Parent Directory | Revision Log
Revision 1.64 - (view) (download)
1 : | edgomez | 1.55 | /************************************************************************** |
2 : | Isibaar | 1.1 | * |
3 : | edgomez | 1.55 | * XVID MPEG-4 VIDEO CODEC |
4 : | * motion estimation | ||
5 : | chl | 1.26 | * |
6 : | edgomez | 1.55 | * This program is an implementation of a part of one or more MPEG-4 |
7 : | * Video tools as specified in ISO/IEC 14496-2 standard. Those intending | ||
8 : | * to use this software module in hardware or software products are | ||
9 : | * advised that its use may infringe existing patents or copyrights, and | ||
10 : | * any such use would be at such party's own risk. The original | ||
11 : | * developer of this software module and his/her company, and subsequent | ||
12 : | * editors and their companies, will have no liability for use of this | ||
13 : | * software or modifications or derivatives thereof. | ||
14 : | chl | 1.45 | * |
15 : | edgomez | 1.55 | * This program is free software; you can redistribute it and/or modify |
16 : | * it under the terms of the GNU General Public License as published by | ||
17 : | * the Free Software Foundation; either version 2 of the License, or | ||
18 : | * (at your option) any later version. | ||
19 : | chl | 1.45 | * |
20 : | edgomez | 1.55 | * This program is distributed in the hope that it will be useful, |
21 : | * but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
22 : | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
23 : | * GNU General Public License for more details. | ||
24 : | chl | 1.45 | * |
25 : | edgomez | 1.55 | * You should have received a copy of the GNU General Public License |
26 : | * along with this program; if not, write to the Free Software | ||
27 : | * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. | ||
28 : | chl | 1.26 | * |
29 : | *************************************************************************/ | ||
30 : | Isibaar | 1.1 | |
31 : | #include <assert.h> | ||
32 : | #include <stdio.h> | ||
33 : | chl | 1.3 | #include <stdlib.h> |
34 : | edgomez | 1.55 | #include <string.h> // memcpy |
35 : | #include <math.h> // lrint | ||
36 : | Isibaar | 1.1 | |
37 : | #include "../encoder.h" | ||
38 : | #include "../utils/mbfunctions.h" | ||
39 : | #include "../prediction/mbprediction.h" | ||
40 : | #include "../global.h" | ||
41 : | #include "../utils/timer.h" | ||
42 : | edgomez | 1.55 | #include "../image/interpolate8x8.h" |
43 : | #include "motion_est.h" | ||
44 : | suxen_drol | 1.6 | #include "motion.h" |
45 : | Isibaar | 1.1 | #include "sad.h" |
46 : | edgomez | 1.55 | #include "../utils/emms.h" |
47 : | #include "../dct/fdct.h" | ||
48 : | Isibaar | 1.1 | |
49 : | edgomez | 1.57 | /***************************************************************************** |
50 : | * Modified rounding tables -- declared in motion.h | ||
51 : | * Original tables see ISO spec tables 7-6 -> 7-9 | ||
52 : | ****************************************************************************/ | ||
53 : | |||
54 : | const uint32_t roundtab[16] = | ||
55 : | {0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2 }; | ||
56 : | |||
57 : | /* K = 4 */ | ||
58 : | const uint32_t roundtab_76[16] = | ||
59 : | { 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1 }; | ||
60 : | |||
61 : | /* K = 2 */ | ||
62 : | const uint32_t roundtab_78[8] = | ||
63 : | { 0, 0, 1, 1, 0, 0, 0, 1 }; | ||
64 : | |||
65 : | /* K = 1 */ | ||
66 : | const uint32_t roundtab_79[4] = | ||
67 : | { 0, 1, 0, 0 }; | ||
68 : | |||
69 : | edgomez | 1.55 | #define INITIAL_SKIP_THRESH (10) |
70 : | #define FINAL_SKIP_THRESH (50) | ||
71 : | #define MAX_SAD00_FOR_SKIP (20) | ||
72 : | #define MAX_CHROMA_SAD_FOR_SKIP (22) | ||
73 : | edgomez | 1.20 | |
74 : | edgomez | 1.55 | #define CHECK_CANDIDATE(X,Y,D) { \ |
75 : | CheckCandidate((X),(Y), (D), &iDirection, data ); } | ||
76 : | edgomez | 1.57 | |
77 : | /***************************************************************************** | ||
78 : | * Code | ||
79 : | ****************************************************************************/ | ||
80 : | edgomez | 1.20 | |
81 : | edgomez | 1.55 | static __inline uint32_t |
82 : | d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv) | ||
83 : | { | ||
84 : | syskin | 1.63 | int bits; |
85 : | const int q = (1 << (iFcode - 1)) - 1; | ||
86 : | |||
87 : | x <<= qpel; | ||
88 : | y <<= qpel; | ||
89 : | edgomez | 1.55 | if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); } |
90 : | |||
91 : | x -= pred.x; | ||
92 : | syskin | 1.63 | bits = (x != 0 ? iFcode:0); |
93 : | x = abs(x); | ||
94 : | x += q; | ||
95 : | x >>= (iFcode - 1); | ||
96 : | bits += mvtab[x]; | ||
97 : | |||
98 : | edgomez | 1.55 | y -= pred.y; |
99 : | syskin | 1.63 | bits += (y != 0 ? iFcode:0); |
100 : | y = abs(y); | ||
101 : | y += q; | ||
102 : | y >>= (iFcode - 1); | ||
103 : | bits += mvtab[y]; | ||
104 : | edgomez | 1.55 | |
105 : | syskin | 1.63 | return bits; |
106 : | edgomez | 1.55 | } |
107 : | chl | 1.10 | |
108 : | syskin | 1.64 | static int32_t ChromaSAD2(const int fx, const int fy, const int bx, const int by, |
109 : | const SearchData * const data) | ||
110 : | edgomez | 1.55 | { |
111 : | int sad; | ||
112 : | const uint32_t stride = data->iEdgedWidth/2; | ||
113 : | uint8_t * f_refu = data->RefQ, | ||
114 : | * f_refv = data->RefQ + 8, | ||
115 : | * b_refu = data->RefQ + 16, | ||
116 : | * b_refv = data->RefQ + 24; | ||
117 : | syskin | 1.64 | int offset = (fx>>1) + (fy>>1)*stride; |
118 : | edgomez | 1.55 | |
119 : | switch (((fx & 1) << 1) | (fy & 1)) { | ||
120 : | case 0: | ||
121 : | syskin | 1.64 | f_refu = (uint8_t*)data->RefP[4] + offset; |
122 : | f_refv = (uint8_t*)data->RefP[5] + offset; | ||
123 : | edgomez | 1.55 | break; |
124 : | case 1: | ||
125 : | syskin | 1.64 | interpolate8x8_halfpel_v(f_refu, data->RefP[4] + offset, stride, data->rounding); |
126 : | interpolate8x8_halfpel_v(f_refv, data->RefP[5] + offset, stride, data->rounding); | ||
127 : | edgomez | 1.55 | break; |
128 : | case 2: | ||
129 : | syskin | 1.64 | interpolate8x8_halfpel_h(f_refu, data->RefP[4] + offset, stride, data->rounding); |
130 : | interpolate8x8_halfpel_h(f_refv, data->RefP[5] + offset, stride, data->rounding); | ||
131 : | edgomez | 1.55 | break; |
132 : | default: | ||
133 : | syskin | 1.64 | interpolate8x8_halfpel_hv(f_refu, data->RefP[4] + offset, stride, data->rounding); |
134 : | interpolate8x8_halfpel_hv(f_refv, data->RefP[5] + offset, stride, data->rounding); | ||
135 : | edgomez | 1.55 | break; |
136 : | } | ||
137 : | |||
138 : | syskin | 1.64 | offset = (bx>>1) + (by>>1)*stride; |
139 : | edgomez | 1.55 | switch (((bx & 1) << 1) | (by & 1)) { |
140 : | case 0: | ||
141 : | syskin | 1.64 | b_refu = (uint8_t*)data->b_RefP[4] + offset; |
142 : | b_refv = (uint8_t*)data->b_RefP[5] + offset; | ||
143 : | edgomez | 1.55 | break; |
144 : | case 1: | ||
145 : | syskin | 1.64 | interpolate8x8_halfpel_v(b_refu, data->b_RefP[4] + offset, stride, data->rounding); |
146 : | interpolate8x8_halfpel_v(b_refv, data->b_RefP[5] + offset, stride, data->rounding); | ||
147 : | edgomez | 1.55 | break; |
148 : | case 2: | ||
149 : | syskin | 1.64 | interpolate8x8_halfpel_h(b_refu, data->b_RefP[4] + offset, stride, data->rounding); |
150 : | interpolate8x8_halfpel_h(b_refv, data->b_RefP[5] + offset, stride, data->rounding); | ||
151 : | edgomez | 1.55 | break; |
152 : | default: | ||
153 : | syskin | 1.64 | interpolate8x8_halfpel_hv(b_refu, data->b_RefP[4] + offset, stride, data->rounding); |
154 : | interpolate8x8_halfpel_hv(b_refv, data->b_RefP[5] + offset, stride, data->rounding); | ||
155 : | edgomez | 1.55 | break; |
156 : | } | ||
157 : | |||
158 : | sad = sad8bi(data->CurU, b_refu, f_refu, stride); | ||
159 : | sad += sad8bi(data->CurV, b_refv, f_refv, stride); | ||
160 : | chl | 1.10 | |
161 : | edgomez | 1.55 | return sad; |
162 : | } | ||
163 : | chl | 1.10 | |
164 : | edgomez | 1.55 | static int32_t |
165 : | syskin | 1.64 | ChromaSAD(const int dx, const int dy, const SearchData * const data) |
166 : | Isibaar | 1.1 | { |
167 : | edgomez | 1.55 | int sad; |
168 : | const uint32_t stride = data->iEdgedWidth/2; | ||
169 : | syskin | 1.64 | int offset = (dx>>1) + (dy>>1)*stride; |
170 : | Isibaar | 1.1 | |
171 : | edgomez | 1.55 | if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently |
172 : | data->temp[5] = dx; data->temp[6] = dy; // backup | ||
173 : | Isibaar | 1.1 | |
174 : | edgomez | 1.55 | switch (((dx & 1) << 1) | (dy & 1)) { |
175 : | case 0: | ||
176 : | syskin | 1.64 | sad = sad8(data->CurU, data->RefP[4] + offset, stride); |
177 : | sad += sad8(data->CurV, data->RefP[5] + offset, stride); | ||
178 : | edgomez | 1.55 | break; |
179 : | case 1: | ||
180 : | syskin | 1.64 | sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + stride, stride); |
181 : | sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + stride, stride); | ||
182 : | edgomez | 1.55 | break; |
183 : | case 2: | ||
184 : | syskin | 1.64 | sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + 1, stride); |
185 : | sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + 1, stride); | ||
186 : | edgomez | 1.55 | break; |
187 : | default: | ||
188 : | syskin | 1.64 | interpolate8x8_halfpel_hv(data->RefQ, data->RefP[4] + offset, stride, data->rounding); |
189 : | edgomez | 1.55 | sad = sad8(data->CurU, data->RefQ, stride); |
190 : | |||
191 : | syskin | 1.64 | interpolate8x8_halfpel_hv(data->RefQ, data->RefP[5] + offset, stride, data->rounding); |
192 : | edgomez | 1.55 | sad += sad8(data->CurV, data->RefQ, stride); |
193 : | break; | ||
194 : | } | ||
195 : | data->temp[7] = sad; //backup, part 2 | ||
196 : | return sad; | ||
197 : | } | ||
198 : | Isibaar | 1.1 | |
199 : | edgomez | 1.55 | static __inline const uint8_t * |
200 : | GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data) | ||
201 : | { | ||
202 : | // dir : 0 = forward, 1 = backward | ||
203 : | syskin | 1.64 | const uint8_t* const *direction = ( dir == 0 ? data->RefP : data->b_RefP ); |
204 : | const int picture = ((x&1)<<1) | (y&1); | ||
205 : | const int offset = (x>>1) + (y>>1)*data->iEdgedWidth; | ||
206 : | return direction[picture] + offset; | ||
207 : | edgomez | 1.55 | } |
208 : | Isibaar | 1.1 | |
209 : | edgomez | 1.55 | // this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate |
210 : | static __inline const uint8_t * | ||
211 : | GetReference(const int x, const int y, const SearchData * const data) | ||
212 : | { | ||
213 : | syskin | 1.64 | const int picture = ((x&1)<<1) | (y&1); |
214 : | const int offset = (x>>1) + (y>>1)*data->iEdgedWidth; | ||
215 : | return data->RefP[picture] + offset; | ||
216 : | Isibaar | 1.1 | } |
217 : | |||
218 : | edgomez | 1.55 | static uint8_t * |
219 : | Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data) | ||
220 : | Isibaar | 1.1 | { |
221 : | edgomez | 1.55 | // create or find a qpel-precision reference picture; return pointer to it |
222 : | uint8_t * Reference = data->RefQ + 16*dir; | ||
223 : | const uint32_t iEdgedWidth = data->iEdgedWidth; | ||
224 : | const uint32_t rounding = data->rounding; | ||
225 : | const int halfpel_x = x/2; | ||
226 : | const int halfpel_y = y/2; | ||
227 : | const uint8_t *ref1, *ref2, *ref3, *ref4; | ||
228 : | |||
229 : | ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); | ||
230 : | ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; | ||
231 : | switch( ((x&1)<<1) + (y&1) ) { | ||
232 : | edgomez | 1.62 | case 3: // x and y in qpel resolution - the "corners" (top left/right and |
233 : | // bottom left/right) during qpel refinement | ||
234 : | ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); | ||
235 : | ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); | ||
236 : | ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); | ||
237 : | ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; | ||
238 : | ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; | ||
239 : | ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; | ||
240 : | interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); | ||
241 : | edgomez | 1.55 | break; |
242 : | |||
243 : | case 1: // x halfpel, y qpel - top or bottom during qpel refinement | ||
244 : | ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); | ||
245 : | ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; | ||
246 : | interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); | ||
247 : | break; | ||
248 : | |||
249 : | case 2: // x qpel, y halfpel - left or right during qpel refinement | ||
250 : | ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); | ||
251 : | ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; | ||
252 : | interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); | ||
253 : | break; | ||
254 : | |||
255 : | edgomez | 1.62 | default: // pure halfpel position |
256 : | return (uint8_t *) ref1; | ||
257 : | syskin | 1.64 | |
258 : | edgomez | 1.55 | } |
259 : | return Reference; | ||
260 : | Isibaar | 1.1 | } |
261 : | |||
262 : | edgomez | 1.55 | static uint8_t * |
263 : | Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data) | ||
264 : | Isibaar | 1.1 | { |
265 : | edgomez | 1.55 | // create or find a qpel-precision reference picture; return pointer to it |
266 : | uint8_t * Reference = data->RefQ + 16*dir; | ||
267 : | const uint32_t iEdgedWidth = data->iEdgedWidth; | ||
268 : | const uint32_t rounding = data->rounding; | ||
269 : | const int halfpel_x = x/2; | ||
270 : | const int halfpel_y = y/2; | ||
271 : | const uint8_t *ref1, *ref2, *ref3, *ref4; | ||
272 : | |||
273 : | ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); | ||
274 : | switch( ((x&1)<<1) + (y&1) ) { | ||
275 : | case 3: // x and y in qpel resolution - the "corners" (top left/right and | ||
276 : | syskin | 1.64 | // bottom left/right) during qpel refinement |
277 : | edgomez | 1.55 | ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
278 : | ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); | ||
279 : | ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); | ||
280 : | interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); | ||
281 : | interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); | ||
282 : | interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); | ||
283 : | interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); | ||
284 : | break; | ||
285 : | |||
286 : | case 1: // x halfpel, y qpel - top or bottom during qpel refinement | ||
287 : | ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); | ||
288 : | interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); | ||
289 : | interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); | ||
290 : | interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); | ||
291 : | interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); | ||
292 : | break; | ||
293 : | |||
294 : | case 2: // x qpel, y halfpel - left or right during qpel refinement | ||
295 : | ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); | ||
296 : | interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); | ||
297 : | interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); | ||
298 : | interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); | ||
299 : | interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); | ||
300 : | break; | ||
301 : | |||
302 : | edgomez | 1.62 | default: // pure halfpel position |
303 : | edgomez | 1.55 | return (uint8_t *) ref1; |
304 : | } | ||
305 : | return Reference; | ||
306 : | Isibaar | 1.1 | } |
307 : | suxen_drol | 1.8 | |
308 : | edgomez | 1.55 | /* CHECK_CANDIATE FUNCTIONS START */ |
309 : | suxen_drol | 1.8 | |
310 : | edgomez | 1.55 | static void |
311 : | CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) | ||
312 : | { | ||
313 : | int xc, yc; | ||
314 : | const uint8_t * Reference; | ||
315 : | VECTOR * current; | ||
316 : | int32_t sad; uint32_t t; | ||
317 : | chl | 1.19 | |
318 : | edgomez | 1.55 | if ( (x > data->max_dx) || (x < data->min_dx) |
319 : | || (y > data->max_dy) || (y < data->min_dy) ) return; | ||
320 : | edgomez | 1.20 | |
321 : | edgomez | 1.55 | if (!data->qpel_precision) { |
322 : | Reference = GetReference(x, y, data); | ||
323 : | current = data->currentMV; | ||
324 : | xc = x; yc = y; | ||
325 : | } else { // x and y are in 1/4 precision | ||
326 : | Reference = Interpolate16x16qpel(x, y, 0, data); | ||
327 : | xc = x/2; yc = y/2; //for chroma sad | ||
328 : | current = data->currentQMV; | ||
329 : | } | ||
330 : | edgomez | 1.20 | |
331 : | edgomez | 1.55 | sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
332 : | t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); | ||
333 : | edgomez | 1.20 | |
334 : | edgomez | 1.55 | sad += (data->lambda16 * t * sad)>>10; |
335 : | data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; | ||
336 : | edgomez | 1.20 | |
337 : | syskin | 1.64 | if (data->chroma) sad += ChromaSAD( (xc >> 1) + roundtab_79[xc & 0x3], |
338 : | (yc >> 1) + roundtab_79[yc & 0x3], data); | ||
339 : | edgomez | 1.20 | |
340 : | edgomez | 1.55 | if (sad < data->iMinSAD[0]) { |
341 : | data->iMinSAD[0] = sad; | ||
342 : | current[0].x = x; current[0].y = y; | ||
343 : | *dir = Direction; | ||
344 : | } | ||
345 : | edgomez | 1.20 | |
346 : | edgomez | 1.55 | if (data->temp[1] < data->iMinSAD[1]) { |
347 : | data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y = y; } | ||
348 : | if (data->temp[2] < data->iMinSAD[2]) { | ||
349 : | data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; } | ||
350 : | if (data->temp[3] < data->iMinSAD[3]) { | ||
351 : | data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; } | ||
352 : | if (data->temp[4] < data->iMinSAD[4]) { | ||
353 : | data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; } | ||
354 : | chl | 1.17 | |
355 : | Isibaar | 1.1 | } |
356 : | |||
357 : | edgomez | 1.55 | static void |
358 : | CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) | ||
359 : | { | ||
360 : | int32_t sad; uint32_t t; | ||
361 : | const uint8_t * Reference; | ||
362 : | syskin | 1.63 | VECTOR * current; |
363 : | edgomez | 1.55 | |
364 : | if ( (x > data->max_dx) || (x < data->min_dx) | ||
365 : | || (y > data->max_dy) || (y < data->min_dy) ) return; | ||
366 : | chl | 1.35 | |
367 : | syskin | 1.63 | if (!data->qpel_precision) { |
368 : | Reference = GetReference(x, y, data); | ||
369 : | current = data->currentMV; | ||
370 : | } else { // x and y are in 1/4 precision | ||
371 : | Reference = Interpolate8x8qpel(x, y, 0, 0, data); | ||
372 : | current = data->currentQMV; | ||
373 : | } | ||
374 : | Isibaar | 1.1 | |
375 : | edgomez | 1.55 | sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
376 : | t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); | ||
377 : | chl | 1.3 | |
378 : | edgomez | 1.55 | sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10; |
379 : | edgomez | 1.20 | |
380 : | edgomez | 1.55 | if (sad < *(data->iMinSAD)) { |
381 : | *(data->iMinSAD) = sad; | ||
382 : | syskin | 1.63 | current->x = x; current->y = y; |
383 : | edgomez | 1.55 | *dir = Direction; |
384 : | } | ||
385 : | } | ||
386 : | edgomez | 1.20 | |
387 : | |||
388 : | edgomez | 1.55 | static void |
389 : | CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) | ||
390 : | { | ||
391 : | uint32_t t; | ||
392 : | const uint8_t * Reference; | ||
393 : | edgomez | 1.20 | |
394 : | edgomez | 1.62 | if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero even value |
395 : | edgomez | 1.55 | (x > data->max_dx) || (x < data->min_dx) |
396 : | || (y > data->max_dy) || (y < data->min_dy) ) return; | ||
397 : | |||
398 : | Reference = GetReference(x, y, data); | ||
399 : | t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1); | ||
400 : | |||
401 : | data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); | ||
402 : | |||
403 : | data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10; | ||
404 : | data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; | ||
405 : | |||
406 : | if (data->temp[0] < data->iMinSAD[0]) { | ||
407 : | data->iMinSAD[0] = data->temp[0]; | ||
408 : | data->currentMV[0].x = x; data->currentMV[0].y = y; | ||
409 : | *dir = Direction; } | ||
410 : | |||
411 : | if (data->temp[1] < data->iMinSAD[1]) { | ||
412 : | data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } | ||
413 : | if (data->temp[2] < data->iMinSAD[2]) { | ||
414 : | data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } | ||
415 : | if (data->temp[3] < data->iMinSAD[3]) { | ||
416 : | data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } | ||
417 : | if (data->temp[4] < data->iMinSAD[4]) { | ||
418 : | data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } | ||
419 : | } | ||
420 : | edgomez | 1.20 | |
421 : | edgomez | 1.55 | static void |
422 : | CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) | ||
423 : | { | ||
424 : | int32_t sad, xc, yc; | ||
425 : | const uint8_t * Reference; | ||
426 : | uint32_t t; | ||
427 : | VECTOR * current; | ||
428 : | |||
429 : | edgomez | 1.62 | if ( (x > data->max_dx) || ( x < data->min_dx) |
430 : | || (y > data->max_dy) || (y < data->min_dy) ) return; | ||
431 : | edgomez | 1.55 | |
432 : | if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; //non-zero even value | ||
433 : | |||
434 : | if (data->qpel_precision) { // x and y are in 1/4 precision | ||
435 : | Reference = Interpolate16x16qpel(x, y, 0, data); | ||
436 : | current = data->currentQMV; | ||
437 : | xc = x/2; yc = y/2; | ||
438 : | } else { | ||
439 : | Reference = GetReference(x, y, data); | ||
440 : | current = data->currentMV; | ||
441 : | xc = x; yc = y; | ||
442 : | } | ||
443 : | t = d_mv_bits(x, y, data->predMV, data->iFcode, | ||
444 : | data->qpel^data->qpel_precision, data->rrv); | ||
445 : | edgomez | 1.20 | |
446 : | edgomez | 1.55 | sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
447 : | sad += (data->lambda16 * t * sad)>>10; | ||
448 : | edgomez | 1.20 | |
449 : | edgomez | 1.55 | if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
450 : | (yc >> 1) + roundtab_79[yc & 0x3], data); | ||
451 : | edgomez | 1.20 | |
452 : | edgomez | 1.55 | if (sad < *(data->iMinSAD)) { |
453 : | *(data->iMinSAD) = sad; | ||
454 : | current->x = x; current->y = y; | ||
455 : | *dir = Direction; | ||
456 : | edgomez | 1.20 | } |
457 : | chl | 1.3 | } |
458 : | |||
459 : | edgomez | 1.55 | static void |
460 : | CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) | ||
461 : | { | ||
462 : | // maximum speed - for P/B/I decision | ||
463 : | int32_t sad; | ||
464 : | |||
465 : | if ( (x > data->max_dx) || (x < data->min_dx) | ||
466 : | || (y > data->max_dy) || (y < data->min_dy) ) return; | ||
467 : | chl | 1.3 | |
468 : | syskin | 1.64 | sad = sad32v_c(data->Cur, data->RefP[0] + x/2 + (y/2)*(data->iEdgedWidth), |
469 : | edgomez | 1.55 | data->iEdgedWidth, data->temp+1); |
470 : | chl | 1.18 | |
471 : | edgomez | 1.55 | if (sad < *(data->iMinSAD)) { |
472 : | *(data->iMinSAD) = sad; | ||
473 : | data->currentMV[0].x = x; data->currentMV[0].y = y; | ||
474 : | *dir = Direction; | ||
475 : | } | ||
476 : | if (data->temp[1] < data->iMinSAD[1]) { | ||
477 : | data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } | ||
478 : | if (data->temp[2] < data->iMinSAD[2]) { | ||
479 : | data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } | ||
480 : | if (data->temp[3] < data->iMinSAD[3]) { | ||
481 : | data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } | ||
482 : | if (data->temp[4] < data->iMinSAD[4]) { | ||
483 : | data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } | ||
484 : | chl | 1.18 | |
485 : | edgomez | 1.55 | } |
486 : | chl | 1.18 | |
487 : | edgomez | 1.55 | static void |
488 : | CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) | ||
489 : | { | ||
490 : | int32_t sad, xb, yb, xcf, ycf, xcb, ycb; | ||
491 : | uint32_t t; | ||
492 : | const uint8_t *ReferenceF, *ReferenceB; | ||
493 : | VECTOR *current; | ||
494 : | |||
495 : | edgomez | 1.62 | if ((xf > data->max_dx) || (xf < data->min_dx) || |
496 : | (yf > data->max_dy) || (yf < data->min_dy)) | ||
497 : | return; | ||
498 : | edgomez | 1.55 | |
499 : | if (!data->qpel_precision) { | ||
500 : | ReferenceF = GetReference(xf, yf, data); | ||
501 : | xb = data->currentMV[1].x; yb = data->currentMV[1].y; | ||
502 : | ReferenceB = GetReferenceB(xb, yb, 1, data); | ||
503 : | current = data->currentMV; | ||
504 : | xcf = xf; ycf = yf; | ||
505 : | xcb = xb; ycb = yb; | ||
506 : | edgomez | 1.20 | } else { |
507 : | edgomez | 1.55 | ReferenceF = Interpolate16x16qpel(xf, yf, 0, data); |
508 : | xb = data->currentQMV[1].x; yb = data->currentQMV[1].y; | ||
509 : | current = data->currentQMV; | ||
510 : | ReferenceB = Interpolate16x16qpel(xb, yb, 1, data); | ||
511 : | xcf = xf/2; ycf = yf/2; | ||
512 : | xcb = xb/2; ycb = yb/2; | ||
513 : | } | ||
514 : | chl | 1.18 | |
515 : | edgomez | 1.55 | t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0) |
516 : | syskin | 1.64 | + d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
517 : | chl | 1.18 | |
518 : | edgomez | 1.55 | sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
519 : | sad += (data->lambda16 * t * sad)>>10; | ||
520 : | chl | 1.18 | |
521 : | edgomez | 1.55 | if (data->chroma) sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3], |
522 : | (ycf >> 1) + roundtab_79[ycf & 0x3], | ||
523 : | (xcb >> 1) + roundtab_79[xcb & 0x3], | ||
524 : | (ycb >> 1) + roundtab_79[ycb & 0x3], data); | ||
525 : | chl | 1.18 | |
526 : | edgomez | 1.55 | if (sad < *(data->iMinSAD)) { |
527 : | *(data->iMinSAD) = sad; | ||
528 : | current->x = xf; current->y = yf; | ||
529 : | *dir = Direction; | ||
530 : | chl | 1.18 | } |
531 : | } | ||
532 : | |||
533 : | edgomez | 1.55 | static void |
534 : | CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) | ||
535 : | chl | 1.18 | { |
536 : | edgomez | 1.55 | int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0; |
537 : | uint32_t k; | ||
538 : | const uint8_t *ReferenceF; | ||
539 : | const uint8_t *ReferenceB; | ||
540 : | VECTOR mvs, b_mvs; | ||
541 : | |||
542 : | syskin | 1.60 | if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
543 : | edgomez | 1.55 | |
544 : | for (k = 0; k < 4; k++) { | ||
545 : | mvs.x = data->directmvF[k].x + x; | ||
546 : | b_mvs.x = ((x == 0) ? | ||
547 : | data->directmvB[k].x | ||
548 : | : mvs.x - data->referencemv[k].x); | ||
549 : | chl | 1.18 | |
550 : | edgomez | 1.55 | mvs.y = data->directmvF[k].y + y; |
551 : | b_mvs.y = ((y == 0) ? | ||
552 : | data->directmvB[k].y | ||
553 : | : mvs.y - data->referencemv[k].y); | ||
554 : | chl | 1.18 | |
555 : | edgomez | 1.62 | if ((mvs.x > data->max_dx) || (mvs.x < data->min_dx) || |
556 : | (mvs.y > data->max_dy) || (mvs.y < data->min_dy) || | ||
557 : | (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx) || | ||
558 : | (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) ) | ||
559 : | return; | ||
560 : | chl | 1.18 | |
561 : | edgomez | 1.55 | if (data->qpel) { |
562 : | xcf += mvs.x/2; ycf += mvs.y/2; | ||
563 : | xcb += b_mvs.x/2; ycb += b_mvs.y/2; | ||
564 : | } else { | ||
565 : | xcf += mvs.x; ycf += mvs.y; | ||
566 : | xcb += b_mvs.x; ycb += b_mvs.y; | ||
567 : | mvs.x *= 2; mvs.y *= 2; //we move to qpel precision anyway | ||
568 : | b_mvs.x *= 2; b_mvs.y *= 2; | ||
569 : | } | ||
570 : | edgomez | 1.20 | |
571 : | edgomez | 1.55 | ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); |
572 : | ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); | ||
573 : | chl | 1.18 | |
574 : | edgomez | 1.55 | sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
575 : | ReferenceF, ReferenceB, data->iEdgedWidth); | ||
576 : | if (sad > *(data->iMinSAD)) return; | ||
577 : | } | ||
578 : | chl | 1.18 | |
579 : | edgomez | 1.55 | sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
580 : | chl | 1.18 | |
581 : | edgomez | 1.55 | if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
582 : | (ycf >> 3) + roundtab_76[ycf & 0xf], | ||
583 : | (xcb >> 3) + roundtab_76[xcb & 0xf], | ||
584 : | (ycb >> 3) + roundtab_76[ycb & 0xf], data); | ||
585 : | chl | 1.18 | |
586 : | edgomez | 1.55 | if (sad < *(data->iMinSAD)) { |
587 : | *(data->iMinSAD) = sad; | ||
588 : | data->currentMV->x = x; data->currentMV->y = y; | ||
589 : | *dir = Direction; | ||
590 : | chl | 1.18 | } |
591 : | } | ||
592 : | |||
593 : | edgomez | 1.55 | static void |
594 : | CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) | ||
595 : | Isibaar | 1.1 | { |
596 : | edgomez | 1.55 | int32_t sad, xcf, ycf, xcb, ycb; |
597 : | const uint8_t *ReferenceF; | ||
598 : | const uint8_t *ReferenceB; | ||
599 : | VECTOR mvs, b_mvs; | ||
600 : | |||
601 : | edgomez | 1.62 | if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
602 : | edgomez | 1.55 | |
603 : | mvs.x = data->directmvF[0].x + x; | ||
604 : | b_mvs.x = ((x == 0) ? | ||
605 : | data->directmvB[0].x | ||
606 : | : mvs.x - data->referencemv[0].x); | ||
607 : | |||
608 : | mvs.y = data->directmvF[0].y + y; | ||
609 : | b_mvs.y = ((y == 0) ? | ||
610 : | data->directmvB[0].y | ||
611 : | : mvs.y - data->referencemv[0].y); | ||
612 : | |||
613 : | edgomez | 1.62 | if ( (mvs.x > data->max_dx) || (mvs.x < data->min_dx) |
614 : | || (mvs.y > data->max_dy) || (mvs.y < data->min_dy) | ||
615 : | || (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx) | ||
616 : | || (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) ) return; | ||
617 : | edgomez | 1.55 | |
618 : | if (data->qpel) { | ||
619 : | xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2); | ||
620 : | xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2); | ||
621 : | ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data); | ||
622 : | ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); | ||
623 : | } else { | ||
624 : | xcf = 4*mvs.x; ycf = 4*mvs.y; | ||
625 : | xcb = 4*b_mvs.x; ycb = 4*b_mvs.y; | ||
626 : | ReferenceF = GetReference(mvs.x, mvs.y, data); | ||
627 : | ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data); | ||
628 : | } | ||
629 : | Isibaar | 1.1 | |
630 : | edgomez | 1.55 | sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
631 : | sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; | ||
632 : | Isibaar | 1.1 | |
633 : | edgomez | 1.55 | if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
634 : | (ycf >> 3) + roundtab_76[ycf & 0xf], | ||
635 : | (xcb >> 3) + roundtab_76[xcb & 0xf], | ||
636 : | (ycb >> 3) + roundtab_76[ycb & 0xf], data); | ||
637 : | edgomez | 1.20 | |
638 : | edgomez | 1.55 | if (sad < *(data->iMinSAD)) { |
639 : | *(data->iMinSAD) = sad; | ||
640 : | data->currentMV->x = x; data->currentMV->y = y; | ||
641 : | *dir = Direction; | ||
642 : | } | ||
643 : | } | ||
644 : | edgomez | 1.20 | |
645 : | Isibaar | 1.1 | |
646 : | edgomez | 1.55 | static void |
647 : | CheckCandidateBits16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) | ||
648 : | { | ||
649 : | edgomez | 1.20 | |
650 : | edgomez | 1.61 | int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; |
651 : | edgomez | 1.55 | int32_t bits = 0, sum; |
652 : | VECTOR * current; | ||
653 : | const uint8_t * ptr; | ||
654 : | int i, cbp = 0, t, xc, yc; | ||
655 : | |||
656 : | if ( (x > data->max_dx) || (x < data->min_dx) | ||
657 : | || (y > data->max_dy) || (y < data->min_dy) ) return; | ||
658 : | |||
659 : | if (!data->qpel_precision) { | ||
660 : | ptr = GetReference(x, y, data); | ||
661 : | current = data->currentMV; | ||
662 : | xc = x; yc = y; | ||
663 : | } else { // x and y are in 1/4 precision | ||
664 : | ptr = Interpolate16x16qpel(x, y, 0, data); | ||
665 : | current = data->currentQMV; | ||
666 : | xc = x/2; yc = y/2; | ||
667 : | } | ||
668 : | |||
669 : | for(i = 0; i < 4; i++) { | ||
670 : | int s = 8*((i&1) + (i>>1)*data->iEdgedWidth); | ||
671 : | transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth); | ||
672 : | fdct(in); | ||
673 : | if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); | ||
674 : | else sum = quant4_inter(coeff, in, data->lambda16); | ||
675 : | if (sum > 0) { | ||
676 : | cbp |= 1 << (5 - i); | ||
677 : | bits += data->temp[i] = CodeCoeffInter_CalcBits(coeff, scan_tables[0]); | ||
678 : | } else data->temp[i] = 0; | ||
679 : | } | ||
680 : | |||
681 : | bits += t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); | ||
682 : | |||
683 : | if (bits < data->iMinSAD[0]) { // there is still a chance, adding chroma | ||
684 : | xc = (xc >> 1) + roundtab_79[xc & 0x3]; | ||
685 : | yc = (yc >> 1) + roundtab_79[yc & 0x3]; | ||
686 : | |||
687 : | //chroma U | ||
688 : | syskin | 1.64 | ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefP[4], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); |
689 : | edgomez | 1.55 | transfer_8to16subro(in, ptr, data->CurU, data->iEdgedWidth/2); |
690 : | fdct(in); | ||
691 : | if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); | ||
692 : | else sum = quant4_inter(coeff, in, data->lambda16); | ||
693 : | if (sum > 0) { | ||
694 : | cbp |= 1 << (5 - 4); | ||
695 : | bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); | ||
696 : | } | ||
697 : | |||
698 : | if (bits < data->iMinSAD[0]) { | ||
699 : | //chroma V | ||
700 : | syskin | 1.64 | ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefP[5], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); |
701 : | edgomez | 1.55 | transfer_8to16subro(in, ptr, data->CurV, data->iEdgedWidth/2); |
702 : | fdct(in); | ||
703 : | if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); | ||
704 : | else sum = quant4_inter(coeff, in, data->lambda16); | ||
705 : | if (sum > 0) { | ||
706 : | cbp |= 1 << (5 - 5); | ||
707 : | bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); | ||
708 : | } | ||
709 : | } | ||
710 : | } | ||
711 : | chl | 1.18 | |
712 : | chl | 1.59 | bits += xvid_cbpy_tab[15-(cbp>>2)].len; |
713 : | edgomez | 1.55 | bits += mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; |
714 : | edgomez | 1.20 | |
715 : | edgomez | 1.55 | if (bits < data->iMinSAD[0]) { |
716 : | data->iMinSAD[0] = bits; | ||
717 : | current[0].x = x; current[0].y = y; | ||
718 : | *dir = Direction; | ||
719 : | } | ||
720 : | |||
721 : | if (data->temp[0] + t < data->iMinSAD[1]) { | ||
722 : | data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; } | ||
723 : | if (data->temp[1] < data->iMinSAD[2]) { | ||
724 : | data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; } | ||
725 : | if (data->temp[2] < data->iMinSAD[3]) { | ||
726 : | data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; } | ||
727 : | if (data->temp[3] < data->iMinSAD[4]) { | ||
728 : | data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; } | ||
729 : | edgomez | 1.20 | |
730 : | edgomez | 1.55 | } |
731 : | static void | ||
732 : | CheckCandidateBits8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) | ||
733 : | { | ||
734 : | Isibaar | 1.1 | |
735 : | edgomez | 1.61 | int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; |
736 : | edgomez | 1.55 | int32_t sum, bits; |
737 : | VECTOR * current; | ||
738 : | const uint8_t * ptr; | ||
739 : | int cbp; | ||
740 : | Isibaar | 1.1 | |
741 : | edgomez | 1.55 | if ( (x > data->max_dx) || (x < data->min_dx) |
742 : | || (y > data->max_dy) || (y < data->min_dy) ) return; | ||
743 : | Isibaar | 1.1 | |
744 : | edgomez | 1.55 | if (!data->qpel_precision) { |
745 : | ptr = GetReference(x, y, data); | ||
746 : | current = data->currentMV; | ||
747 : | } else { // x and y are in 1/4 precision | ||
748 : | ptr = Interpolate8x8qpel(x, y, 0, 0, data); | ||
749 : | current = data->currentQMV; | ||
750 : | edgomez | 1.20 | } |
751 : | |||
752 : | edgomez | 1.55 | transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth); |
753 : | fdct(in); | ||
754 : | if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); | ||
755 : | else sum = quant4_inter(coeff, in, data->lambda16); | ||
756 : | if (sum > 0) { | ||
757 : | bits = CodeCoeffInter_CalcBits(coeff, scan_tables[0]); | ||
758 : | cbp = 1; | ||
759 : | } else cbp = bits = 0; | ||
760 : | chl | 1.35 | |
761 : | edgomez | 1.55 | bits += sum = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
762 : | edgomez | 1.20 | |
763 : | edgomez | 1.55 | if (bits < data->iMinSAD[0]) { |
764 : | data->temp[0] = cbp; | ||
765 : | data->iMinSAD[0] = bits; | ||
766 : | current[0].x = x; current[0].y = y; | ||
767 : | *dir = Direction; | ||
768 : | edgomez | 1.20 | } |
769 : | edgomez | 1.55 | } |
770 : | edgomez | 1.20 | |
771 : | edgomez | 1.55 | /* CHECK_CANDIATE FUNCTIONS END */ |
772 : | edgomez | 1.20 | |
773 : | edgomez | 1.55 | /* MAINSEARCH FUNCTIONS START */ |
774 : | chl | 1.13 | |
775 : | edgomez | 1.55 | static void |
776 : | AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection) | ||
777 : | { | ||
778 : | Isibaar | 1.1 | |
779 : | edgomez | 1.55 | /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
780 : | chl | 1.13 | |
781 : | edgomez | 1.55 | int iDirection; |
782 : | chl | 1.13 | |
783 : | edgomez | 1.55 | for(;;) { //forever |
784 : | iDirection = 0; | ||
785 : | if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); | ||
786 : | if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); | ||
787 : | if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); | ||
788 : | if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); | ||
789 : | chl | 1.13 | |
790 : | edgomez | 1.55 | /* now we're doing diagonal checks near our candidate */ |
791 : | chl | 1.13 | |
792 : | edgomez | 1.55 | if (iDirection) { //if anything found |
793 : | bDirection = iDirection; | ||
794 : | iDirection = 0; | ||
795 : | x = data->currentMV->x; y = data->currentMV->y; | ||
796 : | if (bDirection & 3) { //our candidate is left or right | ||
797 : | CHECK_CANDIDATE(x, y + iDiamondSize, 8); | ||
798 : | CHECK_CANDIDATE(x, y - iDiamondSize, 4); | ||
799 : | } else { // what remains here is up or down | ||
800 : | CHECK_CANDIDATE(x + iDiamondSize, y, 2); | ||
801 : | CHECK_CANDIDATE(x - iDiamondSize, y, 1); | ||
802 : | } | ||
803 : | chl | 1.13 | |
804 : | edgomez | 1.55 | if (iDirection) { |
805 : | bDirection += iDirection; | ||
806 : | x = data->currentMV->x; y = data->currentMV->y; | ||
807 : | } | ||
808 : | } else { //about to quit, eh? not so fast.... | ||
809 : | switch (bDirection) { | ||
810 : | case 2: | ||
811 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
812 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
813 : | break; | ||
814 : | case 1: | ||
815 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
816 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
817 : | break; | ||
818 : | case 2 + 4: | ||
819 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
820 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
821 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
822 : | break; | ||
823 : | case 4: | ||
824 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
825 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
826 : | break; | ||
827 : | case 8: | ||
828 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
829 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
830 : | break; | ||
831 : | case 1 + 4: | ||
832 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
833 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
834 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
835 : | break; | ||
836 : | case 2 + 8: | ||
837 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
838 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
839 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
840 : | break; | ||
841 : | case 1 + 8: | ||
842 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
843 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
844 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
845 : | break; | ||
846 : | default: //1+2+4+8 == we didn't find anything at all | ||
847 : | CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); | ||
848 : | CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); | ||
849 : | CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); | ||
850 : | CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); | ||
851 : | break; | ||
852 : | } | ||
853 : | if (!iDirection) break; //ok, the end. really | ||
854 : | bDirection = iDirection; | ||
855 : | x = data->currentMV->x; y = data->currentMV->y; | ||
856 : | } | ||
857 : | } | ||
858 : | } | ||
859 : | chl | 1.13 | |
860 : | edgomez | 1.55 | static void |
861 : | SquareSearch(int x, int y, const SearchData * const data, int bDirection) | ||
862 : | { | ||
863 : | int iDirection; | ||
864 : | Isibaar | 1.1 | |
865 : | edgomez | 1.55 | do { |
866 : | iDirection = 0; | ||
867 : | if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64); | ||
868 : | if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128); | ||
869 : | if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32); | ||
870 : | if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128); | ||
871 : | if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64); | ||
872 : | if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128); | ||
873 : | if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128); | ||
874 : | if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128); | ||
875 : | |||
876 : | bDirection = iDirection; | ||
877 : | x = data->currentMV->x; y = data->currentMV->y; | ||
878 : | } while (iDirection); | ||
879 : | } | ||
880 : | Isibaar | 1.1 | |
881 : | edgomez | 1.55 | static void |
882 : | DiamondSearch(int x, int y, const SearchData * const data, int bDirection) | ||
883 : | { | ||
884 : | Isibaar | 1.1 | |
885 : | edgomez | 1.55 | /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
886 : | chl | 1.13 | |
887 : | edgomez | 1.55 | int iDirection; |
888 : | edgomez | 1.20 | |
889 : | edgomez | 1.55 | do { |
890 : | iDirection = 0; | ||
891 : | if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); | ||
892 : | if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); | ||
893 : | if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); | ||
894 : | if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); | ||
895 : | chl | 1.13 | |
896 : | edgomez | 1.55 | /* now we're doing diagonal checks near our candidate */ |
897 : | Isibaar | 1.1 | |
898 : | edgomez | 1.55 | if (iDirection) { //checking if anything found |
899 : | bDirection = iDirection; | ||
900 : | iDirection = 0; | ||
901 : | x = data->currentMV->x; y = data->currentMV->y; | ||
902 : | if (bDirection & 3) { //our candidate is left or right | ||
903 : | CHECK_CANDIDATE(x, y + iDiamondSize, 8); | ||
904 : | CHECK_CANDIDATE(x, y - iDiamondSize, 4); | ||
905 : | } else { // what remains here is up or down | ||
906 : | CHECK_CANDIDATE(x + iDiamondSize, y, 2); | ||
907 : | CHECK_CANDIDATE(x - iDiamondSize, y, 1); | ||
908 : | edgomez | 1.20 | } |
909 : | edgomez | 1.55 | bDirection += iDirection; |
910 : | x = data->currentMV->x; y = data->currentMV->y; | ||
911 : | } | ||
912 : | } | ||
913 : | while (iDirection); | ||
914 : | } | ||
915 : | edgomez | 1.20 | |
916 : | edgomez | 1.55 | /* MAINSEARCH FUNCTIONS END */ |
917 : | Isibaar | 1.1 | |
918 : | edgomez | 1.55 | static void |
919 : | SubpelRefine(const SearchData * const data) | ||
920 : | { | ||
921 : | /* Do a half-pel or q-pel refinement */ | ||
922 : | const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV; | ||
923 : | int iDirection; //only needed because macro expects it | ||
924 : | |||
925 : | CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0); | ||
926 : | CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0); | ||
927 : | CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0); | ||
928 : | CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0); | ||
929 : | CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0); | ||
930 : | CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0); | ||
931 : | CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0); | ||
932 : | CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0); | ||
933 : | } | ||
934 : | |||
935 : | static __inline int | ||
936 : | SkipDecisionP(const IMAGE * current, const IMAGE * reference, | ||
937 : | const int x, const int y, | ||
938 : | const uint32_t stride, const uint32_t iQuant, int rrv) | ||
939 : | edgomez | 1.20 | |
940 : | edgomez | 1.55 | { |
941 : | edgomez | 1.62 | int offset = (x + y*stride)*8; |
942 : | edgomez | 1.55 | if(!rrv) { |
943 : | edgomez | 1.62 | uint32_t sadC = sad8(current->u + offset, |
944 : | reference->u + offset, stride); | ||
945 : | edgomez | 1.55 | if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
946 : | edgomez | 1.62 | sadC += sad8(current->v + offset, |
947 : | reference->v + offset, stride); | ||
948 : | edgomez | 1.55 | if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
949 : | return 1; | ||
950 : | Isibaar | 1.1 | |
951 : | edgomez | 1.55 | } else { |
952 : | edgomez | 1.62 | uint32_t sadC = sad16(current->u + 2*offset, |
953 : | reference->u + 2*offset, stride, 256*4096); | ||
954 : | edgomez | 1.55 | if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
955 : | edgomez | 1.62 | sadC += sad16(current->v + 2*offset, |
956 : | reference->v + 2*offset, stride, 256*4096); | ||
957 : | edgomez | 1.55 | if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
958 : | return 1; | ||
959 : | edgomez | 1.2 | } |
960 : | edgomez | 1.55 | } |
961 : | Isibaar | 1.1 | |
962 : | edgomez | 1.55 | static __inline void |
963 : | SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) | ||
964 : | { | ||
965 : | pMB->mode = MODE_NOT_CODED; | ||
966 : | pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV; | ||
967 : | pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV; | ||
968 : | pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; | ||
969 : | } | ||
970 : | Isibaar | 1.1 | |
971 : | edgomez | 1.55 | bool |
972 : | MotionEstimation(MBParam * const pParam, | ||
973 : | syskin | 1.64 | FRAMEINFO * const current, |
974 : | FRAMEINFO * const reference, | ||
975 : | const IMAGE * const pRefH, | ||
976 : | const IMAGE * const pRefV, | ||
977 : | const IMAGE * const pRefHV, | ||
978 : | const uint32_t iLimit) | ||
979 : | edgomez | 1.55 | { |
980 : | MACROBLOCK *const pMBs = current->mbs; | ||
981 : | const IMAGE *const pCurrent = ¤t->image; | ||
982 : | const IMAGE *const pRef = &reference->image; | ||
983 : | Isibaar | 1.1 | |
984 : | edgomez | 1.55 | uint32_t mb_width = pParam->mb_width; |
985 : | uint32_t mb_height = pParam->mb_height; | ||
986 : | const uint32_t iEdgedWidth = pParam->edged_width; | ||
987 : | const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->global_flags); | ||
988 : | chl | 1.18 | |
989 : | edgomez | 1.55 | uint32_t x, y; |
990 : | uint32_t iIntra = 0; | ||
991 : | int32_t quant = current->quant, sad00; | ||
992 : | syskin | 1.64 | int skip_thresh = INITIAL_SKIP_THRESH * |
993 : | (current->global_flags & XVID_REDUCED ? 4:1) * | ||
994 : | syskin | 1.60 | (current->global_flags & XVID_MODEDECISION_BITS ? 2:1); |
995 : | Isibaar | 1.1 | |
996 : | edgomez | 1.55 | // some pre-initialized thingies for SearchP |
997 : | int32_t temp[8]; | ||
998 : | VECTOR currentMV[5]; | ||
999 : | VECTOR currentQMV[5]; | ||
1000 : | int32_t iMinSAD[5]; | ||
1001 : | edgomez | 1.61 | DECLARE_ALIGNED_MATRIX(dct_space, 2, 64, int16_t, CACHE_LINE); |
1002 : | edgomez | 1.55 | SearchData Data; |
1003 : | memset(&Data, 0, sizeof(SearchData)); | ||
1004 : | Data.iEdgedWidth = iEdgedWidth; | ||
1005 : | Data.currentMV = currentMV; | ||
1006 : | Data.currentQMV = currentQMV; | ||
1007 : | Data.iMinSAD = iMinSAD; | ||
1008 : | Data.temp = temp; | ||
1009 : | Data.iFcode = current->fcode; | ||
1010 : | Data.rounding = pParam->m_rounding_type; | ||
1011 : | Data.qpel = pParam->m_quarterpel; | ||
1012 : | Data.chroma = MotionFlags & PMV_CHROMA16; | ||
1013 : | Data.rrv = current->global_flags & XVID_REDUCED; | ||
1014 : | edgomez | 1.61 | Data.dctSpace = dct_space; |
1015 : | edgomez | 1.55 | |
1016 : | if ((current->global_flags & XVID_REDUCED)) { | ||
1017 : | mb_width = (pParam->width + 31) / 32; | ||
1018 : | mb_height = (pParam->height + 31) / 32; | ||
1019 : | Data.qpel = 0; | ||
1020 : | } | ||
1021 : | |||
1022 : | Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose) | ||
1023 : | if (sadInit) (*sadInit) (); | ||
1024 : | |||
1025 : | for (y = 0; y < mb_height; y++) { | ||
1026 : | for (x = 0; x < mb_width; x++) { | ||
1027 : | MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; | ||
1028 : | |||
1029 : | if (!Data.rrv) pMB->sad16 = | ||
1030 : | sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16, | ||
1031 : | pRef->y + (x + y * iEdgedWidth) * 16, | ||
1032 : | pParam->edged_width, pMB->sad8 ); | ||
1033 : | |||
1034 : | else pMB->sad16 = | ||
1035 : | sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32, | ||
1036 : | pRef->y + (x + y * iEdgedWidth) * 32, | ||
1037 : | pParam->edged_width, pMB->sad8 ); | ||
1038 : | |||
1039 : | if (Data.chroma) { | ||
1040 : | Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8, | ||
1041 : | pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2) | ||
1042 : | + sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8, | ||
1043 : | pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); | ||
1044 : | pMB->sad16 += Data.temp[7]; | ||
1045 : | } | ||
1046 : | chl | 1.26 | |
1047 : | edgomez | 1.55 | sad00 = pMB->sad16; |
1048 : | edgomez | 1.20 | |
1049 : | edgomez | 1.55 | if (!(current->global_flags & XVID_LUMIMASKING)) { |
1050 : | pMB->dquant = NO_CHANGE; | ||
1051 : | } else { | ||
1052 : | if (pMB->dquant != NO_CHANGE) { | ||
1053 : | quant += DQtab[pMB->dquant]; | ||
1054 : | if (quant > 31) quant = 31; | ||
1055 : | else if (quant < 1) quant = 1; | ||
1056 : | } | ||
1057 : | } | ||
1058 : | pMB->quant = current->quant; | ||
1059 : | Isibaar | 1.1 | |
1060 : | edgomez | 1.55 | //initial skip decision |
1061 : | /* no early skip for GMC (global vector = skip vector is unknown!) */ | ||
1062 : | if (!(current->global_flags & XVID_GMC)) { /* no fast SKIP for S(GMC)-VOPs */ | ||
1063 : | syskin | 1.60 | if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * skip_thresh) |
1064 : | edgomez | 1.55 | if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) { |
1065 : | SkipMacroblockP(pMB, sad00); | ||
1066 : | continue; | ||
1067 : | } | ||
1068 : | } | ||
1069 : | Isibaar | 1.1 | |
1070 : | edgomez | 1.55 | SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
1071 : | y, MotionFlags, current->global_flags, pMB->quant, | ||
1072 : | &Data, pParam, pMBs, reference->mbs, | ||
1073 : | current->global_flags & XVID_INTER4V, pMB); | ||
1074 : | |||
1075 : | /* final skip decision, a.k.a. "the vector you found, really that good?" */ | ||
1076 : | syskin | 1.60 | if (!(current->global_flags & XVID_GMC || current->global_flags & XVID_MODEDECISION_BITS)) { |
1077 : | edgomez | 1.55 | if ( pMB->dquant == NO_CHANGE && sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) { |
1078 : | syskin | 1.60 | if ( (100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1) ) |
1079 : | if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) | ||
1080 : | SkipMacroblockP(pMB, sad00); | ||
1081 : | edgomez | 1.55 | } |
1082 : | edgomez | 1.20 | } |
1083 : | edgomez | 1.55 | if (pMB->mode == MODE_INTRA) |
1084 : | if (++iIntra > iLimit) return 1; | ||
1085 : | Isibaar | 1.1 | } |
1086 : | edgomez | 1.55 | } |
1087 : | |||
1088 : | if (current->global_flags & XVID_GMC ) /* GMC only for S(GMC)-VOPs */ | ||
1089 : | { | ||
1090 : | current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV); | ||
1091 : | } | ||
1092 : | return 0; | ||
1093 : | } | ||
1094 : | Isibaar | 1.1 | |
1095 : | edgomez | 1.20 | |
1096 : | edgomez | 1.55 | static __inline int |
1097 : | make_mask(const VECTOR * const pmv, const int i) | ||
1098 : | { | ||
1099 : | int mask = 255, j; | ||
1100 : | for (j = 0; j < i; j++) { | ||
1101 : | if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already | ||
1102 : | if (pmv[i].x == pmv[j].x) { | ||
1103 : | if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4; | ||
1104 : | else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8; | ||
1105 : | } else | ||
1106 : | if (pmv[i].y == pmv[j].y) { | ||
1107 : | if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1; | ||
1108 : | else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2; | ||
1109 : | edgomez | 1.20 | } |
1110 : | Isibaar | 1.1 | } |
1111 : | edgomez | 1.55 | return mask; |
1112 : | } | ||
1113 : | Isibaar | 1.1 | |
1114 : | edgomez | 1.55 | static __inline void |
1115 : | PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, | ||
1116 : | int iHcount, const MACROBLOCK * const prevMB, int rrv) | ||
1117 : | { | ||
1118 : | |||
1119 : | //this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself | ||
1120 : | if (rrv) { iWcount /= 2; iHcount /= 2; } | ||
1121 : | |||
1122 : | if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour | ||
1123 : | pmv[5].x = EVEN(pmv[3].x); | ||
1124 : | pmv[5].y = EVEN(pmv[3].y); | ||
1125 : | } else pmv[5].x = pmv[5].y = 0; | ||
1126 : | Isibaar | 1.1 | |
1127 : | edgomez | 1.55 | if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour |
1128 : | else pmv[3].x = pmv[3].y = 0; | ||
1129 : | Isibaar | 1.1 | |
1130 : | edgomez | 1.55 | if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour |
1131 : | else pmv[4].x = pmv[4].y = 0; | ||
1132 : | Isibaar | 1.1 | |
1133 : | edgomez | 1.55 | // [1] median prediction |
1134 : | pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); | ||
1135 : | Isibaar | 1.1 | |
1136 : | edgomez | 1.55 | pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask |
1137 : | Isibaar | 1.1 | |
1138 : | edgomez | 1.55 | pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
1139 : | pmv[2].y = EVEN(prevMB->mvs[0].y); | ||
1140 : | Isibaar | 1.1 | |
1141 : | edgomez | 1.55 | if ((x < iWcount-1) && (y < iHcount-1)) { |
1142 : | pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame | ||
1143 : | pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); | ||
1144 : | } else pmv[6].x = pmv[6].y = 0; | ||
1145 : | Isibaar | 1.1 | |
1146 : | edgomez | 1.55 | if (rrv) { |
1147 : | int i; | ||
1148 : | for (i = 0; i < 7; i++) { | ||
1149 : | pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); | ||
1150 : | pmv[i].y = RRV_MV_SCALEUP(pmv[i].y); | ||
1151 : | chl | 1.38 | } |
1152 : | edgomez | 1.2 | } |
1153 : | Isibaar | 1.1 | } |
1154 : | |||
1155 : | edgomez | 1.55 | static int |
1156 : | ModeDecision(const uint32_t iQuant, SearchData * const Data, | ||
1157 : | int inter4v, | ||
1158 : | MACROBLOCK * const pMB, | ||
1159 : | const MACROBLOCK * const pMBs, | ||
1160 : | const int x, const int y, | ||
1161 : | const MBParam * const pParam, | ||
1162 : | const uint32_t MotionFlags, | ||
1163 : | const uint32_t GlobalFlags) | ||
1164 : | { | ||
1165 : | chl | 1.39 | |
1166 : | edgomez | 1.55 | int mode = MODE_INTER; |
1167 : | chl | 1.39 | |
1168 : | edgomez | 1.55 | if (!(GlobalFlags & XVID_MODEDECISION_BITS)) { //normal, fast, SAD-based mode decision |
1169 : | int sad; | ||
1170 : | int InterBias = MV16_INTER_BIAS; | ||
1171 : | if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + | ||
1172 : | Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) { | ||
1173 : | edgomez | 1.62 | mode = MODE_INTER; |
1174 : | sad = Data->iMinSAD[0]; | ||
1175 : | edgomez | 1.55 | } else { |
1176 : | mode = MODE_INTER4V; | ||
1177 : | sad = Data->iMinSAD[1] + Data->iMinSAD[2] + | ||
1178 : | Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant; | ||
1179 : | Data->iMinSAD[0] = sad; | ||
1180 : | } | ||
1181 : | |||
1182 : | /* intra decision */ | ||
1183 : | |||
1184 : | if (iQuant > 8) InterBias += 100 * (iQuant - 8); // to make high quants work | ||
1185 : | if (y != 0) | ||
1186 : | if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; | ||
1187 : | if (x != 0) | ||
1188 : | if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; | ||
1189 : | |||
1190 : | if (Data->chroma) InterBias += 50; // to compensate bigger SAD | ||
1191 : | if (Data->rrv) InterBias *= 4; | ||
1192 : | |||
1193 : | if (InterBias < pMB->sad16) { | ||
1194 : | int32_t deviation; | ||
1195 : | if (!Data->rrv) deviation = dev16(Data->Cur, Data->iEdgedWidth); | ||
1196 : | else deviation = dev16(Data->Cur, Data->iEdgedWidth) + | ||
1197 : | dev16(Data->Cur+8, Data->iEdgedWidth) + | ||
1198 : | dev16(Data->Cur + 8*Data->iEdgedWidth, Data->iEdgedWidth) + | ||
1199 : | dev16(Data->Cur+8+8*Data->iEdgedWidth, Data->iEdgedWidth); | ||
1200 : | chl | 1.39 | |
1201 : | edgomez | 1.62 | if (deviation < (sad - InterBias)) return MODE_INTRA; |
1202 : | edgomez | 1.55 | } |
1203 : | return mode; | ||
1204 : | chl | 1.39 | |
1205 : | edgomez | 1.55 | } else { |
1206 : | chl | 1.39 | |
1207 : | edgomez | 1.55 | int bits, intra, i; |
1208 : | VECTOR backup[5], *v; | ||
1209 : | Data->lambda16 = iQuant; | ||
1210 : | Data->lambda8 = pParam->m_quant_type; | ||
1211 : | chl | 1.39 | |
1212 : | edgomez | 1.55 | v = Data->qpel ? Data->currentQMV : Data->currentMV; |
1213 : | for (i = 0; i < 5; i++) { | ||
1214 : | Data->iMinSAD[i] = 256*4096; | ||
1215 : | backup[i] = v[i]; | ||
1216 : | } | ||
1217 : | chl | 1.39 | |
1218 : | edgomez | 1.55 | bits = CountMBBitsInter(Data, pMBs, x, y, pParam, MotionFlags); |
1219 : | if (bits == 0) return MODE_INTER; // quick stop | ||
1220 : | chl | 1.39 | |
1221 : | edgomez | 1.55 | if (inter4v) { |
1222 : | edgomez | 1.62 | int bits_inter4v = CountMBBitsInter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup); |
1223 : | if (bits_inter4v < bits) { Data->iMinSAD[0] = bits = bits_inter4v; mode = MODE_INTER4V; } | ||
1224 : | edgomez | 1.55 | } |
1225 : | chl | 1.39 | |
1226 : | |||
1227 : | edgomez | 1.55 | intra = CountMBBitsIntra(Data); |
1228 : | chl | 1.39 | |
1229 : | edgomez | 1.55 | if (intra < bits) { *Data->iMinSAD = bits = intra; return MODE_INTRA; } |
1230 : | chl | 1.39 | |
1231 : | edgomez | 1.55 | return mode; |
1232 : | chl | 1.39 | } |
1233 : | } | ||
1234 : | |||
1235 : | edgomez | 1.55 | static void |
1236 : | SearchP(const IMAGE * const pRef, | ||
1237 : | const uint8_t * const pRefH, | ||
1238 : | const uint8_t * const pRefV, | ||
1239 : | const uint8_t * const pRefHV, | ||
1240 : | const IMAGE * const pCur, | ||
1241 : | const int x, | ||
1242 : | const int y, | ||
1243 : | const uint32_t MotionFlags, | ||
1244 : | const uint32_t GlobalFlags, | ||
1245 : | const uint32_t iQuant, | ||
1246 : | SearchData * const Data, | ||
1247 : | const MBParam * const pParam, | ||
1248 : | const MACROBLOCK * const pMBs, | ||
1249 : | const MACROBLOCK * const prevMBs, | ||
1250 : | int inter4v, | ||
1251 : | MACROBLOCK * const pMB) | ||
1252 : | { | ||
1253 : | |||
1254 : | int i, iDirection = 255, mask, threshA; | ||
1255 : | VECTOR pmv[7]; | ||
1256 : | chl | 1.39 | |
1257 : | edgomez | 1.55 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1258 : | pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); | ||
1259 : | chl | 1.39 | |
1260 : | edgomez | 1.55 | get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); |
1261 : | |||
1262 : | Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache | ||
1263 : | i = Data->rrv ? 2 : 1; | ||
1264 : | Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i; | ||
1265 : | Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; | ||
1266 : | Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; | ||
1267 : | |||
1268 : | syskin | 1.64 | Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16*i; |
1269 : | Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16*i; | ||
1270 : | Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16*i; | ||
1271 : | Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16*i; | ||
1272 : | Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; | ||
1273 : | Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; | ||
1274 : | edgomez | 1.55 | |
1275 : | Data->lambda16 = lambda_vec16[iQuant]; | ||
1276 : | Data->lambda8 = lambda_vec8[iQuant]; | ||
1277 : | Data->qpel_precision = 0; | ||
1278 : | |||
1279 : | if (pMB->dquant != NO_CHANGE) inter4v = 0; | ||
1280 : | |||
1281 : | edgomez | 1.62 | memset(Data->currentMV, 0, 5*sizeof(VECTOR)); |
1282 : | edgomez | 1.55 | |
1283 : | if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); | ||
1284 : | else Data->predMV = pmv[0]; | ||
1285 : | |||
1286 : | i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0); | ||
1287 : | Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10); | ||
1288 : | Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10); | ||
1289 : | Data->iMinSAD[2] = pMB->sad8[1]; | ||
1290 : | Data->iMinSAD[3] = pMB->sad8[2]; | ||
1291 : | Data->iMinSAD[4] = pMB->sad8[3]; | ||
1292 : | |||
1293 : | if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) || (x | y)) { | ||
1294 : | threshA = Data->temp[0]; // that's where we keep this SAD atm | ||
1295 : | if (threshA < 512) threshA = 512; | ||
1296 : | else if (threshA > 1024) threshA = 1024; | ||
1297 : | } else | ||
1298 : | threshA = 512; | ||
1299 : | chl | 1.39 | |
1300 : | edgomez | 1.55 | PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
1301 : | prevMBs + x + y * pParam->mb_width, Data->rrv); | ||
1302 : | chl | 1.39 | |
1303 : | edgomez | 1.55 | if (!Data->rrv) { |
1304 : | if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16; | ||
1305 : | else CheckCandidate = CheckCandidate16no4v; //for extra speed | ||
1306 : | } else CheckCandidate = CheckCandidate32; | ||
1307 : | |||
1308 : | /* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/ | ||
1309 : | |||
1310 : | for (i = 1; i < 7; i++) { | ||
1311 : | if (!(mask = make_mask(pmv, i)) ) continue; | ||
1312 : | CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data); | ||
1313 : | if (Data->iMinSAD[0] <= threshA) break; | ||
1314 : | } | ||
1315 : | |||
1316 : | if ((Data->iMinSAD[0] <= threshA) || | ||
1317 : | (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && | ||
1318 : | (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { | ||
1319 : | if (!(GlobalFlags & XVID_MODEDECISION_BITS)) inter4v = 0; } | ||
1320 : | else { | ||
1321 : | |||
1322 : | MainSearchFunc * MainSearchPtr; | ||
1323 : | if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; | ||
1324 : | else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; | ||
1325 : | else MainSearchPtr = DiamondSearch; | ||
1326 : | |||
1327 : | MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection); | ||
1328 : | |||
1329 : | /* extended search, diamond starting in 0,0 and in prediction. | ||
1330 : | note that this search is/might be done in halfpel positions, | ||
1331 : | which makes it more different than the diamond above */ | ||
1332 : | |||
1333 : | if (MotionFlags & PMV_EXTSEARCH16) { | ||
1334 : | int32_t bSAD; | ||
1335 : | VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; | ||
1336 : | if (Data->rrv) { | ||
1337 : | startMV.x = RRV_MV_SCALEUP(startMV.x); | ||
1338 : | startMV.y = RRV_MV_SCALEUP(startMV.y); | ||
1339 : | } | ||
1340 : | if (!(MVequal(startMV, backupMV))) { | ||
1341 : | bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; | ||
1342 : | |||
1343 : | CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data); | ||
1344 : | MainSearchPtr(startMV.x, startMV.y, Data, 255); | ||
1345 : | if (bSAD < Data->iMinSAD[0]) { | ||
1346 : | Data->currentMV[0] = backupMV; | ||
1347 : | Data->iMinSAD[0] = bSAD; } | ||
1348 : | } | ||
1349 : | edgomez | 1.20 | |
1350 : | edgomez | 1.55 | backupMV = Data->currentMV[0]; |
1351 : | startMV.x = startMV.y = 1; | ||
1352 : | if (!(MVequal(startMV, backupMV))) { | ||
1353 : | bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; | ||
1354 : | |||
1355 : | CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data); | ||
1356 : | MainSearchPtr(startMV.x, startMV.y, Data, 255); | ||
1357 : | if (bSAD < Data->iMinSAD[0]) { | ||
1358 : | Data->currentMV[0] = backupMV; | ||
1359 : | Data->iMinSAD[0] = bSAD; } | ||
1360 : | } | ||
1361 : | } | ||
1362 : | } | ||
1363 : | |||
1364 : | if (MotionFlags & PMV_HALFPELREFINE16) | ||
1365 : | if ((!(MotionFlags & HALFPELREFINE16_BITS)) || Data->iMinSAD[0] < 200*(int)iQuant) | ||
1366 : | SubpelRefine(Data); | ||
1367 : | |||
1368 : | for(i = 0; i < 5; i++) { | ||
1369 : | Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors | ||
1370 : | Data->currentQMV[i].y = 2 * Data->currentMV[i].y; | ||
1371 : | } | ||
1372 : | Isibaar | 1.1 | |
1373 : | syskin | 1.60 | if (MotionFlags & PMV_QUARTERPELREFINE16) { |
1374 : | syskin | 1.64 | |
1375 : | syskin | 1.60 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1376 : | pParam->width, pParam->height, Data->iFcode, 1, 0); | ||
1377 : | |||
1378 : | edgomez | 1.55 | if ((!(MotionFlags & QUARTERPELREFINE16_BITS)) || (Data->iMinSAD[0] < 200*(int)iQuant)) { |
1379 : | Data->qpel_precision = 1; | ||
1380 : | SubpelRefine(Data); | ||
1381 : | } | ||
1382 : | syskin | 1.60 | } |
1383 : | Isibaar | 1.1 | |
1384 : | edgomez | 1.55 | if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) && (Data->iMinSAD[0] < (int32_t)iQuant * 30)) inter4v = 0; |
1385 : | chl | 1.18 | |
1386 : | edgomez | 1.55 | if (inter4v && (!(GlobalFlags & XVID_MODEDECISION_BITS) || |
1387 : | (!(MotionFlags & QUARTERPELREFINE8_BITS)) || (!(MotionFlags & HALFPELREFINE8_BITS)) || | ||
1388 : | ((!(MotionFlags & EXTSEARCH_BITS)) && (!(MotionFlags&PMV_EXTSEARCH8)) ))) { | ||
1389 : | // if decision is BITS-based and all refinement steps will be done in BITS domain, there is no reason to call this loop | ||
1390 : | chl | 1.18 | |
1391 : | edgomez | 1.55 | SearchData Data8; |
1392 : | memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data | ||
1393 : | chl | 1.18 | |
1394 : | edgomez | 1.55 | Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
1395 : | Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); | ||
1396 : | Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); | ||
1397 : | Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); | ||
1398 : | edgomez | 1.20 | |
1399 : | edgomez | 1.55 | if ((Data->chroma) && (!(GlobalFlags & XVID_MODEDECISION_BITS))) { |
1400 : | // chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, there is no reason to compute it | ||
1401 : | int sumx = 0, sumy = 0; | ||
1402 : | const int div = 1 + Data->qpel; | ||
1403 : | const VECTOR * const mv = Data->qpel ? pMB->qmvs : pMB->mvs; | ||
1404 : | chl | 1.13 | |
1405 : | edgomez | 1.55 | for (i = 0; i < 4; i++) { |
1406 : | sumx += mv[i].x / div; | ||
1407 : | sumy += mv[i].y / div; | ||
1408 : | } | ||
1409 : | chl | 1.13 | |
1410 : | edgomez | 1.55 | Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf], |
1411 : | (sumy >> 3) + roundtab_76[sumy & 0xf], Data); | ||
1412 : | } | ||
1413 : | } | ||
1414 : | chl | 1.13 | |
1415 : | edgomez | 1.55 | inter4v = ModeDecision(iQuant, Data, inter4v, pMB, pMBs, x, y, pParam, MotionFlags, GlobalFlags); |
1416 : | chl | 1.13 | |
1417 : | edgomez | 1.55 | if (Data->rrv) { |
1418 : | Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); | ||
1419 : | Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); | ||
1420 : | } | ||
1421 : | chl | 1.13 | |
1422 : | edgomez | 1.55 | if (inter4v == MODE_INTER) { |
1423 : | pMB->mode = MODE_INTER; | ||
1424 : | pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; | ||
1425 : | pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; | ||
1426 : | chl | 1.3 | |
1427 : | edgomez | 1.55 | if(Data->qpel) { |
1428 : | pMB->qmvs[0] = pMB->qmvs[1] | ||
1429 : | = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; | ||
1430 : | pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; | ||
1431 : | pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; | ||
1432 : | } else { | ||
1433 : | pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; | ||
1434 : | pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; | ||
1435 : | } | ||
1436 : | Isibaar | 1.1 | |
1437 : | edgomez | 1.55 | } else if (inter4v == MODE_INTER4V) { |
1438 : | pMB->mode = MODE_INTER4V; | ||
1439 : | pMB->sad16 = Data->iMinSAD[0]; | ||
1440 : | } else { // INTRA mode | ||
1441 : | SkipMacroblockP(pMB, 0); // not skip, but similar enough | ||
1442 : | pMB->mode = MODE_INTRA; | ||
1443 : | } | ||
1444 : | chl | 1.13 | |
1445 : | edgomez | 1.55 | } |
1446 : | Isibaar | 1.1 | |
1447 : | edgomez | 1.55 | static void |
1448 : | Search8(const SearchData * const OldData, | ||
1449 : | const int x, const int y, | ||
1450 : | const uint32_t MotionFlags, | ||
1451 : | const MBParam * const pParam, | ||
1452 : | MACROBLOCK * const pMB, | ||
1453 : | const MACROBLOCK * const pMBs, | ||
1454 : | const int block, | ||
1455 : | SearchData * const Data) | ||
1456 : | { | ||
1457 : | int i = 0; | ||
1458 : | Data->iMinSAD = OldData->iMinSAD + 1 + block; | ||
1459 : | Data->currentMV = OldData->currentMV + 1 + block; | ||
1460 : | Data->currentQMV = OldData->currentQMV + 1 + block; | ||
1461 : | |||
1462 : | if(Data->qpel) { | ||
1463 : | Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); | ||
1464 : | if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y, | ||
1465 : | Data->predMV, Data->iFcode, 0, 0); | ||
1466 : | } else { | ||
1467 : | Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); | ||
1468 : | if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y, | ||
1469 : | Data->predMV, Data->iFcode, 0, Data->rrv); | ||
1470 : | } | ||
1471 : | Isibaar | 1.1 | |
1472 : | edgomez | 1.55 | *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10; |
1473 : | chl | 1.9 | |
1474 : | edgomez | 1.55 | if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8|PMV_QUARTERPELREFINE8)) { |
1475 : | chl | 1.9 | |
1476 : | syskin | 1.64 | if (Data->rrv) i = 16; else i = 8; |
1477 : | |||
1478 : | Data->RefP[0] = OldData->RefP[0] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); | ||
1479 : | Data->RefP[1] = OldData->RefP[1] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); | ||
1480 : | Data->RefP[2] = OldData->RefP[2] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); | ||
1481 : | Data->RefP[3] = OldData->RefP[3] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); | ||
1482 : | Isibaar | 1.1 | |
1483 : | syskin | 1.64 | Data->Cur = OldData->Cur + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1484 : | edgomez | 1.55 | Data->qpel_precision = 0; |
1485 : | Isibaar | 1.1 | |
1486 : | edgomez | 1.55 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
1487 : | pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); | ||
1488 : | Isibaar | 1.1 | |
1489 : | edgomez | 1.55 | if (!Data->rrv) CheckCandidate = CheckCandidate8; |
1490 : | else CheckCandidate = CheckCandidate16no4v; | ||
1491 : | edgomez | 1.20 | |
1492 : | edgomez | 1.55 | if (MotionFlags & PMV_EXTSEARCH8 && (!(MotionFlags & EXTSEARCH_BITS))) { |
1493 : | int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD | ||
1494 : | Isibaar | 1.1 | |
1495 : | edgomez | 1.55 | MainSearchFunc *MainSearchPtr; |
1496 : | if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; | ||
1497 : | else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; | ||
1498 : | else MainSearchPtr = DiamondSearch; | ||
1499 : | Isibaar | 1.1 | |
1500 : | edgomez | 1.55 | MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1501 : | edgomez | 1.20 | |
1502 : | edgomez | 1.55 | if(*(Data->iMinSAD) < temp_sad) { |
1503 : | Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector | ||
1504 : | Data->currentQMV->y = 2 * Data->currentMV->y; | ||
1505 : | edgomez | 1.20 | } |
1506 : | Isibaar | 1.1 | } |
1507 : | |||
1508 : | edgomez | 1.55 | if (MotionFlags & PMV_HALFPELREFINE8) { |
1509 : | int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD | ||
1510 : | edgomez | 1.20 | |
1511 : | edgomez | 1.55 | SubpelRefine(Data); // perform halfpel refine of current best vector |
1512 : | |||
1513 : | if(*(Data->iMinSAD) < temp_sad) { // we have found a better match | ||
1514 : | Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector | ||
1515 : | Data->currentQMV->y = 2 * Data->currentMV->y; | ||
1516 : | edgomez | 1.20 | } |
1517 : | Isibaar | 1.1 | } |
1518 : | edgomez | 1.55 | |
1519 : | if (Data->qpel && MotionFlags & PMV_QUARTERPELREFINE8) { | ||
1520 : | Data->qpel_precision = 1; | ||
1521 : | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, | ||
1522 : | pParam->width, pParam->height, Data->iFcode, 1, 0); | ||
1523 : | SubpelRefine(Data); | ||
1524 : | } | ||
1525 : | } | ||
1526 : | |||
1527 : | if (Data->rrv) { | ||
1528 : | Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x); | ||
1529 : | Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y); | ||
1530 : | Isibaar | 1.1 | } |
1531 : | |||
1532 : | edgomez | 1.55 | if(Data->qpel) { |
1533 : | pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x; | ||
1534 : | pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y; | ||
1535 : | pMB->qmvs[block] = *Data->currentQMV; | ||
1536 : | } else { | ||
1537 : | pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; | ||
1538 : | pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; | ||
1539 : | } | ||
1540 : | |||
1541 : | pMB->mvs[block] = *Data->currentMV; | ||
1542 : | pMB->sad8[block] = 4 * *Data->iMinSAD; | ||
1543 : | } | ||
1544 : | |||
1545 : | /* motion estimation for B-frames */ | ||
1546 : | |||
1547 : | static __inline VECTOR | ||
1548 : | ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) | ||
1549 : | { | ||
1550 : | /* the stupidiest function ever */ | ||
1551 : | return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]); | ||
1552 : | } | ||
1553 : | |||
1554 : | static void __inline | ||
1555 : | PreparePredictionsBF(VECTOR * const pmv, const int x, const int y, | ||
1556 : | const uint32_t iWcount, | ||
1557 : | const MACROBLOCK * const pMB, | ||
1558 : | const uint32_t mode_curr) | ||
1559 : | { | ||
1560 : | |||
1561 : | // [0] is prediction | ||
1562 : | pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); | ||
1563 : | |||
1564 : | pmv[1].x = pmv[1].y = 0; // [1] is zero | ||
1565 : | edgomez | 1.20 | |
1566 : | edgomez | 1.55 | pmv[2] = ChoosePred(pMB, mode_curr); |
1567 : | pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); | ||
1568 : | chl | 1.3 | |
1569 : | edgomez | 1.55 | if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour |
1570 : | pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); | ||
1571 : | pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); | ||
1572 : | } else pmv[3].x = pmv[3].y = 0; | ||
1573 : | chl | 1.3 | |
1574 : | edgomez | 1.55 | if (y != 0) { |
1575 : | pmv[4] = ChoosePred(pMB-iWcount, mode_curr); | ||
1576 : | pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y); | ||
1577 : | } else pmv[4].x = pmv[4].y = 0; | ||
1578 : | chl | 1.35 | |
1579 : | edgomez | 1.55 | if (x != 0) { |
1580 : | pmv[5] = ChoosePred(pMB-1, mode_curr); | ||
1581 : | pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); | ||
1582 : | } else pmv[5].x = pmv[5].y = 0; | ||
1583 : | chl | 1.3 | |
1584 : | edgomez | 1.55 | if (x != 0 && y != 0) { |
1585 : | pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); | ||
1586 : | pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y); | ||
1587 : | } else pmv[6].x = pmv[6].y = 0; | ||
1588 : | } | ||
1589 : | chl | 1.3 | |
1590 : | |||
1591 : | edgomez | 1.55 | /* search backward or forward */ |
1592 : | static void | ||
1593 : | SearchBF( const IMAGE * const pRef, | ||
1594 : | const uint8_t * const pRefH, | ||
1595 : | const uint8_t * const pRefV, | ||
1596 : | const uint8_t * const pRefHV, | ||
1597 : | const IMAGE * const pCur, | ||
1598 : | const int x, const int y, | ||
1599 : | const uint32_t MotionFlags, | ||
1600 : | const uint32_t iFcode, | ||
1601 : | const MBParam * const pParam, | ||
1602 : | MACROBLOCK * const pMB, | ||
1603 : | const VECTOR * const predMV, | ||
1604 : | int32_t * const best_sad, | ||
1605 : | const int32_t mode_current, | ||
1606 : | SearchData * const Data) | ||
1607 : | { | ||
1608 : | |||
1609 : | int i, iDirection = 255, mask; | ||
1610 : | VECTOR pmv[7]; | ||
1611 : | MainSearchFunc *MainSearchPtr; | ||
1612 : | *Data->iMinSAD = MV_MAX_ERROR; | ||
1613 : | Data->iFcode = iFcode; | ||
1614 : | Data->qpel_precision = 0; | ||
1615 : | Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; // reset chroma-sad cache | ||
1616 : | |||
1617 : | syskin | 1.64 | Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16; |
1618 : | Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16; | ||
1619 : | Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16; | ||
1620 : | Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16; | ||
1621 : | Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; | ||
1622 : | Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; | ||
1623 : | edgomez | 1.55 | |
1624 : | Data->predMV = *predMV; | ||
1625 : | |||
1626 : | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, | ||
1627 : | pParam->width, pParam->height, iFcode - Data->qpel, 0, 0); | ||
1628 : | |||
1629 : | pmv[0] = Data->predMV; | ||
1630 : | if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } | ||
1631 : | |||
1632 : | PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); | ||
1633 : | |||
1634 : | Data->currentMV->x = Data->currentMV->y = 0; | ||
1635 : | CheckCandidate = CheckCandidate16no4v; | ||
1636 : | |||
1637 : | // main loop. checking all predictions | ||
1638 : | for (i = 0; i < 7; i++) { | ||
1639 : | if (!(mask = make_mask(pmv, i)) ) continue; | ||
1640 : | CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); | ||
1641 : | } | ||
1642 : | |||
1643 : | if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; | ||
1644 : | else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; | ||
1645 : | else MainSearchPtr = DiamondSearch; | ||
1646 : | |||
1647 : | MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection); | ||
1648 : | |||
1649 : | SubpelRefine(Data); | ||
1650 : | |||
1651 : | if (Data->qpel && *Data->iMinSAD < *best_sad + 300) { | ||
1652 : | Data->currentQMV->x = 2*Data->currentMV->x; | ||
1653 : | Data->currentQMV->y = 2*Data->currentMV->y; | ||
1654 : | Data->qpel_precision = 1; | ||
1655 : | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, | ||
1656 : | pParam->width, pParam->height, iFcode, 1, 0); | ||
1657 : | SubpelRefine(Data); | ||
1658 : | } | ||
1659 : | |||
1660 : | // three bits are needed to code backward mode. four for forward | ||
1661 : | |||
1662 : | if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; | ||
1663 : | else *Data->iMinSAD += 3 * Data->lambda16; | ||
1664 : | |||
1665 : | if (*Data->iMinSAD < *best_sad) { | ||
1666 : | *best_sad = *Data->iMinSAD; | ||
1667 : | pMB->mode = mode_current; | ||
1668 : | if (Data->qpel) { | ||
1669 : | pMB->pmvs[0].x = Data->currentQMV->x - predMV->x; | ||
1670 : | pMB->pmvs[0].y = Data->currentQMV->y - predMV->y; | ||
1671 : | if (mode_current == MODE_FORWARD) | ||
1672 : | pMB->qmvs[0] = *Data->currentQMV; | ||
1673 : | else | ||
1674 : | pMB->b_qmvs[0] = *Data->currentQMV; | ||
1675 : | } else { | ||
1676 : | pMB->pmvs[0].x = Data->currentMV->x - predMV->x; | ||
1677 : | pMB->pmvs[0].y = Data->currentMV->y - predMV->y; | ||
1678 : | } | ||
1679 : | if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; | ||
1680 : | else pMB->b_mvs[0] = *Data->currentMV; | ||
1681 : | } | ||
1682 : | |||
1683 : | if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; | ||
1684 : | else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search | ||
1685 : | } | ||
1686 : | |||
1687 : | static void | ||
1688 : | SkipDecisionB(const IMAGE * const pCur, | ||
1689 : | const IMAGE * const f_Ref, | ||
1690 : | const IMAGE * const b_Ref, | ||
1691 : | MACROBLOCK * const pMB, | ||
1692 : | const uint32_t x, const uint32_t y, | ||
1693 : | const SearchData * const Data) | ||
1694 : | { | ||
1695 : | int dx = 0, dy = 0, b_dx = 0, b_dy = 0; | ||
1696 : | int32_t sum; | ||
1697 : | const int div = 1 + Data->qpel; | ||
1698 : | int k; | ||
1699 : | const uint32_t stride = Data->iEdgedWidth/2; | ||
1700 : | //this is not full chroma compensation, only it's fullpel approximation. should work though | ||
1701 : | |||
1702 : | for (k = 0; k < 4; k++) { | ||
1703 : | dy += Data->directmvF[k].y / div; | ||
1704 : | edgomez | 1.62 | dx += Data->directmvF[k].x / div; |
1705 : | b_dy += Data->directmvB[k].y / div; | ||
1706 : | b_dx += Data->directmvB[k].x / div; | ||
1707 : | edgomez | 1.55 | } |
1708 : | |||
1709 : | dy = (dy >> 3) + roundtab_76[dy & 0xf]; | ||
1710 : | dx = (dx >> 3) + roundtab_76[dx & 0xf]; | ||
1711 : | b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf]; | ||
1712 : | b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf]; | ||
1713 : | |||
1714 : | sum = sad8bi(pCur->u + 8 * x + 8 * y * stride, | ||
1715 : | f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2, | ||
1716 : | b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, | ||
1717 : | stride); | ||
1718 : | |||
1719 : | if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; //no skip | ||
1720 : | |||
1721 : | sum += sad8bi(pCur->v + 8*x + 8 * y * stride, | ||
1722 : | f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2, | ||
1723 : | b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, | ||
1724 : | stride); | ||
1725 : | |||
1726 : | edgomez | 1.62 | if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) { |
1727 : | pMB->mode = MODE_DIRECT_NONE_MV; //skipped | ||
1728 : | syskin | 1.64 | for (k = 0; k < 4; k++) { |
1729 : | edgomez | 1.62 | pMB->qmvs[k] = pMB->mvs[k]; |
1730 : | pMB->b_qmvs[k] = pMB->b_mvs[k]; | ||
1731 : | } | ||
1732 : | } | ||
1733 : | edgomez | 1.55 | } |
1734 : | chl | 1.3 | |
1735 : | edgomez | 1.55 | static __inline uint32_t |
1736 : | SearchDirect(const IMAGE * const f_Ref, | ||
1737 : | const uint8_t * const f_RefH, | ||
1738 : | const uint8_t * const f_RefV, | ||
1739 : | const uint8_t * const f_RefHV, | ||
1740 : | const IMAGE * const b_Ref, | ||
1741 : | const uint8_t * const b_RefH, | ||
1742 : | const uint8_t * const b_RefV, | ||
1743 : | const uint8_t * const b_RefHV, | ||
1744 : | const IMAGE * const pCur, | ||
1745 : | const int x, const int y, | ||
1746 : | const uint32_t MotionFlags, | ||
1747 : | const int32_t TRB, const int32_t TRD, | ||
1748 : | const MBParam * const pParam, | ||
1749 : | MACROBLOCK * const pMB, | ||
1750 : | const MACROBLOCK * const b_mb, | ||
1751 : | int32_t * const best_sad, | ||
1752 : | SearchData * const Data) | ||
1753 : | chl | 1.3 | |
1754 : | edgomez | 1.55 | { |
1755 : | int32_t skip_sad; | ||
1756 : | int k = (x + Data->iEdgedWidth*y) * 16; | ||
1757 : | MainSearchFunc *MainSearchPtr; | ||
1758 : | |||
1759 : | *Data->iMinSAD = 256*4096; | ||
1760 : | syskin | 1.64 | Data->RefP[0] = f_Ref->y + k; |
1761 : | Data->RefP[2] = f_RefH + k; | ||
1762 : | Data->RefP[1] = f_RefV + k; | ||
1763 : | Data->RefP[3] = f_RefHV + k; | ||
1764 : | Data->b_RefP[0] = b_Ref->y + k; | ||
1765 : | Data->b_RefP[2] = b_RefH + k; | ||
1766 : | Data->b_RefP[1] = b_RefV + k; | ||
1767 : | Data->b_RefP[3] = b_RefHV + k; | ||
1768 : | Data->RefP[4] = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; | ||
1769 : | Data->RefP[5] = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; | ||
1770 : | Data->b_RefP[4] = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; | ||
1771 : | Data->b_RefP[5] = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; | ||
1772 : | edgomez | 1.55 | |
1773 : | k = Data->qpel ? 4 : 2; | ||
1774 : | Data->max_dx = k * (pParam->width - x * 16); | ||
1775 : | Data->max_dy = k * (pParam->height - y * 16); | ||
1776 : | Data->min_dx = -k * (16 + x * 16); | ||
1777 : | Data->min_dy = -k * (16 + y * 16); | ||
1778 : | |||
1779 : | Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs; | ||
1780 : | Data->qpel_precision = 0; | ||
1781 : | |||
1782 : | for (k = 0; k < 4; k++) { | ||
1783 : | pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); | ||
1784 : | pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD; | ||
1785 : | pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); | ||
1786 : | pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; | ||
1787 : | |||
1788 : | if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx) | ||
1789 : | | (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) { | ||
1790 : | |||
1791 : | *best_sad = 256*4096; // in that case, we won't use direct mode | ||
1792 : | pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV" | ||
1793 : | pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; | ||
1794 : | return 256*4096; | ||
1795 : | } | ||
1796 : | if (b_mb->mode != MODE_INTER4V) { | ||
1797 : | pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; | ||
1798 : | pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0]; | ||
1799 : | Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0]; | ||
1800 : | Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0]; | ||
1801 : | break; | ||
1802 : | } | ||
1803 : | } | ||
1804 : | |||
1805 : | CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v; | ||
1806 : | |||
1807 : | CheckCandidate(0, 0, 255, &k, Data); | ||
1808 : | |||
1809 : | // initial (fast) skip decision | ||
1810 : | if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (2 + Data->chroma?1:0)) { | ||
1811 : | //possible skip | ||
1812 : | if (Data->chroma) { | ||
1813 : | pMB->mode = MODE_DIRECT_NONE_MV; | ||
1814 : | return *Data->iMinSAD; // skip. | ||
1815 : | } else { | ||
1816 : | SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data); | ||
1817 : | if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip. | ||
1818 : | } | ||
1819 : | } | ||
1820 : | |||
1821 : | syskin | 1.60 | *Data->iMinSAD += Data->lambda16; |
1822 : | edgomez | 1.55 | skip_sad = *Data->iMinSAD; |
1823 : | |||
1824 : | // DIRECT MODE DELTA VECTOR SEARCH. | ||
1825 : | // This has to be made more effective, but at the moment I'm happy it's running at all | ||
1826 : | |||
1827 : | if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; | ||
1828 : | else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; | ||
1829 : | else MainSearchPtr = DiamondSearch; | ||
1830 : | |||
1831 : | MainSearchPtr(0, 0, Data, 255); | ||
1832 : | |||
1833 : | SubpelRefine(Data); | ||
1834 : | |||
1835 : | *best_sad = *Data->iMinSAD; | ||
1836 : | |||
1837 : | if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT; | ||
1838 : | else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation | ||
1839 : | |||
1840 : | pMB->pmvs[3] = *Data->currentMV; | ||
1841 : | |||
1842 : | for (k = 0; k < 4; k++) { | ||
1843 : | pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x; | ||
1844 : | pMB->b_mvs[k].x = ( (Data->currentMV->x == 0) | ||
1845 : | ? Data->directmvB[k].x | ||
1846 : | :pMB->mvs[k].x - Data->referencemv[k].x); | ||
1847 : | pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y); | ||
1848 : | pMB->b_mvs[k].y = ((Data->currentMV->y == 0) | ||
1849 : | ? Data->directmvB[k].y | ||
1850 : | : pMB->mvs[k].y - Data->referencemv[k].y); | ||
1851 : | if (Data->qpel) { | ||
1852 : | pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2; | ||
1853 : | pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2; | ||
1854 : | pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2; | ||
1855 : | pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2; | ||
1856 : | } | ||
1857 : | |||
1858 : | if (b_mb->mode != MODE_INTER4V) { | ||
1859 : | pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; | ||
1860 : | pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; | ||
1861 : | pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0]; | ||
1862 : | pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0]; | ||
1863 : | break; | ||
1864 : | } | ||
1865 : | } | ||
1866 : | return skip_sad; | ||
1867 : | } | ||
1868 : | |||
1869 : | static void | ||
1870 : | SearchInterpolate(const IMAGE * const f_Ref, | ||
1871 : | const uint8_t * const f_RefH, | ||
1872 : | const uint8_t * const f_RefV, | ||
1873 : | const uint8_t * const f_RefHV, | ||
1874 : | const IMAGE * const b_Ref, | ||
1875 : | const uint8_t * const b_RefH, | ||
1876 : | const uint8_t * const b_RefV, | ||
1877 : | const uint8_t * const b_RefHV, | ||
1878 : | const IMAGE * const pCur, | ||
1879 : | const int x, const int y, | ||
1880 : | const uint32_t fcode, | ||
1881 : | const uint32_t bcode, | ||
1882 : | const uint32_t MotionFlags, | ||
1883 : | const MBParam * const pParam, | ||
1884 : | const VECTOR * const f_predMV, | ||
1885 : | const VECTOR * const b_predMV, | ||
1886 : | MACROBLOCK * const pMB, | ||
1887 : | int32_t * const best_sad, | ||
1888 : | SearchData * const fData) | ||
1889 : | chl | 1.3 | |
1890 : | edgomez | 1.55 | { |
1891 : | chl | 1.3 | |
1892 : | edgomez | 1.55 | int iDirection, i, j; |
1893 : | SearchData bData; | ||
1894 : | edgomez | 1.20 | |
1895 : | edgomez | 1.55 | fData->qpel_precision = 0; |
1896 : | memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data | ||
1897 : | *fData->iMinSAD = 4096*256; | ||
1898 : | bData.currentMV++; bData.currentQMV++; | ||
1899 : | fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; | ||
1900 : | |||
1901 : | i = (x + y * fData->iEdgedWidth) * 16; | ||
1902 : | |||
1903 : | syskin | 1.64 | bData.b_RefP[0] = fData->RefP[0] = f_Ref->y + i; |
1904 : | bData.b_RefP[2] = fData->RefP[2] = f_RefH + i; | ||
1905 : | bData.b_RefP[1] = fData->RefP[1] = f_RefV + i; | ||
1906 : | bData.b_RefP[3] = fData->RefP[3] = f_RefHV + i; | ||
1907 : | bData.RefP[0] = fData->b_RefP[0] = b_Ref->y + i; | ||
1908 : | bData.RefP[2] = fData->b_RefP[2] = b_RefH + i; | ||
1909 : | bData.RefP[1] = fData->b_RefP[1] = b_RefV + i; | ||
1910 : | bData.RefP[3] = fData->b_RefP[3] = b_RefHV + i; | ||
1911 : | bData.b_RefP[4] = fData->RefP[4] = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; | ||
1912 : | bData.b_RefP[5] = fData->RefP[5] = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; | ||
1913 : | bData.RefP[4] = fData->b_RefP[4] = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; | ||
1914 : | bData.RefP[5] = fData->b_RefP[5] = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; | ||
1915 : | edgomez | 1.55 | |
1916 : | bData.bpredMV = fData->predMV = *f_predMV; | ||
1917 : | fData->bpredMV = bData.predMV = *b_predMV; | ||
1918 : | fData->currentMV[0] = fData->currentMV[2]; | ||
1919 : | |||
1920 : | get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode - fData->qpel, 0, 0); | ||
1921 : | get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode - fData->qpel, 0, 0); | ||
1922 : | |||
1923 : | if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; | ||
1924 : | if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx; | ||
1925 : | if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy; | ||
1926 : | if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy; | ||
1927 : | |||
1928 : | if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx; | ||
1929 : | if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx; | ||
1930 : | if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy; | ||
1931 : | if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy; | ||
1932 : | |||
1933 : | CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); | ||
1934 : | |||
1935 : | //diamond | ||
1936 : | do { | ||
1937 : | iDirection = 255; | ||
1938 : | // forward MV moves | ||
1939 : | i = fData->currentMV[0].x; j = fData->currentMV[0].y; | ||
1940 : | |||
1941 : | CheckCandidateInt(i + 1, j, 0, &iDirection, fData); | ||
1942 : | CheckCandidateInt(i, j + 1, 0, &iDirection, fData); | ||
1943 : | CheckCandidateInt(i - 1, j, 0, &iDirection, fData); | ||
1944 : | CheckCandidateInt(i, j - 1, 0, &iDirection, fData); | ||
1945 : | |||
1946 : | // backward MV moves | ||
1947 : | i = fData->currentMV[1].x; j = fData->currentMV[1].y; | ||
1948 : | fData->currentMV[2] = fData->currentMV[0]; | ||
1949 : | CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); | ||
1950 : | CheckCandidateInt(i, j + 1, 0, &iDirection, &bData); | ||
1951 : | CheckCandidateInt(i - 1, j, 0, &iDirection, &bData); | ||
1952 : | CheckCandidateInt(i, j - 1, 0, &iDirection, &bData); | ||
1953 : | |||
1954 : | } while (!(iDirection)); | ||
1955 : | |||
1956 : | //qpel refinement | ||
1957 : | if (fData->qpel) { | ||
1958 : | if (*fData->iMinSAD > *best_sad + 500) return; | ||
1959 : | CheckCandidate = CheckCandidateInt; | ||
1960 : | fData->qpel_precision = bData.qpel_precision = 1; | ||
1961 : | get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0); | ||
1962 : | get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0); | ||
1963 : | fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x; | ||
1964 : | fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y; | ||
1965 : | fData->currentQMV[1].x = 2 * fData->currentMV[1].x; | ||
1966 : | fData->currentQMV[1].y = 2 * fData->currentMV[1].y; | ||
1967 : | SubpelRefine(fData); | ||
1968 : | if (*fData->iMinSAD > *best_sad + 300) return; | ||
1969 : | fData->currentQMV[2] = fData->currentQMV[0]; | ||
1970 : | SubpelRefine(&bData); | ||
1971 : | } | ||
1972 : | |||
1973 : | *fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode. | ||
1974 : | |||
1975 : | if (*fData->iMinSAD < *best_sad) { | ||
1976 : | *best_sad = *fData->iMinSAD; | ||
1977 : | pMB->mvs[0] = fData->currentMV[0]; | ||
1978 : | pMB->b_mvs[0] = fData->currentMV[1]; | ||
1979 : | pMB->mode = MODE_INTERPOLATE; | ||
1980 : | if (fData->qpel) { | ||
1981 : | pMB->qmvs[0] = fData->currentQMV[0]; | ||
1982 : | pMB->b_qmvs[0] = fData->currentQMV[1]; | ||
1983 : | pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x; | ||
1984 : | pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y; | ||
1985 : | pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x; | ||
1986 : | pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y; | ||
1987 : | } else { | ||
1988 : | pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; | ||
1989 : | pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; | ||
1990 : | pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; | ||
1991 : | pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; | ||
1992 : | } | ||
1993 : | } | ||
1994 : | } | ||
1995 : | |||
1996 : | void | ||
1997 : | MotionEstimationBVOP(MBParam * const pParam, | ||
1998 : | syskin | 1.64 | FRAMEINFO * const frame, |
1999 : | const int32_t time_bp, | ||
2000 : | const int32_t time_pp, | ||
2001 : | // forward (past) reference | ||
2002 : | const MACROBLOCK * const f_mbs, | ||
2003 : | const IMAGE * const f_ref, | ||
2004 : | const IMAGE * const f_refH, | ||
2005 : | const IMAGE * const f_refV, | ||
2006 : | const IMAGE * const f_refHV, | ||
2007 : | // backward (future) reference | ||
2008 : | const FRAMEINFO * const b_reference, | ||
2009 : | const IMAGE * const b_ref, | ||
2010 : | const IMAGE * const b_refH, | ||
2011 : | const IMAGE * const b_refV, | ||
2012 : | const IMAGE * const b_refHV) | ||
2013 : | edgomez | 1.55 | { |
2014 : | uint32_t i, j; | ||
2015 : | int32_t best_sad; | ||
2016 : | uint32_t skip_sad; | ||
2017 : | int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; | ||
2018 : | const MACROBLOCK * const b_mbs = b_reference->mbs; | ||
2019 : | |||
2020 : | VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ | ||
2021 : | |||
2022 : | const int32_t TRB = time_pp - time_bp; | ||
2023 : | const int32_t TRD = time_pp; | ||
2024 : | |||
2025 : | // some pre-inintialized data for the rest of the search | ||
2026 : | |||
2027 : | SearchData Data; | ||
2028 : | int32_t iMinSAD; | ||
2029 : | VECTOR currentMV[3]; | ||
2030 : | VECTOR currentQMV[3]; | ||
2031 : | int32_t temp[8]; | ||
2032 : | memset(&Data, 0, sizeof(SearchData)); | ||
2033 : | Data.iEdgedWidth = pParam->edged_width; | ||
2034 : | Data.currentMV = currentMV; Data.currentQMV = currentQMV; | ||
2035 : | Data.iMinSAD = &iMinSAD; | ||
2036 : | Data.lambda16 = lambda_vec16[frame->quant]; | ||
2037 : | Data.qpel = pParam->m_quarterpel; | ||
2038 : | Data.rounding = 0; | ||
2039 : | Data.chroma = frame->motion_flags & PMV_CHROMA8; | ||
2040 : | Data.temp = temp; | ||
2041 : | |||
2042 : | Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose) | ||
2043 : | // note: i==horizontal, j==vertical | ||
2044 : | for (j = 0; j < pParam->mb_height; j++) { | ||
2045 : | |||
2046 : | f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ | ||
2047 : | |||
2048 : | for (i = 0; i < pParam->mb_width; i++) { | ||
2049 : | MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; | ||
2050 : | const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; | ||
2051 : | |||
2052 : | /* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */ | ||
2053 : | if (b_reference->coding_type != S_VOP) | ||
2054 : | if (b_mb->mode == MODE_NOT_CODED) { | ||
2055 : | pMB->mode = MODE_NOT_CODED; | ||
2056 : | continue; | ||
2057 : | } | ||
2058 : | edgomez | 1.20 | |
2059 : | edgomez | 1.55 | Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
2060 : | Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8; | ||
2061 : | Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8; | ||
2062 : | pMB->quant = frame->quant; | ||
2063 : | |||
2064 : | /* direct search comes first, because it (1) checks for SKIP-mode | ||
2065 : | and (2) sets very good predictions for forward and backward search */ | ||
2066 : | skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y, | ||
2067 : | b_ref, b_refH->y, b_refV->y, b_refHV->y, | ||
2068 : | &frame->image, | ||
2069 : | i, j, | ||
2070 : | frame->motion_flags, | ||
2071 : | TRB, TRD, | ||
2072 : | pParam, | ||
2073 : | pMB, b_mb, | ||
2074 : | &best_sad, | ||
2075 : | &Data); | ||
2076 : | |||
2077 : | if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } | ||
2078 : | |||
2079 : | // forward search | ||
2080 : | SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y, | ||
2081 : | &frame->image, i, j, | ||
2082 : | frame->motion_flags, | ||
2083 : | frame->fcode, pParam, | ||
2084 : | pMB, &f_predMV, &best_sad, | ||
2085 : | MODE_FORWARD, &Data); | ||
2086 : | |||
2087 : | // backward search | ||
2088 : | SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y, | ||
2089 : | &frame->image, i, j, | ||
2090 : | frame->motion_flags, | ||
2091 : | frame->bcode, pParam, | ||
2092 : | pMB, &b_predMV, &best_sad, | ||
2093 : | MODE_BACKWARD, &Data); | ||
2094 : | |||
2095 : | // interpolate search comes last, because it uses data from forward and backward as prediction | ||
2096 : | SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y, | ||
2097 : | b_ref, b_refH->y, b_refV->y, b_refHV->y, | ||
2098 : | &frame->image, | ||
2099 : | i, j, | ||
2100 : | frame->fcode, frame->bcode, | ||
2101 : | frame->motion_flags, | ||
2102 : | pParam, | ||
2103 : | &f_predMV, &b_predMV, | ||
2104 : | pMB, &best_sad, | ||
2105 : | &Data); | ||
2106 : | |||
2107 : | // final skip decision | ||
2108 : | if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2) | ||
2109 : | && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) | ||
2110 : | SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data); | ||
2111 : | |||
2112 : | switch (pMB->mode) { | ||
2113 : | case MODE_FORWARD: | ||
2114 : | f_count++; | ||
2115 : | f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; | ||
2116 : | break; | ||
2117 : | case MODE_BACKWARD: | ||
2118 : | b_count++; | ||
2119 : | b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; | ||
2120 : | break; | ||
2121 : | case MODE_INTERPOLATE: | ||
2122 : | i_count++; | ||
2123 : | f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; | ||
2124 : | b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; | ||
2125 : | break; | ||
2126 : | case MODE_DIRECT: | ||
2127 : | case MODE_DIRECT_NO4V: | ||
2128 : | d_count++; | ||
2129 : | default: | ||
2130 : | break; | ||
2131 : | chl | 1.3 | } |
2132 : | } | ||
2133 : | } | ||
2134 : | edgomez | 1.55 | } |
2135 : | |||
2136 : | static __inline void | ||
2137 : | MEanalyzeMB ( const uint8_t * const pRef, | ||
2138 : | const uint8_t * const pCur, | ||
2139 : | const int x, | ||
2140 : | const int y, | ||
2141 : | const MBParam * const pParam, | ||
2142 : | MACROBLOCK * const pMBs, | ||
2143 : | SearchData * const Data) | ||
2144 : | { | ||
2145 : | chl | 1.3 | |
2146 : | edgomez | 1.55 | int i, mask; |
2147 : | VECTOR pmv[3]; | ||
2148 : | syskin | 1.60 | MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width]; |
2149 : | chl | 1.3 | |
2150 : | edgomez | 1.55 | for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
2151 : | chl | 1.3 | |
2152 : | edgomez | 1.55 | //median is only used as prediction. it doesn't have to be real |
2153 : | if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; | ||
2154 : | else | ||
2155 : | if (x == 1) //left macroblock does not have any vector now | ||
2156 : | Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median | ||
2157 : | else if (y == 1) // top macroblock doesn't have it's vector | ||
2158 : | Data->predMV = (pMB - 1)->mvs[0]; // left instead of median | ||
2159 : | else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median | ||
2160 : | chl | 1.3 | |
2161 : | edgomez | 1.55 | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2162 : | syskin | 1.60 | pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, 0); |
2163 : | chl | 1.3 | |
2164 : | edgomez | 1.55 | Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
2165 : | syskin | 1.64 | Data->RefP[0] = pRef + (x + y * pParam->edged_width) * 16; |
2166 : | chl | 1.3 | |
2167 : | edgomez | 1.55 | pmv[1].x = EVEN(pMB->mvs[0].x); |
2168 : | pmv[1].y = EVEN(pMB->mvs[0].y); | ||
2169 : | pmv[2].x = EVEN(Data->predMV.x); | ||
2170 : | pmv[2].y = EVEN(Data->predMV.y); | ||
2171 : | pmv[0].x = pmv[0].y = 0; | ||
2172 : | chl | 1.3 | |
2173 : | edgomez | 1.55 | CheckCandidate32I(0, 0, 255, &i, Data); |
2174 : | chl | 1.3 | |
2175 : | syskin | 1.60 | if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) { |
2176 : | chl | 1.3 | |
2177 : | edgomez | 1.55 | if (!(mask = make_mask(pmv, 1))) |
2178 : | CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data); | ||
2179 : | if (!(mask = make_mask(pmv, 2))) | ||
2180 : | CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data); | ||
2181 : | chl | 1.3 | |
2182 : | syskin | 1.60 | if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) // diamond only if needed |
2183 : | edgomez | 1.55 | DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
2184 : | syskin | 1.60 | } |
2185 : | chl | 1.3 | |
2186 : | syskin | 1.60 | for (i = 0; i < 4; i++) { |
2187 : | MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width]; | ||
2188 : | MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i]; | ||
2189 : | MB->mode = MODE_INTER; | ||
2190 : | MB->sad16 = Data->iMinSAD[i+1]; | ||
2191 : | edgomez | 1.20 | } |
2192 : | edgomez | 1.55 | } |
2193 : | chl | 1.3 | |
2194 : | syskin | 1.60 | #define INTRA_THRESH 2400 |
2195 : | syskin | 1.64 | #define INTER_THRESH 1300 |
2196 : | chl | 1.3 | |
2197 : | edgomez | 1.55 | int |
2198 : | MEanalysis( const IMAGE * const pRef, | ||
2199 : | syskin | 1.60 | const FRAMEINFO * const Current, |
2200 : | const MBParam * const pParam, | ||
2201 : | const int maxIntra, //maximum number if non-I frames | ||
2202 : | const int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame | ||
2203 : | const int bCount, // number of B frames in a row | ||
2204 : | const int b_thresh) | ||
2205 : | edgomez | 1.55 | { |
2206 : | uint32_t x, y, intra = 0; | ||
2207 : | int sSAD = 0; | ||
2208 : | MACROBLOCK * const pMBs = Current->mbs; | ||
2209 : | const IMAGE * const pCurrent = &Current->image; | ||
2210 : | syskin | 1.60 | int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + 10*b_thresh; |
2211 : | int s = 0, blocks = 0; | ||
2212 : | edgomez | 1.55 | |
2213 : | int32_t iMinSAD[5], temp[5]; | ||
2214 : | VECTOR currentMV[5]; | ||
2215 : | SearchData Data; | ||
2216 : | Data.iEdgedWidth = pParam->edged_width; | ||
2217 : | Data.currentMV = currentMV; | ||
2218 : | Data.iMinSAD = iMinSAD; | ||
2219 : | Data.iFcode = Current->fcode; | ||
2220 : | Data.temp = temp; | ||
2221 : | CheckCandidate = CheckCandidate32I; | ||
2222 : | chl | 1.3 | |
2223 : | edgomez | 1.55 | if (intraCount != 0 && intraCount < 10) // we're right after an I frame |
2224 : | syskin | 1.60 | IntraThresh += 8 * (intraCount - 10) * (intraCount - 10); |
2225 : | chl | 1.19 | else |
2226 : | edgomez | 1.55 | if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec |
2227 : | IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; | ||
2228 : | chl | 1.19 | |
2229 : | syskin | 1.60 | InterThresh -= (350 - 8*b_thresh) * bCount; |
2230 : | if (InterThresh < 300 + 5*b_thresh) InterThresh = 300 + 5*b_thresh; | ||
2231 : | chl | 1.3 | |
2232 : | edgomez | 1.55 | if (sadInit) (*sadInit) (); |
2233 : | edgomez | 1.20 | |
2234 : | edgomez | 1.55 | for (y = 1; y < pParam->mb_height-1; y += 2) { |
2235 : | for (x = 1; x < pParam->mb_width-1; x += 2) { | ||
2236 : | int i; | ||
2237 : | syskin | 1.60 | blocks += 4; |
2238 : | edgomez | 1.55 | |
2239 : | if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; | ||
2240 : | syskin | 1.60 | else { //extrapolation of the vector found for last frame |
2241 : | syskin | 1.64 | pMBs[x + y * pParam->mb_width].mvs[0].x = |
2242 : | syskin | 1.60 | (pMBs[x + y * pParam->mb_width].mvs[0].x * (bCount+1) ) / bCount; |
2243 : | syskin | 1.64 | pMBs[x + y * pParam->mb_width].mvs[0].y = |
2244 : | syskin | 1.60 | (pMBs[x + y * pParam->mb_width].mvs[0].y * (bCount+1) ) / bCount; |
2245 : | } | ||
2246 : | edgomez | 1.55 | |
2247 : | MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data); | ||
2248 : | |||
2249 : | for (i = 0; i < 4; i++) { | ||
2250 : | int dev; | ||
2251 : | MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width]; | ||
2252 : | if (pMB->sad16 > IntraThresh) { | ||
2253 : | dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16, | ||
2254 : | pParam->edged_width); | ||
2255 : | if (dev + IntraThresh < pMB->sad16) { | ||
2256 : | pMB->mode = MODE_INTRA; | ||
2257 : | syskin | 1.60 | if (++intra > ((pParam->mb_height-2)*(pParam->mb_width-2))/2) return I_VOP; |
2258 : | edgomez | 1.55 | } |
2259 : | } | ||
2260 : | syskin | 1.60 | if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0) s++; |
2261 : | |||
2262 : | edgomez | 1.55 | sSAD += pMB->sad16; |
2263 : | } | ||
2264 : | } | ||
2265 : | chl | 1.3 | } |
2266 : | syskin | 1.60 | |
2267 : | sSAD /= blocks; | ||
2268 : | s = (10*s) / blocks; | ||
2269 : | |||
2270 : | syskin | 1.64 | if (s > 4) sSAD += (s - 3) * (300 - 2*b_thresh); //static block - looks bad when in bframe... |
2271 : | syskin | 1.60 | |
2272 : | edgomez | 1.55 | if (sSAD > InterThresh ) return P_VOP; |
2273 : | emms(); | ||
2274 : | return B_VOP; | ||
2275 : | } | ||
2276 : | |||
2277 : | chl | 1.3 | |
2278 : | edgomez | 1.55 | static WARPPOINTS |
2279 : | GlobalMotionEst(const MACROBLOCK * const pMBs, | ||
2280 : | const MBParam * const pParam, | ||
2281 : | const FRAMEINFO * const current, | ||
2282 : | const FRAMEINFO * const reference, | ||
2283 : | const IMAGE * const pRefH, | ||
2284 : | const IMAGE * const pRefV, | ||
2285 : | const IMAGE * const pRefHV ) | ||
2286 : | { | ||
2287 : | chl | 1.3 | |
2288 : | edgomez | 1.55 | const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs |
2289 : | const int deltay=8; | ||
2290 : | const int grad=512; // lower bound for deviation in MB | ||
2291 : | chl | 1.3 | |
2292 : | edgomez | 1.55 | WARPPOINTS gmc; |
2293 : | edgomez | 1.20 | |
2294 : | edgomez | 1.55 | uint32_t mx, my; |
2295 : | edgomez | 1.20 | |
2296 : | edgomez | 1.55 | int MBh = pParam->mb_height; |
2297 : | int MBw = pParam->mb_width; | ||
2298 : | edgomez | 1.20 | |
2299 : | edgomez | 1.55 | int *MBmask= calloc(MBh*MBw,sizeof(int)); |
2300 : | double DtimesF[4] = { 0.,0., 0., 0. }; | ||
2301 : | double sol[4] = { 0., 0., 0., 0. }; | ||
2302 : | double a,b,c,n,denom; | ||
2303 : | double meanx,meany; | ||
2304 : | int num,oldnum; | ||
2305 : | chl | 1.3 | |
2306 : | syskin | 1.64 | if (!MBmask) { fprintf(stderr,"Mem error\n"); |
2307 : | gmc.duv[0].x= gmc.duv[0].y = | ||
2308 : | gmc.duv[1].x= gmc.duv[1].y = | ||
2309 : | chl | 1.56 | gmc.duv[2].x= gmc.duv[2].y = 0; |
2310 : | return gmc; } | ||
2311 : | chl | 1.3 | |
2312 : | edgomez | 1.55 | // filter mask of all blocks |
2313 : | chl | 1.3 | |
2314 : | edgomez | 1.58 | for (my = 1; my < (uint32_t)MBh-1; my++) |
2315 : | for (mx = 1; mx < (uint32_t)MBw-1; mx++) | ||
2316 : | edgomez | 1.55 | { |
2317 : | const int mbnum = mx + my * MBw; | ||
2318 : | const MACROBLOCK *pMB = &pMBs[mbnum]; | ||
2319 : | const VECTOR mv = pMB->mvs[0]; | ||
2320 : | chl | 1.3 | |
2321 : | edgomez | 1.55 | if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) |
2322 : | continue; | ||
2323 : | edgomez | 1.20 | |
2324 : | edgomez | 1.55 | if ( ( (ABS(mv.x - (pMB-1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-1)->mvs[0].y) < deltay) ) |
2325 : | && ( (ABS(mv.x - (pMB+1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+1)->mvs[0].y) < deltay) ) | ||
2326 : | && ( (ABS(mv.x - (pMB-MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-MBw)->mvs[0].y) < deltay) ) | ||
2327 : | && ( (ABS(mv.x - (pMB+MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) ) | ||
2328 : | MBmask[mbnum]=1; | ||
2329 : | } | ||
2330 : | chl | 1.3 | |
2331 : | edgomez | 1.58 | for (my = 1; my < (uint32_t)MBh-1; my++) |
2332 : | for (mx = 1; mx < (uint32_t)MBw-1; mx++) | ||
2333 : | edgomez | 1.55 | { |
2334 : | const uint8_t *const pCur = current->image.y + 16*my*pParam->edged_width + 16*mx; | ||
2335 : | chl | 1.3 | |
2336 : | edgomez | 1.55 | const int mbnum = mx + my * MBw; |
2337 : | if (!MBmask[mbnum]) | ||
2338 : | continue; | ||
2339 : | chl | 1.10 | |
2340 : | edgomez | 1.58 | if (sad16 ( pCur, pCur+1 , pParam->edged_width, 65536) <= (uint32_t)grad ) |
2341 : | edgomez | 1.55 | MBmask[mbnum] = 0; |
2342 : | edgomez | 1.58 | if (sad16 ( pCur, pCur+pParam->edged_width, pParam->edged_width, 65536) <= (uint32_t)grad ) |
2343 : | edgomez | 1.55 | MBmask[mbnum] = 0; |
2344 : | edgomez | 1.20 | |
2345 : | edgomez | 1.55 | } |
2346 : | chl | 1.3 | |
2347 : | edgomez | 1.55 | emms(); |
2348 : | chl | 1.3 | |
2349 : | edgomez | 1.55 | do { /* until convergence */ |
2350 : | edgomez | 1.20 | |
2351 : | edgomez | 1.55 | a = b = c = n = 0; |
2352 : | DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.; | ||
2353 : | edgomez | 1.58 | for (my = 0; my < (uint32_t)MBh; my++) |
2354 : | for (mx = 0; mx < (uint32_t)MBw; mx++) | ||
2355 : | edgomez | 1.55 | { |
2356 : | const int mbnum = mx + my * MBw; | ||
2357 : | const MACROBLOCK *pMB = &pMBs[mbnum]; | ||
2358 : | const VECTOR mv = pMB->mvs[0]; | ||
2359 : | |||
2360 : | if (!MBmask[mbnum]) | ||
2361 : | continue; | ||
2362 : | |||
2363 : | n++; | ||
2364 : | a += 16*mx+8; | ||
2365 : | b += 16*my+8; | ||
2366 : | c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8); | ||
2367 : | |||
2368 : | DtimesF[0] += (double)mv.x; | ||
2369 : | DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8); | ||
2370 : | DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8); | ||
2371 : | DtimesF[3] += (double)mv.y; | ||
2372 : | } | ||
2373 : | |||
2374 : | denom = a*a+b*b-c*n; | ||
2375 : | |||
2376 : | syskin | 1.64 | /* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */ |
2377 : | edgomez | 1.55 | /* D'*E*F has been calculated in the same loop as matrix */ |
2378 : | |||
2379 : | sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2]; | ||
2380 : | syskin | 1.64 | sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3]; |
2381 : | sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3]; | ||
2382 : | sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3]; | ||
2383 : | edgomez | 1.55 | |
2384 : | sol[0] /= denom; | ||
2385 : | sol[1] /= denom; | ||
2386 : | sol[2] /= denom; | ||
2387 : | sol[3] /= denom; | ||
2388 : | |||
2389 : | meanx = meany = 0.; | ||
2390 : | oldnum = 0; | ||
2391 : | edgomez | 1.58 | for (my = 0; my < (uint32_t)MBh; my++) |
2392 : | for (mx = 0; mx < (uint32_t)MBw; mx++) | ||
2393 : | edgomez | 1.55 | { |
2394 : | const int mbnum = mx + my * MBw; | ||
2395 : | const MACROBLOCK *pMB = &pMBs[mbnum]; | ||
2396 : | const VECTOR mv = pMB->mvs[0]; | ||
2397 : | edgomez | 1.20 | |
2398 : | edgomez | 1.55 | if (!MBmask[mbnum]) |
2399 : | continue; | ||
2400 : | chl | 1.3 | |
2401 : | edgomez | 1.55 | oldnum++; |
2402 : | meanx += ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ); | ||
2403 : | meany += ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ); | ||
2404 : | } | ||
2405 : | chl | 1.3 | |
2406 : | edgomez | 1.55 | if (4*meanx > oldnum) /* better fit than 0.25 is useless */ |
2407 : | meanx /= oldnum; | ||
2408 : | else | ||
2409 : | meanx = 0.25; | ||
2410 : | edgomez | 1.20 | |
2411 : | edgomez | 1.55 | if (4*meany > oldnum) |
2412 : | meany /= oldnum; | ||
2413 : | else | ||
2414 : | meany = 0.25; | ||
2415 : | chl | 1.3 | |
2416 : | edgomez | 1.55 | /* fprintf(stderr,"sol = (%8.5f, %8.5f, %8.5f, %8.5f)\n",sol[0],sol[1],sol[2],sol[3]); |
2417 : | fprintf(stderr,"meanx = %8.5f meany = %8.5f %d\n",meanx,meany, oldnum); | ||
2418 : | */ | ||
2419 : | num = 0; | ||
2420 : | edgomez | 1.58 | for (my = 0; my < (uint32_t)MBh; my++) |
2421 : | for (mx = 0; mx < (uint32_t)MBw; mx++) | ||
2422 : | edgomez | 1.55 | { |
2423 : | const int mbnum = mx + my * MBw; | ||
2424 : | const MACROBLOCK *pMB = &pMBs[mbnum]; | ||
2425 : | const VECTOR mv = pMB->mvs[0]; | ||
2426 : | chl | 1.3 | |
2427 : | edgomez | 1.55 | if (!MBmask[mbnum]) |
2428 : | continue; | ||
2429 : | chl | 1.3 | |
2430 : | edgomez | 1.55 | if ( ( ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ) > meanx ) |
2431 : | syskin | 1.64 | || ( ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ) > meany ) ) |
2432 : | edgomez | 1.55 | MBmask[mbnum]=0; |
2433 : | else | ||
2434 : | num++; | ||
2435 : | } | ||
2436 : | chl | 1.3 | |
2437 : | edgomez | 1.55 | } while ( (oldnum != num) && (num>=4) ); |
2438 : | chl | 1.3 | |
2439 : | edgomez | 1.55 | if (num < 4) |
2440 : | { | ||
2441 : | 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; | ||
2442 : | } else { | ||
2443 : | chl | 1.3 | |
2444 : | edgomez | 1.55 | gmc.duv[0].x=(int)(sol[0]+0.5); |
2445 : | gmc.duv[0].y=(int)(sol[3]+0.5); | ||
2446 : | chl | 1.3 | |
2447 : | edgomez | 1.55 | gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5); |
2448 : | gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5); | ||
2449 : | edgomez | 1.20 | |
2450 : | edgomez | 1.55 | gmc.duv[2].x=0; |
2451 : | gmc.duv[2].y=0; | ||
2452 : | chl | 1.3 | } |
2453 : | edgomez | 1.55 | // fprintf(stderr,"wp1 = ( %4d, %4d) wp2 = ( %4d, %4d) \n", gmc.duv[0].x, gmc.duv[0].y, gmc.duv[1].x, gmc.duv[1].y); |
2454 : | chl | 1.3 | |
2455 : | edgomez | 1.55 | free(MBmask); |
2456 : | edgomez | 1.20 | |
2457 : | edgomez | 1.55 | return gmc; |
2458 : | } | ||
2459 : | edgomez | 1.20 | |
2460 : | edgomez | 1.55 | // functions which perform BITS-based search/bitcount |
2461 : | edgomez | 1.20 | |
2462 : | edgomez | 1.55 | static int |
2463 : | CountMBBitsInter(SearchData * const Data, | ||
2464 : | const MACROBLOCK * const pMBs, const int x, const int y, | ||
2465 : | const MBParam * const pParam, | ||
2466 : | const uint32_t MotionFlags) | ||
2467 : | { | ||
2468 : | int i, iDirection; | ||
2469 : | int32_t bsad[5]; | ||
2470 : | edgomez | 1.20 | |
2471 : | edgomez | 1.55 | CheckCandidate = CheckCandidateBits16; |
2472 : | edgomez | 1.20 | |
2473 : | edgomez | 1.55 | if (Data->qpel) { |
2474 : | for(i = 0; i < 5; i++) { | ||
2475 : | Data->currentMV[i].x = Data->currentQMV[i].x/2; | ||
2476 : | Data->currentMV[i].y = Data->currentQMV[i].y/2; | ||
2477 : | } | ||
2478 : | Data->qpel_precision = 1; | ||
2479 : | CheckCandidateBits16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data); | ||
2480 : | chl | 1.3 | |
2481 : | edgomez | 1.55 | //checking if this vector is perfect. if it is, we stop. |
2482 : | if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0) | ||
2483 : | return 0; //quick stop | ||
2484 : | chl | 1.3 | |
2485 : | edgomez | 1.55 | if (MotionFlags & (HALFPELREFINE16_BITS | EXTSEARCH_BITS)) { //we have to prepare for halfpixel-precision search |
2486 : | for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i]; | ||
2487 : | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, | ||
2488 : | pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); | ||
2489 : | Data->qpel_precision = 0; | ||
2490 : | if (Data->currentQMV->x & 1 || Data->currentQMV->y & 1) | ||
2491 : | CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); | ||
2492 : | } | ||
2493 : | chl | 1.10 | |
2494 : | edgomez | 1.55 | } else { // not qpel |
2495 : | chl | 1.10 | |
2496 : | edgomez | 1.55 | CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); |
2497 : | //checking if this vector is perfect. if it is, we stop. | ||
2498 : | if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0) { | ||
2499 : | return 0; //inter | ||
2500 : | chl | 1.10 | } |
2501 : | } | ||
2502 : | edgomez | 1.20 | |
2503 : | edgomez | 1.55 | if (MotionFlags&EXTSEARCH_BITS) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
2504 : | |||
2505 : | if (MotionFlags&HALFPELREFINE16_BITS) SubpelRefine(Data); | ||
2506 : | |||
2507 : | if (Data->qpel) { | ||
2508 : | if (MotionFlags&(EXTSEARCH_BITS | HALFPELREFINE16_BITS)) { // there was halfpel-precision search | ||
2509 : | for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) { | ||
2510 : | Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // we have found a better match | ||
2511 : | Data->currentQMV[i].y = 2 * Data->currentMV[i].y; | ||
2512 : | chl | 1.10 | } |
2513 : | edgomez | 1.55 | |
2514 : | // preparing for qpel-precision search | ||
2515 : | Data->qpel_precision = 1; | ||
2516 : | get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, | ||
2517 : | pParam->width, pParam->height, Data->iFcode, 1, 0); | ||
2518 : | chl | 1.10 | } |
2519 : | edgomez | 1.55 | if (MotionFlags&QUARTERPELREFINE16_BITS) SubpelRefine(Data); |
2520 : | chl | 1.10 | } |
2521 : | |||
2522 : | edgomez | 1.55 | if (MotionFlags&CHECKPREDICTION_BITS) { //let's check vector equal to prediction |
2523 : | VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV; | ||
2524 : | if (!(Data->predMV.x == v->x && Data->predMV.y == v->y)) | ||
2525 : | CheckCandidateBits16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data); | ||
2526 : | } | ||
2527 : | return Data->iMinSAD[0]; | ||
2528 : | } | ||
2529 : | chl | 1.10 | |
2530 : | chl | 1.3 | |
2531 : | edgomez | 1.55 | static int |
2532 : | CountMBBitsInter4v(const SearchData * const Data, | ||
2533 : | MACROBLOCK * const pMB, const MACROBLOCK * const pMBs, | ||
2534 : | const int x, const int y, | ||
2535 : | const MBParam * const pParam, const uint32_t MotionFlags, | ||
2536 : | const VECTOR * const backup) | ||
2537 : | { | ||
2538 : | chl | 1.3 | |
2539 : | edgomez | 1.55 | int cbp = 0, bits = 0, t = 0, i, iDirection; |
2540 : | SearchData Data2, *Data8 = &Data2; | ||
2541 : | int sumx = 0, sumy = 0; | ||
2542 : | edgomez | 1.61 | int16_t *in = Data->dctSpace, *coeff = Data->dctSpace + 64; |
2543 : | edgomez | 1.55 | |
2544 : | memcpy(Data8, Data, sizeof(SearchData)); | ||
2545 : | CheckCandidate = CheckCandidateBits8; | ||
2546 : | |||
2547 : | for (i = 0; i < 4; i++) { | ||
2548 : | Data8->iMinSAD = Data->iMinSAD + i + 1; | ||
2549 : | Data8->currentMV = Data->currentMV + i + 1; | ||
2550 : | Data8->currentQMV = Data->currentQMV + i + 1; | ||
2551 : | Data8->Cur = Data->Cur + 8*((i&1) + (i>>1)*Data->iEdgedWidth); | ||
2552 : | syskin | 1.64 | Data8->RefP[0] = Data->RefP[0] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2553 : | Data8->RefP[2] = Data->RefP[2] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); | ||
2554 : | Data8->RefP[1] = Data->RefP[1] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); | ||
2555 : | Data8->RefP[3] = Data->RefP[3] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); | ||
2556 : | edgomez | 1.55 | |
2557 : | if(Data->qpel) { | ||
2558 : | Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i); | ||
2559 : | if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y, | ||
2560 : | Data8->predMV, Data8->iFcode, 0, 0); | ||
2561 : | } else { | ||
2562 : | Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i); | ||
2563 : | if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y, | ||
2564 : | Data8->predMV, Data8->iFcode, 0, 0); | ||
2565 : | } | ||
2566 : | |||
2567 : | get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, | ||
2568 : | pParam->width, pParam->height, Data8->iFcode, Data8->qpel, 0); | ||
2569 : | |||
2570 : | *Data8->iMinSAD += t; | ||
2571 : | |||
2572 : | Data8->qpel_precision = Data8->qpel; | ||
2573 : | // checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far) | ||
2574 : | if (Data8->qpel) { | ||
2575 : | if (!(Data8->currentQMV->x == backup[i+1].x && Data8->currentQMV->y == backup[i+1].y)) | ||
2576 : | CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8); | ||
2577 : | } else { | ||
2578 : | if (!(Data8->currentMV->x == backup[i+1].x && Data8->currentMV->y == backup[i+1].y)) | ||
2579 : | CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8); | ||
2580 : | } | ||
2581 : | |||
2582 : | if (Data8->qpel) { | ||
2583 : | if (MotionFlags&HALFPELREFINE8_BITS || (MotionFlags&PMV_EXTSEARCH8 && MotionFlags&EXTSEARCH_BITS)) { // halfpixel motion search follows | ||
2584 : | int32_t s = *Data8->iMinSAD; | ||
2585 : | Data8->currentMV->x = Data8->currentQMV->x/2; | ||
2586 : | Data8->currentMV->y = Data8->currentQMV->y/2; | ||
2587 : | Data8->qpel_precision = 0; | ||
2588 : | get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, | ||
2589 : | pParam->width, pParam->height, Data8->iFcode - 1, 0, 0); | ||
2590 : | |||
2591 : | if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1) | ||
2592 : | CheckCandidateBits8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8); | ||
2593 : | |||
2594 : | if (MotionFlags & PMV_EXTSEARCH8 && MotionFlags & EXTSEARCH_BITS) | ||
2595 : | SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255); | ||
2596 : | |||
2597 : | if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8); | ||
2598 : | |||
2599 : | if(s > *Data8->iMinSAD) { //we have found a better match | ||
2600 : | Data8->currentQMV->x = 2*Data8->currentMV->x; | ||
2601 : | Data8->currentQMV->y = 2*Data8->currentMV->y; | ||
2602 : | } | ||
2603 : | chl | 1.3 | |
2604 : | edgomez | 1.55 | Data8->qpel_precision = 1; |
2605 : | get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, | ||
2606 : | pParam->width, pParam->height, Data8->iFcode, 1, 0); | ||
2607 : | chl | 1.3 | |
2608 : | edgomez | 1.55 | } |
2609 : | if (MotionFlags & QUARTERPELREFINE8_BITS) SubpelRefine(Data8); | ||
2610 : | chl | 1.3 | |
2611 : | edgomez | 1.55 | } else // not qpel |
2612 : | if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8); //halfpel mode, halfpel refinement | ||
2613 : | chl | 1.3 | |
2614 : | edgomez | 1.55 | //checking vector equal to predicion |
2615 : | if (i != 0 && MotionFlags & CHECKPREDICTION_BITS) { | ||
2616 : | const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV; | ||
2617 : | if (!(Data8->predMV.x == v->x && Data8->predMV.y == v->y)) | ||
2618 : | CheckCandidateBits8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8); | ||
2619 : | } | ||
2620 : | |||
2621 : | bits += *Data8->iMinSAD; | ||
2622 : | if (bits >= Data->iMinSAD[0]) break; // no chances for INTER4V | ||
2623 : | |||
2624 : | // MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else | ||
2625 : | if(Data->qpel) { | ||
2626 : | pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x; | ||
2627 : | pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y; | ||
2628 : | pMB->qmvs[i] = *Data8->currentQMV; | ||
2629 : | sumx += Data8->currentQMV->x/2; | ||
2630 : | sumy += Data8->currentQMV->y/2; | ||
2631 : | } else { | ||
2632 : | pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x; | ||
2633 : | pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y; | ||
2634 : | sumx += Data8->currentMV->x; | ||
2635 : | sumy += Data8->currentMV->y; | ||
2636 : | } | ||
2637 : | pMB->mvs[i] = *Data8->currentMV; | ||
2638 : | pMB->sad8[i] = 4 * *Data8->iMinSAD; | ||
2639 : | if (Data8->temp[0]) cbp |= 1 << (5 - i); | ||
2640 : | } | ||
2641 : | |||
2642 : | if (bits < *Data->iMinSAD) { // there is still a chance for inter4v mode. let's check chroma | ||
2643 : | const uint8_t * ptr; | ||
2644 : | sumx = (sumx >> 3) + roundtab_76[sumx & 0xf]; | ||
2645 : | sumy = (sumy >> 3) + roundtab_76[sumy & 0xf]; | ||
2646 : | |||
2647 : | //chroma U | ||
2648 : | syskin | 1.64 | ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[4], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
2649 : | edgomez | 1.55 | transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2); |
2650 : | fdct(in); | ||
2651 : | if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16); | ||
2652 : | else i = quant4_inter(coeff, in, Data->lambda16); | ||
2653 : | if (i > 0) { | ||
2654 : | bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); | ||
2655 : | cbp |= 1 << (5 - 4); | ||
2656 : | } | ||
2657 : | |||
2658 : | if (bits < *Data->iMinSAD) { // still possible | ||
2659 : | //chroma V | ||
2660 : | syskin | 1.64 | ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[5], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
2661 : | edgomez | 1.55 | transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2); |
2662 : | fdct(in); | ||
2663 : | if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16); | ||
2664 : | else i = quant4_inter(coeff, in, Data->lambda16); | ||
2665 : | if (i > 0) { | ||
2666 : | bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); | ||
2667 : | cbp |= 1 << (5 - 5); | ||
2668 : | } | ||
2669 : | chl | 1.59 | bits += xvid_cbpy_tab[15-(cbp>>2)].len; |
2670 : | edgomez | 1.55 | bits += mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len; |
2671 : | } | ||
2672 : | } | ||
2673 : | edgomez | 1.20 | |
2674 : | edgomez | 1.55 | return bits; |
2675 : | } | ||
2676 : | chl | 1.10 | |
2677 : | chl | 1.3 | |
2678 : | edgomez | 1.55 | static int |
2679 : | CountMBBitsIntra(const SearchData * const Data) | ||
2680 : | { | ||
2681 : | int bits = 1; //this one is ac/dc prediction flag. always 1. | ||
2682 : | syskin | 1.64 | int cbp = 0, i, t, dc = 1024, b_dc; |
2683 : | edgomez | 1.55 | const uint32_t iQuant = Data->lambda16; |
2684 : | edgomez | 1.61 | int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64; |
2685 : | syskin | 1.64 | uint32_t iDcScaler = get_dc_scaler(iQuant, 1);; |
2686 : | chl | 1.18 | |
2687 : | edgomez | 1.55 | for(i = 0; i < 4; i++) { |
2688 : | int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth); | ||
2689 : | transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth); | ||
2690 : | fdct(in); | ||
2691 : | syskin | 1.64 | b_dc = in[0]; |
2692 : | in[0] -= dc; | ||
2693 : | dc = b_dc; | ||
2694 : | if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler); | ||
2695 : | else quant4_intra(coeff, in, iQuant, iDcScaler); | ||
2696 : | chl | 1.3 | |
2697 : | edgomez | 1.55 | bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcy_tab[coeff[0] + 255].len;; |
2698 : | Data->temp[i] = t; | ||
2699 : | if (t != 0) cbp |= 1 << (5 - i); | ||
2700 : | if (bits >= Data->iMinSAD[0]) break; | ||
2701 : | chl | 1.3 | } |
2702 : | |||
2703 : | edgomez | 1.55 | if (bits < Data->iMinSAD[0]) { // INTRA still looks good, let's add chroma |
2704 : | syskin | 1.64 | iDcScaler = get_dc_scaler(iQuant, 0); |
2705 : | edgomez | 1.55 | //chroma U |
2706 : | transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2); | ||
2707 : | fdct(in); | ||
2708 : | in[0] -= 1024; | ||
2709 : | if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler); | ||
2710 : | else quant4_intra(coeff, in, iQuant, iDcScaler); | ||
2711 : | chl | 1.3 | |
2712 : | edgomez | 1.55 | bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len; |
2713 : | if (t != 0) cbp |= 1 << (5 - 4); | ||
2714 : | edgomez | 1.20 | |
2715 : | edgomez | 1.55 | if (bits < Data->iMinSAD[0]) { |
2716 : | //chroma V | ||
2717 : | transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2); | ||
2718 : | fdct(in); | ||
2719 : | in[0] -= 1024; | ||
2720 : | if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler); | ||
2721 : | else quant4_intra(coeff, in, iQuant, iDcScaler); | ||
2722 : | chl | 1.3 | |
2723 : | edgomez | 1.55 | bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len; |
2724 : | if (t != 0) cbp |= 1 << (5 - 5); | ||
2725 : | edgomez | 1.20 | |
2726 : | syskin | 1.60 | bits += xvid_cbpy_tab[cbp>>2].len; |
2727 : | bits += mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len; | ||
2728 : | edgomez | 1.55 | } |
2729 : | } | ||
2730 : | return bits; | ||
2731 : | chl | 1.3 | } |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |