qpel refinement | ||
267 : | */ | ||
268 : | edgomez | 851 | ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
269 : | ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); | ||
270 : | ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); | ||
271 : | interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); | ||
272 : | interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); | ||
273 : | interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); | ||
274 : | interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); | ||
275 : | break; | ||
276 : | chl | 184 | |
277 : | edgomez | 1053 | case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */ |
278 : | edgomez | 851 | ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
279 : | interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); | ||
280 : | interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); | ||
281 : | interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); | ||
282 : | interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); | ||
283 : | break; | ||
284 : | chl | 184 | |
285 : | edgomez | 1053 | case 2: /* x qpel, y halfpel - left or right during qpel refinement */ |
286 : | edgomez | 851 | ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
287 : | interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); | ||
288 : | interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); | ||
289 : | interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); | ||
290 : | interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); | ||
291 : | break; | ||
292 : | edgomez | 195 | |
293 : | syskin | 1129 | |
294 : | edgomez | 1053 | default: /* pure halfpel position */ |
295 : | edgomez | 851 | return (uint8_t *) ref1; |
296 : | } | ||
297 : | return Reference; | ||
298 : | } | ||
299 : | edgomez | 195 | |
300 : | edgomez | 851 | /* CHECK_CANDIATE FUNCTIONS START */ |
301 : | edgomez | 195 | |
302 : | edgomez | 851 | static void |
303 : | syskin | 1132 | CheckCandidate16(const int x, const int y, const SearchData * const data, const int Direction) |
304 : | edgomez | 851 | { |
305 : | int xc, yc; | ||
306 : | const uint8_t * Reference; | ||
307 : | VECTOR * current; | ||
308 : | int32_t sad; uint32_t t; | ||
309 : | edgomez | 195 | |
310 : | edgomez | 851 | if ( (x > data->max_dx) || (x < data->min_dx) |
311 : | || (y > data->max_dy) || (y < data->min_dy) ) return; | ||
312 : | edgomez | 195 | |
313 : | edgomez | 851 | if (!data->qpel_precision) { |
314 : | Reference = GetReference(x, y, data); | ||
315 : | current = data->currentMV; | ||
316 : | xc = x; yc = y; | ||
317 : | edgomez | 1053 | } else { /* x and y are in 1/4 precision */ |
318 : | edgomez | 851 | Reference = Interpolate16x16qpel(x, y, 0, data); |
319 : | edgomez | 1053 | xc = x/2; yc = y/2; /* for chroma sad */ |
320 : | edgomez | 851 | current = data->currentQMV; |
321 : | } | ||
322 : | edgomez | 195 | |
323 : | syskin | 1132 | sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp); |
324 : | edgomez | 851 | t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
325 : | edgomez | 195 | |
326 : | edgomez | 851 | sad += (data->lambda16 * t * sad)>>10; |
327 : | syskin | 1132 | data->temp[0] += (data->lambda8 * t * (data->temp[0] + NEIGH_8X8_BIAS))>>10; |
328 : | chl | 175 | |
329 : | syskin | 1084 | if (data->chroma && sad < data->iMinSAD[0]) |
330 : | sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], | ||
331 : | (yc >> 1) + roundtab_79[yc & 0x3], data); | ||
332 : | edgomez | 195 | |
333 : | edgomez | 851 | if (sad < data->iMinSAD[0]) { |
334 : | data->iMinSAD[0] = sad; | ||
335 : | current[0].x = x; current[0].y = y; | ||
336 : | syskin | 1132 | *data->dir = Direction; |
337 : | edgomez | 851 | } |
338 : | edgomez | 195 | |
339 : | syskin | 1132 | if (data->temp[0] < data->iMinSAD[1]) { |
340 : | data->iMinSAD[1] = data->temp[0]; current[1].x = x; current[1].y = y; } | ||
341 : | if (data->temp[1] < data->iMinSAD[2]) { | ||
342 : | data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; } | ||
343 : | if (data->temp[2] < data->iMinSAD[3]) { | ||
344 : | data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; } | ||
345 : | if (data->temp[3] < data->iMinSAD[4]) { | ||
346 : | data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; } | ||
347 : | Isibaar | 3 | } |
348 : | |||
349 : | edgomez | 851 | static void |
350 : | Isibaar | 1136 | CheckCandidate16_subpel(const int x, const int y, const SearchData * const data, const int Direction) |
351 : | { | ||
352 : | int xc, yc; | ||
353 : | const uint8_t *Reference; | ||
354 : | VECTOR *current, *current2; | ||
355 : | int32_t sad; uint32_t t; | ||
356 : | |||
357 : | if ( (x > data->max_dx) || (x < data->min_dx) | ||
358 : | || (y > data->max_dy) || (y < data->min_dy) ) return; | ||
359 : | |||
360 : | if (!data->qpel_precision) { | ||
361 : | Reference = GetReference(x, y, data); | ||
362 : | current = data->currentMV; | ||
363 : | current2 = data->currentMV2; | ||
364 : | xc = x; yc = y; | ||
365 : | } else { /* x and y are in 1/4 precision */ | ||
366 : | Reference = Interpolate16x16qpel(x, y, 0, data); | ||
367 : | xc = x/2; yc = y/2; /* for chroma sad */ | ||
368 : | current = data->currentQMV; | ||
369 : | current2 = data->currentQMV2; | ||
370 : | } | ||
371 : | |||
372 : | sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp); | ||
373 : | t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); | ||
374 : | |||
375 : | sad += (data->lambda16 * t * sad)>>10; | ||
376 : | data->temp[0] += (data->lambda8 * t * (data->temp[0] + NEIGH_8X8_BIAS))>>10; | ||
377 : | |||
378 : | if (data->chroma && sad < data->iMinSAD[0]) | ||
379 : | sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], | ||
380 : | (yc >> 1) + roundtab_79[yc & 0x3], data); | ||
381 : | |||
382 : | if (data->temp[0] < data->iMinSAD[1]) { | ||
383 : | data->iMinSAD[1] = data->temp[0]; current[1].x = x; current[1].y = y; } | ||
384 : | if (data->temp[1] < data->iMinSAD[2]) { | ||
385 : | data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; } | ||
386 : | if (data->temp[2] < data->iMinSAD[3]) { | ||
387 : | data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; } | ||
388 : | if (data->temp[3] < data->iMinSAD[4]) { | ||
389 : | data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; } | ||
390 : | |||
391 : | if (sad < data->iMinSAD[0]) { | ||
392 : | *(data->iMinSAD2) = *(data->iMinSAD); | ||
393 : | current2->x = current->x; current2->y = current->y; | ||
394 : | |||
395 : | data->iMinSAD[0] = sad; | ||
396 : | current[0].x = x; current[0].y = y; | ||
397 : | *data->dir = Direction; | ||
398 : | return; | ||
399 : | } | ||
400 : | |||
401 : | if (sad < *(data->iMinSAD2)) { | ||
402 : | *(data->iMinSAD2) = sad; | ||
403 : | current2->x = x; current2->y = y; | ||
404 : | *data->dir = Direction; | ||
405 : | } | ||
406 : | } | ||
407 : | |||
408 : | static void | ||
409 : | syskin | 1132 | CheckCandidate8(const int x, const int y, const SearchData * const data, const int Direction) |
410 : | edgomez | 851 | { |
411 : | int32_t sad; uint32_t t; | ||
412 : | const uint8_t * Reference; | ||
413 : | edgomez | 974 | VECTOR * current; |
414 : | chl | 326 | |
415 : | edgomez | 851 | if ( (x > data->max_dx) || (x < data->min_dx) |
416 : | || (y > data->max_dy) || (y < data->min_dy) ) return; | ||
417 : | Isibaar | 3 | |
418 : | edgomez | 974 | if (!data->qpel_precision) { |
419 : | Reference = GetReference(x, y, data); | ||
420 : | current = data->currentMV; | ||
421 : | edgomez | 1053 | } else { /* x and y are in 1/4 precision */ |
422 : | edgomez | 974 | Reference = Interpolate8x8qpel(x, y, 0, 0, data); |
423 : | current = data->currentQMV; | ||
424 : | } | ||
425 : | Isibaar | 3 | |
426 : | edgomez | 851 | sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
427 : | t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); | ||
428 : | Isibaar | 3 | |
429 : | edgomez | 851 | sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10; |
430 : | Isibaar | 3 | |
431 : | edgomez | 851 | if (sad < *(data->iMinSAD)) { |
432 : | *(data->iMinSAD) = sad; | ||
433 : | edgomez | 974 | current->x = x; current->y = y; |
434 : | syskin | 1132 | *data->dir = Direction; |
435 : | edgomez | 851 | } |
436 : | Isibaar | 3 | } |
437 : | |||
438 : | edgomez | 851 | static void |
439 : | syskin | 1132 | CheckCandidate32(const int x, const int y, const SearchData * const data, const int Direction) |
440 : | edgomez | 851 | { |
441 : | uint32_t t; | ||
442 : | const uint8_t * Reference; | ||
443 : | syskin | 1132 | int sad; |
444 : | edgomez | 851 | |
445 : | edgomez | 1053 | if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || /* non-zero even value */ |
446 : | edgomez | 851 | (x > data->max_dx) || (x < data->min_dx) |
447 : | || (y > data->max_dy) || (y < data->min_dy) ) return; | ||
448 : | |||
449 : | Reference = GetReference(x, y, data); | ||
450 : | t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1); | ||
451 : | |||
452 : | syskin | 1132 | sad = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp); |
453 : | edgomez | 851 | |
454 : | syskin | 1132 | sad += (data->lambda16 * t * sad) >> 10; |
455 : | data->temp[0] += (data->lambda8 * t * (data->temp[0] + NEIGH_8X8_BIAS))>>10; | ||
456 : | edgomez | 851 | |
457 : | syskin | 1132 | if (sad < data->iMinSAD[0]) { |
458 : | data->iMinSAD[0] = sad; | ||
459 : | edgomez | 851 | data->currentMV[0].x = x; data->currentMV[0].y = y; |
460 : | syskin | 1132 | *data->dir = Direction; |
461 : | } | ||
462 : | edgomez | 851 | |
463 : | syskin | 1132 | if (data->temp[0] < data->iMinSAD[1]) { |
464 : | data->iMinSAD[1] = data->temp[0]; data->currentMV[1].x = x; data->currentMV[1].y = y; } | ||
465 : | if (data->temp[1] < data->iMinSAD[2]) { | ||
466 : | data->iMinSAD[2] = data->temp[1]; data->currentMV[2].x = x; data->currentMV[2].y = y; } | ||
467 : | if (data->temp[2] < data->iMinSAD[3]) { | ||
468 : | data->iMinSAD[3] = data->temp[2]; data->currentMV[3].x = x; data->currentMV[3].y = y; } | ||
469 : | if (data->temp[3] < data->iMinSAD[4]) { | ||
470 : | data->iMinSAD[4] = data->temp[3]; data->currentMV[4].x = x; data->currentMV[4].y = y; } | ||
471 : | Isibaar | 3 | } |
472 : | edgomez | 195 | |
473 : | edgomez | 851 | static void |
474 : | syskin | 1132 | CheckCandidate16no4v(const int x, const int y, const SearchData * const data, const int Direction) |
475 : | edgomez | 851 | { |
476 : | int32_t sad, xc, yc; | ||
477 : | const uint8_t * Reference; | ||
478 : | uint32_t t; | ||
479 : | VECTOR * current; | ||
480 : | |||
481 : | edgomez | 963 | if ( (x > data->max_dx) || ( x < data->min_dx) |
482 : | || (y > data->max_dy) || (y < data->min_dy) ) return; | ||
483 : | edgomez | 851 | |
484 : | edgomez | 1053 | if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; /* non-zero even value */ |
485 : | edgomez | 851 | |
486 : | edgomez | 1053 | if (data->qpel_precision) { /* x and y are in 1/4 precision */ |
487 : | edgomez | 851 | Reference = Interpolate16x16qpel(x, y, 0, data); |
488 : | current = data->currentQMV; | ||
489 : | xc = x/2; yc = y/2; | ||
490 : | } else { | ||
491 : | Reference = GetReference(x, y, data); | ||
492 : | current = data->currentMV; | ||
493 : | xc = x; yc = y; | ||
494 : | } | ||
495 : | t = d_mv_bits(x, y, data->predMV, data->iFcode, | ||
496 : | data->qpel^data->qpel_precision, data->rrv); | ||
497 : | |||
498 : | sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); | ||
499 : | sad += (data->lambda16 * t * sad)>>10; | ||
500 : | |||
501 : | syskin | 1084 | if (data->chroma && sad < *data->iMinSAD) |
502 : | sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], | ||
503 : | (yc >> 1) + roundtab_79[yc & 0x3], data); | ||
504 : | edgomez | 851 | |
505 : | if (sad < *(data->iMinSAD)) { | ||
506 : | *(data->iMinSAD) = sad; | ||
507 : | current->x = x; current->y = y; | ||
508 : | syskin | 1132 | *data->dir = Direction; |
509 : | edgomez | 851 | } |
510 : | chl | 96 | } |
511 : | Isibaar | 3 | |
512 : | edgomez | 851 | static void |
513 : | syskin | 1132 | CheckCandidate16I(const int x, const int y, const SearchData * const data, const int Direction) |
514 : | chl | 1077 | { |
515 : | int sad; | ||
516 : | // int xc, yc; | ||
517 : | const uint8_t * Reference; | ||
518 : | // VECTOR * current; | ||
519 : | |||
520 : | if ( (x > data->max_dx) || ( x < data->min_dx) | ||
521 : | || (y > data->max_dy) || (y < data->min_dy) ) return; | ||
522 : | |||
523 : | Reference = GetReference(x, y, data); | ||
524 : | // xc = x; yc = y; | ||
525 : | |||
526 : | sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); | ||
527 : | // sad += d_mv_bits(x, y, data->predMV, data->iFcode, 0, 0); | ||
528 : | |||
529 : | /* if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], | ||
530 : | (yc >> 1) + roundtab_79[yc & 0x3], data); | ||
531 : | */ | ||
532 : | |||
533 : | if (sad < data->iMinSAD[0]) { | ||
534 : | data->iMinSAD[0] = sad; | ||
535 : | data->currentMV[0].x = x; data->currentMV[0].y = y; | ||
536 : | syskin | 1132 | *data->dir = Direction; |
537 : | chl | 1077 | } |
538 : | } | ||
539 : | |||
540 : | static void | ||
541 : | syskin | 1132 | CheckCandidate32I(const int x, const int y, const SearchData * const data, const int Direction) |
542 : | edgomez | 851 | { |
543 : | edgomez | 1053 | /* maximum speed - for P/B/I decision */ |
544 : | edgomez | 851 | int32_t sad; |
545 : | |||
546 : | if ( (x > data->max_dx) || (x < data->min_dx) | ||
547 : | || (y > data->max_dy) || (y < data->min_dy) ) return; | ||
548 : | |||
549 : | syskin | 1135 | sad = sad32v_c(data->Cur, data->RefP[0] + x + y*((int)data->iEdgedWidth), |
550 : | syskin | 1132 | data->iEdgedWidth, data->temp); |
551 : | edgomez | 851 | |
552 : | if (sad < *(data->iMinSAD)) { | ||
553 : | *(data->iMinSAD) = sad; | ||
554 : | data->currentMV[0].x = x; data->currentMV[0].y = y; | ||
555 : | syskin | 1132 | *data->dir = Direction; |
556 : | edgomez | 851 | } |
557 : | syskin | 1132 | if (data->temp[0] < data->iMinSAD[1]) { |
558 : | data->iMinSAD[1] = data->temp[0]; data->currentMV[1].x = x; data->currentMV[1].y = y; } | ||
559 : | if (data->temp[1] < data->iMinSAD[2]) { | ||
560 : | data->iMinSAD[2] = data->temp[1]; data->currentMV[2].x = x; data->currentMV[2].y = y; } | ||
561 : | if (data->temp[2] < data->iMinSAD[3]) { | ||
562 : | data->iMinSAD[3] = data->temp[2]; data->currentMV[3].x = x; data->currentMV[3].y = y; } | ||
563 : | if (data->temp[3] < data->iMinSAD[4]) { | ||
564 : | data->iMinSAD[4] = data->temp[3]; data->currentMV[4].x = x; data->currentMV[4].y = y; } | ||
565 : | edgomez | 851 | |
566 : | Isibaar | 3 | } |
567 : | |||
568 : | edgomez | 851 | static void |
569 : | syskin | 1132 | CheckCandidateInt(const int xf, const int yf, const SearchData * const data, const int Direction) |
570 : | edgomez | 851 | { |
571 : | int32_t sad, xb, yb, xcf, ycf, xcb, ycb; | ||
572 : | uint32_t t; | ||
573 : | const uint8_t *ReferenceF, *ReferenceB; | ||
574 : | VECTOR *current; | ||
575 : | |||
576 : | edgomez | 953 | if ((xf > data->max_dx) || (xf < data->min_dx) || |
577 : | (yf > data->max_dy) || (yf < data->min_dy)) | ||
578 : | return; | ||
579 : | edgomez | 851 | |
580 : | if (!data->qpel_precision) { | ||
581 : | ReferenceF = GetReference(xf, yf, data); | ||
582 : | xb = data->currentMV[1].x; yb = data->currentMV[1].y; | ||
583 : | ReferenceB = GetReferenceB(xb, yb, 1, data); | ||
584 : | current = data->currentMV; | ||
585 : | xcf = xf; ycf = yf; | ||
586 : | xcb = xb; ycb = yb; | ||
587 : | } else { | ||
588 : | ReferenceF = Interpolate16x16qpel(xf, yf, 0, data); | ||
589 : | xb = data->currentQMV[1].x; yb = data->currentQMV[1].y; | ||
590 : | current = data->currentQMV; | ||
591 : | ReferenceB = Interpolate16x16qpel(xb, yb, 1, data); | ||
592 : | xcf = xf/2; ycf = yf/2; | ||
593 : | xcb = xb/2; ycb = yb/2; | ||
594 : | } | ||
595 : | |||
596 : | t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0) | ||
597 : | + d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0); | ||
598 : | |||
599 : | sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); | ||
600 : | sad += (data->lambda16 * t * sad)>>10; | ||
601 : | |||
602 : | syskin | 1084 | if (data->chroma && sad < *data->iMinSAD) |
603 : | sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3], | ||
604 : | (ycf >> 1) + roundtab_79[ycf & 0x3], | ||
605 : | (xcb >> 1) + roundtab_79[xcb & 0x3], | ||
606 : | (ycb >> 1) + roundtab_79[ycb & 0x3], data); | ||
607 : | edgomez | 851 | |
608 : | if (sad < *(data->iMinSAD)) { | ||
609 : | *(data->iMinSAD) = sad; | ||
610 : | current->x = xf; current->y = yf; | ||
611 : | syskin | 1132 | *data->dir = Direction; |
612 : | edgomez | 851 | } |
613 : | Isibaar | 3 | } |
614 : | |||
615 : | edgomez | 851 | static void |
616 : | syskin | 1132 | CheckCandidateDirect(const int x, const int y, const SearchData * const data, const int Direction) |
617 : | edgomez | 851 | { |
618 : | int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0; | ||
619 : | uint32_t k; | ||
620 : | const uint8_t *ReferenceF; | ||
621 : | const uint8_t *ReferenceB; | ||
622 : | VECTOR mvs, b_mvs; | ||
623 : | |||
624 : | edgomez | 953 | if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
625 : | edgomez | 851 | |
626 : | for (k = 0; k < 4; k++) { | ||
627 : | mvs.x = data->directmvF[k].x + x; | ||
628 : | b_mvs.x = ((x == 0) ? | ||
629 : | data->directmvB[k].x | ||
630 : | : mvs.x - data->referencemv[k].x); | ||
631 : | |||
632 : | mvs.y = data->directmvF[k].y + y; | ||
633 : | b_mvs.y = ((y == 0) ? | ||
634 : | data->directmvB[k].y | ||
635 : | : mvs.y - data->referencemv[k].y); | ||
636 : | |||
637 : | edgomez | 953 | if ((mvs.x > data->max_dx) || (mvs.x < data->min_dx) || |
638 : | (mvs.y > data->max_dy) || (mvs.y < data->min_dy) || | ||
639 : | (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx) || | ||
640 : | (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) ) | ||
641 : | return; | ||
642 : | edgomez | 851 | |
643 : | if (data->qpel) { | ||
644 : | xcf += mvs.x/2; ycf += mvs.y/2; | ||
645 : | xcb += b_mvs.x/2; ycb += b_mvs.y/2; | ||
646 : | } else { | ||
647 : | xcf += mvs.x; ycf += mvs.y; | ||
648 : | xcb += b_mvs.x; ycb += b_mvs.y; | ||
649 : | edgomez | 1053 | mvs.x *= 2; mvs.y *= 2; /* we move to qpel precision anyway */ |
650 : | edgomez | 851 | b_mvs.x *= 2; b_mvs.y *= 2; |
651 : | } | ||
652 : | |||
653 : | ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); | ||
654 : | ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); | ||
655 : | |||
656 : | sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), | ||
657 : | ReferenceF, ReferenceB, data->iEdgedWidth); | ||
658 : | if (sad > *(data->iMinSAD)) return; | ||
659 : | } | ||
660 : | |||
661 : | sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; | ||
662 : | |||
663 : | syskin | 1084 | if (data->chroma && sad < *data->iMinSAD) |
664 : | sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], | ||
665 : | (ycf >> 3) + roundtab_76[ycf & 0xf], | ||
666 : | (xcb >> 3) + roundtab_76[xcb & 0xf], | ||
667 : | (ycb >> 3) + roundtab_76[ycb & 0xf], data); | ||
668 : | edgomez | 851 | |
669 : | if (sad < *(data->iMinSAD)) { | ||
670 : | *(data->iMinSAD) = sad; | ||
671 : | data->currentMV->x = x; data->currentMV->y = y; | ||
672 : | syskin | 1132 | *data->dir = Direction; |
673 : | edgomez | 851 | } |
674 : | Isibaar | 3 | } |
675 : | |||
676 : | edgomez | 851 | static void |
677 : | syskin | 1132 | CheckCandidateDirectno4v(const int x, const int y, const SearchData * const data, const int Direction) |
678 : | Isibaar | 3 | { |
679 : | edgomez | 851 | int32_t sad, xcf, ycf, xcb, ycb; |
680 : | const uint8_t *ReferenceF; | ||
681 : | const uint8_t *ReferenceB; | ||
682 : | VECTOR mvs, b_mvs; | ||
683 : | Isibaar | 3 | |
684 : | edgomez | 963 | if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
685 : | Isibaar | 3 | |
686 : | edgomez | 851 | mvs.x = data->directmvF[0].x + x; |
687 : | b_mvs.x = ((x == 0) ? | ||
688 : | data->directmvB[0].x | ||
689 : | : mvs.x - data->referencemv[0].x); | ||
690 : | Isibaar | 3 | |
691 : | edgomez | 851 | mvs.y = data->directmvF[0].y + y; |
692 : | b_mvs.y = ((y == 0) ? | ||
693 : | data->directmvB[0].y | ||
694 : | : mvs.y - data->referencemv[0].y); | ||
695 : | Isibaar | 3 | |
696 : | edgomez | 963 | if ( (mvs.x > data->max_dx) || (mvs.x < data->min_dx) |
697 : | || (mvs.y > data->max_dy) || (mvs.y < data->min_dy) | ||
698 : | || (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx) | ||
699 : | || (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) ) return; | ||
700 : | Isibaar | 3 | |
701 : | edgomez | 851 | if (data->qpel) { |
702 : | xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2); | ||
703 : | xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2); | ||
704 : | ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data); | ||
705 : | ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); | ||
706 : | } else { | ||
707 : | xcf = 4*mvs.x; ycf = 4*mvs.y; | ||
708 : | xcb = 4*b_mvs.x; ycb = 4*b_mvs.y; | ||
709 : | ReferenceF = GetReference(mvs.x, mvs.y, data); | ||
710 : | ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data); | ||
711 : | } | ||
712 : | |||
713 : | sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); | ||
714 : | sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; | ||
715 : | |||
716 : | syskin | 1084 | if (data->chroma && sad < *data->iMinSAD) |
717 : | sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], | ||
718 : | (ycf >> 3) + roundtab_76[ycf & 0xf], | ||
719 : | (xcb >> 3) + roundtab_76[xcb & 0xf], | ||
720 : | (ycb >> 3) + roundtab_76[ycb & 0xf], data); | ||
721 : | edgomez | 851 | |
722 : | if (sad < *(data->iMinSAD)) { | ||
723 : | *(data->iMinSAD) = sad; | ||
724 : | data->currentMV->x = x; data->currentMV->y = y; | ||
725 : | syskin | 1132 | *data->dir = Direction; |
726 : | edgomez | 851 | } |
727 : | Isibaar | 3 | } |
728 : | |||
729 : | edgomez | 851 | |
730 : | static void | ||
731 : | syskin | 1132 | CheckCandidateRD16(const int x, const int y, const SearchData * const data, const int Direction) |
732 : | Isibaar | 3 | { |
733 : | |||
734 : | edgomez | 962 | int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; |
735 : | syskin | 1108 | int32_t rd = 0; |
736 : | edgomez | 851 | VECTOR * current; |
737 : | const uint8_t * ptr; | ||
738 : | int i, cbp = 0, t, xc, yc; | ||
739 : | edgomez | 195 | |
740 : | edgomez | 851 | if ( (x > data->max_dx) || (x < data->min_dx) |
741 : | || (y > data->max_dy) || (y < data->min_dy) ) return; | ||
742 : | edgomez | 195 | |
743 : | edgomez | 851 | if (!data->qpel_precision) { |
744 : | ptr = GetReference(x, y, data); | ||
745 : | current = data->currentMV; | ||
746 : | xc = x; yc = y; | ||
747 : | edgomez | 1053 | } else { /* x and y are in 1/4 precision */ |
748 : | edgomez | 851 | ptr = Interpolate16x16qpel(x, y, 0, data); |
749 : | current = data->currentQMV; | ||
750 : | xc = x/2; yc = y/2; | ||
751 : | } | ||
752 : | Isibaar | 3 | |
753 : | edgomez | 851 | for(i = 0; i < 4; i++) { |
754 : | int s = 8*((i&1) + (i>>1)*data->iEdgedWidth); | ||
755 : | transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth); | ||
756 : | syskin | 1108 | rd += data->temp[i] = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, i); |
757 : | edgomez | 851 | } |
758 : | Isibaar | 3 | |
759 : | syskin | 1108 | rd += t = BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
760 : | edgomez | 195 | |
761 : | syskin | 1071 | if (data->temp[0] + t < data->iMinSAD[1]) { |
762 : | chl | 1115 | data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; data->cbp[1] = (data->cbp[1]&~32) | (cbp&32); } |
763 : | syskin | 1071 | if (data->temp[1] < data->iMinSAD[2]) { |
764 : | chl | 1115 | data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; data->cbp[1] = (data->cbp[1]&~16) | (cbp&16); } |
765 : | syskin | 1071 | if (data->temp[2] < data->iMinSAD[3]) { |
766 : | chl | 1115 | data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; data->cbp[1] = (data->cbp[1]&~8) | (cbp&8); } |
767 : | syskin | 1071 | if (data->temp[3] < data->iMinSAD[4]) { |
768 : | chl | 1115 | data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; data->cbp[1] = (data->cbp[1]&~4) | (cbp&4); } |
769 : | syskin | 1071 | |
770 : | syskin | 1108 | rd += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; |
771 : | edgomez | 851 | |
772 : | syskin | 1108 | if (rd >= data->iMinSAD[0]) return; |
773 : | edgomez | 851 | |
774 : | edgomez | 1053 | /* chroma */ |
775 : | edgomez | 1022 | xc = (xc >> 1) + roundtab_79[xc & 0x3]; |
776 : | yc = (yc >> 1) + roundtab_79[yc & 0x3]; | ||
777 : | edgomez | 851 | |
778 : | edgomez | 1053 | /* chroma U */ |
779 : | syskin | 1084 | ptr = interpolate8x8_switch2(data->RefQ, data->RefP[4], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); |
780 : | transfer_8to16subro(in, data->CurU, ptr, data->iEdgedWidth/2); | ||
781 : | syskin | 1108 | rd += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 4); |
782 : | if (rd >= data->iMinSAD[0]) return; | ||
783 : | edgomez | 1022 | |
784 : | edgomez | 1053 | /* chroma V */ |
785 : | syskin | 1084 | ptr = interpolate8x8_switch2(data->RefQ, data->RefP[5], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); |
786 : | transfer_8to16subro(in, data->CurV, ptr, data->iEdgedWidth/2); | ||
787 : | syskin | 1108 | rd += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5); |
788 : | edgomez | 1022 | |
789 : | syskin | 1108 | rd += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; |
790 : | edgomez | 851 | |
791 : | syskin | 1108 | if (rd < data->iMinSAD[0]) { |
792 : | data->iMinSAD[0] = rd; | ||
793 : | edgomez | 851 | current[0].x = x; current[0].y = y; |
794 : | syskin | 1132 | *data->dir = Direction; |
795 : | syskin | 1084 | *data->cbp = cbp; |
796 : | edgomez | 851 | } |
797 : | syskin | 1071 | } |
798 : | edgomez | 851 | |
799 : | static void | ||
800 : | syskin | 1132 | CheckCandidateRD8(const int x, const int y, const SearchData * const data, const int Direction) |
801 : | chl | 96 | { |
802 : | |||
803 : | edgomez | 962 | int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; |
804 : | syskin | 1108 | int32_t rd; |
805 : | edgomez | 851 | VECTOR * current; |
806 : | const uint8_t * ptr; | ||
807 : | edgomez | 1022 | int cbp = 0; |
808 : | edgomez | 195 | |
809 : | edgomez | 851 | if ( (x > data->max_dx) || (x < data->min_dx) |
810 : | || (y > data->max_dy) || (y < data->min_dy) ) return; | ||
811 : | edgomez | 195 | |
812 : | edgomez | 851 | if (!data->qpel_precision) { |
813 : | ptr = GetReference(x, y, data); | ||
814 : | current = data->currentMV; | ||
815 : | edgomez | 1053 | } else { /* x and y are in 1/4 precision */ |
816 : | edgomez | 851 | ptr = Interpolate8x8qpel(x, y, 0, 0, data); |
817 : | current = data->currentQMV; | ||
818 : | } | ||
819 : | chl | 96 | |
820 : | edgomez | 851 | transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth); |
821 : | syskin | 1108 | rd = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5); |
822 : | rd += BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); | ||
823 : | chl | 96 | |
824 : | syskin | 1108 | if (rd < data->iMinSAD[0]) { |
825 : | syskin | 1084 | *data->cbp = cbp; |
826 : | syskin | 1108 | data->iMinSAD[0] = rd; |
827 : | edgomez | 851 | current[0].x = x; current[0].y = y; |
828 : | syskin | 1132 | *data->dir = Direction; |
829 : | edgomez | 851 | } |
830 : | } | ||
831 : | chl | 96 | |
832 : | edgomez | 851 | /* CHECK_CANDIATE FUNCTIONS END */ |
833 : | edgomez | 195 | |
834 : | edgomez | 851 | /* MAINSEARCH FUNCTIONS START */ |
835 : | chl | 96 | |
836 : | edgomez | 851 | static void |
837 : | syskin | 1132 | AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection, CheckFunc * const CheckCandidate) |
838 : | edgomez | 851 | { |
839 : | edgomez | 195 | |
840 : | edgomez | 851 | /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
841 : | edgomez | 195 | |
842 : | syskin | 1132 | unsigned int * const iDirection = data->dir; |
843 : | edgomez | 195 | |
844 : | edgomez | 1053 | for(;;) { /* forever */ |
845 : | syskin | 1132 | *iDirection = 0; |
846 : | edgomez | 851 | if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
847 : | if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); | ||
848 : | if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); | ||
849 : | if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); | ||
850 : | edgomez | 195 | |
851 : | edgomez | 851 | /* now we're doing diagonal checks near our candidate */ |
852 : | edgomez | 195 | |
853 : | syskin | 1132 | if (*iDirection) { /* if anything found */ |
854 : | bDirection = *iDirection; | ||
855 : | *iDirection = 0; | ||
856 : | edgomez | 851 | x = data->currentMV->x; y = data->currentMV->y; |
857 : | edgomez | 1053 | if (bDirection & 3) { /* our candidate is left or right */ |
858 : | edgomez | 851 | CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
859 : | CHECK_CANDIDATE(x, y - iDiamondSize, 4); | ||
860 : | edgomez | 1053 | } else { /* what remains here is up or down */ |
861 : | edgomez | 851 | CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
862 : | CHECK_CANDIDATE(x - iDiamondSize, y, 1); | ||
863 : | } | ||
864 : | edgomez | 195 | |
865 : | syskin | 1132 | if (*iDirection) { |
866 : | bDirection += *iDirection; | ||
867 : | edgomez | 851 | x = data->currentMV->x; y = data->currentMV->y; |
868 : | } | ||
869 : | edgomez | 1053 | } else { /* about to quit, eh? not so fast.... */ |
870 : | edgomez | 851 | switch (bDirection) { |
871 : | case 2: | ||
872 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
873 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
874 : | edgomez | 195 | break; |
875 : | edgomez | 851 | case 1: |
876 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
877 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
878 : | edgomez | 195 | break; |
879 : | edgomez | 851 | case 2 + 4: |
880 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
881 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
882 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
883 : | break; | ||
884 : | case 4: | ||
885 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
886 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
887 : | break; | ||
888 : | edgomez | 195 | case 8: |
889 : | edgomez | 851 | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
890 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
891 : | edgomez | 195 | break; |
892 : | edgomez | 851 | case 1 + 4: |
893 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
894 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
895 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
896 : | edgomez | 195 | break; |
897 : | edgomez | 851 | case 2 + 8: |
898 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
899 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
900 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
901 : | break; | ||
902 : | case 1 + 8: | ||
903 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
904 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
905 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
906 : | break; | ||
907 : | edgomez | 1053 | default: /* 1+2+4+8 == we didn't find anything at all */ |
908 : | edgomez | 851 | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
909 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
910 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
911 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
912 : | break; | ||
913 : | chl | 96 | } |
914 : | syskin | 1132 | if (!*iDirection) break; /* ok, the end. really */ |
915 : | bDirection = *iDirection; | ||
916 : | edgomez | 851 | x = data->currentMV->x; y = data->currentMV->y; |
917 : | chl | 345 | } |
918 : | edgomez | 195 | } |
919 : | chl | 96 | } |
920 : | |||
921 : | edgomez | 851 | static void |
922 : | syskin | 1132 | SquareSearch(int x, int y, const SearchData * const data, int bDirection, CheckFunc * const CheckCandidate) |
923 : | chl | 96 | { |
924 : | syskin | 1132 | unsigned int * const iDirection = data->dir; |
925 : | edgomez | 195 | |
926 : | edgomez | 851 | do { |
927 : | syskin | 1132 | *iDirection = 0; |
928 : | edgomez | 851 | if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64); |
929 : | if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128); | ||
930 : | if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32); | ||
931 : | if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128); | ||
932 : | if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64); | ||
933 : | if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128); | ||
934 : | if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128); | ||
935 : | if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128); | ||
936 : | chl | 96 | |
937 : | syskin | 1132 | bDirection = *iDirection; |
938 : | edgomez | 851 | x = data->currentMV->x; y = data->currentMV->y; |
939 : | syskin | 1132 | } while (*iDirection); |
940 : | chl | 96 | } |
941 : | |||
942 : | edgomez | 851 | static void |
943 : | syskin | 1132 | DiamondSearch(int x, int y, const SearchData * const data, int bDirection, CheckFunc * const CheckCandidate) |
944 : | chl | 181 | { |
945 : | |||
946 : | /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ | ||
947 : | |||
948 : | syskin | 1132 | unsigned int * const iDirection = data->dir; |
949 : | edgomez | 195 | |
950 : | edgomez | 851 | do { |
951 : | syskin | 1132 | *iDirection = 0; |
952 : | edgomez | 851 | if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
953 : | if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); | ||
954 : | if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); | ||
955 : | if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); | ||
956 : | |||
957 : | /* now we're doing diagonal checks near our candidate */ | ||
958 : | |||
959 : | syskin | 1132 | if (*iDirection) { /* checking if anything found */ |
960 : | bDirection = *iDirection; | ||
961 : | *iDirection = 0; | ||
962 : | edgomez | 851 | x = data->currentMV->x; y = data->currentMV->y; |
963 : | edgomez | 1053 | if (bDirection & 3) { /* our candidate is left or right */ |
964 : | edgomez | 851 | CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
965 : | CHECK_CANDIDATE(x, y - iDiamondSize, 4); | ||
966 : | edgomez | 1053 | } else { /* what remains here is up or down */ |
967 : | edgomez | 851 | CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
968 : | CHECK_CANDIDATE(x - iDiamondSize, y, 1); | ||
969 : | } | ||
970 : | syskin | 1132 | bDirection += *iDirection; |
971 : | edgomez | 851 | x = data->currentMV->x; y = data->currentMV->y; |
972 : | } | ||
973 : | } | ||
974 : | syskin | 1132 | while (*iDirection); |
975 : | edgomez | 851 | } |
976 : | chl | 181 | |
977 : | edgomez | 851 | /* MAINSEARCH FUNCTIONS END */ |
978 : | chl | 181 | |
979 : | edgomez | 851 | static void |
980 : | Isibaar | 1136 | SubpelRefine_Fast(SearchData * data, CheckFunc * CheckCandidate) |
981 : | { | ||
982 : | /* Do a half-pel or q-pel refinement */ | ||
983 : | VECTOR centerMV; | ||
984 : | VECTOR second_best; | ||
985 : | int best_sad = *data->iMinSAD; | ||
986 : | int xo, yo, xo2, yo2; | ||
987 : | int size = 2; | ||
988 : | CheckFunc *backupFunc = CheckCandidate; | ||
989 : | |||
990 : | if(data->qpel_precision) | ||
991 : | size = 1; | ||
992 : | |||
993 : | centerMV = *data->currentMV; | ||
994 : | *data->iMinSAD = 256 * 4096; | ||
995 : | |||
996 : | CHECK_CANDIDATE(centerMV.x, centerMV.y - size, 0); | ||
997 : | CHECK_CANDIDATE(centerMV.x + size, centerMV.y - size, 0); | ||
998 : | CHECK_CANDIDATE(centerMV.x + size, centerMV.y, 0); | ||
999 : | CHECK_CANDIDATE(centerMV.x + size, centerMV.y + size, 0); | ||
1000 : | |||
1001 : | CHECK_CANDIDATE(centerMV.x, centerMV.y + size, 0); | ||
1002 : | CHECK_CANDIDATE(centerMV.x - size, centerMV.y + size, 0); | ||
1003 : | CHECK_CANDIDATE(centerMV.x - size, centerMV.y, 0); | ||
1004 : | CHECK_CANDIDATE(centerMV.x - size, centerMV.y - size, 0); | ||
1005 : | |||
1006 : | second_best = *data->currentMV; | ||
1007 : | |||
1008 : | Isibaar | 1137 | if(data->qpel_precision) { |
1009 : | Isibaar | 1136 | second_best.x *= 2; second_best.y *= 2; |
1010 : | Isibaar | 1137 | } |
1011 : | |||
1012 : | Isibaar | 1136 | data->currentMV[0] = centerMV; |
1013 : | *data->iMinSAD = best_sad; | ||
1014 : | |||
1015 : | centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV; | ||
1016 : | |||
1017 : | xo = centerMV.x; | ||
1018 : | yo = centerMV.y; | ||
1019 : | xo2 = second_best.x; | ||
1020 : | yo2 = second_best.y; | ||
1021 : | |||
1022 : | CheckCandidate = CheckCandidate16_subpel; | ||
1023 : | *data->iMinSAD2 = 256 * 4096; | ||
1024 : | |||
1025 : | if (yo == yo2) | ||
1026 : | { | ||
1027 : | CHECK_CANDIDATE((xo+xo2)>>1, yo, 0); | ||
1028 : | CHECK_CANDIDATE(xo, yo-1, 0); | ||
1029 : | CHECK_CANDIDATE(xo, yo+1, 0); | ||
1030 : | |||
1031 : | if(best_sad <= *data->iMinSAD2) | ||
1032 : | goto ende; | ||
1033 : | |||
1034 : | if(data->currentQMV[0].x == data->currentQMV2[0].x) { | ||
1035 : | CHECK_CANDIDATE((xo+xo2)>>1, yo-1, 0); | ||
1036 : | CHECK_CANDIDATE((xo+xo2)>>1, yo+1, 0); | ||
1037 : | goto ende; | ||
1038 : | } | ||
1039 : | else { | ||
1040 : | CHECK_CANDIDATE((xo+xo2)>>1, | ||
1041 : | (data->currentQMV[0].x == xo) ? data->currentQMV[0].y : data->currentQMV2[0].y, | ||
1042 : | 0); | ||
1043 : | goto ende; | ||
1044 : | } | ||
1045 : | } | ||
1046 : | |||
1047 : | if (xo == xo2) | ||
1048 : | { | ||
1049 : | CHECK_CANDIDATE(xo, (yo+yo2)>>1, 0); | ||
1050 : | CHECK_CANDIDATE(xo-1, yo, 0); | ||
1051 : | CHECK_CANDIDATE(xo+1, yo, 0); | ||
1052 : | |||
1053 : | if(best_sad < *data->iMinSAD2) | ||
1054 : | goto ende; | ||
1055 : | |||
1056 : | if(data->currentQMV[0].y == data->currentQMV2[0].y) { | ||
1057 : | CHECK_CANDIDATE(xo-1, (yo+yo2)>>1, 0); | ||
1058 : | CHECK_CANDIDATE(xo+1, (yo+yo2)>>1, 0); | ||
1059 : | goto ende; | ||
1060 : | } | ||
1061 : | else { | ||
1062 : | CHECK_CANDIDATE((data->currentQMV[0].y == yo) ? data->currentQMV[0].x : data->currentQMV2[0].x, (yo+yo2)>>1, 0); | ||
1063 : | goto ende; | ||
1064 : | } | ||
1065 : | } | ||
1066 : | |||
1067 : | CHECK_CANDIDATE(xo, (yo+yo2)>>1, 0); | ||
1068 : | CHECK_CANDIDATE((xo+xo2)>>1, yo, 0); | ||
1069 : | |||
1070 : | if(best_sad <= *data->iMinSAD2) | ||
1071 : | goto ende; | ||
1072 : | |||
1073 : | CHECK_CANDIDATE((xo+xo2)>>1, (yo+yo2)>>1, 0); | ||
1074 : | |||
1075 : | ende: | ||
1076 : | CheckCandidate = backupFunc; | ||
1077 : | } | ||
1078 : | |||
1079 : | static void | ||
1080 : | syskin | 1132 | SubpelRefine(const SearchData * const data, CheckFunc * const CheckCandidate) |
1081 : | edgomez | 851 | { |
1082 : | /* Do a half-pel or q-pel refinement */ | ||
1083 : | const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV; | ||
1084 : | chl | 181 | |
1085 : | edgomez | 851 | CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0); |
1086 : | CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0); | ||
1087 : | CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0); | ||
1088 : | CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0); | ||
1089 : | CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0); | ||
1090 : | CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0); | ||
1091 : | CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0); | ||
1092 : | CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0); | ||
1093 : | } | ||
1094 : | chl | 181 | |
1095 : | edgomez | 851 | static __inline int |
1096 : | SkipDecisionP(const IMAGE * current, const IMAGE * reference, | ||
1097 : | const int x, const int y, | ||
1098 : | const uint32_t stride, const uint32_t iQuant, int rrv) | ||
1099 : | chl | 181 | |
1100 : | edgomez | 851 | { |
1101 : | edgomez | 963 | int offset = (x + y*stride)*8; |
1102 : | edgomez | 851 | if(!rrv) { |
1103 : | edgomez | 963 | uint32_t sadC = sad8(current->u + offset, |
1104 : | reference->u + offset, stride); | ||
1105 : | edgomez | 851 | if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
1106 : | edgomez | 963 | sadC += sad8(current->v + offset, |
1107 : | reference->v + offset, stride); | ||
1108 : | edgomez | 851 | if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
1109 : | return 1; | ||
1110 : | chl | 181 | |
1111 : | edgomez | 851 | } else { |
1112 : | edgomez | 963 | uint32_t sadC = sad16(current->u + 2*offset, |
1113 : | reference->u + 2*offset, stride, 256*4096); | ||
1114 : | edgomez | 851 | if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
1115 : | edgomez | 963 | sadC += sad16(current->v + 2*offset, |
1116 : | reference->v + 2*offset, stride, 256*4096); | ||
1117 : | edgomez | 851 | if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
1118 : | return 1; | ||
1119 : | chl | 181 | } |
1120 : | edgomez | 851 | } |
1121 : | edgomez | 504 | |
1122 : | edgomez | 851 | static __inline void |
1123 : | Isibaar | 1095 | ZeroMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
1124 : | { | ||
1125 : | pMB->mode = MODE_INTER; | ||
1126 : | pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV; | ||
1127 : | pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV; | ||
1128 : | pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; | ||
1129 : | } | ||
1130 : | |||
1131 : | static __inline void | ||
1132 : | Isibaar | 1136 | ModeDecision_Fast(SearchData * const Data, |
1133 : | MACROBLOCK * const pMB, | ||
1134 : | const MACROBLOCK * const pMBs, | ||
1135 : | const int x, const int y, | ||
1136 : | const MBParam * const pParam, | ||
1137 : | const uint32_t MotionFlags, | ||
1138 : | const uint32_t VopFlags, | ||
1139 : | const uint32_t VolFlags, | ||
1140 : | const IMAGE * const pCurrent, | ||
1141 : | const IMAGE * const pRef, | ||
1142 : | const IMAGE * const vGMC, | ||
1143 : | const int coding_type) | ||
1144 : | { | ||
1145 : | int mode = MODE_INTER; | ||
1146 : | int mcsel = 0; | ||
1147 : | int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0); | ||
1148 : | const uint32_t iQuant = pMB->quant; | ||
1149 : | const int skip_possible = (coding_type == P_VOP) && (pMB->dquant == 0); | ||
1150 : | int sad; | ||
1151 : | int min_rd = -1, intra_rd, i, cbp, c[2] = {0, 0}; | ||
1152 : | VECTOR backup[5], *v; | ||
1153 : | int sad_backup[5]; | ||
1154 : | int InterBias = MV16_INTER_BIAS; | ||
1155 : | int thresh = 0; | ||
1156 : | int count = 0; | ||
1157 : | int top = 0, top_right = 0, left = 0; | ||
1158 : | |||
1159 : | pMB->mcsel = 0; | ||
1160 : | |||
1161 : | /* INTER <-> INTER4V decision */ | ||
1162 : | Isibaar | 1137 | if ((Data->iMinSAD[0] + 50 < Data->iMinSAD[1] + |
1163 : | Isibaar | 1136 | Data->iMinSAD[2] + Data->iMinSAD[3] + Data->iMinSAD[4])) { /* normal, fast, SAD-based mode decision */ |
1164 : | if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + | ||
1165 : | Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) { | ||
1166 : | mode = MODE_INTER; | ||
1167 : | sad = Data->iMinSAD[0]; | ||
1168 : | } else { | ||
1169 : | mode = MODE_INTER4V; | ||
1170 : | sad = Data->iMinSAD[1] + Data->iMinSAD[2] + | ||
1171 : | Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant; | ||
1172 : | Data->iMinSAD[0] = sad; | ||
1173 : | } | ||
1174 : | |||
1175 : | /* final skip decision, a.k.a. "the vector you found, really that good?" */ | ||
1176 : | if (skip_possible && (pMB->sad16 < (int)iQuant * MAX_SAD00_FOR_SKIP)) | ||
1177 : | if ( (100*sad)/(pMB->sad16+1) > FINAL_SKIP_THRESH) | ||
1178 : | if (Data->chroma || SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant, Data->rrv)) { | ||
1179 : | mode = MODE_NOT_CODED; | ||
1180 : | goto early_out; | ||
1181 : | } | ||
1182 : | |||
1183 : | /* mcsel */ | ||
1184 : | if (coding_type == S_VOP) { | ||
1185 : | |||
1186 : | int32_t iSAD = sad16(Data->Cur, | ||
1187 : | vGMC->y + 16*y*Data->iEdgedWidth + 16*x, Data->iEdgedWidth, 65536); | ||
1188 : | |||
1189 : | if (Data->chroma) { | ||
1190 : | iSAD += sad8(Data->CurU, vGMC->u + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); | ||
1191 : | iSAD += sad8(Data->CurV, vGMC->v + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); | ||
1192 : | } | ||
1193 : | |||
1194 : | if (iSAD <= sad) { /* mode decision GMC */ | ||
1195 : | mode = MODE_INTER; | ||
1196 : | mcsel = 1; | ||
1197 : | sad = iSAD; | ||
1198 : | } | ||
1199 : | |||
1200 : | } | ||
1201 : | } else { /* Rate-Distortion INTER<->INTER4V */ | ||
1202 : | Data->iQuant = iQuant; | ||
1203 : | Data->cbp = c; | ||
1204 : | v = Data->qpel ? Data->currentQMV : Data->currentMV; | ||
1205 : | |||
1206 : | /* final skip decision, a.k.a. "the vector you found, really that good?" */ | ||
1207 : | if (skip_possible && (pMB->sad16 < (int)iQuant * MAX_SAD00_FOR_SKIP)) | ||
1208 : | if ( (100*Data->iMinSAD[0])/(pMB->sad16+1) > FINAL_SKIP_THRESH) | ||
1209 : | if (Data->chroma || SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant, Data->rrv)) { | ||
1210 : | mode = MODE_NOT_CODED; | ||
1211 : | goto early_out; | ||
1212 : | } | ||
1213 : | |||
1214 : | for (i = 0; i < 5; i++) { | ||
1215 : | sad_backup[i] = Data->iMinSAD[i]; | ||
1216 : | Data->iMinSAD[i] = 256*4096; | ||
1217 : | backup[i] = v[i]; | ||
1218 : | } | ||
1219 : | |||
1220 : | min_rd = findRDinter(Data, pMBs, x, y, pParam, MotionFlags); | ||
1221 : | cbp = *Data->cbp; | ||
1222 : | sad = sad_backup[0]; | ||
1223 : | |||
1224 : | if (coding_type == S_VOP) { | ||
1225 : | int gmc_rd; | ||
1226 : | int32_t iSAD = sad16(Data->Cur, | ||
1227 : | vGMC->y + 16*y*Data->iEdgedWidth + 16*x, Data->iEdgedWidth, 65536); | ||
1228 : | |||
1229 : | if (Data->chroma) { | ||
1230 : | iSAD += sad8(Data->CurU, vGMC->u + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); | ||
1231 : | iSAD += sad8(Data->CurV, vGMC->v + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); | ||
1232 : | } | ||
1233 : | |||
1234 : | *Data->iMinSAD = min_rd += BITS_MULT*1; /* mcsel */ | ||
1235 : | gmc_rd = findRDgmc(Data, vGMC, x, y); | ||
1236 : | if (gmc_rd < min_rd) { | ||
1237 : | mcsel = 1; | ||
1238 : | *Data->iMinSAD = min_rd = gmc_rd; | ||
1239 : | mode = MODE_INTER; | ||
1240 : | cbp = *Data->cbp; | ||
1241 : | sad = iSAD; | ||
1242 : | } | ||
1243 : | } | ||
1244 : | |||
1245 : | if (inter4v) { | ||
1246 : | int v4_rd; | ||
1247 : | v4_rd = findRDinter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup); | ||
1248 : | if (v4_rd < min_rd) { | ||
1249 : | Data->iMinSAD[0] = min_rd = v4_rd; | ||
1250 : | mode = MODE_INTER4V; | ||
1251 : | cbp = *Data->cbp; | ||
1252 : | sad = sad_backup[1] + sad_backup[2] + | ||
1253 : | sad_backup[3] + sad_backup[4] + IMV16X16 * (int32_t)iQuant; | ||
1254 : | } | ||
1255 : | } | ||
1256 : | } | ||
1257 : | |||
1258 : | left = top = top_right = -1; | ||
1259 : | thresh = 0; | ||
1260 : | |||
1261 : | if(x > 0 && y > 0 && x < pParam->mb_width) { | ||
1262 : | left = (&pMBs[(x-1) + y * pParam->mb_width])->sad16; // left | ||
1263 : | top = (&pMBs[x + (y-1) * pParam->mb_width])->sad16; // top | ||
1264 : | top_right = (&pMBs[(x+1) + (y-1) * pParam->mb_width])->sad16; // top right | ||
1265 : | |||
1266 : | if(((&pMBs[(x-1) + y * pParam->mb_width])->mode != MODE_INTRA) && | ||
1267 : | ((&pMBs[x + (y-1) * pParam->mb_width])->mode != MODE_INTRA) && | ||
1268 : | ((&pMBs[(x+1) + (y-1) * pParam->mb_width])->mode != MODE_INTRA)) | ||
1269 : | thresh = MAX(MAX(top, left), top_right); | ||
1270 : | else | ||
1271 : | thresh = MIN(MIN(top, left), top_right); | ||
1272 : | } | ||
1273 : | |||
1274 : | /* INTRA <-> INTER decision */ | ||
1275 : | if (sad < thresh) { /* normal, fast, SAD-based mode decision */ | ||
1276 : | /* intra decision */ | ||
1277 : | |||
1278 : | if (iQuant > 8) InterBias += 100 * (iQuant - 8); /* to make high quants work */ | ||
1279 : | if (y != 0) | ||
1280 : | if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; | ||
1281 : | if (x != 0) | ||
1282 : | if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; | ||
1283 : | |||
1284 : | if (Data->chroma) InterBias += 50; /* dev8(chroma) ??? <-- yes, we need dev8 (no big difference though) */ | ||
1285 : | if (Data->rrv) InterBias *= 4; | ||
1286 : | |||
1287 : | if (InterBias < sad) { | ||
1288 : | int32_t deviation; | ||
1289 : | if (!Data->rrv) | ||
1290 : | deviation = dev16(Data->Cur, Data->iEdgedWidth); | ||
1291 : | else | ||
1292 : | deviation = dev16(Data->Cur, Data->iEdgedWidth) + /* dev32() */ | ||
1293 : | dev16(Data->Cur+16, Data->iEdgedWidth) + | ||
1294 : | dev16(Data->Cur + 16*Data->iEdgedWidth, Data->iEdgedWidth) + | ||
1295 : | dev16(Data->Cur+16+16*Data->iEdgedWidth, Data->iEdgedWidth); | ||
1296 : | |||
1297 : | if (deviation < (sad - InterBias)) mode = MODE_INTRA; | ||
1298 : | } | ||
1299 : | |||
1300 : | pMB->cbp = 63; | ||
1301 : | } else { /* Rate-Distortion INTRA<->INTER */ | ||
1302 : | if(min_rd < 0) { | ||
1303 : | Data->iQuant = iQuant; | ||
1304 : | Data->cbp = c; | ||
1305 : | v = Data->qpel ? Data->currentQMV : Data->currentMV; | ||
1306 : | |||
1307 : | for (i = 0; i < 5; i++) { | ||
1308 : | Data->iMinSAD[i] = 256*4096; | ||
1309 : | backup[i] = v[i]; | ||
1310 : | } | ||
1311 : | |||
1312 : | if(mode == MODE_INTER) { | ||
1313 : | min_rd = findRDinter(Data, pMBs, x, y, pParam, MotionFlags); | ||
1314 : | cbp = *Data->cbp; | ||
1315 : | |||
1316 : | if (coding_type == S_VOP) { | ||
1317 : | int gmc_rd; | ||
1318 : | |||
1319 : | *Data->iMinSAD = min_rd += BITS_MULT*1; /* mcsel */ | ||
1320 : | gmc_rd = findRDgmc(Data, vGMC, x, y); | ||
1321 : | if (gmc_rd < min_rd) { | ||
1322 : | mcsel = 1; | ||
1323 : | *Data->iMinSAD = min_rd = gmc_rd; | ||
1324 : | mode = MODE_INTER; | ||
1325 : | cbp = *Data->cbp; | ||
1326 : | } | ||
1327 : | } | ||
1328 : | } | ||
1329 : | |||
1330 : | if(mode == MODE_INTER4V) { | ||
1331 : | int v4_rd; | ||
1332 : | v4_rd = findRDinter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup); | ||
1333 : | if (v4_rd < min_rd) { | ||
1334 : | Data->iMinSAD[0] = min_rd = v4_rd; | ||
1335 : | mode = MODE_INTER4V; | ||
1336 : | cbp = *Data->cbp; | ||
1337 : | } | ||
1338 : | } | ||
1339 : | } | ||
1340 : | |||
1341 : | intra_rd = findRDintra(Data); | ||
1342 : | if (intra_rd < min_rd) { | ||
1343 : | *Data->iMinSAD = min_rd = intra_rd; | ||
1344 : | mode = MODE_INTRA; | ||
1345 : | } | ||
1346 : | |||
1347 : | pMB->cbp = cbp; | ||
1348 : | } | ||
1349 : | |||
1350 : | early_out: | ||
1351 : | pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; | ||
1352 : | |||
1353 : | if (Data->rrv) { | ||
1354 : | Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); | ||
1355 : | Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); | ||
1356 : | } | ||
1357 : | |||
1358 : | if (mode == MODE_INTER && mcsel == 0) { | ||
1359 : | pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; | ||
1360 : | |||
1361 : | if(Data->qpel) { | ||
1362 : | pMB->qmvs[0] = pMB->qmvs[1] | ||
1363 : | = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; | ||
1364 : | pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; | ||
1365 : | pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; | ||
1366 : | } else { | ||
1367 : | pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; | ||
1368 : | pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; | ||
1369 : | } | ||
1370 : | |||
1371 : | } else if (mode == MODE_INTER ) { // but mcsel == 1 | ||
1372 : | |||
1373 : | pMB->mcsel = 1; | ||
1374 : | if (Data->qpel) { | ||
1375 : | pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = pMB->amv; | ||
1376 : | pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = pMB->amv.x/2; | ||
1377 : | pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = pMB->amv.y/2; | ||
1378 : | } else | ||
1379 : | pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->amv; | ||
1380 : | |||
1381 : | } else | ||
1382 : | if (mode == MODE_INTER4V) ; /* anything here? */ | ||
1383 : | else /* INTRA, NOT_CODED */ | ||
1384 : | ZeroMacroblockP(pMB, 0); | ||
1385 : | |||
1386 : | pMB->mode = mode; | ||
1387 : | } | ||
1388 : | |||
1389 : | static __inline void | ||
1390 : | edgomez | 1022 | ModeDecision(SearchData * const Data, |
1391 : | MACROBLOCK * const pMB, | ||
1392 : | const MACROBLOCK * const pMBs, | ||
1393 : | const int x, const int y, | ||
1394 : | const MBParam * const pParam, | ||
1395 : | const uint32_t MotionFlags, | ||
1396 : | const uint32_t VopFlags, | ||
1397 : | const uint32_t VolFlags, | ||
1398 : | const IMAGE * const pCurrent, | ||
1399 : | syskin | 1084 | const IMAGE * const pRef, |
1400 : | const IMAGE * const vGMC, | ||
1401 : | const int coding_type) | ||
1402 : | edgomez | 1022 | { |
1403 : | int mode = MODE_INTER; | ||
1404 : | syskin | 1084 | int mcsel = 0; |
1405 : | edgomez | 1022 | int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0); |
1406 : | const uint32_t iQuant = pMB->quant; | ||
1407 : | |||
1408 : | syskin | 1084 | const int skip_possible = (coding_type == P_VOP) && (pMB->dquant == 0); |
1409 : | edgomez | 1022 | |
1410 : | syskin | 1084 | pMB->mcsel = 0; |
1411 : | |||
1412 : | edgomez | 1107 | if (!(VopFlags & XVID_VOP_MODEDECISION_RD)) { /* normal, fast, SAD-based mode decision */ |
1413 : | edgomez | 1022 | int sad; |
1414 : | int InterBias = MV16_INTER_BIAS; | ||
1415 : | if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + | ||
1416 : | Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) { | ||
1417 : | mode = MODE_INTER; | ||
1418 : | sad = Data->iMinSAD[0]; | ||
1419 : | } else { | ||
1420 : | mode = MODE_INTER4V; | ||
1421 : | sad = Data->iMinSAD[1] + Data->iMinSAD[2] + | ||
1422 : | Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant; | ||
1423 : | Data->iMinSAD[0] = sad; | ||
1424 : | } | ||
1425 : | |||
1426 : | /* final skip decision, a.k.a. "the vector you found, really that good?" */ | ||
1427 : | if (skip_possible && (pMB->sad16 < (int)iQuant * MAX_SAD00_FOR_SKIP)) | ||
1428 : | if ( (100*sad)/(pMB->sad16+1) > FINAL_SKIP_THRESH) | ||
1429 : | if (Data->chroma || SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant, Data->rrv)) { | ||
1430 : | mode = MODE_NOT_CODED; | ||
1431 : | sad = 0; | ||
1432 : | } | ||
1433 : | |||
1434 : | syskin | 1084 | /* mcsel */ |
1435 : | if (coding_type == S_VOP) { | ||
1436 : | |||
1437 : | int32_t iSAD = sad16(Data->Cur, | ||
1438 : | vGMC->y + 16*y*Data->iEdgedWidth + 16*x, Data->iEdgedWidth, 65536); | ||
1439 : | |||
1440 : | if (Data->chroma) { | ||
1441 : | iSAD += sad8(Data->CurU, vGMC->u + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); | ||
1442 : | iSAD += sad8(Data->CurV, vGMC->v + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); | ||
1443 : | } | ||
1444 : | |||
1445 : | if (iSAD <= sad) { /* mode decision GMC */ | ||
1446 : | mode = MODE_INTER; | ||
1447 : | mcsel = 1; | ||
1448 : | sad = iSAD; | ||
1449 : | } | ||
1450 : | |||
1451 : | } | ||
1452 : | |||
1453 : | edgomez | 1022 | /* intra decision */ |
1454 : | |||
1455 : | edgomez | 1053 | if (iQuant > 8) InterBias += 100 * (iQuant - 8); /* to make high quants work */ |
1456 : | edgomez | 1022 | if (y != 0) |
1457 : | if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; | ||
1458 : | if (x != 0) | ||
1459 : | if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; | ||
1460 : | |||
1461 : | syskin | 1084 | if (Data->chroma) InterBias += 50; /* dev8(chroma) ??? <-- yes, we need dev8 (no big difference though) */ |
1462 : | edgomez | 1022 | if (Data->rrv) InterBias *= 4; |
1463 : | |||
1464 : | syskin | 1084 | if (InterBias < sad) { |
1465 : | edgomez | 1022 | int32_t deviation; |
1466 : | syskin | 1084 | if (!Data->rrv) |
1467 : | deviation = dev16(Data->Cur, Data->iEdgedWidth); | ||
1468 : | else | ||
1469 : | deviation = dev16(Data->Cur, Data->iEdgedWidth) + /* dev32() */ | ||
1470 : | dev16(Data->Cur+16, Data->iEdgedWidth) + | ||
1471 : | dev16(Data->Cur + 16*Data->iEdgedWidth, Data->iEdgedWidth) + | ||
1472 : | dev16(Data->Cur+16+16*Data->iEdgedWidth, Data->iEdgedWidth); | ||
1473 : | edgomez | 1022 | |
1474 : | if (deviation < (sad - InterBias)) mode = MODE_INTRA; | ||
1475 : | } | ||
1476 : | |||
1477 : | syskin | 1084 | pMB->cbp = 63; |
1478 : | pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; | ||
1479 : | |||
1480 : | syskin | 1108 | } else { /* Rate-Distortion */ |
1481 : | edgomez | 1022 | |
1482 : | syskin | 1108 | int min_rd, intra_rd, i, cbp, c[2] = {0, 0}; |
1483 : | edgomez | 1022 | VECTOR backup[5], *v; |
1484 : | Data->iQuant = iQuant; | ||
1485 : | syskin | 1084 | Data->cbp = c; |
1486 : | edgomez | 1022 | |
1487 : | v = Data->qpel ? Data->currentQMV : Data->currentMV; | ||
1488 : | for (i = 0; i < 5; i++) { | ||
1489 : | Data->iMinSAD[i] = 256*4096; | ||
1490 : | backup[i] = v[i]; | ||
1491 : | } | ||
1492 : | |||
1493 : | syskin | 1108 | min_rd = findRDinter(Data, pMBs, x, y, pParam, MotionFlags); |
1494 : | syskin | 1084 | cbp = *Data->cbp; |
1495 : | |||
1496 : | if (coding_type == S_VOP) { | ||
1497 : | syskin | 1108 | int gmc_rd; |
1498 : | *Data->iMinSAD = min_rd += BITS_MULT*1; /* mcsel */ | ||
1499 : | gmc_rd = findRDgmc(Data, vGMC, x, y); | ||
1500 : | if (gmc_rd < min_rd) { | ||
1501 : | syskin | 1084 | mcsel = 1; |
1502 : | syskin | 1108 | *Data->iMinSAD = min_rd = gmc_rd; |
1503 : | syskin | 1084 | mode = MODE_INTER; |
1504 : | cbp = *Data->cbp; | ||
1505 : | edgomez | 1022 | } |
1506 : | syskin | 1084 | } |
1507 : | edgomez | 1022 | |
1508 : | syskin | 1084 | if (inter4v) { |
1509 : | syskin | 1108 | int v4_rd; |
1510 : | v4_rd = findRDinter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup); | ||
1511 : | if (v4_rd < min_rd) { | ||
1512 : | Data->iMinSAD[0] = min_rd = v4_rd; | ||
1513 : | syskin | 1084 | mode = MODE_INTER4V; |
1514 : | cbp = *Data->cbp; | ||
1515 : | } | ||
1516 : | } | ||
1517 : | edgomez | 1022 | |
1518 : | syskin | 1108 | intra_rd = findRDintra(Data); |
1519 : | if (intra_rd < min_rd) { | ||
1520 : | *Data->iMinSAD = min_rd = intra_rd; | ||
1521 : | syskin | 1084 | mode = MODE_INTRA; |
1522 : | edgomez | 1022 | } |
1523 : | syskin | 1084 | |
1524 : | pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = 0; | ||
1525 : | pMB->cbp = cbp; | ||
1526 : | edgomez | 1022 | } |
1527 : | |||
1528 : | if (Data->rrv) { | ||
1529 : | Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); | ||
1530 : | Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); | ||
1531 : | } | ||
1532 : | |||
1533 : | syskin | 1084 | if (mode == MODE_INTER && mcsel == 0) { |
1534 : | edgomez | 1022 | pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
1535 : | |||
1536 : | if(Data->qpel) { | ||
1537 : | pMB->qmvs[0] = pMB->qmvs[1] | ||
1538 : | = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; | ||
1539 : | pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; | ||
1540 : | pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; | ||
1541 : | } else { | ||
1542 : | pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; | ||
1543 : | pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; | ||
1544 : | } | ||
1545 : | |||
1546 : | syskin | 1084 | } else if (mode == MODE_INTER ) { // but mcsel == 1 |
1547 : | |||
1548 : | pMB->mcsel = 1; | ||
1549 : | if (Data->qpel) { | ||
1550 : | pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = pMB->amv; | ||
1551 : | pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = pMB->amv.x/2; | ||
1552 : | pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = pMB->amv.y/2; | ||
1553 : | } else | ||
1554 : | pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->amv; | ||
1555 : | |||
1556 : | } else | ||
1557 : | if (mode == MODE_INTER4V) ; /* anything here? */ | ||
1558 : | else /* INTRA, NOT_CODED */ | ||
1559 : | syskin | 1108 | ZeroMacroblockP(pMB, 0); |
1560 : | edgomez | 1022 | |
1561 : | pMB->mode = mode; | ||
1562 : | } | ||
1563 : | |||
1564 : | edgomez | 851 | bool |
1565 : | MotionEstimation(MBParam * const pParam, | ||
1566 : | FRAMEINFO * const current, | ||
1567 : | FRAMEINFO * const reference, | ||
1568 : | const IMAGE * const pRefH, | ||
1569 : | const IMAGE * const pRefV, | ||
1570 : | const IMAGE * const pRefHV, | ||
1571 : | chl | 1077 | const IMAGE * const pGMC, |
1572 : | edgomez | 851 | const uint32_t iLimit) |
1573 : | chl | 181 | { |
1574 : | edgomez | 851 | MACROBLOCK *const pMBs = current->mbs; |
1575 : | const IMAGE *const pCurrent = ¤t->image; | ||
1576 : | const IMAGE *const pRef = &reference->image; | ||
1577 : | chl | 181 | |
1578 : | edgomez | 851 | uint32_t mb_width = pParam->mb_width; |
1579 : | uint32_t mb_height = pParam->mb_height; | ||
1580 : | const uint32_t iEdgedWidth = pParam->edged_width; | ||
1581 : | suxen_drol | 890 | const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->vop_flags, current->vol_flags); |
1582 : | Isibaar | 1136 | int stat_thresh = 0; |
1583 : | chl | 181 | |
1584 : | edgomez | 851 | uint32_t x, y; |
1585 : | uint32_t iIntra = 0; | ||
1586 : | syskin | 1108 | int32_t sad00; |
1587 : | chl | 1077 | int skip_thresh = INITIAL_SKIP_THRESH * \ |
1588 : | edgomez | 953 | (current->vop_flags & XVID_VOP_REDUCED ? 4:1) * \ |
1589 : | edgomez | 1107 | (current->vop_flags & XVID_VOP_MODEDECISION_RD ? 2:1); |
1590 : | chl | 181 | |
1591 : | edgomez | 1053 | /* some pre-initialized thingies for SearchP */ |
1592 : | syskin | 1132 | int32_t temp[8]; uint32_t dir; |
1593 : | edgomez | 851 | VECTOR currentMV[5]; |
1594 : | VECTOR currentQMV[5]; | ||
1595 : | Isibaar | 1136 | VECTOR currentMV2[5]; |
1596 : | VECTOR currentQMV2[5]; | ||
1597 : | edgomez | 851 | int32_t iMinSAD[5]; |
1598 : | Isibaar | 1136 | int32_t iMinSAD2[5]; |
1599 : | syskin | 1075 | DECLARE_ALIGNED_MATRIX(dct_space, 3, 64, int16_t, CACHE_LINE); |
1600 : | edgomez | 851 | SearchData Data; |
1601 : | memset(&Data, 0, sizeof(SearchData)); | ||
1602 : | Data.iEdgedWidth = iEdgedWidth; | ||
1603 : | Data.currentMV = currentMV; | ||
1604 : | Data.currentQMV = currentQMV; | ||
1605 : | Isibaar | 1136 | Data.currentMV2 = currentMV2; |
1606 : | Data.currentQMV2 = currentQMV2; | ||
1607 : | edgomez | 851 | Data.iMinSAD = iMinSAD; |
1608 : | Isibaar | 1136 | Data.iMinSAD2 = iMinSAD2; |
1609 : | edgomez | 851 | Data.temp = temp; |
1610 : | syskin | 1132 | Data.dir = &dir; |
1611 : | edgomez | 851 | Data.iFcode = current->fcode; |
1612 : | Data.rounding = pParam->m_rounding_type; | ||
1613 : | edgomez | 963 | Data.qpel = (current->vol_flags & XVID_VOL_QUARTERPEL ? 1:0); |
1614 : | edgomez | 1107 | Data.chroma = MotionFlags & XVID_ME_CHROMA_PVOP; |
1615 : | chl | 1077 | Data.rrv = (current->vop_flags & XVID_VOP_REDUCED) ? 1:0; |
1616 : | edgomez | 962 | Data.dctSpace = dct_space; |
1617 : | edgomez | 1022 | Data.quant_type = !(pParam->vol_flags & XVID_VOL_MPEGQUANT); |
1618 : | edgomez | 195 | |
1619 : | edgomez | 949 | if ((current->vop_flags & XVID_VOP_REDUCED)) { |
1620 : | edgomez | 851 | mb_width = (pParam->width + 31) / 32; |
1621 : | mb_height = (pParam->height + 31) / 32; | ||
1622 : | Data.qpel = 0; | ||
1623 : | } | ||
1624 : | chl | 181 | |
1625 : | edgomez | 1053 | Data.RefQ = pRefV->u; /* a good place, also used in MC (for similar purpose) */ |
1626 : | edgomez | 851 | if (sadInit) (*sadInit) (); |
1627 : | chl | 181 | |
1628 : | edgomez | 851 | for (y = 0; y < mb_height; y++) { |
1629 : | for (x = 0; x < mb_width; x++) { | ||
1630 : | MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; | ||
1631 : | Isibaar | 1136 | MACROBLOCK *prevMB = &reference->mbs[x + y * pParam->mb_width]; |
1632 : | chl | 181 | |
1633 : | edgomez | 851 | if (!Data.rrv) pMB->sad16 = |
1634 : | sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16, | ||
1635 : | pRef->y + (x + y * iEdgedWidth) * 16, | ||
1636 : | pParam->edged_width, pMB->sad8 ); | ||
1637 : | chl | 181 | |
1638 : | edgomez | 851 | else pMB->sad16 = |
1639 : | sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32, | ||
1640 : | pRef->y + (x + y * iEdgedWidth) * 32, | ||
1641 : | pParam->edged_width, pMB->sad8 ); | ||
1642 : | chl | 181 | |
1643 : | edgomez | 851 | if (Data.chroma) { |
1644 : | Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8, | ||
1645 : | pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2) | ||
1646 : | + sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8, | ||
1647 : | pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); | ||
1648 : | pMB->sad16 += Data.temp[7]; | ||
1649 : | } | ||
1650 : | |||
1651 : | sad00 = pMB->sad16; | ||
1652 : | |||
1653 : | edgomez | 1053 | /* initial skip decision */ |
1654 : | /* no early skip for GMC (global vector = skip vector is unknown!) */ | ||
1655 : | syskin | 1084 | if (current->coding_type != S_VOP) { /* no fast SKIP for S(GMC)-VOPs */ |
1656 : | edgomez | 953 | if (pMB->dquant == 0 && sad00 < pMB->quant * skip_thresh) |
1657 : | edgomez | 851 | if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) { |
1658 : | syskin | 1108 | ZeroMacroblockP(pMB, sad00); |
1659 : | pMB->mode = MODE_NOT_CODED; | ||
1660 : | edgomez | 851 | continue; |
1661 : | } | ||
1662 : | } | ||
1663 : | |||
1664 : | Isibaar | 1136 | if(MotionFlags & XVID_ME_DETECT_STATIC_MOTION) { |
1665 : | if(x > 0 && y > 0 && x < pParam->mb_width) | ||
1666 : | if(MVequal((&pMBs[(x-1) + y * pParam->mb_width])->mvs[0], zeroMV) && | ||
1667 : | MVequal((&pMBs[x + (y-1) * pParam->mb_width])->mvs[0], zeroMV) && | ||
1668 : | MVequal((&pMBs[(x+1) + (y-1) * pParam->mb_width])->mvs[0], zeroMV) && | ||
1669 : | MVequal(prevMB->mvs[0], zeroMV)) { | ||
1670 : | stat_thresh = MAX((&pMBs[(x-1) + y * pParam->mb_width])->sad16, | ||
1671 : | MAX((&pMBs[x + (y-1) * pParam->mb_width])->sad16, | ||
1672 : | MAX((&pMBs[(x+1) + (y-1) * pParam->mb_width])->sad16, | ||
1673 : | prevMB->sad16))); | ||
1674 : | } | ||
1675 : | else | ||
1676 : | stat_thresh = MIN((&pMBs[(x-1) + y * pParam->mb_width])->sad16, | ||
1677 : | MIN((&pMBs[x + (y-1) * pParam->mb_width])->sad16, | ||
1678 : | MIN((&pMBs[(x+1) + (y-1) * pParam->mb_width])->sad16, | ||
1679 : | prevMB->sad16))); | ||
1680 : | } | ||
1681 : | |||
1682 : | Isibaar | 1095 | if ((current->vop_flags & XVID_VOP_CARTOON) && |
1683 : | Isibaar | 1136 | (sad00 < pMB->quant * 4 * skip_thresh) || (sad00 < stat_thresh)) { /* favorize (0,0) vector for cartoons */ |
1684 : | Isibaar | 1095 | ZeroMacroblockP(pMB, sad00); |
1685 : | continue; | ||
1686 : | } | ||
1687 : | |||
1688 : | edgomez | 851 | SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
1689 : | edgomez | 1022 | y, MotionFlags, current->vop_flags, current->vol_flags, |
1690 : | &Data, pParam, pMBs, reference->mbs, pMB); | ||
1691 : | edgomez | 851 | |
1692 : | Isibaar | 1136 | if(current->vop_flags & XVID_VOP_FAST_MODEDECISION_RD) { |
1693 : | ModeDecision_Fast(&Data, pMB, pMBs, x, y, pParam, | ||
1694 : | MotionFlags, current->vop_flags, current->vol_flags, | ||
1695 : | pCurrent, pRef, pGMC, current->coding_type); | ||
1696 : | } | ||
1697 : | else { | ||
1698 : | ModeDecision(&Data, pMB, pMBs, x, y, pParam, | ||
1699 : | MotionFlags, current->vop_flags, current->vol_flags, | ||
1700 : | pCurrent, pRef, pGMC, current->coding_type); | ||
1701 : | } | ||
1702 : | edgomez | 1022 | |
1703 : | edgomez | 851 | if (pMB->mode == MODE_INTRA) |
1704 : | if (++iIntra > iLimit) return 1; | ||
1705 : | chl | 181 | } |
1706 : | } | ||
1707 : | edgomez | 851 | |
1708 : | return 0; | ||
1709 : | chl | 181 | } |
1710 : | |||
1711 : | syskin | 1132 | /* check if given vector is equal to any vector checked before */ |
1712 : | static __inline int | ||
1713 : | vector_repeats(const VECTOR * const pmv, const int i) | ||
1714 : | { | ||
1715 : | unsigned int j; | ||
1716 : | for (j = 0; j < i; j++) | ||
1717 : | if (MVequal(pmv[i], pmv[j])) return 1; /* same vector has been checked already */ | ||
1718 : | return 0; | ||
1719 : | } | ||
1720 : | chl | 181 | |
1721 : | syskin | 1132 | /* make a binary mask that prevents diamonds/squares |
1722 : | from checking a vector which has been checked as a prediction */ | ||
1723 : | edgomez | 851 | static __inline int |
1724 : | syskin | 1132 | make_mask(const VECTOR * const pmv, const int i, const int current) |
1725 : | chl | 96 | { |
1726 : | syskin | 1132 | unsigned int mask = 255, j; |
1727 : | edgomez | 851 | for (j = 0; j < i; j++) { |
1728 : | syskin | 1132 | if (pmv[current].x == pmv[j].x) { |
1729 : | if (pmv[current].y == pmv[j].y + iDiamondSize) mask &= ~4; | ||
1730 : | else if (pmv[current].y == pmv[j].y - iDiamondSize) mask &= ~8; | ||
1731 : | edgomez | 851 | } else |
1732 : | syskin | 1132 | if (pmv[current].y == pmv[j].y) { |
1733 : | if (pmv[current].x == pmv[j].x + iDiamondSize) mask &= ~1; | ||
1734 : | else if (pmv[current].x == pmv[j].x - iDiamondSize) mask &= ~2; | ||
1735 : | edgomez | 851 | } |
1736 : | } | ||
1737 : | return mask; | ||
1738 : | } | ||
1739 : | edgomez | 195 | |
1740 : | edgomez | 851 | static __inline void |
1741 : | PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, | ||
1742 : | int iHcount, const MACROBLOCK * const prevMB, int rrv) | ||
1743 : | { | ||
1744 : | edgomez | 1053 | /* this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself */ |
1745 : | edgomez | 851 | if (rrv) { iWcount /= 2; iHcount /= 2; } |
1746 : | edgomez | 195 | |
1747 : | edgomez | 1053 | if ( (y != 0) && (x < (iWcount-1)) ) { /* [5] top-right neighbour */ |
1748 : | edgomez | 851 | pmv[5].x = EVEN(pmv[3].x); |
1749 : | pmv[5].y = EVEN(pmv[3].y); | ||
1750 : | } else pmv[5].x = pmv[5].y = 0; | ||
1751 : | chl | 96 | |
1752 : | edgomez | 1053 | if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }/* pmv[3] is left neighbour */ |
1753 : | edgomez | 851 | else pmv[3].x = pmv[3].y = 0; |
1754 : | chl | 96 | |
1755 : | edgomez | 1053 | if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }/* [4] top neighbour */ |
1756 : | edgomez | 851 | else pmv[4].x = pmv[4].y = 0; |
1757 : | Isibaar | 3 | |
1758 : | edgomez | 1053 | /* [1] median prediction */ |
1759 : | edgomez | 851 | pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); |
1760 : | edgomez | 195 | |
1761 : | edgomez | 1053 | pmv[0].x = pmv[0].y = 0; /* [0] is zero; not used in the loop (checked before) but needed here for make_mask */ |
1762 : | edgomez | 195 | |
1763 : | edgomez | 1053 | pmv[2].x = EVEN(prevMB->mvs[0].x); /* [2] is last frame */ |
1764 : | edgomez | 851 | pmv[2].y = EVEN(prevMB->mvs[0].y); |
1765 : | |||
1766 : | if ((x < iWcount-1) && (y < iHcount-1)) { | ||
1767 : | edgomez | 1053 | pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); /* [6] right-down neighbour in last frame */ |
1768 : | edgomez | 851 | pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
1769 : | } else pmv[6].x = pmv[6].y = 0; | ||
1770 : | |||
1771 : | if (rrv) { | ||
1772 : | int i; | ||
1773 : | for (i = 0; i < 7; i++) { | ||
1774 : | pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); | ||
1775 : | pmv[i].y = RRV_MV_SCALEUP(pmv[i].y); | ||
1776 : | } | ||
1777 : | } | ||
1778 : | Isibaar | 3 | } |
1779 : | |||
1780 : | edgomez | 851 | static void |
1781 : | SearchP(const IMAGE * const pRef, | ||
1782 : | const uint8_t * const pRefH, | ||
1783 : | const uint8_t * const pRefV, | ||
1784 : | const uint8_t * const pRefHV, | ||
1785 : | const IMAGE * const pCur, | ||
1786 : | const int x, | ||
1787 : | const int y, | ||
1788 : | const uint32_t MotionFlags, | ||
1789 : | suxen_drol | 890 | const uint32_t VopFlags, |
1790 : | edgomez | 1022 | const uint32_t VolFlags, |
1791 : | edgomez | 851 | SearchData * const Data, |
1792 : | const MBParam * const pParam, | ||
1793 : | const MACROBLOCK * const pMBs, | ||
1794 : | const MACROBLOCK * const prevMBs, | ||
1795 : | MACROBLOCK * const pMB) | ||
1796 : | { | ||
1797 : | Isibaar | 3 | |
1798 : | syskin | 1132 | int i, threshA; |
1799 : | edgomez | 851 | VECTOR pmv[7]; |
1800 : | edgomez | 1022 | int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0); |
1801 : | syskin | 1132 | CheckFunc * CheckCandidate; |
1802 : | Isibaar | 3 | |
1803 : | syskin | 1129 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
1804 : | pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, Data->rrv); | ||
1805 : | edgomez | 195 | |
1806 : | edgomez | 851 | get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); |
1807 : | Isibaar | 3 | |
1808 : | edgomez | 1053 | Data->temp[5] = Data->temp[6] = 0; /* chroma-sad cache */ |
1809 : | edgomez | 851 | i = Data->rrv ? 2 : 1; |
1810 : | Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i; | ||
1811 : | Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; | ||
1812 : | Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; | ||
1813 : | chl | 326 | |
1814 : | edgomez | 978 | Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16*i; |
1815 : | Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16*i; | ||
1816 : | Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16*i; | ||
1817 : | Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16*i; | ||
1818 : | Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; | ||
1819 : | Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; | ||
1820 : | edgomez | 195 | |
1821 : | edgomez | 1022 | Data->lambda16 = lambda_vec16[pMB->quant]; |
1822 : | Data->lambda8 = lambda_vec8[pMB->quant]; | ||
1823 : | edgomez | 851 | Data->qpel_precision = 0; |
1824 : | syskin | 1132 | *Data->dir = 0; |
1825 : | edgomez | 195 | |
1826 : | edgomez | 963 | memset(Data->currentMV, 0, 5*sizeof(VECTOR)); |
1827 : | chl | 169 | |
1828 : | edgomez | 851 | if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
1829 : | else Data->predMV = pmv[0]; | ||
1830 : | Isibaar | 3 | |
1831 : | edgomez | 851 | i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0); |
1832 : | Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10); | ||
1833 : | Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10); | ||
1834 : | Data->iMinSAD[2] = pMB->sad8[1]; | ||
1835 : | Data->iMinSAD[3] = pMB->sad8[2]; | ||
1836 : | Data->iMinSAD[4] = pMB->sad8[3]; | ||
1837 : | chl | 169 | |
1838 : | edgomez | 1107 | if ((!(VopFlags & XVID_VOP_MODEDECISION_RD)) && (x | y)) { |
1839 : | edgomez | 1053 | threshA = Data->temp[0]; /* that's where we keep this SAD atm */ |
1840 : | edgomez | 851 | if (threshA < 512) threshA = 512; |
1841 : | else if (threshA > 1024) threshA = 1024; | ||
1842 : | } else | ||
1843 : | threshA = 512; | ||
1844 : | chl | 169 | |
1845 : | edgomez | 851 | PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
1846 : | prevMBs + x + y * pParam->mb_width, Data->rrv); | ||
1847 : | chl | 169 | |
1848 : | edgomez | 851 | if (!Data->rrv) { |
1849 : | syskin | 1132 | if (inter4v) CheckCandidate = CheckCandidate16; |
1850 : | edgomez | 1053 | else CheckCandidate = CheckCandidate16no4v; /* for extra speed */ |
1851 : | edgomez | 851 | } else CheckCandidate = CheckCandidate32; |
1852 : | chl | 169 | |
1853 : | edgomez | 851 | /* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/ |
1854 : | chl | 169 | |
1855 : | syskin | 1132 | for (i = 1; i < 7; i++) |
1856 : | if (!vector_repeats(pmv, i)) { | ||
1857 : | CheckCandidate(pmv[i].x, pmv[i].y, Data, i); | ||
1858 : | if (Data->iMinSAD[0] <= threshA) { i++; break; } | ||
1859 : | } | ||
1860 : | chl | 169 | |
1861 : | edgomez | 851 | if ((Data->iMinSAD[0] <= threshA) || |
1862 : | (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && | ||
1863 : | edgomez | 1022 | (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) |
1864 : | inter4v = 0; | ||
1865 : | edgomez | 851 | else { |
1866 : | Isibaar | 3 | |
1867 : | edgomez | 851 | MainSearchFunc * MainSearchPtr; |
1868 : | syskin | 1132 | int mask = make_mask(pmv, i, *Data->dir); // all vectors pmv[0..i-1] have been checked |
1869 : | |||
1870 : | edgomez | 949 | if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch; |
1871 : | else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; | ||
1872 : | edgomez | 851 | else MainSearchPtr = DiamondSearch; |
1873 : | Isibaar | 3 | |
1874 : | syskin | 1132 | MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate); |
1875 : | Isibaar | 3 | |
1876 : | edgomez | 851 | /* extended search, diamond starting in 0,0 and in prediction. |
1877 : | note that this search is/might be done in halfpel positions, | ||
1878 : | which makes it more different than the diamond above */ | ||
1879 : | chl | 169 | |
1880 : | edgomez | 949 | if (MotionFlags & XVID_ME_EXTSEARCH16) { |
1881 : | edgomez | 851 | int32_t bSAD; |
1882 : | VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; | ||
1883 : | if (Data->rrv) { | ||
1884 : | startMV.x = RRV_MV_SCALEUP(startMV.x); | ||
1885 : | startMV.y = RRV_MV_SCALEUP(startMV.y); | ||
1886 : | } | ||
1887 : | if (!(MVequal(startMV, backupMV))) { | ||
1888 : | bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; | ||
1889 : | edgomez | 195 | |
1890 : | syskin | 1132 | CheckCandidate(startMV.x, startMV.y, Data, 255); |
1891 : | MainSearchPtr(startMV.x, startMV.y, Data, 255, CheckCandidate); | ||
1892 : | edgomez | 851 | if (bSAD < Data->iMinSAD[0]) { |
1893 : | Data->currentMV[0] = backupMV; | ||
1894 : | Data->iMinSAD[0] = bSAD; } | ||
1895 : | } | ||
1896 : | chl | 169 | |
1897 : | edgomez | 851 | backupMV = Data->currentMV[0]; |
1898 : | startMV.x = startMV.y = 1; | ||
1899 : | if (!(MVequal(startMV, backupMV))) { | ||
1900 : | bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; | ||
1901 : | chl | 169 | |
1902 : | syskin | 1132 | CheckCandidate(startMV.x, startMV.y, Data, 255); |
1903 : | MainSearchPtr(startMV.x, startMV.y, Data, 255, CheckCandidate); | ||
1904 : | edgomez | 851 | if (bSAD < Data->iMinSAD[0]) { |
1905 : | Data->currentMV[0] = backupMV; | ||
1906 : | syskin | 1132 | Data->iMinSAD[0] = bSAD; |
1907 : | } | ||
1908 : | edgomez | 195 | } |
1909 : | edgomez | 851 | } |
1910 : | } | ||
1911 : | edgomez | 195 | |
1912 : | edgomez | 949 | if (MotionFlags & XVID_ME_HALFPELREFINE16) |
1913 : | syskin | 1132 | SubpelRefine(Data, CheckCandidate); |
1914 : | edgomez | 195 | |
1915 : | edgomez | 851 | for(i = 0; i < 5; i++) { |
1916 : | edgomez | 1053 | Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* initialize qpel vectors */ |
1917 : | edgomez | 851 | Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
1918 : | } | ||
1919 : | Isibaar | 3 | |
1920 : | edgomez | 1023 | if (Data->qpel) { |
1921 : | syskin | 1129 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
1922 : | pParam->width, pParam->height, Data->iFcode, 2, 0); | ||
1923 : | edgomez | 987 | Data->qpel_precision = 1; |
1924 : | edgomez | 1023 | if (MotionFlags & XVID_ME_QUARTERPELREFINE16) |
1925 : | Isibaar | 1136 | if(MotionFlags & XVID_ME_FASTREFINE16) |
1926 : | SubpelRefine_Fast(Data, CheckCandidate); | ||
1927 : | else | ||
1928 : | SubpelRefine(Data, CheckCandidate); | ||
1929 : | edgomez | 953 | } |
1930 : | Isibaar | 3 | |
1931 : | edgomez | 1023 | if (Data->iMinSAD[0] < (int32_t)pMB->quant * 30) |
1932 : | edgomez | 1022 | inter4v = 0; |
1933 : | Isibaar | 3 | |
1934 : | edgomez | 987 | if (inter4v) { |
1935 : | edgomez | 851 | SearchData Data8; |
1936 : | edgomez | 1053 | memcpy(&Data8, Data, sizeof(SearchData)); /* quick copy of common data */ |
1937 : | Isibaar | 3 | |
1938 : | edgomez | 851 | Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
1939 : | Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); | ||
1940 : | Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); | ||
1941 : | Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); | ||
1942 : | chl | 181 | |
1943 : | edgomez | 1107 | if ((Data->chroma) && (!(VopFlags & XVID_VOP_MODEDECISION_RD))) { |
1944 : | edgomez | 1053 | /* chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, it will not be used */ |
1945 : | edgomez | 851 | int sumx = 0, sumy = 0; |
1946 : | Isibaar | 3 | |
1947 : | edgomez | 1022 | if (Data->qpel) |
1948 : | for (i = 1; i < 5; i++) { | ||
1949 : | sumx += Data->currentQMV[i].x/2; | ||
1950 : | sumy += Data->currentQMV[i].y/2; | ||
1951 : | } | ||
1952 : | else | ||
1953 : | for (i = 1; i < 5; i++) { | ||
1954 : | sumx += Data->currentMV[i].x; | ||
1955 : | sumy += Data->currentMV[i].y; | ||
1956 : | } | ||
1957 : | chl | 259 | |
1958 : | edgomez | 851 | Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf], |
1959 : | (sumy >> 3) + roundtab_76[sumy & 0xf], Data); | ||
1960 : | } | ||
1961 : | edgomez | 1022 | } else Data->iMinSAD[1] = 4096*256; |
1962 : | edgomez | 851 | } |
1963 : | |||
1964 : | static void | ||
1965 : | Search8(const SearchData * const OldData, | ||
1966 : | const int x, const int y, | ||
1967 : | const uint32_t MotionFlags, | ||
1968 : | const MBParam * const pParam, | ||
1969 : | MACROBLOCK * const pMB, | ||
1970 : | const MACROBLOCK * const pMBs, | ||
1971 : | const int block, | ||
1972 : | SearchData * const Data) | ||
1973 : | { | ||
1974 : | int i = 0; | ||
1975 : | syskin | 1132 | CheckFunc * CheckCandidate; |
1976 : | edgomez | 851 | Data->iMinSAD = OldData->iMinSAD + 1 + block; |
1977 : | Data->currentMV = OldData->currentMV + 1 + block; | ||
1978 : | Data->currentQMV = OldData->currentQMV + 1 + block; | ||
1979 : | |||
1980 : | if(Data->qpel) { | ||
1981 : | Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); | ||
1982 : | if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y, | ||
1983 : | Data->predMV, Data->iFcode, 0, 0); | ||
1984 : | } else { | ||
1985 : | Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); | ||
1986 : | if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y, | ||
1987 : | Data->predMV, Data->iFcode, 0, Data->rrv); | ||
1988 : | Isibaar | 3 | } |
1989 : | |||
1990 : | edgomez | 851 | *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10; |
1991 : | Isibaar | 3 | |
1992 : | edgomez | 949 | if (MotionFlags & (XVID_ME_EXTSEARCH8|XVID_ME_HALFPELREFINE8|XVID_ME_QUARTERPELREFINE8)) { |
1993 : | Isibaar | 3 | |
1994 : | edgomez | 978 | if (Data->rrv) i = 16; else i = 8; |
1995 : | Isibaar | 3 | |
1996 : | edgomez | 978 | Data->RefP[0] = OldData->RefP[0] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1997 : | Data->RefP[1] = OldData->RefP[1] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); | ||
1998 : | Data->RefP[2] = OldData->RefP[2] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); | ||
1999 : | Data->RefP[3] = OldData->RefP[3] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); | ||
2000 : | |||
2001 : | Data->Cur = OldData->Cur + i * ((block&1) + Data->iEdgedWidth*(block>>1)); | ||
2002 : | edgomez | 851 | Data->qpel_precision = 0; |
2003 : | Isibaar | 3 | |
2004 : | syskin | 1129 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 3, |
2005 : | pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, Data->rrv); | ||
2006 : | Isibaar | 3 | |
2007 : | edgomez | 851 | if (!Data->rrv) CheckCandidate = CheckCandidate8; |
2008 : | else CheckCandidate = CheckCandidate16no4v; | ||
2009 : | Isibaar | 3 | |
2010 : | edgomez | 1107 | if (MotionFlags & XVID_ME_EXTSEARCH8 && (!(MotionFlags & XVID_ME_EXTSEARCH_RD))) { |
2011 : | edgomez | 1053 | int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */ |
2012 : | Isibaar | 3 | |
2013 : | edgomez | 851 | MainSearchFunc *MainSearchPtr; |
2014 : | edgomez | 949 | if (MotionFlags & XVID_ME_USESQUARES8) MainSearchPtr = SquareSearch; |
2015 : | else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; | ||
2016 : | edgomez | 851 | else MainSearchPtr = DiamondSearch; |
2017 : | Isibaar | 3 | |
2018 : | syskin | 1132 | MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255, CheckCandidate); |
2019 : | Isibaar | 3 | |
2020 : | edgomez | 851 | if(*(Data->iMinSAD) < temp_sad) { |
2021 : | edgomez | 1053 | Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */ |
2022 : | edgomez | 851 | Data->currentQMV->y = 2 * Data->currentMV->y; |
2023 : | } | ||
2024 : | } | ||
2025 : | edgomez | 195 | |
2026 : | edgomez | 949 | if (MotionFlags & XVID_ME_HALFPELREFINE8) { |
2027 : | edgomez | 1053 | int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */ |
2028 : | edgomez | 195 | |
2029 : | syskin | 1132 | SubpelRefine(Data, CheckCandidate); /* perform halfpel refine of current best vector */ |
2030 : | Isibaar | 3 | |
2031 : | edgomez | 1053 | if(*(Data->iMinSAD) < temp_sad) { /* we have found a better match */ |
2032 : | Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */ | ||
2033 : | edgomez | 851 | Data->currentQMV->y = 2 * Data->currentMV->y; |
2034 : | } | ||
2035 : | } | ||
2036 : | Isibaar | 3 | |
2037 : | edgomez | 949 | if (Data->qpel && MotionFlags & XVID_ME_QUARTERPELREFINE8) { |
2038 : | edgomez | 851 | Data->qpel_precision = 1; |
2039 : | syskin | 1129 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 3, |
2040 : | pParam->width, pParam->height, Data->iFcode, 2, 0); | ||
2041 : | syskin | 1132 | SubpelRefine(Data, CheckCandidate); |
2042 : | edgomez | 851 | } |
2043 : | } | ||
2044 : | edgomez | 195 | |
2045 : | edgomez | 851 | if (Data->rrv) { |
2046 : | Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x); | ||
2047 : | Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y); | ||
2048 : | } | ||
2049 : | |||
2050 : | if(Data->qpel) { | ||
2051 : | pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x; | ||
2052 : | pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y; | ||
2053 : | pMB->qmvs[block] = *Data->currentQMV; | ||
2054 : | edgomez | 195 | } else { |
2055 : | edgomez | 851 | pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
2056 : | pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; | ||
2057 : | edgomez | 78 | } |
2058 : | edgomez | 851 | |
2059 : | pMB->mvs[block] = *Data->currentMV; | ||
2060 : | pMB->sad8[block] = 4 * *Data->iMinSAD; | ||
2061 : | Isibaar | 3 | } |
2062 : | |||
2063 : | edgomez | 851 | /* motion estimation for B-frames */ |
2064 : | chl | 345 | |
2065 : | edgomez | 851 | static __inline VECTOR |
2066 : | ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) | ||
2067 : | { | ||
2068 : | /* the stupidiest function ever */ | ||
2069 : | return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]); | ||
2070 : | } | ||
2071 : | chl | 345 | |
2072 : | edgomez | 851 | static void __inline |
2073 : | PreparePredictionsBF(VECTOR * const pmv, const int x, const int y, | ||
2074 : | const uint32_t iWcount, | ||
2075 : | const MACROBLOCK * const pMB, | ||
2076 : | const uint32_t mode_curr) | ||
2077 : | chl | 345 | { |
2078 : | |||
2079 : | edgomez | 1053 | /* [0] is prediction */ |
2080 : | edgomez | 851 | pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); |
2081 : | chl | 345 | |
2082 : | edgomez | 1053 | pmv[1].x = pmv[1].y = 0; /* [1] is zero */ |
2083 : | chl | 345 | |
2084 : | edgomez | 851 | pmv[2] = ChoosePred(pMB, mode_curr); |
2085 : | pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); | ||
2086 : | chl | 345 | |
2087 : | edgomez | 1053 | if ((y != 0)&&(x != (int)(iWcount+1))) { /* [3] top-right neighbour */ |
2088 : | edgomez | 851 | pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); |
2089 : | pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); | ||
2090 : | } else pmv[3].x = pmv[3].y = 0; | ||
2091 : | chl | 345 | |
2092 : | edgomez | 851 | if (y != 0) { |
2093 : | pmv[4] = ChoosePred(pMB-iWcount, mode_curr); | ||
2094 : | pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y); | ||
2095 : | } else pmv[4].x = pmv[4].y = 0; | ||
2096 : | chl | 345 | |
2097 : | edgomez | 851 | if (x != 0) { |
2098 : | pmv[5] = ChoosePred(pMB-1, mode_curr); | ||
2099 : | pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); | ||
2100 : | } else pmv[5].x = pmv[5].y = 0; | ||
2101 : | chl | 345 | |
2102 : | edgomez | 851 | if (x != 0 && y != 0) { |
2103 : | pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); | ||
2104 : | pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y); | ||
2105 : | } else pmv[6].x = pmv[6].y = 0; | ||
2106 : | } | ||
2107 : | chl | 345 | |
2108 : | |||
2109 : | edgomez | 851 | /* search backward or forward */ |
2110 : | static void | ||
2111 : | SearchBF( const IMAGE * const pRef, | ||
2112 : | const uint8_t * const pRefH, | ||
2113 : | const uint8_t * const pRefV, | ||
2114 : | const uint8_t * const pRefHV, | ||
2115 : | const IMAGE * const pCur, | ||
2116 : | const int x, const int y, | ||
2117 : | const uint32_t MotionFlags, | ||
2118 : | const uint32_t iFcode, | ||
2119 : | const MBParam * const pParam, | ||
2120 : | MACROBLOCK * const pMB, | ||
2121 : | const VECTOR * const predMV, | ||
2122 : | int32_t * const best_sad, | ||
2123 : | const int32_t mode_current, | ||
2124 : | SearchData * const Data) | ||
2125 : | { | ||
2126 : | chl | 345 | |
2127 : | syskin | 1135 | int i; |
2128 : | edgomez | 851 | VECTOR pmv[7]; |
2129 : | MainSearchFunc *MainSearchPtr; | ||
2130 : | *Data->iMinSAD = MV_MAX_ERROR; | ||
2131 : | Data->iFcode = iFcode; | ||
2132 : | Data->qpel_precision = 0; | ||
2133 : | edgomez | 1053 | Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; /* reset chroma-sad cache */ |
2134 : | chl | 345 | |
2135 : | edgomez | 978 | Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16; |
2136 : | Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16; | ||
2137 : | Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16; | ||
2138 : | Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16; | ||
2139 : | Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; | ||
2140 : | Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; | ||
2141 : | chl | 345 | |
2142 : | edgomez | 851 | Data->predMV = *predMV; |
2143 : | chl | 345 | |
2144 : | syskin | 1129 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
2145 : | pParam->width, pParam->height, iFcode - Data->qpel, 1, 0); | ||
2146 : | chl | 345 | |
2147 : | edgomez | 851 | pmv[0] = Data->predMV; |
2148 : | if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } | ||
2149 : | chl | 345 | |
2150 : | edgomez | 851 | PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
2151 : | chl | 345 | |
2152 : | edgomez | 851 | Data->currentMV->x = Data->currentMV->y = 0; |
2153 : | chl | 345 | |
2154 : | edgomez | 1053 | /* main loop. checking all predictions */ |
2155 : | syskin | 1132 | for (i = 0; i < 7; i++) |
2156 : | if (!vector_repeats(pmv, i) ) | ||
2157 : | CheckCandidate16no4v(pmv[i].x, pmv[i].y, Data, i); | ||
2158 : | chl | 345 | |
2159 : | edgomez | 949 | if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch; |
2160 : | else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; | ||
2161 : | edgomez | 851 | else MainSearchPtr = DiamondSearch; |
2162 : | chl | 345 | |
2163 : | syskin | 1135 | if (*Data->iMinSAD > 512) { |
2164 : | unsigned int mask = make_mask(pmv, 7, *Data->dir); | ||
2165 : | MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate16no4v); | ||
2166 : | } | ||
2167 : | chl | 345 | |
2168 : | syskin | 1132 | SubpelRefine(Data, CheckCandidate16no4v); |
2169 : | |||
2170 : | edgomez | 851 | if (Data->qpel && *Data->iMinSAD < *best_sad + 300) { |
2171 : | Data->currentQMV->x = 2*Data->currentMV->x; | ||
2172 : | Data->currentQMV->y = 2*Data->currentMV->y; | ||
2173 : | Data->qpel_precision = 1; | ||
2174 : | syskin | 1129 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
2175 : | pParam->width, pParam->height, iFcode, 2, 0); | ||
2176 : | syskin | 1132 | SubpelRefine(Data, CheckCandidate16no4v); |
2177 : | edgomez | 851 | } |
2178 : | chl | 345 | |
2179 : | edgomez | 1053 | /* three bits are needed to code backward mode. four for forward */ |
2180 : | Isibaar | 3 | |
2181 : | edgomez | 851 | if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
2182 : | else *Data->iMinSAD += 3 * Data->lambda16; | ||
2183 : | edgomez | 195 | |
2184 : | edgomez | 851 | if (*Data->iMinSAD < *best_sad) { |
2185 : | *best_sad = *Data->iMinSAD; | ||
2186 : | pMB->mode = mode_current; | ||
2187 : | if (Data->qpel) { | ||
2188 : | pMB->pmvs[0].x = Data->currentQMV->x - predMV->x; | ||
2189 : | pMB->pmvs[0].y = Data->currentQMV->y - predMV->y; | ||
2190 : | if (mode_current == MODE_FORWARD) | ||
2191 : | pMB->qmvs[0] = *Data->currentQMV; | ||
2192 : | else | ||
2193 : | pMB->b_qmvs[0] = *Data->currentQMV; | ||
2194 : | } else { | ||
2195 : | pMB->pmvs[0].x = Data->currentMV->x - predMV->x; | ||
2196 : | pMB->pmvs[0].y = Data->currentMV->y - predMV->y; | ||
2197 : | } | ||
2198 : | if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; | ||
2199 : | else pMB->b_mvs[0] = *Data->currentMV; | ||
2200 : | } | ||
2201 : | edgomez | 195 | |
2202 : | edgomez | 851 | if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; |
2203 : | edgomez | 1053 | else *(Data->currentMV+1) = *Data->currentMV; /* we store currmv for interpolate search */ |
2204 : | Isibaar | 3 | } |
2205 : | |||
2206 : | edgomez | 851 | static void |
2207 : | SkipDecisionB(const IMAGE * const pCur, | ||
2208 : | const IMAGE * const f_Ref, | ||
2209 : | const IMAGE * const b_Ref, | ||
2210 : | MACROBLOCK * const pMB, | ||
2211 : | const uint32_t x, const uint32_t y, | ||
2212 : | const SearchData * const Data) | ||
2213 : | { | ||
2214 : | int dx = 0, dy = 0, b_dx = 0, b_dy = 0; | ||
2215 : | int32_t sum; | ||
2216 : | int k; | ||
2217 : | const uint32_t stride = Data->iEdgedWidth/2; | ||
2218 : | edgomez | 1053 | /* this is not full chroma compensation, only it's fullpel approximation. should work though */ |
2219 : | Isibaar | 3 | |
2220 : | edgomez | 851 | for (k = 0; k < 4; k++) { |
2221 : | syskin | 1135 | dy += Data->directmvF[k].y >> Data->qpel; |
2222 : | dx += Data->directmvF[k].x >> Data->qpel; | ||
2223 : | b_dy += Data->directmvB[k].y >> Data->qpel; | ||
2224 : | b_dx += Data->directmvB[k].x >> Data->qpel; | ||
2225 : | edgomez | 851 | } |
2226 : | Isibaar | 3 | |
2227 : | edgomez | 851 | dy = (dy >> 3) + roundtab_76[dy & 0xf]; |
2228 : | dx = (dx >> 3) + roundtab_76[dx & 0xf]; | ||
2229 : | b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf]; | ||
2230 : | b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf]; | ||
2231 : | Isibaar | 3 | |
2232 : | edgomez | 851 | sum = sad8bi(pCur->u + 8 * x + 8 * y * stride, |
2233 : | f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2, | ||
2234 : | b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, | ||
2235 : | stride); | ||
2236 : | Isibaar | 3 | |
2237 : | syskin | 1135 | if (sum >= MAX_CHROMA_SAD_FOR_SKIP * Data->iQuant) return; /* no skip */ |
2238 : | Isibaar | 3 | |
2239 : | edgomez | 851 | sum += sad8bi(pCur->v + 8*x + 8 * y * stride, |
2240 : | f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2, | ||
2241 : | b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, | ||
2242 : | stride); | ||
2243 : | edgomez | 195 | |
2244 : | syskin | 1135 | if (sum < MAX_CHROMA_SAD_FOR_SKIP * Data->iQuant) { |
2245 : | edgomez | 1053 | pMB->mode = MODE_DIRECT_NONE_MV; /* skipped */ |
2246 : | edgomez | 978 | for (k = 0; k < 4; k++) { |
2247 : | edgomez | 963 | pMB->qmvs[k] = pMB->mvs[k]; |
2248 : | pMB->b_qmvs[k] = pMB->b_mvs[k]; | ||
2249 : | } | ||
2250 : | } | ||
2251 : | edgomez | 851 | } |
2252 : | Isibaar | 3 | |
2253 : | edgomez | 851 | static __inline uint32_t |
2254 : | SearchDirect(const IMAGE * const f_Ref, | ||
2255 : | const uint8_t * const f_RefH, | ||
2256 : | const uint8_t * const f_RefV, | ||
2257 : | const uint8_t * const f_RefHV, | ||
2258 : | const IMAGE * const b_Ref, | ||
2259 : | const uint8_t * const b_RefH, | ||
2260 : | const uint8_t * const b_RefV, | ||
2261 : | const uint8_t * const b_RefHV, | ||
2262 : | const IMAGE * const pCur, | ||
2263 : | const int x, const int y, | ||
2264 : | const uint32_t MotionFlags, | ||
2265 : | const int32_t TRB, const int32_t TRD, | ||
2266 : | const MBParam * const pParam, | ||
2267 : | MACROBLOCK * const pMB, | ||
2268 : | const MACROBLOCK * const b_mb, | ||
2269 : | int32_t * const best_sad, | ||
2270 : | SearchData * const Data) | ||
2271 : | Isibaar | 3 | |
2272 : | edgomez | 851 | { |
2273 : | int32_t skip_sad; | ||
2274 : | int k = (x + Data->iEdgedWidth*y) * 16; | ||
2275 : | MainSearchFunc *MainSearchPtr; | ||
2276 : | syskin | 1132 | CheckFunc * CheckCandidate; |
2277 : | Isibaar | 3 | |
2278 : | edgomez | 851 | *Data->iMinSAD = 256*4096; |
2279 : | edgomez | 978 | Data->RefP[0] = f_Ref->y + k; |
2280 : | Data->RefP[2] = f_RefH + k; | ||
2281 : | Data->RefP[1] = f_RefV + k; | ||
2282 : | Data->RefP[3] = f_RefHV + k; | ||
2283 : | Data->b_RefP[0] = b_Ref->y + k; | ||
2284 : | Data->b_RefP[2] = b_RefH + k; | ||
2285 : | Data->b_RefP[1] = b_RefV + k; | ||
2286 : | Data->b_RefP[3] = b_RefHV + k; | ||
2287 : | Data->RefP[4] = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; | ||
2288 : | Data->RefP[5] = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; | ||
2289 : | Data->b_RefP[4] = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; | ||
2290 : | Data->b_RefP[5] = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; | ||
2291 : | edgomez | 195 | |
2292 : | edgomez | 851 | k = Data->qpel ? 4 : 2; |
2293 : | Data->max_dx = k * (pParam->width - x * 16); | ||
2294 : | Data->max_dy = k * (pParam->height - y * 16); | ||
2295 : | Data->min_dx = -k * (16 + x * 16); | ||
2296 : | Data->min_dy = -k * (16 + y * 16); | ||
2297 : | chl | 181 | |
2298 : | edgomez | 851 | Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs; |
2299 : | Data->qpel_precision = 0; | ||
2300 : | edgomez | 170 | |
2301 : | edgomez | 851 | for (k = 0; k < 4; k++) { |
2302 : | pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); | ||
2303 : | pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD; | ||
2304 : | pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); | ||
2305 : | pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; | ||
2306 : | Isibaar | 3 | |
2307 : | edgomez | 851 | if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx) |
2308 : | | (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) { | ||
2309 : | |||
2310 : | edgomez | 1053 | *best_sad = 256*4096; /* in that case, we won't use direct mode */ |
2311 : | pMB->mode = MODE_DIRECT; /* just to make sure it doesn't say "MODE_DIRECT_NONE_MV" */ | ||
2312 : | edgomez | 851 | pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
2313 : | return 256*4096; | ||
2314 : | } | ||
2315 : | if (b_mb->mode != MODE_INTER4V) { | ||
2316 : | pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; | ||
2317 : | pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0]; | ||
2318 : | Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0]; | ||
2319 : | Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0]; | ||
2320 : | break; | ||
2321 : | } | ||
2322 : | edgomez | 195 | } |
2323 : | Isibaar | 3 | |
2324 : | edgomez | 851 | CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v; |
2325 : | Isibaar | 3 | |
2326 : | syskin | 1132 | CheckCandidate(0, 0, Data, 255); |
2327 : | edgomez | 195 | |
2328 : | edgomez | 1053 | /* initial (fast) skip decision */ |
2329 : | syskin | 1135 | if (*Data->iMinSAD < Data->iQuant * INITIAL_SKIP_THRESH * (Data->chroma?3:2)) { |
2330 : | edgomez | 1053 | /* possible skip */ |
2331 : | edgomez | 851 | if (Data->chroma) { |
2332 : | pMB->mode = MODE_DIRECT_NONE_MV; | ||
2333 : | edgomez | 1053 | return *Data->iMinSAD; /* skip. */ |
2334 : | edgomez | 851 | } else { |
2335 : | SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data); | ||
2336 : | edgomez | 1053 | if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; /* skip. */ |
2337 : | edgomez | 851 | } |
2338 : | Isibaar | 3 | } |
2339 : | |||
2340 : | edgomez | 953 | *Data->iMinSAD += Data->lambda16; |
2341 : | edgomez | 851 | skip_sad = *Data->iMinSAD; |
2342 : | syskin | 1084 | |
2343 : | edgomez | 1053 | /* |
2344 : | * DIRECT MODE DELTA VECTOR SEARCH. | ||
2345 : | * This has to be made more effective, but at the moment I'm happy it's running at all | ||
2346 : | */ | ||
2347 : | edgomez | 195 | |
2348 : | edgomez | 949 | if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch; |
2349 : | else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; | ||
2350 : | edgomez | 851 | else MainSearchPtr = DiamondSearch; |
2351 : | Isibaar | 3 | |
2352 : | syskin | 1132 | MainSearchPtr(0, 0, Data, 255, CheckCandidate); |
2353 : | Isibaar | 3 | |
2354 : | syskin | 1132 | SubpelRefine(Data, CheckCandidate); |
2355 : | chl | 181 | |
2356 : | edgomez | 851 | *best_sad = *Data->iMinSAD; |
2357 : | chl | 181 | |
2358 : | edgomez | 851 | if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT; |
2359 : | edgomez | 1053 | else pMB->mode = MODE_DIRECT_NO4V; /* for faster compensation */ |
2360 : | chl | 181 | |
2361 : | edgomez | 851 | pMB->pmvs[3] = *Data->currentMV; |
2362 : | edgomez | 195 | |
2363 : | edgomez | 851 | for (k = 0; k < 4; k++) { |
2364 : | pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x; | ||
2365 : | pMB->b_mvs[k].x = ( (Data->currentMV->x == 0) | ||
2366 : | ? Data->directmvB[k].x | ||
2367 : | :pMB->mvs[k].x - Data->referencemv[k].x); | ||
2368 : | pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y); | ||
2369 : | pMB->b_mvs[k].y = ((Data->currentMV->y == 0) | ||
2370 : | ? Data->directmvB[k].y | ||
2371 : | : pMB->mvs[k].y - Data->referencemv[k].y); | ||
2372 : | if (Data->qpel) { | ||
2373 : | pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2; | ||
2374 : | pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2; | ||
2375 : | pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2; | ||
2376 : | pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2; | ||
2377 : | } | ||
2378 : | edgomez | 195 | |
2379 : | edgomez | 851 | if (b_mb->mode != MODE_INTER4V) { |
2380 : | pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; | ||
2381 : | pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; | ||
2382 : | pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0]; | ||
2383 : | pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0]; | ||
2384 : | break; | ||
2385 : | } | ||
2386 : | edgomez | 78 | } |
2387 : | edgomez | 851 | return skip_sad; |
2388 : | } | ||
2389 : | Isibaar | 3 | |
2390 : | edgomez | 851 | static void |
2391 : | SearchInterpolate(const IMAGE * const f_Ref, | ||
2392 : | const uint8_t * const f_RefH, | ||
2393 : | const uint8_t * const f_RefV, | ||
2394 : | const uint8_t * const f_RefHV, | ||
2395 : | const IMAGE * const b_Ref, | ||
2396 : | const uint8_t * const b_RefH, | ||
2397 : | const uint8_t * const b_RefV, | ||
2398 : | const uint8_t * const b_RefHV, | ||
2399 : | const IMAGE * const pCur, | ||
2400 : | const int x, const int y, | ||
2401 : | const uint32_t fcode, | ||
2402 : | const uint32_t bcode, | ||
2403 : | const uint32_t MotionFlags, | ||
2404 : | const MBParam * const pParam, | ||
2405 : | const VECTOR * const f_predMV, | ||
2406 : | const VECTOR * const b_predMV, | ||
2407 : | MACROBLOCK * const pMB, | ||
2408 : | int32_t * const best_sad, | ||
2409 : | SearchData * const fData) | ||
2410 : | chl | 96 | |
2411 : | edgomez | 851 | { |
2412 : | chl | 169 | |
2413 : | syskin | 1132 | int i, j; |
2414 : | edgomez | 851 | SearchData bData; |
2415 : | chl | 169 | |
2416 : | edgomez | 851 | fData->qpel_precision = 0; |
2417 : | edgomez | 1053 | memcpy(&bData, fData, sizeof(SearchData)); /* quick copy of common data */ |
2418 : | edgomez | 851 | *fData->iMinSAD = 4096*256; |
2419 : | bData.currentMV++; bData.currentQMV++; | ||
2420 : | fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; | ||
2421 : | chl | 169 | |
2422 : | edgomez | 851 | i = (x + y * fData->iEdgedWidth) * 16; |
2423 : | chl | 169 | |
2424 : | edgomez | 978 | bData.b_RefP[0] = fData->RefP[0] = f_Ref->y + i; |
2425 : | bData.b_RefP[2] = fData->RefP[2] = f_RefH + i; | ||
2426 : | bData.b_RefP[1] = fData->RefP[1] = f_RefV + i; | ||
2427 : | bData.b_RefP[3] = fData->RefP[3] = f_RefHV + i; | ||
2428 : | bData.RefP[0] = fData->b_RefP[0] = b_Ref->y + i; | ||
2429 : | bData.RefP[2] = fData->b_RefP[2] = b_RefH + i; | ||
2430 : | bData.RefP[1] = fData->b_RefP[1] = b_RefV + i; | ||
2431 : | bData.RefP[3] = fData->b_RefP[3] = b_RefHV + i; | ||
2432 : | bData.b_RefP[4] = fData->RefP[4] = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; | ||
2433 : | bData.b_RefP[5] = fData->RefP[5] = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; | ||
2434 : | bData.RefP[4] = fData->b_RefP[4] = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; | ||
2435 : | bData.RefP[5] = fData->b_RefP[5] = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; | ||
2436 : | syskin | 1132 | bData.dir = fData->dir; |
2437 : | chl | 169 | |
2438 : | edgomez | 851 | bData.bpredMV = fData->predMV = *f_predMV; |
2439 : | fData->bpredMV = bData.predMV = *b_predMV; | ||
2440 : | fData->currentMV[0] = fData->currentMV[2]; | ||
2441 : | Isibaar | 3 | |
2442 : | syskin | 1129 | get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 4, pParam->width, pParam->height, fcode - fData->qpel, 1, 0); |
2443 : | get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 4, pParam->width, pParam->height, bcode - fData->qpel, 1, 0); | ||
2444 : | Isibaar | 3 | |
2445 : | edgomez | 851 | if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; |
2446 : | if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx; | ||
2447 : | if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy; | ||
2448 : | if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy; | ||
2449 : | chl | 169 | |
2450 : | edgomez | 851 | if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx; |
2451 : | if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx; | ||
2452 : | if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy; | ||
2453 : | if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy; | ||
2454 : | Isibaar | 3 | |
2455 : | syskin | 1132 | CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, fData, 255); |
2456 : | chl | 169 | |
2457 : | edgomez | 1053 | /* diamond */ |
2458 : | edgomez | 851 | do { |
2459 : | syskin | 1132 | *fData->dir = 255; |
2460 : | edgomez | 1053 | /* forward MV moves */ |
2461 : | edgomez | 851 | i = fData->currentMV[0].x; j = fData->currentMV[0].y; |
2462 : | chl | 169 | |
2463 : | syskin | 1132 | CheckCandidateInt(i + 1, j, fData, 0); |
2464 : | CheckCandidateInt(i, j + 1, fData, 0); | ||
2465 : | CheckCandidateInt(i - 1, j, fData, 0); | ||
2466 : | CheckCandidateInt(i, j - 1, fData, 0); | ||
2467 : | edgomez | 195 | |
2468 : | edgomez | 1053 | /* backward MV moves */ |
2469 : | edgomez | 851 | i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
2470 : | fData->currentMV[2] = fData->currentMV[0]; | ||
2471 : | syskin | 1132 | CheckCandidateInt(i + 1, j, &bData, 0); |
2472 : | CheckCandidateInt(i, j + 1, &bData, 0); | ||
2473 : | CheckCandidateInt(i - 1, j, &bData, 0); | ||
2474 : | CheckCandidateInt(i, j - 1, &bData, 0); | ||
2475 : | Isibaar | 3 | |
2476 : | syskin | 1132 | } while (!(*fData->dir)); |
2477 : | chl | 140 | |
2478 : | edgomez | 1053 | /* qpel refinement */ |
2479 : | edgomez | 851 | if (fData->qpel) { |
2480 : | if (*fData->iMinSAD > *best_sad + 500) return; | ||
2481 : | fData->qpel_precision = bData.qpel_precision = 1; | ||
2482 : | syskin | 1129 | get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 4, pParam->width, pParam->height, fcode, 2, 0); |
2483 : | get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 4, pParam->width, pParam->height, bcode, 2, 0); | ||
2484 : | edgomez | 851 | fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x; |
2485 : | fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y; | ||
2486 : | fData->currentQMV[1].x = 2 * fData->currentMV[1].x; | ||
2487 : | fData->currentQMV[1].y = 2 * fData->currentMV[1].y; | ||
2488 : | syskin | 1132 | SubpelRefine(fData, CheckCandidateInt); |
2489 : | edgomez | 851 | if (*fData->iMinSAD > *best_sad + 300) return; |
2490 : | fData->currentQMV[2] = fData->currentQMV[0]; | ||
2491 : | syskin | 1132 | SubpelRefine(&bData, CheckCandidateInt); |
2492 : | edgomez | 851 | } |
2493 : | chl | 140 | |
2494 : | edgomez | 1053 | *fData->iMinSAD += (2+3) * fData->lambda16; /* two bits are needed to code interpolate mode. */ |
2495 : | Isibaar | 3 | |
2496 : | edgomez | 851 | if (*fData->iMinSAD < *best_sad) { |
2497 : | *best_sad = *fData->iMinSAD; | ||
2498 : | pMB->mvs[0] = fData->currentMV[0]; | ||
2499 : | pMB->b_mvs[0] = fData->currentMV[1]; | ||
2500 : | pMB->mode = MODE_INTERPOLATE; | ||
2501 : | if (fData->qpel) { | ||
2502 : | pMB->qmvs[0] = fData->currentQMV[0]; | ||
2503 : | pMB->b_qmvs[0] = fData->currentQMV[1]; | ||
2504 : | pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x; | ||
2505 : | pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y; | ||
2506 : | pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x; | ||
2507 : | pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y; | ||
2508 : | } else { | ||
2509 : | pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; | ||
2510 : | pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; | ||
2511 : | pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; | ||
2512 : | pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; | ||
2513 : | } | ||
2514 : | edgomez | 78 | } |
2515 : | edgomez | 851 | } |
2516 : | Isibaar | 3 | |
2517 : | edgomez | 851 | void |
2518 : | MotionEstimationBVOP(MBParam * const pParam, | ||
2519 : | FRAMEINFO * const frame, | ||
2520 : | const int32_t time_bp, | ||
2521 : | const int32_t time_pp, | ||
2522 : | edgomez | 1053 | /* forward (past) reference */ |
2523 : | edgomez | 851 | const MACROBLOCK * const f_mbs, |
2524 : | const IMAGE * const f_ref, | ||
2525 : | const IMAGE * const f_refH, | ||
2526 : | const IMAGE * const f_refV, | ||
2527 : | const IMAGE * const f_refHV, | ||
2528 : | edgomez | 1053 | /* backward (future) reference */ |
2529 : | edgomez | 851 | const FRAMEINFO * const b_reference, |
2530 : | const IMAGE * const b_ref, | ||
2531 : | const IMAGE * const b_refH, | ||
2532 : | const IMAGE * const b_refV, | ||
2533 : | const IMAGE * const b_refHV) | ||
2534 : | { | ||
2535 : | uint32_t i, j; | ||
2536 : | int32_t best_sad; | ||
2537 : | uint32_t skip_sad; | ||
2538 : | int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; | ||
2539 : | const MACROBLOCK * const b_mbs = b_reference->mbs; | ||
2540 : | Isibaar | 3 | |
2541 : | edgomez | 851 | VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
2542 : | Isibaar | 3 | |
2543 : | edgomez | 851 | const int32_t TRB = time_pp - time_bp; |
2544 : | const int32_t TRD = time_pp; | ||
2545 : | edgomez | 195 | |
2546 : | edgomez | 1053 | /* some pre-inintialized data for the rest of the search */ |
2547 : | Isibaar | 3 | |
2548 : | edgomez | 851 | SearchData Data; |
2549 : | int32_t iMinSAD; | ||
2550 : | syskin | 1132 | uint32_t dir; |
2551 : | edgomez | 851 | VECTOR currentMV[3]; |
2552 : | VECTOR currentQMV[3]; | ||
2553 : | int32_t temp[8]; | ||
2554 : | memset(&Data, 0, sizeof(SearchData)); | ||
2555 : | Data.iEdgedWidth = pParam->edged_width; | ||
2556 : | Data.currentMV = currentMV; Data.currentQMV = currentQMV; | ||
2557 : | Data.iMinSAD = &iMinSAD; | ||
2558 : | syskin | 1135 | Data.lambda16 = lambda_vec16[MAX(frame->quant-2, 2)]; |
2559 : | syskin | 1112 | Data.qpel = pParam->vol_flags & XVID_VOL_QUARTERPEL ? 1 : 0; |
2560 : | edgomez | 851 | Data.rounding = 0; |
2561 : | edgomez | 1107 | Data.chroma = frame->motion_flags & XVID_ME_CHROMA_BVOP; |
2562 : | edgomez | 851 | Data.temp = temp; |
2563 : | syskin | 1132 | Data.dir = &dir; |
2564 : | syskin | 1135 | Data.iQuant = frame->quant; |
2565 : | Isibaar | 3 | |
2566 : | edgomez | 1053 | Data.RefQ = f_refV->u; /* a good place, also used in MC (for similar purpose) */ |
2567 : | |||
2568 : | /* note: i==horizontal, j==vertical */ | ||
2569 : | edgomez | 851 | for (j = 0; j < pParam->mb_height; j++) { |
2570 : | edgomez | 195 | |
2571 : | edgomez | 851 | f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
2572 : | Isibaar | 3 | |
2573 : | edgomez | 851 | for (i = 0; i < pParam->mb_width; i++) { |
2574 : | MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; | ||
2575 : | const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; | ||
2576 : | edgomez | 195 | |
2577 : | edgomez | 851 | /* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */ |
2578 : | if (b_reference->coding_type != S_VOP) | ||
2579 : | if (b_mb->mode == MODE_NOT_CODED) { | ||
2580 : | pMB->mode = MODE_NOT_CODED; | ||
2581 : | continue; | ||
2582 : | } | ||
2583 : | Isibaar | 3 | |
2584 : | edgomez | 851 | Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
2585 : | Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8; | ||
2586 : | Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8; | ||
2587 : | Isibaar | 3 | |
2588 : | edgomez | 851 | /* direct search comes first, because it (1) checks for SKIP-mode |
2589 : | and (2) sets very good predictions for forward and backward search */ | ||
2590 : | skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y, | ||
2591 : | b_ref, b_refH->y, b_refV->y, b_refHV->y, | ||
2592 : | &frame->image, | ||
2593 : | i, j, | ||
2594 : | frame->motion_flags, | ||
2595 : | TRB, TRD, | ||
2596 : | pParam, | ||
2597 : | pMB, b_mb, | ||
2598 : | &best_sad, | ||
2599 : | &Data); | ||
2600 : | Isibaar | 3 | |
2601 : | edgomez | 851 | if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
2602 : | edgomez | 195 | |
2603 : | edgomez | 1053 | /* forward search */ |
2604 : | edgomez | 851 | SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
2605 : | &frame->image, i, j, | ||
2606 : | frame->motion_flags, | ||
2607 : | frame->fcode, pParam, | ||
2608 : | pMB, &f_predMV, &best_sad, | ||
2609 : | MODE_FORWARD, &Data); | ||
2610 : | edgomez | 195 | |
2611 : | edgomez | 1053 | /* backward search */ |
2612 : | edgomez | 851 | SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y, |
2613 : | &frame->image, i, j, | ||
2614 : | frame->motion_flags, | ||
2615 : | frame->bcode, pParam, | ||
2616 : | pMB, &b_predMV, &best_sad, | ||
2617 : | MODE_BACKWARD, &Data); | ||
2618 : | |||
2619 : | edgomez | 1053 | /* interpolate search comes last, because it uses data from forward and backward as prediction */ |
2620 : | edgomez | 851 | SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
2621 : | b_ref, b_refH->y, b_refV->y, b_refHV->y, | ||
2622 : | &frame->image, | ||
2623 : | i, j, | ||
2624 : | frame->fcode, frame->bcode, | ||
2625 : | frame->motion_flags, | ||
2626 : | pParam, | ||
2627 : | &f_predMV, &b_predMV, | ||
2628 : | pMB, &best_sad, | ||
2629 : | &Data); | ||
2630 : | |||
2631 : | edgomez | 1053 | /* final skip decision */ |
2632 : | syskin | 1135 | if ( (skip_sad < Data.iQuant * MAX_SAD00_FOR_SKIP * 2) |
2633 : | edgomez | 851 | && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) |
2634 : | SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data); | ||
2635 : | |||
2636 : | switch (pMB->mode) { | ||
2637 : | case MODE_FORWARD: | ||
2638 : | f_count++; | ||
2639 : | f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; | ||
2640 : | break; | ||
2641 : | case MODE_BACKWARD: | ||
2642 : | b_count++; | ||
2643 : | b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; | ||
2644 : | break; | ||
2645 : | case MODE_INTERPOLATE: | ||
2646 : | i_count++; | ||
2647 : | f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; | ||
2648 : | b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; | ||
2649 : | break; | ||
2650 : | case MODE_DIRECT: | ||
2651 : | case MODE_DIRECT_NO4V: | ||
2652 : | d_count++; | ||
2653 : | default: | ||
2654 : | break; | ||
2655 : | } | ||
2656 : | } | ||
2657 : | } | ||
2658 : | Isibaar | 3 | } |
2659 : | chl | 96 | |
2660 : | edgomez | 851 | static __inline void |
2661 : | MEanalyzeMB ( const uint8_t * const pRef, | ||
2662 : | const uint8_t * const pCur, | ||
2663 : | const int x, | ||
2664 : | const int y, | ||
2665 : | const MBParam * const pParam, | ||
2666 : | MACROBLOCK * const pMBs, | ||
2667 : | SearchData * const Data) | ||
2668 : | chl | 96 | { |
2669 : | |||
2670 : | syskin | 1132 | int i; |
2671 : | edgomez | 851 | VECTOR pmv[3]; |
2672 : | edgomez | 953 | MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width]; |
2673 : | chl | 96 | |
2674 : | syskin | 1135 | unsigned int simplicity = 0; |
2675 : | |||
2676 : | edgomez | 851 | for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
2677 : | chl | 96 | |
2678 : | syskin | 1129 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
2679 : | syskin | 1135 | pParam->width, pParam->height, Data->iFcode - Data->qpel - 1, 0, 0); |
2680 : | edgomez | 195 | |
2681 : | edgomez | 851 | Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
2682 : | edgomez | 978 | Data->RefP[0] = pRef + (x + y * pParam->edged_width) * 16; |
2683 : | chl | 96 | |
2684 : | syskin | 1135 | pmv[0].x = pMB->mvs[0].x; |
2685 : | pmv[0].y = pMB->mvs[0].y; | ||
2686 : | edgomez | 195 | |
2687 : | syskin | 1135 | CheckCandidate32I(pmv[0].x, pmv[0].y, Data, 0); |
2688 : | edgomez | 195 | |
2689 : | syskin | 1135 | if (*Data->iMinSAD > 200) { |
2690 : | chl | 96 | |
2691 : | syskin | 1135 | pmv[1].x = pmv[1].y = 0; |
2692 : | |||
2693 : | /* median is only used as prediction. it doesn't have to be real */ | ||
2694 : | if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; | ||
2695 : | else | ||
2696 : | if (x == 1) /* left macroblock does not have any vector now */ | ||
2697 : | Data->predMV = (pMB - pParam->mb_width)->mvs[0]; /* top instead of median */ | ||
2698 : | else if (y == 1) /* top macroblock doesn't have it's vector */ | ||
2699 : | Data->predMV = (pMB - 1)->mvs[0]; /* left instead of median */ | ||
2700 : | else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); /* else median */ | ||
2701 : | |||
2702 : | pmv[2].x = Data->predMV.x; | ||
2703 : | pmv[2].y = Data->predMV.y; | ||
2704 : | |||
2705 : | syskin | 1133 | if (!vector_repeats(pmv, 1)) |
2706 : | syskin | 1132 | CheckCandidate32I(pmv[1].x, pmv[1].y, Data, 1); |
2707 : | syskin | 1133 | if (!vector_repeats(pmv, 2)) |
2708 : | syskin | 1132 | CheckCandidate32I(pmv[2].x, pmv[2].y, Data, 2); |
2709 : | chl | 96 | |
2710 : | syskin | 1135 | if (*Data->iMinSAD > 500) { /* diamond only if needed */ |
2711 : | syskin | 1133 | unsigned int mask = make_mask(pmv, 3, *Data->dir); |
2712 : | syskin | 1134 | DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate32I); |
2713 : | syskin | 1135 | } else simplicity++; |
2714 : | edgomez | 851 | |
2715 : | syskin | 1135 | if (*Data->iMinSAD > 500) /* refinement from 2-pixel to 1-pixel */ |
2716 : | SubpelRefine(Data, CheckCandidate32I); | ||
2717 : | else simplicity++; | ||
2718 : | } else simplicity++; | ||
2719 : | |||
2720 : | edgomez | 953 | for (i = 0; i < 4; i++) { |
2721 : | MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width]; | ||
2722 : | MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i]; | ||
2723 : | MB->mode = MODE_INTER; | ||
2724 : | syskin | 1135 | /* if we skipped some search steps, we have to assume that SAD would be lower with them */ |
2725 : | MB->sad16 = Data->iMinSAD[i+1] - (simplicity<<7); | ||
2726 : | chl | 96 | } |
2727 : | edgomez | 851 | } |
2728 : | chl | 96 | |
2729 : | syskin | 1075 | #define INTRA_THRESH 2200 |
2730 : | syskin | 1135 | #define INTER_THRESH 40 |
2731 : | syskin | 1075 | #define INTRA_THRESH2 95 |
2732 : | chl | 96 | |
2733 : | edgomez | 851 | int |
2734 : | MEanalysis( const IMAGE * const pRef, | ||
2735 : | edgomez | 953 | const FRAMEINFO * const Current, |
2736 : | const MBParam * const pParam, | ||
2737 : | syskin | 1075 | const int maxIntra, //maximum number if non-I frames |
2738 : | const int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame | ||
2739 : | syskin | 1084 | const int bCount, // number of B frames in a row |
2740 : | edgomez | 953 | const int b_thresh) |
2741 : | edgomez | 851 | { |
2742 : | uint32_t x, y, intra = 0; | ||
2743 : | int sSAD = 0; | ||
2744 : | MACROBLOCK * const pMBs = Current->mbs; | ||
2745 : | const IMAGE * const pCurrent = &Current->image; | ||
2746 : | syskin | 1075 | int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + b_thresh; |
2747 : | chl | 1077 | int blocks = 0; |
2748 : | syskin | 1075 | int complexity = 0; |
2749 : | chl | 96 | |
2750 : | edgomez | 851 | int32_t iMinSAD[5], temp[5]; |
2751 : | syskin | 1132 | uint32_t dir; |
2752 : | edgomez | 851 | VECTOR currentMV[5]; |
2753 : | SearchData Data; | ||
2754 : | Data.iEdgedWidth = pParam->edged_width; | ||
2755 : | Data.currentMV = currentMV; | ||
2756 : | Data.iMinSAD = iMinSAD; | ||
2757 : | Data.iFcode = Current->fcode; | ||
2758 : | Data.temp = temp; | ||
2759 : | syskin | 1132 | Data.dir = &dir; |
2760 : | Data.qpel = (pParam->vol_flags & XVID_VOL_QUARTERPEL)? 1: 0; | ||
2761 : | syskin | 1135 | Data.qpel_precision = 0; |
2762 : | chl | 96 | |
2763 : | edgomez | 1023 | if (intraCount != 0) { |
2764 : | syskin | 1075 | if (intraCount < 10) // we're right after an I frame |
2765 : | edgomez | 1023 | IntraThresh += 15* (intraCount - 10) * (intraCount - 10); |
2766 : | else | ||
2767 : | syskin | 1075 | if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec |
2768 : | edgomez | 1023 | IntraThresh -= (IntraThresh * (maxIntra - 8*(maxIntra - intraCount)))/maxIntra; |
2769 : | } | ||
2770 : | chl | 96 | |
2771 : | syskin | 1135 | InterThresh -= 20 * bCount; |
2772 : | if (InterThresh < 10 + b_thresh) InterThresh = 10 + b_thresh; | ||
2773 : | chl | 326 | |
2774 : | edgomez | 851 | if (sadInit) (*sadInit) (); |
2775 : | chl | 96 | |
2776 : | edgomez | 851 | for (y = 1; y < pParam->mb_height-1; y += 2) { |
2777 : | for (x = 1; x < pParam->mb_width-1; x += 2) { | ||
2778 : | int i; | ||
2779 : | syskin | 1075 | blocks += 10; |
2780 : | edgomez | 195 | |
2781 : | edgomez | 851 | if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; |
2782 : | syskin | 1075 | else { //extrapolation of the vector found for last frame |
2783 : | edgomez | 978 | pMBs[x + y * pParam->mb_width].mvs[0].x = |
2784 : | edgomez | 953 | (pMBs[x + y * pParam->mb_width].mvs[0].x * (bCount+1) ) / bCount; |
2785 : | edgomez | 978 | pMBs[x + y * pParam->mb_width].mvs[0].y = |
2786 : | edgomez | 953 | (pMBs[x + y * pParam->mb_width].mvs[0].y * (bCount+1) ) / bCount; |
2787 : | } | ||
2788 : | edgomez | 195 | |
2789 : | edgomez | 851 | MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data); |
2790 : | edgomez | 195 | |
2791 : | edgomez | 851 | for (i = 0; i < 4; i++) { |
2792 : | int dev; | ||
2793 : | MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width]; | ||
2794 : | syskin | 1075 | dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16, |
2795 : | pParam->edged_width); | ||
2796 : | syskin | 1084 | |
2797 : | syskin | 1108 | complexity += MAX(dev, 300); |
2798 : | syskin | 1075 | if (dev + IntraThresh < pMB->sad16) { |
2799 : | pMB->mode = MODE_INTRA; | ||
2800 : | if (++intra > ((pParam->mb_height-2)*(pParam->mb_width-2))/2) return I_VOP; | ||
2801 : | edgomez | 851 | } |
2802 : | edgomez | 953 | |
2803 : | syskin | 1084 | if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0) |
2804 : | syskin | 1135 | if (dev > 1000 && pMB->sad16 < 1000) |
2805 : | syskin | 1075 | sSAD += 1000; |
2806 : | |||
2807 : | syskin | 1135 | sSAD += (dev < 4000) ? pMB->sad16 : pMB->sad16/2; /* blocks with big contrast differences usually have large SAD - while they look very good in b-frames */ |
2808 : | edgomez | 851 | } |
2809 : | } | ||
2810 : | edgomez | 195 | } |
2811 : | syskin | 1075 | complexity >>= 7; |
2812 : | edgomez | 953 | |
2813 : | syskin | 1075 | sSAD /= complexity + 4*blocks; |
2814 : | edgomez | 953 | |
2815 : | syskin | 1108 | if (intraCount > 80 && sSAD > INTRA_THRESH2 ) return I_VOP; |
2816 : | edgomez | 851 | if (sSAD > InterThresh ) return P_VOP; |
2817 : | emms(); | ||
2818 : | return B_VOP; | ||
2819 : | } | ||
2820 : | chl | 96 | |
2821 : | |||
2822 : | edgomez | 1053 | /* functions which perform BITS-based search/bitcount */ |
2823 : | chl | 96 | |
2824 : | edgomez | 851 | static int |
2825 : | syskin | 1108 | findRDinter(SearchData * const Data, |
2826 : | const MACROBLOCK * const pMBs, const int x, const int y, | ||
2827 : | const MBParam * const pParam, | ||
2828 : | const uint32_t MotionFlags) | ||
2829 : | chl | 96 | { |
2830 : | syskin | 1132 | int i; |
2831 : | edgomez | 851 | int32_t bsad[5]; |
2832 : | chl | 141 | |
2833 : | edgomez | 851 | if (Data->qpel) { |
2834 : | for(i = 0; i < 5; i++) { | ||
2835 : | Data->currentMV[i].x = Data->currentQMV[i].x/2; | ||
2836 : | Data->currentMV[i].y = Data->currentQMV[i].y/2; | ||
2837 : | } | ||
2838 : | Data->qpel_precision = 1; | ||
2839 : | syskin | 1132 | CheckCandidateRD16(Data->currentQMV[0].x, Data->currentQMV[0].y, Data, 255); |
2840 : | chl | 96 | |
2841 : | edgomez | 1107 | if (MotionFlags & (XVID_ME_HALFPELREFINE16_RD | XVID_ME_EXTSEARCH_RD)) { /* we have to prepare for halfpixel-precision search */ |
2842 : | edgomez | 851 | for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i]; |
2843 : | syskin | 1129 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
2844 : | pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, Data->rrv); | ||
2845 : | edgomez | 851 | Data->qpel_precision = 0; |
2846 : | if (Data->currentQMV->x & 1 || Data->currentQMV->y & 1) | ||
2847 : | syskin | 1132 | CheckCandidateRD16(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255); |
2848 : | edgomez | 851 | } |
2849 : | edgomez | 195 | |
2850 : | edgomez | 1053 | } else { /* not qpel */ |
2851 : | edgomez | 195 | |
2852 : | syskin | 1132 | CheckCandidateRD16(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255); |
2853 : | edgomez | 851 | } |
2854 : | chl | 96 | |
2855 : | syskin | 1132 | if (MotionFlags&XVID_ME_EXTSEARCH_RD) |
2856 : | SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, 255, CheckCandidateRD16); | ||
2857 : | chl | 96 | |
2858 : | syskin | 1132 | if (MotionFlags&XVID_ME_HALFPELREFINE16_RD) |
2859 : | SubpelRefine(Data, CheckCandidateRD16); | ||
2860 : | chl | 96 | |
2861 : | edgomez | 851 | if (Data->qpel) { |
2862 : | edgomez | 1107 | if (MotionFlags&(XVID_ME_EXTSEARCH_RD | XVID_ME_HALFPELREFINE16_RD)) { /* there was halfpel-precision search */ |
2863 : | edgomez | 851 | for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) { |
2864 : | edgomez | 1053 | Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* we have found a better match */ |
2865 : | edgomez | 851 | Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
2866 : | } | ||
2867 : | edgomez | 195 | |
2868 : | edgomez | 1053 | /* preparing for qpel-precision search */ |
2869 : | edgomez | 851 | Data->qpel_precision = 1; |
2870 : | syskin | 1129 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
2871 : | pParam->width, pParam->height, Data->iFcode, 2, 0); | ||
2872 : | edgomez | 851 | } |
2873 : | syskin | 1132 | if (MotionFlags&XVID_ME_QUARTERPELREFINE16_RD) |
2874 : | SubpelRefine(Data, CheckCandidateRD16); | ||
2875 : | edgomez | 851 | } |
2876 : | chl | 96 | |
2877 : | edgomez | 1107 | if (MotionFlags&XVID_ME_CHECKPREDICTION_RD) { /* let's check vector equal to prediction */ |
2878 : | edgomez | 851 | VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV; |
2879 : | if (!(Data->predMV.x == v->x && Data->predMV.y == v->y)) | ||
2880 : | syskin | 1132 | CheckCandidateRD16(Data->predMV.x, Data->predMV.y, Data, 255); |
2881 : | edgomez | 851 | } |
2882 : | return Data->iMinSAD[0]; | ||
2883 : | } | ||
2884 : | chl | 96 | |
2885 : | edgomez | 851 | static int |
2886 : | syskin | 1108 | findRDinter4v(const SearchData * const Data, |
2887 : | MACROBLOCK * const pMB, const MACROBLOCK * const pMBs, | ||
2888 : | const int x, const int y, | ||
2889 : | const MBParam * const pParam, const uint32_t MotionFlags, | ||
2890 : | const VECTOR * const backup) | ||
2891 : | edgomez | 851 | { |
2892 : | chl | 96 | |
2893 : | syskin | 1132 | int cbp = 0, bits = 0, t = 0, i; |
2894 : | edgomez | 851 | SearchData Data2, *Data8 = &Data2; |
2895 : | int sumx = 0, sumy = 0; | ||
2896 : | edgomez | 962 | int16_t *in = Data->dctSpace, *coeff = Data->dctSpace + 64; |
2897 : | edgomez | 1022 | uint8_t * ptr; |
2898 : | chl | 96 | |
2899 : | edgomez | 851 | memcpy(Data8, Data, sizeof(SearchData)); |
2900 : | syskin | 1132 | |
2901 : | edgomez | 1053 | for (i = 0; i < 4; i++) { /* for all luma blocks */ |
2902 : | edgomez | 1022 | |
2903 : | edgomez | 851 | Data8->iMinSAD = Data->iMinSAD + i + 1; |
2904 : | Data8->currentMV = Data->currentMV + i + 1; | ||
2905 : | Data8->currentQMV = Data->currentQMV + i + 1; | ||
2906 : | Data8->Cur = Data->Cur + 8*((i&1) + (i>>1)*Data->iEdgedWidth); | ||
2907 : | edgomez | 978 | Data8->RefP[0] = Data->RefP[0] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2908 : | Data8->RefP[2] = Data->RefP[2] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); | ||
2909 : | Data8->RefP[1] = Data->RefP[1] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); | ||
2910 : | Data8->RefP[3] = Data->RefP[3] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); | ||
2911 : | syskin | 1084 | *Data8->cbp = (Data->cbp[1] & (1<<(5-i))) ? 1:0; // copy corresponding cbp bit |
2912 : | edgomez | 195 | |
2913 : | edgomez | 851 | if(Data->qpel) { |
2914 : | Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i); | ||
2915 : | if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y, | ||
2916 : | Data8->predMV, Data8->iFcode, 0, 0); | ||
2917 : | } else { | ||
2918 : | Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i); | ||
2919 : | if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y, | ||
2920 : | Data8->predMV, Data8->iFcode, 0, 0); | ||
2921 : | } | ||
2922 : | chl | 96 | |
2923 : | syskin | 1129 | get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 3, |
2924 : | pParam->width, pParam->height, Data8->iFcode, Data8->qpel+1, 0); | ||
2925 : | chl | 96 | |
2926 : | edgomez | 1023 | *Data8->iMinSAD += BITS_MULT*t; |
2927 : | chl | 96 | |
2928 : | edgomez | 851 | Data8->qpel_precision = Data8->qpel; |
2929 : | edgomez | 1053 | /* checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far) */ |
2930 : | edgomez | 987 | { |
2931 : | VECTOR *v = Data8->qpel ? Data8->currentQMV : Data8->currentMV; | ||
2932 : | edgomez | 1022 | if (!MVequal (*v, backup[i+1]) ) |
2933 : | syskin | 1132 | CheckCandidateRD8(backup[i+1].x, backup[i+1].y, Data8, 255); |
2934 : | edgomez | 851 | } |
2935 : | edgomez | 195 | |
2936 : | edgomez | 851 | if (Data8->qpel) { |
2937 : | edgomez | 1107 | if (MotionFlags&XVID_ME_HALFPELREFINE8_RD || (MotionFlags&XVID_ME_EXTSEARCH8 && MotionFlags&XVID_ME_EXTSEARCH_RD)) { /* halfpixel motion search follows */ |
2938 : | edgomez | 851 | int32_t s = *Data8->iMinSAD; |
2939 : | Data8->currentMV->x = Data8->currentQMV->x/2; | ||
2940 : | Data8->currentMV->y = Data8->currentQMV->y/2; | ||
2941 : | Data8->qpel_precision = 0; | ||
2942 : | syskin | 1129 | get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 3, |
2943 : | pParam->width, pParam->height, Data8->iFcode - 1, 1, 0); | ||
2944 : | edgomez | 195 | |
2945 : | edgomez | 851 | if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1) |
2946 : | syskin | 1132 | CheckCandidateRD8(Data8->currentMV->x, Data8->currentMV->y, Data8, 255); |
2947 : | edgomez | 195 | |
2948 : | edgomez | 1107 | if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_RD) |
2949 : | syskin | 1132 | SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255, CheckCandidateRD8); |
2950 : | chl | 96 | |
2951 : | edgomez | 1107 | if (MotionFlags & XVID_ME_HALFPELREFINE8_RD) |
2952 : | syskin | 1132 | SubpelRefine(Data8, CheckCandidateRD8); |
2953 : | chl | 96 | |
2954 : | edgomez | 1053 | if(s > *Data8->iMinSAD) { /* we have found a better match */ |
2955 : | edgomez | 851 | Data8->currentQMV->x = 2*Data8->currentMV->x; |
2956 : | Data8->currentQMV->y = 2*Data8->currentMV->y; | ||
2957 : | } | ||
2958 : | chl | 96 | |
2959 : | edgomez | 851 | Data8->qpel_precision = 1; |
2960 : | syskin | 1129 | get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 3, |
2961 : | pParam->width, pParam->height, Data8->iFcode, 2, 0); | ||
2962 : | chl | 96 | |
2963 : | edgomez | 851 | } |
2964 : | syskin | 1132 | if (MotionFlags & XVID_ME_QUARTERPELREFINE8_RD) |
2965 : | SubpelRefine(Data8, CheckCandidateRD8); | ||
2966 : | edgomez | 195 | |
2967 : | edgomez | 1053 | } else { /* not qpel */ |
2968 : | edgomez | 851 | |
2969 : | edgomez | 1107 | if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_RD) /* extsearch */ |
2970 : | syskin | 1132 | SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255, CheckCandidateRD8); |
2971 : | edgomez | 1022 | |
2972 : | edgomez | 1107 | if (MotionFlags & XVID_ME_HALFPELREFINE8_RD) |
2973 : | syskin | 1132 | SubpelRefine(Data8, CheckCandidateRD8); /* halfpel refinement */ |
2974 : | edgomez | 1022 | } |
2975 : | |||
2976 : | edgomez | 1053 | /* checking vector equal to predicion */ |
2977 : | edgomez | 1107 | if (i != 0 && MotionFlags & XVID_ME_CHECKPREDICTION_RD) { |
2978 : | edgomez | 851 | const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV; |
2979 : | edgomez | 1022 | if (!MVequal(*v, Data8->predMV)) |
2980 : | syskin | 1132 | CheckCandidateRD8(Data8->predMV.x, Data8->predMV.y, Data8, 255); |
2981 : | chl | 141 | } |
2982 : | edgomez | 195 | |
2983 : | edgomez | 851 | bits += *Data8->iMinSAD; |
2984 : | edgomez | 1053 | if (bits >= Data->iMinSAD[0]) return bits; /* no chances for INTER4V */ |
2985 : | edgomez | 851 | |
2986 : | edgomez | 1053 | /* MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else */ |
2987 : | edgomez | 851 | if(Data->qpel) { |
2988 : | pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x; | ||
2989 : | pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y; | ||
2990 : | pMB->qmvs[i] = *Data8->currentQMV; | ||
2991 : | sumx += Data8->currentQMV->x/2; | ||
2992 : | sumy += Data8->currentQMV->y/2; | ||
2993 : | } else { | ||
2994 : | pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x; | ||
2995 : | pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y; | ||
2996 : | sumx += Data8->currentMV->x; | ||
2997 : | sumy += Data8->currentMV->y; | ||
2998 : | chl | 141 | } |
2999 : | edgomez | 851 | pMB->mvs[i] = *Data8->currentMV; |
3000 : | pMB->sad8[i] = 4 * *Data8->iMinSAD; | ||
3001 : | syskin | 1084 | if (Data8->cbp[0]) cbp |= 1 << (5 - i); |
3002 : | chl | 141 | |
3003 : | syskin | 1084 | } /* end - for all luma blocks */ |
3004 : | chl | 141 | |
3005 : | edgomez | 1023 | bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; |
3006 : | chl | 141 | |
3007 : | edgomez | 1053 | /* let's check chroma */ |
3008 : | edgomez | 1022 | sumx = (sumx >> 3) + roundtab_76[sumx & 0xf]; |
3009 : | sumy = (sumy >> 3) + roundtab_76[sumy & 0xf]; | ||
3010 : | chl | 96 | |
3011 : | edgomez | 1053 | /* chroma U */ |
3012 : | edgomez | 1022 | ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[4], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
3013 : | transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2); | ||
3014 : | syskin | 1075 | bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4); |
3015 : | edgomez | 1022 | |
3016 : | if (bits >= *Data->iMinSAD) return bits; | ||
3017 : | |||
3018 : | edgomez | 1053 | /* chroma V */ |
3019 : | edgomez | 1022 | ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[5], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
3020 : | transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2); | ||
3021 : | syskin | 1075 | bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5); |
3022 : | edgomez | 1022 | |
3023 : | edgomez | 1023 | bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len; |
3024 : | edgomez | 1022 | |
3025 : | syskin | 1084 | *Data->cbp = cbp; |
3026 : | edgomez | 851 | return bits; |
3027 : | } | ||
3028 : | chl | 96 | |
3029 : | edgomez | 851 | static int |
3030 : | syskin | 1108 | findRDintra(const SearchData * const Data) |
3031 : | edgomez | 851 | { |
3032 : | edgomez | 1053 | int bits = BITS_MULT*1; /* this one is ac/dc prediction flag bit */ |
3033 : | edgomez | 1023 | int cbp = 0, i, dc = 0; |
3034 : | edgomez | 962 | int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64; |
3035 : | edgomez | 195 | |
3036 : | edgomez | 851 | for(i = 0; i < 4; i++) { |
3037 : | int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth); | ||
3038 : | transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth); | ||
3039 : | syskin | 1075 | bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i, &dc); |
3040 : | chl | 141 | |
3041 : | edgomez | 1022 | if (bits >= Data->iMinSAD[0]) return bits; |
3042 : | edgomez | 851 | } |
3043 : | chl | 181 | |
3044 : | edgomez | 1023 | bits += BITS_MULT*xvid_cbpy_tab[cbp>>2].len; |
3045 : | chl | 96 | |
3046 : | edgomez | 1053 | /*chroma U */ |
3047 : | edgomez | 1022 | transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2); |
3048 : | syskin | 1075 | bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4, &dc); |
3049 : | syskin | 1084 | |
3050 : | edgomez | 1022 | if (bits >= Data->iMinSAD[0]) return bits; |
3051 : | |||
3052 : | edgomez | 1053 | /* chroma V */ |
3053 : | edgomez | 1022 | transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2); |
3054 : | syskin | 1075 | bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5, &dc); |
3055 : | edgomez | 1022 | |
3056 : | edgomez | 1023 | bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len; |
3057 : | edgomez | 1022 | |
3058 : | edgomez | 851 | return bits; |
3059 : | chl | 96 | } |
3060 : | chl | 1077 | |
3061 : | syskin | 1084 | static int |
3062 : | syskin | 1108 | findRDgmc(const SearchData * const Data, const IMAGE * const vGMC, const int x, const int y) |
3063 : | syskin | 1084 | { |
3064 : | int bits = BITS_MULT*1; /* this one is mcsel */ | ||
3065 : | int cbp = 0, i; | ||
3066 : | int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64; | ||
3067 : | chl | 1077 | |
3068 : | syskin | 1084 | for(i = 0; i < 4; i++) { |
3069 : | int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth); | ||
3070 : | transfer_8to16subro(in, Data->Cur + s, vGMC->y + s + 16*(x+y*Data->iEdgedWidth), Data->iEdgedWidth); | ||
3071 : | bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i); | ||
3072 : | if (bits >= Data->iMinSAD[0]) return bits; | ||
3073 : | } | ||
3074 : | chl | 1077 | |
3075 : | syskin | 1084 | bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; |
3076 : | chl | 1077 | |
3077 : | syskin | 1084 | /*chroma U */ |
3078 : | transfer_8to16subro(in, Data->CurU, vGMC->u + 8*(x+y*(Data->iEdgedWidth/2)), Data->iEdgedWidth/2); | ||
3079 : | bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4); | ||
3080 : | chl | 1077 | |
3081 : | syskin | 1084 | if (bits >= Data->iMinSAD[0]) return bits; |
3082 : | |||
3083 : | /* chroma V */ | ||
3084 : | transfer_8to16subro(in, Data->CurV , vGMC->v + 8*(x+y*(Data->iEdgedWidth/2)), Data->iEdgedWidth/2); | ||
3085 : | bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5); | ||
3086 : | |||
3087 : | bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; | ||
3088 : | |||
3089 : | *Data->cbp = cbp; | ||
3090 : | |||
3091 : | return bits; | ||
3092 : | } | ||
3093 : | |||
3094 : | |||
3095 : | |||
3096 : | |||
3097 : | chl | 1077 | static __inline void |
3098 : | GMEanalyzeMB ( const uint8_t * const pCur, | ||
3099 : | const uint8_t * const pRef, | ||
3100 : | const uint8_t * const pRefH, | ||
3101 : | const uint8_t * const pRefV, | ||
3102 : | const uint8_t * const pRefHV, | ||
3103 : | const int x, | ||
3104 : | const int y, | ||
3105 : | const MBParam * const pParam, | ||
3106 : | MACROBLOCK * const pMBs, | ||
3107 : | SearchData * const Data) | ||
3108 : | { | ||
3109 : | |||
3110 : | int i=0; | ||
3111 : | MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width]; | ||
3112 : | |||
3113 : | Data->iMinSAD[0] = MV_MAX_ERROR; | ||
3114 : | |||
3115 : | syskin | 1084 | Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
3116 : | chl | 1077 | |
3117 : | syskin | 1129 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
3118 : | pParam->width, pParam->height, 16, 1, 0); | ||
3119 : | chl | 1077 | |
3120 : | Data->Cur = pCur + 16*(x + y * pParam->edged_width); | ||
3121 : | Data->RefP[0] = pRef + 16*(x + y * pParam->edged_width); | ||
3122 : | Data->RefP[1] = pRefV + 16*(x + y * pParam->edged_width); | ||
3123 : | Data->RefP[2] = pRefH + 16*(x + y * pParam->edged_width); | ||
3124 : | Data->RefP[3] = pRefHV + 16*(x + y * pParam->edged_width); | ||
3125 : | |||
3126 : | Data->currentMV[0].x = Data->currentMV[0].y = 0; | ||
3127 : | syskin | 1132 | CheckCandidate16I(0, 0, Data, 255); |
3128 : | chl | 1077 | |
3129 : | if ( (Data->predMV.x !=0) || (Data->predMV.y != 0) ) | ||
3130 : | syskin | 1132 | CheckCandidate16I(Data->predMV.x, Data->predMV.y, Data, 255); |
3131 : | chl | 1077 | |
3132 : | syskin | 1132 | DiamondSearch(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255, CheckCandidate16I); |
3133 : | chl | 1077 | |
3134 : | syskin | 1132 | SubpelRefine(Data, CheckCandidate16I); |
3135 : | syskin | 1084 | |
3136 : | |||
3137 : | chl | 1077 | /* for QPel halfpel positions are worse than in halfpel mode :( */ |
3138 : | /* if (Data->qpel) { | ||
3139 : | Data->currentQMV->x = 2*Data->currentMV->x; | ||
3140 : | Data->currentQMV->y = 2*Data->currentMV->y; | ||
3141 : | Data->qpel_precision = 1; | ||
3142 : | syskin | 1129 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
3143 : | pParam->width, pParam->height, iFcode, 2, 0); | ||
3144 : | chl | 1077 | SubpelRefine(Data); |
3145 : | } | ||
3146 : | */ | ||
3147 : | |||
3148 : | pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; | ||
3149 : | pMB->sad16 = Data->iMinSAD[0]; | ||
3150 : | syskin | 1084 | pMB->mode = MODE_INTER; |
3151 : | pMB->sad16 += 10*d_mv_bits(pMB->mvs[0].x, pMB->mvs[0].y, Data->predMV, Data->iFcode, 0, 0); | ||
3152 : | chl | 1077 | return; |
3153 : | } | ||
3154 : | |||
3155 : | void | ||
3156 : | GMEanalysis(const MBParam * const pParam, | ||
3157 : | const FRAMEINFO * const current, | ||
3158 : | const FRAMEINFO * const reference, | ||
3159 : | const IMAGE * const pRefH, | ||
3160 : | const IMAGE * const pRefV, | ||
3161 : | const IMAGE * const pRefHV) | ||
3162 : | { | ||
3163 : | uint32_t x, y; | ||
3164 : | MACROBLOCK * const pMBs = current->mbs; | ||
3165 : | const IMAGE * const pCurrent = ¤t->image; | ||
3166 : | const IMAGE * const pReference = &reference->image; | ||
3167 : | |||
3168 : | int32_t iMinSAD[5], temp[5]; | ||
3169 : | VECTOR currentMV[5]; | ||
3170 : | syskin | 1132 | uint32_t dir; |
3171 : | chl | 1077 | SearchData Data; |
3172 : | memset(&Data, 0, sizeof(SearchData)); | ||
3173 : | syskin | 1084 | |
3174 : | chl | 1077 | Data.iEdgedWidth = pParam->edged_width; |
3175 : | Data.rounding = pParam->m_rounding_type; | ||
3176 : | syskin | 1084 | |
3177 : | chl | 1077 | Data.currentMV = ¤tMV[0]; |
3178 : | Data.iMinSAD = &iMinSAD[0]; | ||
3179 : | Data.iFcode = current->fcode; | ||
3180 : | Data.temp = temp; | ||
3181 : | syskin | 1132 | Data.dir = &dir; |
3182 : | chl | 1077 | |
3183 : | if (sadInit) (*sadInit) (); | ||
3184 : | |||
3185 : | for (y = 0; y < pParam->mb_height; y ++) { | ||
3186 : | for (x = 0; x < pParam->mb_width; x ++) { | ||
3187 : | GMEanalyzeMB(pCurrent->y, pReference->y, pRefH->y, pRefV->y, pRefHV->y, x, y, pParam, pMBs, &Data); | ||
3188 : | } | ||
3189 : | } | ||
3190 : | return; | ||
3191 : | } | ||
3192 : | |||
3193 : | |||
3194 : | WARPPOINTS | ||
3195 : | GlobalMotionEst(MACROBLOCK * const pMBs, | ||
3196 : | const MBParam * const pParam, | ||
3197 : | const FRAMEINFO * const current, | ||
3198 : | const FRAMEINFO * const reference, | ||
3199 : | const IMAGE * const pRefH, | ||
3200 : | const IMAGE * const pRefV, | ||
3201 : | const IMAGE * const pRefHV) | ||
3202 : | { | ||
3203 : | |||
3204 : | syskin | 1084 | const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs |
3205 : | const int deltay=8; | ||
3206 : | chl | 1077 | const unsigned int gradx=512; // lower bound for gradient in MB (ignore "flat" blocks) |
3207 : | syskin | 1084 | const unsigned int grady=512; |
3208 : | chl | 1077 | |
3209 : | double sol[4] = { 0., 0., 0., 0. }; | ||
3210 : | syskin | 1084 | |
3211 : | chl | 1077 | WARPPOINTS gmc; |
3212 : | |||
3213 : | uint32_t mx, my; | ||
3214 : | |||
3215 : | int MBh = pParam->mb_height; | ||
3216 : | int MBw = pParam->mb_width; | ||
3217 : | const int minblocks = 9; //MBh*MBw/32+3; /* just some reasonable number 3% + 3 */ | ||
3218 : | const int maxblocks = MBh*MBw/4; /* just some reasonable number 3% + 3 */ | ||
3219 : | |||
3220 : | int num=0; | ||
3221 : | int oldnum; | ||
3222 : | |||
3223 : | gmc.duv[0].x = gmc.duv[0].y = gmc.duv[1].x = gmc.duv[1].y = gmc.duv[2].x = gmc.duv[2].y = 0; | ||
3224 : | |||
3225 : | GMEanalysis(pParam,current, reference, pRefH, pRefV, pRefHV); | ||
3226 : | |||
3227 : | /* block based ME isn't done, yet, so do a quick presearch */ | ||
3228 : | |||
3229 : | // filter mask of all blocks | ||
3230 : | |||
3231 : | for (my = 0; my < (uint32_t)MBh; my++) | ||
3232 : | for (mx = 0; mx < (uint32_t)MBw; mx++) | ||
3233 : | { | ||
3234 : | const int mbnum = mx + my * MBw; | ||
3235 : | pMBs[mbnum].mcsel = 0; | ||
3236 : | } | ||
3237 : | syskin | 1084 | |
3238 : | |||
3239 : | for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */ | ||
3240 : | chl | 1077 | for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */ |
3241 : | { | ||
3242 : | const int mbnum = mx + my * MBw; | ||
3243 : | MACROBLOCK *const pMB = &pMBs[mbnum]; | ||
3244 : | const VECTOR mv = pMB->mvs[0]; | ||
3245 : | |||
3246 : | /* don't use object boundaries */ | ||
3247 : | syskin | 1084 | if ( (abs(mv.x - (pMB-1)->mvs[0].x) < deltax) |
3248 : | chl | 1077 | && (abs(mv.y - (pMB-1)->mvs[0].y) < deltay) |
3249 : | syskin | 1084 | && (abs(mv.x - (pMB+1)->mvs[0].x) < deltax) |
3250 : | chl | 1077 | && (abs(mv.y - (pMB+1)->mvs[0].y) < deltay) |
3251 : | syskin | 1084 | && (abs(mv.x - (pMB-MBw)->mvs[0].x) < deltax) |
3252 : | chl | 1077 | && (abs(mv.y - (pMB-MBw)->mvs[0].y) < deltay) |
3253 : | syskin | 1084 | && (abs(mv.x - (pMB+MBw)->mvs[0].x) < deltax) |
3254 : | chl | 1077 | && (abs(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) |
3255 : | { const int iEdgedWidth = pParam->edged_width; | ||
3256 : | const uint8_t *const pCur = current->image.y + 16*(my*iEdgedWidth + mx); | ||
3257 : | if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx ) | ||
3258 : | && (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) ) | ||
3259 : | syskin | 1084 | { pMB->mcsel = 1; |
3260 : | chl | 1077 | num++; |
3261 : | } | ||
3262 : | |||
3263 : | /* only use "structured" blocks */ | ||
3264 : | } | ||
3265 : | } | ||
3266 : | emms(); | ||
3267 : | |||
3268 : | syskin | 1084 | /* further filtering would be possible, but during iteration, remaining |
3269 : | chl | 1077 | outliers usually are removed, too */ |
3270 : | |||
3271 : | if (num>= minblocks) | ||
3272 : | do { /* until convergence */ | ||
3273 : | double DtimesF[4]; | ||
3274 : | double a,b,c,n,invdenom; | ||
3275 : | double meanx,meany; | ||
3276 : | |||
3277 : | a = b = c = n = 0; | ||
3278 : | DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.; | ||
3279 : | for (my = 1; my < (uint32_t)MBh-1; my++) | ||
3280 : | for (mx = 1; mx < (uint32_t)MBw-1; mx++) | ||
3281 : | { | ||
3282 : | const int mbnum = mx + my * MBw; | ||
3283 : | const VECTOR mv = pMBs[mbnum].mvs[0]; | ||
3284 : | |||
3285 : | if (!pMBs[mbnum].mcsel) | ||
3286 : | continue; | ||
3287 : | |||
3288 : | n++; | ||
3289 : | a += 16*mx+8; | ||
3290 : | b += 16*my+8; | ||
3291 : | c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8); | ||
3292 : | |||
3293 : | DtimesF[0] += (double)mv.x; | ||
3294 : | DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8); | ||
3295 : | DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8); | ||
3296 : | DtimesF[3] += (double)mv.y; | ||
3297 : | } | ||
3298 : | syskin | 1084 | |
3299 : | chl | 1077 | invdenom = a*a+b*b-c*n; |
3300 : | |||
3301 : | /* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */ | ||
3302 : | /* D'*E*F has been calculated in the same loop as matrix */ | ||
3303 : | |||
3304 : | sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2]; | ||
3305 : | sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3]; | ||
3306 : | sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3]; | ||
3307 : | sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3]; | ||
3308 : | |||
3309 : | sol[0] /= invdenom; | ||
3310 : | sol[1] /= invdenom; | ||
3311 : | sol[2] /= invdenom; | ||
3312 : | sol[3] /= invdenom; | ||
3313 : | |||
3314 : | meanx = meany = 0.; | ||
3315 : | oldnum = 0; | ||
3316 : | for (my = 1; my < (uint32_t)MBh-1; my++) | ||
3317 : | for (mx = 1; mx < (uint32_t)MBw-1; mx++) | ||
3318 : | { | ||
3319 : | const int mbnum = mx + my * MBw; | ||
3320 : | const VECTOR mv = pMBs[mbnum].mvs[0]; | ||
3321 : | |||
3322 : | if (!pMBs[mbnum].mcsel) | ||
3323 : | continue; | ||
3324 : | |||
3325 : | oldnum++; | ||
3326 : | meanx += fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - (double)mv.x ); | ||
3327 : | meany += fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - (double)mv.y ); | ||
3328 : | } | ||
3329 : | |||
3330 : | if (4*meanx > oldnum) /* better fit than 0.25 (=1/4pel) is useless */ | ||
3331 : | meanx /= oldnum; | ||
3332 : | else | ||
3333 : | meanx = 0.25; | ||
3334 : | |||
3335 : | if (4*meany > oldnum) | ||
3336 : | meany /= oldnum; | ||
3337 : | else | ||
3338 : | meany = 0.25; | ||
3339 : | |||
3340 : | num = 0; | ||
3341 : | for (my = 0; my < (uint32_t)MBh; my++) | ||
3342 : | for (mx = 0; mx < (uint32_t)MBw; mx++) | ||
3343 : | { | ||
3344 : | const int mbnum = mx + my * MBw; | ||
3345 : | const VECTOR mv = pMBs[mbnum].mvs[0]; | ||
3346 : | |||
3347 : | if (!pMBs[mbnum].mcsel) | ||
3348 : | continue; | ||
3349 : | |||
3350 : | if ( ( fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - (double)mv.x ) > meanx ) | ||
3351 : | || ( fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - (double)mv.y ) > meany ) ) | ||
3352 : | pMBs[mbnum].mcsel=0; | ||
3353 : | else | ||
3354 : | num++; | ||
3355 : | } | ||
3356 : | |||
3357 : | } while ( (oldnum != num) && (num>= minblocks) ); | ||
3358 : | |||
3359 : | if (num < minblocks) | ||
3360 : | { | ||
3361 : | const int iEdgedWidth = pParam->edged_width; | ||
3362 : | num = 0; | ||
3363 : | |||
3364 : | /* fprintf(stderr,"Warning! Unreliable GME (%d/%d blocks), falling back to translation.\n",num,MBh*MBw); | ||
3365 : | syskin | 1084 | */ |
3366 : | chl | 1077 | gmc.duv[0].x= gmc.duv[0].y= gmc.duv[1].x= gmc.duv[1].y= gmc.duv[2].x= gmc.duv[2].y=0; |
3367 : | |||
3368 : | edgomez | 1107 | if (!(current->motion_flags & XVID_ME_GME_REFINE)) |
3369 : | chl | 1077 | return gmc; |
3370 : | syskin | 1084 | |
3371 : | for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */ | ||
3372 : | chl | 1077 | for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */ |
3373 : | { | ||
3374 : | const int mbnum = mx + my * MBw; | ||
3375 : | MACROBLOCK *const pMB = &pMBs[mbnum]; | ||
3376 : | const uint8_t *const pCur = current->image.y + 16*(my*iEdgedWidth + mx); | ||
3377 : | if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx ) | ||
3378 : | && (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) ) | ||
3379 : | syskin | 1084 | { pMB->mcsel = 1; |
3380 : | chl | 1077 | gmc.duv[0].x += pMB->mvs[0].x; |
3381 : | gmc.duv[0].y += pMB->mvs[0].y; | ||
3382 : | num++; | ||
3383 : | } | ||
3384 : | } | ||
3385 : | syskin | 1084 | |
3386 : | chl | 1077 | if (gmc.duv[0].x) |
3387 : | gmc.duv[0].x /= num; | ||
3388 : | syskin | 1084 | if (gmc.duv[0].y) |
3389 : | chl | 1077 | gmc.duv[0].y /= num; |
3390 : | } else { | ||
3391 : | |||
3392 : | gmc.duv[0].x=(int)(sol[0]+0.5); | ||
3393 : | gmc.duv[0].y=(int)(sol[3]+0.5); | ||
3394 : | |||
3395 : | gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5); | ||
3396 : | gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5); | ||
3397 : | |||
3398 : | gmc.duv[2].x=-gmc.duv[1].y; /* two warp points only */ | ||
3399 : | gmc.duv[2].y=gmc.duv[1].x; | ||
3400 : | } | ||
3401 : | if (num>maxblocks) | ||
3402 : | { for (my = 1; my < (uint32_t)MBh-1; my++) | ||
3403 : | for (mx = 1; mx < (uint32_t)MBw-1; mx++) | ||
3404 : | { | ||
3405 : | const int mbnum = mx + my * MBw; | ||
3406 : | if (pMBs[mbnum-1].mcsel) | ||
3407 : | pMBs[mbnum].mcsel=0; | ||
3408 : | syskin | 1084 | else |
3409 : | chl | 1077 | if (pMBs[mbnum-MBw].mcsel) |
3410 : | pMBs[mbnum].mcsel=0; | ||
3411 : | } | ||
3412 : | } | ||
3413 : | return gmc; | ||
3414 : | } | ||
3415 : | |||
3416 : | int | ||
3417 : | GlobalMotionEstRefine( | ||
3418 : | WARPPOINTS *const startwp, | ||
3419 : | MACROBLOCK * const pMBs, | ||
3420 : | const MBParam * const pParam, | ||
3421 : | const FRAMEINFO * const current, | ||
3422 : | const FRAMEINFO * const reference, | ||
3423 : | const IMAGE * const pCurr, | ||
3424 : | const IMAGE * const pRef, | ||
3425 : | const IMAGE * const pRefH, | ||
3426 : | const IMAGE * const pRefV, | ||
3427 : | const IMAGE * const pRefHV) | ||
3428 : | { | ||
3429 : | uint8_t* GMCblock = (uint8_t*)malloc(16*pParam->edged_width); | ||
3430 : | WARPPOINTS bestwp=*startwp; | ||
3431 : | WARPPOINTS centerwp,currwp; | ||
3432 : | int gmcminSAD=0; | ||
3433 : | int gmcSAD=0; | ||
3434 : | int direction; | ||
3435 : | // int mx,my; | ||
3436 : | |||
3437 : | /* use many blocks... */ | ||
3438 : | /* for (my = 0; my < (uint32_t)pParam->mb_height; my++) | ||
3439 : | for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++) | ||
3440 : | { | ||
3441 : | const int mbnum = mx + my * pParam->mb_width; | ||
3442 : | syskin | 1084 | pMBs[mbnum].mcsel=1; |
3443 : | chl | 1077 | } |
3444 : | */ | ||
3445 : | |||
3446 : | /* or rather don't use too many blocks... */ | ||
3447 : | /* | ||
3448 : | for (my = 1; my < (uint32_t)MBh-1; my++) | ||
3449 : | for (mx = 1; mx < (uint32_t)MBw-1; mx++) | ||
3450 : | { | ||
3451 : | const int mbnum = mx + my * MBw; | ||
3452 : | if (MBmask[mbnum-1]) | ||
3453 : | MBmask[mbnum-1]=0; | ||
3454 : | syskin | 1084 | else |
3455 : | chl | 1077 | if (MBmask[mbnum-MBw]) |
3456 : | MBmask[mbnum-1]=0; | ||
3457 : | syskin | 1084 | |
3458 : | chl | 1077 | } |
3459 : | */ | ||
3460 : | gmcminSAD = globalSAD(&bestwp, pParam, pMBs, current, pRef, pCurr, GMCblock); | ||
3461 : | |||
3462 : | syskin | 1084 | if ( (reference->coding_type == S_VOP) |
3463 : | && ( (reference->warp.duv[1].x != bestwp.duv[1].x) | ||
3464 : | || (reference->warp.duv[1].y != bestwp.duv[1].y) | ||
3465 : | || (reference->warp.duv[0].x != bestwp.duv[0].x) | ||
3466 : | || (reference->warp.duv[0].y != bestwp.duv[0].y) | ||
3467 : | || (reference->warp.duv[2].x != bestwp.duv[2].x) | ||
3468 : | || (reference->warp.duv[2].y != bestwp.duv[2].y) ) ) | ||
3469 : | chl | 1077 | { |
3470 : | syskin | 1084 | gmcSAD = globalSAD(&reference->warp, pParam, pMBs, |
3471 : | chl | 1077 | current, pRef, pCurr, GMCblock); |
3472 : | syskin | 1084 | |
3473 : | chl | 1077 | if (gmcSAD < gmcminSAD) |
3474 : | { bestwp = reference->warp; | ||
3475 : | gmcminSAD = gmcSAD; | ||
3476 : | } | ||
3477 : | } | ||
3478 : | |||
3479 : | do { | ||
3480 : | direction = 0; | ||
3481 : | centerwp = bestwp; | ||
3482 : | |||
3483 : | syskin | 1084 | currwp = centerwp; |
3484 : | |||
3485 : | chl | 1077 | currwp.duv[0].x--; |
3486 : | gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); | ||
3487 : | if (gmcSAD < gmcminSAD) | ||
3488 : | { bestwp = currwp; | ||
3489 : | gmcminSAD = gmcSAD; | ||
3490 : | syskin | 1084 | direction = 1; |
3491 : | chl | 1077 | } |
3492 : | syskin | 1084 | else |
3493 : | { | ||
3494 : | chl | 1077 | currwp = centerwp; currwp.duv[0].x++; |
3495 : | gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); | ||
3496 : | if (gmcSAD < gmcminSAD) | ||
3497 : | { bestwp = currwp; | ||
3498 : | gmcminSAD = gmcSAD; | ||
3499 : | syskin | 1084 | direction = 2; |
3500 : | chl | 1077 | } |
3501 : | } | ||
3502 : | if (direction) continue; | ||
3503 : | |||
3504 : | currwp = centerwp; currwp.duv[0].y--; | ||
3505 : | gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); | ||
3506 : | if (gmcSAD < gmcminSAD) | ||
3507 : | { bestwp = currwp; | ||
3508 : | gmcminSAD = gmcSAD; | ||
3509 : | syskin | 1084 | direction = 4; |
3510 : | } | ||
3511 : | else | ||
3512 : | chl | 1077 | { |
3513 : | currwp = centerwp; currwp.duv[0].y++; | ||
3514 : | gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); | ||
3515 : | if (gmcSAD < gmcminSAD) | ||
3516 : | { bestwp = currwp; | ||
3517 : | gmcminSAD = gmcSAD; | ||
3518 : | syskin | 1084 | direction = 8; |
3519 : | chl | 1077 | } |
3520 : | syskin | 1084 | } |
3521 : | chl | 1077 | if (direction) continue; |
3522 : | syskin | 1084 | |
3523 : | chl | 1077 | currwp = centerwp; currwp.duv[1].x++; |
3524 : | gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); | ||
3525 : | if (gmcSAD < gmcminSAD) | ||
3526 : | { bestwp = currwp; | ||
3527 : | gmcminSAD = gmcSAD; | ||
3528 : | syskin | 1084 | direction = 32; |
3529 : | } | ||
3530 : | chl | 1077 | currwp.duv[2].y++; |
3531 : | gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); | ||
3532 : | if (gmcSAD < gmcminSAD) | ||
3533 : | { bestwp = currwp; | ||
3534 : | gmcminSAD = gmcSAD; | ||
3535 : | syskin | 1084 | direction = 1024; |
3536 : | } | ||
3537 : | chl | 1077 | |
3538 : | currwp = centerwp; currwp.duv[1].x--; | ||
3539 : | gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); | ||
3540 : | if (gmcSAD < gmcminSAD) | ||
3541 : | { bestwp = currwp; | ||
3542 : | gmcminSAD = gmcSAD; | ||
3543 : | syskin | 1084 | direction = 16; |
3544 : | } | ||
3545 : | else | ||
3546 : | chl | 1077 | { |
3547 : | currwp = centerwp; currwp.duv[1].x++; | ||
3548 : | gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); | ||
3549 : | if (gmcSAD < gmcminSAD) | ||
3550 : | { bestwp = currwp; | ||
3551 : | gmcminSAD = gmcSAD; | ||
3552 : | syskin | 1084 | direction = 32; |
3553 : | chl | 1077 | } |
3554 : | syskin | 1084 | } |
3555 : | chl | 1077 | if (direction) continue; |
3556 : | |||
3557 : | |||
3558 : | currwp = centerwp; currwp.duv[1].y--; | ||
3559 : | gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); | ||
3560 : | if (gmcSAD < gmcminSAD) | ||
3561 : | { bestwp = currwp; | ||
3562 : | gmcminSAD = gmcSAD; | ||
3563 : | syskin | 1084 | direction = 64; |
3564 : | } | ||
3565 : | else | ||
3566 : | chl | 1077 | { |
3567 : | currwp = centerwp; currwp.duv[1].y++; | ||
3568 : | gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); | ||
3569 : | if (gmcSAD < gmcminSAD) | ||
3570 : | { bestwp = currwp; | ||
3571 : | gmcminSAD = gmcSAD; | ||
3572 : | syskin | 1084 | direction = 128; |
3573 : | chl | 1077 | } |
3574 : | } | ||
3575 : | if (direction) continue; | ||
3576 : | syskin | 1084 | |
3577 : | chl | 1077 | currwp = centerwp; currwp.duv[2].x--; |
3578 : | gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); | ||
3579 : | if (gmcSAD < gmcminSAD) | ||
3580 : | { bestwp = currwp; | ||
3581 : | gmcminSAD = gmcSAD; | ||
3582 : | syskin | 1084 | direction = 256; |
3583 : | } | ||
3584 : | else | ||
3585 : | chl | 1077 | { |
3586 : | currwp = centerwp; currwp.duv[2].x++; | ||
3587 : | gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); | ||
3588 : | if (gmcSAD < gmcminSAD) | ||
3589 : | { bestwp = currwp; | ||
3590 : | gmcminSAD = gmcSAD; | ||
3591 : | syskin | 1084 | direction = 512; |
3592 : | chl | 1077 | } |
3593 : | } | ||
3594 : | if (direction) continue; | ||
3595 : | syskin | 1084 | |
3596 : | chl | 1077 | currwp = centerwp; currwp.duv[2].y--; |
3597 : | gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); | ||
3598 : | if (gmcSAD < gmcminSAD) | ||
3599 : | { bestwp = currwp; | ||
3600 : | gmcminSAD = gmcSAD; | ||
3601 : | syskin | 1084 | direction = 1024; |
3602 : | } | ||
3603 : | chl | 1077 | else |
3604 : | syskin | 1084 | { |
3605 : | chl | 1077 | currwp = centerwp; currwp.duv[2].y++; |
3606 : | gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); | ||
3607 : | if (gmcSAD < gmcminSAD) | ||
3608 : | { bestwp = currwp; | ||
3609 : | gmcminSAD = gmcSAD; | ||
3610 : | syskin | 1084 | direction = 2048; |
3611 : | chl | 1077 | } |
3612 : | } | ||
3613 : | } while (direction); | ||
3614 : | free(GMCblock); | ||
3615 : | syskin | 1084 | |
3616 : | chl | 1077 | *startwp = bestwp; |
3617 : | |||
3618 : | return gmcminSAD; | ||
3619 : | } | ||
3620 : | |||
3621 : | int | ||
3622 : | syskin | 1084 | globalSAD(const WARPPOINTS *const wp, |
3623 : | const MBParam * const pParam, | ||
3624 : | chl | 1077 | const MACROBLOCK * const pMBs, |
3625 : | const FRAMEINFO * const current, | ||
3626 : | const IMAGE * const pRef, | ||
3627 : | const IMAGE * const pCurr, | ||
3628 : | uint8_t *const GMCblock) | ||
3629 : | { | ||
3630 : | NEW_GMC_DATA gmc_data; | ||
3631 : | int iSAD, gmcSAD=0; | ||
3632 : | int num=0; | ||
3633 : | unsigned int mx, my; | ||
3634 : | syskin | 1084 | |
3635 : | chl | 1077 | generate_GMCparameters( 3, 3, wp, pParam->width, pParam->height, &gmc_data); |
3636 : | syskin | 1084 | |
3637 : | chl | 1077 | for (my = 0; my < (uint32_t)pParam->mb_height; my++) |
3638 : | for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++) { | ||
3639 : | |||
3640 : | const int mbnum = mx + my * pParam->mb_width; | ||
3641 : | const int iEdgedWidth = pParam->edged_width; | ||
3642 : | syskin | 1084 | |
3643 : | chl | 1077 | if (!pMBs[mbnum].mcsel) |
3644 : | continue; | ||
3645 : | syskin | 1084 | |
3646 : | gmc_data.predict_16x16(&gmc_data, GMCblock, | ||
3647 : | chl | 1077 | pRef->y, |
3648 : | syskin | 1084 | iEdgedWidth, |
3649 : | iEdgedWidth, | ||
3650 : | mx, my, | ||
3651 : | chl | 1077 | pParam->m_rounding_type); |
3652 : | |||
3653 : | iSAD = sad16 ( pCurr->y + 16*(my*iEdgedWidth + mx), | ||
3654 : | syskin | 1084 | GMCblock , iEdgedWidth, 65536); |
3655 : | chl | 1077 | iSAD -= pMBs[mbnum].sad16; |
3656 : | syskin | 1084 | |
3657 : | if (iSAD<0) | ||
3658 : | chl | 1077 | gmcSAD += iSAD; |
3659 : | num++; | ||
3660 : | } | ||
3661 : | return gmcSAD; | ||
3662 : | } | ||
3663 : |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |