1 |
/************************************************************************** |
/***************************************************************************** |
2 |
* |
* |
3 |
* XVID MPEG-4 VIDEO CODEC |
* XVID MPEG-4 VIDEO CODEC |
4 |
* motion estimation |
* - Motion Estimation related code - |
5 |
* |
* |
6 |
* This program is an implementation of a part of one or more MPEG-4 |
* Copyright(C) 2002 Christoph Lampert <gruel@web.de> |
7 |
* Video tools as specified in ISO/IEC 14496-2 standard. Those intending |
* 2002 Michael Militzer <michael@xvid.org> |
8 |
* to use this software module in hardware or software products are |
* 2002-2003 Radoslaw Czyz <xvid@syskin.cjb.net> |
|
* advised that its use may infringe existing patents or copyrights, and |
|
|
* any such use would be at such party's own risk. The original |
|
|
* developer of this software module and his/her company, and subsequent |
|
|
* editors and their companies, will have no liability for use of this |
|
|
* software or modifications or derivatives thereof. |
|
9 |
* |
* |
10 |
* This program is free software; you can redistribute it and/or modify |
* This program is free software; you can redistribute it and/or modify |
11 |
* it under the terms of the GNU General Public License as published by |
* it under the terms of the GNU General Public License as published by |
19 |
* |
* |
20 |
* You should have received a copy of the GNU General Public License |
* You should have received a copy of the GNU General Public License |
21 |
* along with this program; if not, write to the Free Software |
* along with this program; if not, write to the Free Software |
22 |
* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
23 |
* |
* |
24 |
*************************************************************************/ |
* $Id$ |
25 |
|
* |
26 |
|
****************************************************************************/ |
27 |
|
|
28 |
#include <assert.h> |
#include <assert.h> |
29 |
#include <stdio.h> |
#include <stdio.h> |
30 |
#include <stdlib.h> |
#include <stdlib.h> |
31 |
|
#include <string.h> /* memcpy */ |
32 |
|
#include <math.h> /* lrint */ |
33 |
|
|
34 |
#include "../encoder.h" |
#include "../encoder.h" |
35 |
#include "../utils/mbfunctions.h" |
#include "../utils/mbfunctions.h" |
36 |
#include "../prediction/mbprediction.h" |
#include "../prediction/mbprediction.h" |
37 |
#include "../global.h" |
#include "../global.h" |
38 |
#include "../utils/timer.h" |
#include "../utils/timer.h" |
39 |
|
#include "../image/interpolate8x8.h" |
40 |
#include "motion_est.h" |
#include "motion_est.h" |
41 |
#include "motion.h" |
#include "motion.h" |
42 |
#include "sad.h" |
#include "sad.h" |
43 |
|
#include "gmc.h" |
44 |
#include "../utils/emms.h" |
#include "../utils/emms.h" |
45 |
|
#include "../dct/fdct.h" |
46 |
|
|
47 |
|
/***************************************************************************** |
48 |
|
* Modified rounding tables -- declared in motion.h |
49 |
|
* Original tables see ISO spec tables 7-6 -> 7-9 |
50 |
|
****************************************************************************/ |
51 |
|
|
52 |
|
const uint32_t roundtab[16] = |
53 |
|
{0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2 }; |
54 |
|
|
55 |
|
/* K = 4 */ |
56 |
|
const uint32_t roundtab_76[16] = |
57 |
|
{ 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1 }; |
58 |
|
|
59 |
|
/* K = 2 */ |
60 |
|
const uint32_t roundtab_78[8] = |
61 |
|
{ 0, 0, 1, 1, 0, 0, 0, 1 }; |
62 |
|
|
63 |
|
/* K = 1 */ |
64 |
|
const uint32_t roundtab_79[4] = |
65 |
|
{ 0, 1, 0, 0 }; |
66 |
|
|
67 |
#define INITIAL_SKIP_THRESH (10) |
#define INITIAL_SKIP_THRESH (10) |
68 |
#define FINAL_SKIP_THRESH (50) |
#define FINAL_SKIP_THRESH (50) |
69 |
#define MAX_SAD00_FOR_SKIP (20) |
#define MAX_SAD00_FOR_SKIP (20) |
70 |
#define MAX_CHROMA_SAD_FOR_SKIP (22) |
#define MAX_CHROMA_SAD_FOR_SKIP (22) |
|
#define SKIP_THRESH_B (25) |
|
71 |
|
|
72 |
#define CHECK_CANDIDATE(X,Y,D) { \ |
#define CHECK_CANDIDATE(X,Y,D) { \ |
73 |
(*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } |
CheckCandidate((X),(Y), (D), &iDirection, data ); } |
74 |
|
|
|
#define iDiamondSize 2 |
|
75 |
|
|
76 |
static __inline int |
/***************************************************************************** |
77 |
d_mv_bits(int x, int y, const uint32_t iFcode) |
* Code |
78 |
|
****************************************************************************/ |
79 |
|
|
80 |
|
static __inline uint32_t |
81 |
|
d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv) |
82 |
{ |
{ |
83 |
int xb, yb; |
int bits; |
84 |
|
const int q = (1 << (iFcode - 1)) - 1; |
85 |
|
|
86 |
if (x == 0) xb = 1; |
x <<= qpel; |
87 |
else { |
y <<= qpel; |
88 |
if (x < 0) x = -x; |
if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); } |
89 |
x += (1 << (iFcode - 1)) - 1; |
|
90 |
|
x -= pred.x; |
91 |
|
bits = (x != 0 ? iFcode:0); |
92 |
|
x = abs(x); |
93 |
|
x += q; |
94 |
x >>= (iFcode - 1); |
x >>= (iFcode - 1); |
95 |
if (x > 32) x = 32; |
bits += mvtab[x]; |
|
xb = mvtab[x] + iFcode; |
|
|
} |
|
96 |
|
|
97 |
if (y == 0) yb = 1; |
y -= pred.y; |
98 |
else { |
bits += (y != 0 ? iFcode:0); |
99 |
if (y < 0) y = -y; |
y = abs(y); |
100 |
y += (1 << (iFcode - 1)) - 1; |
y += q; |
101 |
y >>= (iFcode - 1); |
y >>= (iFcode - 1); |
102 |
if (y > 32) y = 32; |
bits += mvtab[y]; |
103 |
yb = mvtab[y] + iFcode; |
|
104 |
|
return bits; |
105 |
|
} |
106 |
|
|
107 |
|
static int32_t ChromaSAD2(const int fx, const int fy, const int bx, const int by, |
108 |
|
const SearchData * const data) |
109 |
|
{ |
110 |
|
int sad; |
111 |
|
const uint32_t stride = data->iEdgedWidth/2; |
112 |
|
uint8_t * f_refu = data->RefQ, |
113 |
|
* f_refv = data->RefQ + 8, |
114 |
|
* b_refu = data->RefQ + 16, |
115 |
|
* b_refv = data->RefQ + 24; |
116 |
|
int offset = (fx>>1) + (fy>>1)*stride; |
117 |
|
|
118 |
|
switch (((fx & 1) << 1) | (fy & 1)) { |
119 |
|
case 0: |
120 |
|
f_refu = (uint8_t*)data->RefP[4] + offset; |
121 |
|
f_refv = (uint8_t*)data->RefP[5] + offset; |
122 |
|
break; |
123 |
|
case 1: |
124 |
|
interpolate8x8_halfpel_v(f_refu, data->RefP[4] + offset, stride, data->rounding); |
125 |
|
interpolate8x8_halfpel_v(f_refv, data->RefP[5] + offset, stride, data->rounding); |
126 |
|
break; |
127 |
|
case 2: |
128 |
|
interpolate8x8_halfpel_h(f_refu, data->RefP[4] + offset, stride, data->rounding); |
129 |
|
interpolate8x8_halfpel_h(f_refv, data->RefP[5] + offset, stride, data->rounding); |
130 |
|
break; |
131 |
|
default: |
132 |
|
interpolate8x8_halfpel_hv(f_refu, data->RefP[4] + offset, stride, data->rounding); |
133 |
|
interpolate8x8_halfpel_hv(f_refv, data->RefP[5] + offset, stride, data->rounding); |
134 |
|
break; |
135 |
|
} |
136 |
|
|
137 |
|
offset = (bx>>1) + (by>>1)*stride; |
138 |
|
switch (((bx & 1) << 1) | (by & 1)) { |
139 |
|
case 0: |
140 |
|
b_refu = (uint8_t*)data->b_RefP[4] + offset; |
141 |
|
b_refv = (uint8_t*)data->b_RefP[5] + offset; |
142 |
|
break; |
143 |
|
case 1: |
144 |
|
interpolate8x8_halfpel_v(b_refu, data->b_RefP[4] + offset, stride, data->rounding); |
145 |
|
interpolate8x8_halfpel_v(b_refv, data->b_RefP[5] + offset, stride, data->rounding); |
146 |
|
break; |
147 |
|
case 2: |
148 |
|
interpolate8x8_halfpel_h(b_refu, data->b_RefP[4] + offset, stride, data->rounding); |
149 |
|
interpolate8x8_halfpel_h(b_refv, data->b_RefP[5] + offset, stride, data->rounding); |
150 |
|
break; |
151 |
|
default: |
152 |
|
interpolate8x8_halfpel_hv(b_refu, data->b_RefP[4] + offset, stride, data->rounding); |
153 |
|
interpolate8x8_halfpel_hv(b_refv, data->b_RefP[5] + offset, stride, data->rounding); |
154 |
|
break; |
155 |
|
} |
156 |
|
|
157 |
|
sad = sad8bi(data->CurU, b_refu, f_refu, stride); |
158 |
|
sad += sad8bi(data->CurV, b_refv, f_refv, stride); |
159 |
|
|
160 |
|
return sad; |
161 |
|
} |
162 |
|
|
163 |
|
static int32_t |
164 |
|
ChromaSAD(const int dx, const int dy, const SearchData * const data) |
165 |
|
{ |
166 |
|
int sad; |
167 |
|
const uint32_t stride = data->iEdgedWidth/2; |
168 |
|
int offset = (dx>>1) + (dy>>1)*stride; |
169 |
|
|
170 |
|
if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; /* it has been checked recently */ |
171 |
|
data->temp[5] = dx; data->temp[6] = dy; /* backup */ |
172 |
|
|
173 |
|
switch (((dx & 1) << 1) | (dy & 1)) { |
174 |
|
case 0: |
175 |
|
sad = sad8(data->CurU, data->RefP[4] + offset, stride); |
176 |
|
sad += sad8(data->CurV, data->RefP[5] + offset, stride); |
177 |
|
break; |
178 |
|
case 1: |
179 |
|
sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + stride, stride); |
180 |
|
sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + stride, stride); |
181 |
|
break; |
182 |
|
case 2: |
183 |
|
sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + 1, stride); |
184 |
|
sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + 1, stride); |
185 |
|
break; |
186 |
|
default: |
187 |
|
interpolate8x8_halfpel_hv(data->RefQ, data->RefP[4] + offset, stride, data->rounding); |
188 |
|
sad = sad8(data->CurU, data->RefQ, stride); |
189 |
|
|
190 |
|
interpolate8x8_halfpel_hv(data->RefQ, data->RefP[5] + offset, stride, data->rounding); |
191 |
|
sad += sad8(data->CurV, data->RefQ, stride); |
192 |
|
break; |
193 |
|
} |
194 |
|
data->temp[7] = sad; /* backup, part 2 */ |
195 |
|
return sad; |
196 |
|
} |
197 |
|
|
198 |
|
static __inline const uint8_t * |
199 |
|
GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data) |
200 |
|
{ |
201 |
|
/* dir : 0 = forward, 1 = backward */ |
202 |
|
const uint8_t *const *const direction = ( dir == 0 ? data->RefP : data->b_RefP ); |
203 |
|
const int picture = ((x&1)<<1) | (y&1); |
204 |
|
const int offset = (x>>1) + (y>>1)*data->iEdgedWidth; |
205 |
|
return direction[picture] + offset; |
206 |
|
} |
207 |
|
|
208 |
|
/* this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate */ |
209 |
|
static __inline const uint8_t * |
210 |
|
GetReference(const int x, const int y, const SearchData * const data) |
211 |
|
{ |
212 |
|
const int picture = ((x&1)<<1) | (y&1); |
213 |
|
const int offset = (x>>1) + (y>>1)*data->iEdgedWidth; |
214 |
|
return data->RefP[picture] + offset; |
215 |
|
} |
216 |
|
|
217 |
|
static uint8_t * |
218 |
|
Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data) |
219 |
|
{ |
220 |
|
/* create or find a qpel-precision reference picture; return pointer to it */ |
221 |
|
uint8_t * Reference = data->RefQ + 16*dir; |
222 |
|
const uint32_t iEdgedWidth = data->iEdgedWidth; |
223 |
|
const uint32_t rounding = data->rounding; |
224 |
|
const int halfpel_x = x/2; |
225 |
|
const int halfpel_y = y/2; |
226 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
227 |
|
|
228 |
|
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
229 |
|
ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
230 |
|
switch( ((x&1)<<1) + (y&1) ) { |
231 |
|
case 3: /* x and y in qpel resolution - the "corners" (top left/right and */ |
232 |
|
/* bottom left/right) during qpel refinement */ |
233 |
|
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
234 |
|
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
235 |
|
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
236 |
|
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
237 |
|
ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
238 |
|
ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
239 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
240 |
|
break; |
241 |
|
|
242 |
|
case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */ |
243 |
|
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
244 |
|
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
245 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
246 |
|
break; |
247 |
|
|
248 |
|
case 2: /* x qpel, y halfpel - left or right during qpel refinement */ |
249 |
|
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
250 |
|
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
251 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
252 |
|
break; |
253 |
|
|
254 |
|
default: /* pure halfpel position */ |
255 |
|
return (uint8_t *) ref1; |
256 |
|
|
257 |
} |
} |
258 |
return xb + yb; |
return Reference; |
259 |
} |
} |
260 |
|
|
261 |
|
static uint8_t * |
262 |
|
Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data) |
263 |
|
{ |
264 |
|
/* create or find a qpel-precision reference picture; return pointer to it */ |
265 |
|
uint8_t * Reference = data->RefQ + 16*dir; |
266 |
|
const uint32_t iEdgedWidth = data->iEdgedWidth; |
267 |
|
const uint32_t rounding = data->rounding; |
268 |
|
const int halfpel_x = x/2; |
269 |
|
const int halfpel_y = y/2; |
270 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
271 |
|
|
272 |
|
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
273 |
|
switch( ((x&1)<<1) + (y&1) ) { |
274 |
|
case 3: |
275 |
|
/* |
276 |
|
* x and y in qpel resolution - the "corners" (top left/right and |
277 |
|
* bottom left/right) during qpel refinement |
278 |
|
*/ |
279 |
|
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
280 |
|
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
281 |
|
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
282 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
283 |
|
interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); |
284 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); |
285 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); |
286 |
|
break; |
287 |
|
|
288 |
|
case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */ |
289 |
|
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
290 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
291 |
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
292 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); |
293 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
294 |
|
break; |
295 |
|
|
296 |
|
case 2: /* x qpel, y halfpel - left or right during qpel refinement */ |
297 |
|
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
298 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
299 |
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
300 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); |
301 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
302 |
|
break; |
303 |
|
|
304 |
|
default: /* pure halfpel position */ |
305 |
|
return (uint8_t *) ref1; |
306 |
|
} |
307 |
|
return Reference; |
308 |
|
} |
309 |
|
|
310 |
/* CHECK_CANDIATE FUNCTIONS START */ |
/* CHECK_CANDIATE FUNCTIONS START */ |
311 |
|
|
312 |
static void |
static void |
313 |
CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
314 |
{ |
{ |
315 |
int32_t * const sad = data->temp; |
int xc, yc; |
|
int t; |
|
316 |
const uint8_t * Reference; |
const uint8_t * Reference; |
317 |
|
VECTOR * current; |
318 |
|
int32_t sad; uint32_t t; |
319 |
|
|
320 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if (( x > data->max_dx) || ( x < data->min_dx) |
321 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
322 |
|
|
323 |
switch ( ((x&1)<<1) + (y&1) ) { |
if (!data->qpel_precision) { |
324 |
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
Reference = GetReference(x, y, data); |
325 |
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
current = data->currentMV; |
326 |
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
xc = x; yc = y; |
327 |
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
} else { /* x and y are in 1/4 precision */ |
328 |
|
Reference = Interpolate16x16qpel(x, y, 0, data); |
329 |
|
xc = x/2; yc = y/2; /* for chroma sad */ |
330 |
|
current = data->currentQMV; |
331 |
|
} |
332 |
|
|
333 |
|
sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
334 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
335 |
|
|
336 |
|
sad += (data->lambda16 * t * sad)>>10; |
337 |
|
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; |
338 |
|
|
339 |
|
if (data->chroma && sad < data->iMinSAD[0]) |
340 |
|
sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
341 |
|
(yc >> 1) + roundtab_79[yc & 0x3], data); |
342 |
|
|
343 |
|
if (sad < data->iMinSAD[0]) { |
344 |
|
data->iMinSAD[0] = sad; |
345 |
|
current[0].x = x; current[0].y = y; |
346 |
|
*dir = Direction; |
347 |
|
} |
348 |
|
|
349 |
|
if (data->temp[1] < data->iMinSAD[1]) { |
350 |
|
data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y = y; } |
351 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
352 |
|
data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; } |
353 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
354 |
|
data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; } |
355 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
356 |
|
data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; } |
357 |
|
} |
358 |
|
|
359 |
|
static void |
360 |
|
CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
361 |
|
{ |
362 |
|
int32_t sad; uint32_t t; |
363 |
|
const uint8_t * Reference; |
364 |
|
VECTOR * current; |
365 |
|
|
366 |
|
if ( (x > data->max_dx) || (x < data->min_dx) |
367 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
368 |
|
|
369 |
|
if (!data->qpel_precision) { |
370 |
|
Reference = GetReference(x, y, data); |
371 |
|
current = data->currentMV; |
372 |
|
} else { /* x and y are in 1/4 precision */ |
373 |
|
Reference = Interpolate8x8qpel(x, y, 0, 0, data); |
374 |
|
current = data->currentQMV; |
375 |
|
} |
376 |
|
|
377 |
|
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
378 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
379 |
|
|
380 |
|
sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10; |
381 |
|
|
382 |
|
if (sad < *(data->iMinSAD)) { |
383 |
|
*(data->iMinSAD) = sad; |
384 |
|
current->x = x; current->y = y; |
385 |
|
*dir = Direction; |
386 |
|
} |
387 |
} |
} |
388 |
|
|
389 |
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, sad+1); |
static void |
390 |
|
CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
391 |
|
{ |
392 |
|
uint32_t t; |
393 |
|
const uint8_t * Reference; |
394 |
|
|
395 |
|
if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || /* non-zero even value */ |
396 |
|
(x > data->max_dx) || (x < data->min_dx) |
397 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
398 |
|
|
399 |
|
Reference = GetReference(x, y, data); |
400 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1); |
401 |
|
|
402 |
t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
403 |
data->temp[0] += lambda_vec16[data->iQuant] * t; |
|
404 |
data->temp[1] += lambda_vec8[data->iQuant] * t; |
data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10; |
405 |
|
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; |
406 |
|
|
407 |
if (data->temp[0] < data->iMinSAD[0]) { |
if (data->temp[0] < data->iMinSAD[0]) { |
408 |
data->iMinSAD[0] = data->temp[0]; |
data->iMinSAD[0] = data->temp[0]; |
417 |
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
418 |
if (data->temp[4] < data->iMinSAD[4]) { |
if (data->temp[4] < data->iMinSAD[4]) { |
419 |
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
|
|
|
420 |
} |
} |
421 |
|
|
422 |
static void |
static void |
423 |
CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
424 |
{ |
{ |
425 |
int32_t sad; |
int32_t sad, xc, yc; |
426 |
const uint8_t * Reference; |
const uint8_t * Reference; |
427 |
|
uint32_t t; |
428 |
|
VECTOR * current; |
429 |
|
|
430 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if (( x > data->max_dx) || ( x < data->min_dx) |
431 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
432 |
|
|
433 |
switch ( ((x&1)<<1) + (y&1) ) |
if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; /* non-zero even value */ |
|
{ |
|
|
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
|
|
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
|
|
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
|
|
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
|
|
} |
|
434 |
|
|
435 |
sad = lambda_vec16[data->iQuant] * |
if (data->qpel_precision) { /* x and y are in 1/4 precision */ |
436 |
d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
Reference = Interpolate16x16qpel(x, y, 0, data); |
437 |
sad += sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
current = data->currentQMV; |
438 |
|
xc = x/2; yc = y/2; |
439 |
|
} else { |
440 |
|
Reference = GetReference(x, y, data); |
441 |
|
current = data->currentMV; |
442 |
|
xc = x; yc = y; |
443 |
|
} |
444 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, |
445 |
|
data->qpel^data->qpel_precision, data->rrv); |
446 |
|
|
447 |
|
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
448 |
|
sad += (data->lambda16 * t * sad)>>10; |
449 |
|
|
450 |
|
if (data->chroma && sad < *data->iMinSAD) |
451 |
|
sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
452 |
|
(yc >> 1) + roundtab_79[yc & 0x3], data); |
453 |
|
|
454 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
455 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
456 |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
current->x = x; current->y = y; |
457 |
*dir = Direction; } |
*dir = Direction; |
458 |
|
} |
459 |
} |
} |
460 |
|
|
461 |
static void |
static void |
462 |
CheckCandidate16no4vI(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidate16I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
463 |
{ |
{ |
464 |
int32_t sad; |
int sad; |
465 |
|
// int xc, yc; |
466 |
|
const uint8_t * Reference; |
467 |
|
// VECTOR * current; |
468 |
|
|
469 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if (( x > data->max_dx) || ( x < data->min_dx) |
470 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
471 |
|
|
472 |
sad = lambda_vec16[data->iQuant] * |
Reference = GetReference(x, y, data); |
473 |
d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
// xc = x; yc = y; |
474 |
|
|
475 |
sad += sad16(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), |
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
476 |
data->iEdgedWidth, 256*4096); |
// sad += d_mv_bits(x, y, data->predMV, data->iFcode, 0, 0); |
477 |
|
|
478 |
if (sad < *(data->iMinSAD)) { |
/* if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
479 |
*(data->iMinSAD) = sad; |
(yc >> 1) + roundtab_79[yc & 0x3], data); |
480 |
|
*/ |
481 |
|
|
482 |
|
if (sad < data->iMinSAD[0]) { |
483 |
|
data->iMinSAD[0] = sad; |
484 |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
485 |
*dir = Direction; } |
*dir = Direction; |
486 |
|
} |
487 |
} |
} |
|
|
|
488 |
|
|
489 |
static void |
static void |
490 |
CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
491 |
{ |
{ |
492 |
|
/* maximum speed - for P/B/I decision */ |
493 |
int32_t sad; |
int32_t sad; |
|
const int xb = data->currentMV[1].x; |
|
|
const int yb = data->currentMV[1].y; |
|
|
const uint8_t *ReferenceF, *ReferenceB; |
|
494 |
|
|
495 |
if (( xf > data->max_dx) || ( xf < data->min_dx) |
if ( (x > data->max_dx) || (x < data->min_dx) |
496 |
|| ( yf > data->max_dy) || (yf < data->min_dy)) return; |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
497 |
|
|
498 |
|
sad = sad32v_c(data->Cur, data->RefP[0] + (x>>1) + (y>>1)*((int)data->iEdgedWidth), |
499 |
|
data->iEdgedWidth, data->temp+1); |
500 |
|
|
501 |
switch ( ((xf&1)<<1) + (yf&1) ) { |
if (sad < *(data->iMinSAD)) { |
502 |
case 0 : ReferenceF = data->Ref + xf/2 + (yf/2)*(data->iEdgedWidth); break; |
*(data->iMinSAD) = sad; |
503 |
case 1 : ReferenceF = data->RefV + xf/2 + ((yf-1)/2)*(data->iEdgedWidth); break; |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
504 |
case 2 : ReferenceF = data->RefH + (xf-1)/2 + (yf/2)*(data->iEdgedWidth); break; |
*dir = Direction; |
|
default : ReferenceF = data->RefHV + (xf-1)/2 + ((yf-1)/2)*(data->iEdgedWidth); break; |
|
505 |
} |
} |
506 |
|
if (data->temp[1] < data->iMinSAD[1]) { |
507 |
|
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
508 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
509 |
|
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
510 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
511 |
|
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
512 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
513 |
|
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
514 |
|
|
|
switch ( ((xb&1)<<1) + (yb&1) ) { |
|
|
case 0 : ReferenceB = data->bRef + xb/2 + (yb/2)*(data->iEdgedWidth); break; |
|
|
case 1 : ReferenceB = data->bRefV + xb/2 + ((yb-1)/2)*(data->iEdgedWidth); break; |
|
|
case 2 : ReferenceB = data->bRefH + (xb-1)/2 + (yb/2)*(data->iEdgedWidth); break; |
|
|
default : ReferenceB = data->bRefHV + (xb-1)/2 + ((yb-1)/2)*(data->iEdgedWidth); break; |
|
515 |
} |
} |
516 |
|
|
517 |
sad = lambda_vec16[data->iQuant] * |
static void |
518 |
( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) + |
CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) |
519 |
d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode) ); |
{ |
520 |
|
int32_t sad, xb, yb, xcf, ycf, xcb, ycb; |
521 |
|
uint32_t t; |
522 |
|
const uint8_t *ReferenceF, *ReferenceB; |
523 |
|
VECTOR *current; |
524 |
|
|
525 |
sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
if ((xf > data->max_dx) || (xf < data->min_dx) || |
526 |
|
(yf > data->max_dy) || (yf < data->min_dy)) |
527 |
|
return; |
528 |
|
|
529 |
|
if (!data->qpel_precision) { |
530 |
|
ReferenceF = GetReference(xf, yf, data); |
531 |
|
xb = data->currentMV[1].x; yb = data->currentMV[1].y; |
532 |
|
ReferenceB = GetReferenceB(xb, yb, 1, data); |
533 |
|
current = data->currentMV; |
534 |
|
xcf = xf; ycf = yf; |
535 |
|
xcb = xb; ycb = yb; |
536 |
|
} else { |
537 |
|
ReferenceF = Interpolate16x16qpel(xf, yf, 0, data); |
538 |
|
xb = data->currentQMV[1].x; yb = data->currentQMV[1].y; |
539 |
|
current = data->currentQMV; |
540 |
|
ReferenceB = Interpolate16x16qpel(xb, yb, 1, data); |
541 |
|
xcf = xf/2; ycf = yf/2; |
542 |
|
xcb = xb/2; ycb = yb/2; |
543 |
|
} |
544 |
|
|
545 |
|
t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0) |
546 |
|
+ d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
547 |
|
|
548 |
|
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
549 |
|
sad += (data->lambda16 * t * sad)>>10; |
550 |
|
|
551 |
|
if (data->chroma && sad < *data->iMinSAD) |
552 |
|
sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3], |
553 |
|
(ycf >> 1) + roundtab_79[ycf & 0x3], |
554 |
|
(xcb >> 1) + roundtab_79[xcb & 0x3], |
555 |
|
(ycb >> 1) + roundtab_79[ycb & 0x3], data); |
556 |
|
|
557 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
558 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
559 |
data->currentMV->x = xf; data->currentMV->y = yf; |
current->x = xf; current->y = yf; |
560 |
*dir = Direction; } |
*dir = Direction; |
561 |
|
} |
562 |
} |
} |
563 |
|
|
564 |
static void |
static void |
565 |
CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
566 |
{ |
{ |
567 |
int32_t sad; |
int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0; |
568 |
int k; |
uint32_t k; |
569 |
const uint8_t *ReferenceF; |
const uint8_t *ReferenceF; |
570 |
const uint8_t *ReferenceB; |
const uint8_t *ReferenceB; |
571 |
VECTOR mvs, b_mvs; |
VECTOR mvs, b_mvs; |
572 |
|
|
573 |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
574 |
|
|
|
sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1); |
|
|
|
|
575 |
for (k = 0; k < 4; k++) { |
for (k = 0; k < 4; k++) { |
576 |
mvs.x = data->directmvF[k].x + x; |
mvs.x = data->directmvF[k].x + x; |
577 |
b_mvs.x = ((x == 0) ? |
b_mvs.x = ((x == 0) ? |
583 |
data->directmvB[k].y |
data->directmvB[k].y |
584 |
: mvs.y - data->referencemv[k].y); |
: mvs.y - data->referencemv[k].y); |
585 |
|
|
586 |
if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
if ((mvs.x > data->max_dx) || (mvs.x < data->min_dx) || |
587 |
|| ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) |
(mvs.y > data->max_dy) || (mvs.y < data->min_dy) || |
588 |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
(b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx) || |
589 |
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
(b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) ) |
590 |
|
return; |
591 |
switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
|
592 |
case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
if (data->qpel) { |
593 |
case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
xcf += mvs.x/2; ycf += mvs.y/2; |
594 |
case 2 : ReferenceF = data->RefH + (mvs.x-1)/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
xcb += b_mvs.x/2; ycb += b_mvs.y/2; |
595 |
default : ReferenceF = data->RefHV + (mvs.x-1)/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
} else { |
596 |
|
xcf += mvs.x; ycf += mvs.y; |
597 |
|
xcb += b_mvs.x; ycb += b_mvs.y; |
598 |
|
mvs.x *= 2; mvs.y *= 2; /* we move to qpel precision anyway */ |
599 |
|
b_mvs.x *= 2; b_mvs.y *= 2; |
600 |
} |
} |
601 |
|
|
602 |
switch ( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { |
ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); |
603 |
case 0 : ReferenceB = data->bRef + b_mvs.x/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); |
|
case 1 : ReferenceB = data->bRefV + b_mvs.x/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
|
|
case 2 : ReferenceB = data->bRefH + (b_mvs.x-1)/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
|
|
default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
|
|
} |
|
604 |
|
|
605 |
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
606 |
ReferenceF + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
ReferenceF, ReferenceB, data->iEdgedWidth); |
|
ReferenceB + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
|
|
data->iEdgedWidth); |
|
607 |
if (sad > *(data->iMinSAD)) return; |
if (sad > *(data->iMinSAD)) return; |
608 |
} |
} |
609 |
|
|
610 |
|
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
611 |
|
|
612 |
|
if (data->chroma && sad < *data->iMinSAD) |
613 |
|
sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
614 |
|
(ycf >> 3) + roundtab_76[ycf & 0xf], |
615 |
|
(xcb >> 3) + roundtab_76[xcb & 0xf], |
616 |
|
(ycb >> 3) + roundtab_76[ycb & 0xf], data); |
617 |
|
|
618 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
619 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
620 |
data->currentMV->x = x; data->currentMV->y = y; |
data->currentMV->x = x; data->currentMV->y = y; |
621 |
*dir = Direction; } |
*dir = Direction; |
622 |
|
} |
623 |
} |
} |
624 |
|
|
625 |
static void |
static void |
626 |
CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
627 |
{ |
{ |
628 |
int32_t sad; |
int32_t sad, xcf, ycf, xcb, ycb; |
629 |
const uint8_t *ReferenceF; |
const uint8_t *ReferenceF; |
630 |
const uint8_t *ReferenceB; |
const uint8_t *ReferenceB; |
631 |
VECTOR mvs, b_mvs; |
VECTOR mvs, b_mvs; |
632 |
|
|
633 |
if (( x > 31) || ( x < -31) || ( y > 31) || (y < -31)) return; |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
|
|
|
|
sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1); |
|
634 |
|
|
635 |
mvs.x = data->directmvF[0].x + x; |
mvs.x = data->directmvF[0].x + x; |
636 |
b_mvs.x = ((x == 0) ? |
b_mvs.x = ((x == 0) ? |
647 |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
648 |
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
649 |
|
|
650 |
switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
if (data->qpel) { |
651 |
case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2); |
652 |
case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2); |
653 |
case 2 : ReferenceF = data->RefH + (mvs.x-1)/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data); |
654 |
default : ReferenceF = data->RefHV + (mvs.x-1)/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); |
655 |
} |
} else { |
656 |
|
xcf = 4*mvs.x; ycf = 4*mvs.y; |
657 |
switch ( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { |
xcb = 4*b_mvs.x; ycb = 4*b_mvs.y; |
658 |
case 0 : ReferenceB = data->bRef + b_mvs.x/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
ReferenceF = GetReference(mvs.x, mvs.y, data); |
659 |
case 1 : ReferenceB = data->bRefV + b_mvs.x/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data); |
|
case 2 : ReferenceB = data->bRefH + (b_mvs.x-1)/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
|
|
default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
|
660 |
} |
} |
661 |
|
|
662 |
sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
663 |
|
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
664 |
|
|
665 |
|
if (data->chroma && sad < *data->iMinSAD) |
666 |
|
sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
667 |
|
(ycf >> 3) + roundtab_76[ycf & 0xf], |
668 |
|
(xcb >> 3) + roundtab_76[xcb & 0xf], |
669 |
|
(ycb >> 3) + roundtab_76[ycb & 0xf], data); |
670 |
|
|
671 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
672 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
673 |
data->currentMV->x = x; data->currentMV->y = y; |
data->currentMV->x = x; data->currentMV->y = y; |
674 |
*dir = Direction; } |
*dir = Direction; |
675 |
|
} |
676 |
} |
} |
677 |
|
|
678 |
|
|
679 |
static void |
static void |
680 |
CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidateRD16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
681 |
{ |
{ |
682 |
int32_t sad; |
|
683 |
const uint8_t * Reference; |
int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; |
684 |
|
int32_t rd = 0; |
685 |
|
VECTOR * current; |
686 |
|
const uint8_t * ptr; |
687 |
|
int i, cbp = 0, t, xc, yc; |
688 |
|
|
689 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if (( x > data->max_dx) || ( x < data->min_dx) |
690 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
691 |
|
|
692 |
switch ( ((x&1)<<1) + (y&1) ) |
if (!data->qpel_precision) { |
693 |
{ |
ptr = GetReference(x, y, data); |
694 |
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
current = data->currentMV; |
695 |
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
xc = x; yc = y; |
696 |
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
} else { /* x and y are in 1/4 precision */ |
697 |
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
ptr = Interpolate16x16qpel(x, y, 0, data); |
698 |
|
current = data->currentQMV; |
699 |
|
xc = x/2; yc = y/2; |
700 |
|
} |
701 |
|
|
702 |
|
for(i = 0; i < 4; i++) { |
703 |
|
int s = 8*((i&1) + (i>>1)*data->iEdgedWidth); |
704 |
|
transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth); |
705 |
|
rd += data->temp[i] = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, i); |
706 |
|
} |
707 |
|
|
708 |
|
rd += t = BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
709 |
|
|
710 |
|
if (data->temp[0] + t < data->iMinSAD[1]) { |
711 |
|
data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; data->cbp[1] = (data->cbp[1]&~32) | cbp&32; } |
712 |
|
if (data->temp[1] < data->iMinSAD[2]) { |
713 |
|
data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; data->cbp[1] = (data->cbp[1]&~16) | cbp&16; } |
714 |
|
if (data->temp[2] < data->iMinSAD[3]) { |
715 |
|
data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; data->cbp[1] = (data->cbp[1]&~8) | cbp&8; } |
716 |
|
if (data->temp[3] < data->iMinSAD[4]) { |
717 |
|
data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; data->cbp[1] = (data->cbp[1]&~4) | cbp&4; } |
718 |
|
|
719 |
|
rd += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; |
720 |
|
|
721 |
|
if (rd >= data->iMinSAD[0]) return; |
722 |
|
|
723 |
|
/* chroma */ |
724 |
|
xc = (xc >> 1) + roundtab_79[xc & 0x3]; |
725 |
|
yc = (yc >> 1) + roundtab_79[yc & 0x3]; |
726 |
|
|
727 |
|
/* chroma U */ |
728 |
|
ptr = interpolate8x8_switch2(data->RefQ, data->RefP[4], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); |
729 |
|
transfer_8to16subro(in, data->CurU, ptr, data->iEdgedWidth/2); |
730 |
|
rd += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 4); |
731 |
|
if (rd >= data->iMinSAD[0]) return; |
732 |
|
|
733 |
|
/* chroma V */ |
734 |
|
ptr = interpolate8x8_switch2(data->RefQ, data->RefP[5], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); |
735 |
|
transfer_8to16subro(in, data->CurV, ptr, data->iEdgedWidth/2); |
736 |
|
rd += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5); |
737 |
|
|
738 |
|
rd += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; |
739 |
|
|
740 |
|
if (rd < data->iMinSAD[0]) { |
741 |
|
data->iMinSAD[0] = rd; |
742 |
|
current[0].x = x; current[0].y = y; |
743 |
|
*dir = Direction; |
744 |
|
*data->cbp = cbp; |
745 |
|
} |
746 |
} |
} |
747 |
|
|
748 |
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
static void |
749 |
sad += lambda_vec8[data->iQuant] * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
CheckCandidateRD8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
750 |
|
{ |
751 |
|
|
752 |
if (sad < *(data->iMinSAD)) { |
int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; |
753 |
*(data->iMinSAD) = sad; |
int32_t rd; |
754 |
data->currentMV->x = x; data->currentMV->y = y; |
VECTOR * current; |
755 |
*dir = Direction; } |
const uint8_t * ptr; |
756 |
|
int cbp = 0; |
757 |
|
|
758 |
|
if ( (x > data->max_dx) || (x < data->min_dx) |
759 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
760 |
|
|
761 |
|
if (!data->qpel_precision) { |
762 |
|
ptr = GetReference(x, y, data); |
763 |
|
current = data->currentMV; |
764 |
|
} else { /* x and y are in 1/4 precision */ |
765 |
|
ptr = Interpolate8x8qpel(x, y, 0, 0, data); |
766 |
|
current = data->currentQMV; |
767 |
|
} |
768 |
|
|
769 |
|
transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth); |
770 |
|
rd = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5); |
771 |
|
rd += BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
772 |
|
|
773 |
|
if (rd < data->iMinSAD[0]) { |
774 |
|
*data->cbp = cbp; |
775 |
|
data->iMinSAD[0] = rd; |
776 |
|
current[0].x = x; current[0].y = y; |
777 |
|
*dir = Direction; |
778 |
|
} |
779 |
} |
} |
780 |
|
|
781 |
/* CHECK_CANDIATE FUNCTIONS END */ |
/* CHECK_CANDIATE FUNCTIONS END */ |
790 |
|
|
791 |
int iDirection; |
int iDirection; |
792 |
|
|
793 |
do { |
for(;;) { /* forever */ |
794 |
iDirection = 0; |
iDirection = 0; |
795 |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
796 |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
799 |
|
|
800 |
/* now we're doing diagonal checks near our candidate */ |
/* now we're doing diagonal checks near our candidate */ |
801 |
|
|
802 |
if (iDirection) { //checking if anything found |
if (iDirection) { /* if anything found */ |
803 |
bDirection = iDirection; |
bDirection = iDirection; |
804 |
iDirection = 0; |
iDirection = 0; |
805 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
806 |
if (bDirection & 3) { //our candidate is left or right |
if (bDirection & 3) { /* our candidate is left or right */ |
807 |
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
808 |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
809 |
} else { // what remains here is up or down |
} else { /* what remains here is up or down */ |
810 |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
811 |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); } |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
812 |
|
} |
813 |
|
|
814 |
if (iDirection) { |
if (iDirection) { |
815 |
bDirection += iDirection; |
bDirection += iDirection; |
816 |
x = data->currentMV->x; y = data->currentMV->y; } |
x = data->currentMV->x; y = data->currentMV->y; |
817 |
} else { //about to quit, eh? not so fast.... |
} |
818 |
|
} else { /* about to quit, eh? not so fast.... */ |
819 |
switch (bDirection) { |
switch (bDirection) { |
820 |
case 2: |
case 2: |
821 |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
853 |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
854 |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
855 |
break; |
break; |
856 |
default: //1+2+4+8 == we didn't find anything at all |
default: /* 1+2+4+8 == we didn't find anything at all */ |
857 |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
858 |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
859 |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
860 |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
861 |
break; |
break; |
862 |
} |
} |
863 |
if (!iDirection) break; //ok, the end. really |
if (!iDirection) break; /* ok, the end. really */ |
864 |
bDirection = iDirection; |
bDirection = iDirection; |
865 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
866 |
} |
} |
867 |
} |
} |
|
while (1); //forever |
|
868 |
} |
} |
869 |
|
|
870 |
static void |
static void |
905 |
|
|
906 |
/* now we're doing diagonal checks near our candidate */ |
/* now we're doing diagonal checks near our candidate */ |
907 |
|
|
908 |
if (iDirection) { //checking if anything found |
if (iDirection) { /* checking if anything found */ |
909 |
bDirection = iDirection; |
bDirection = iDirection; |
910 |
iDirection = 0; |
iDirection = 0; |
911 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
912 |
if (bDirection & 3) { //our candidate is left or right |
if (bDirection & 3) { /* our candidate is left or right */ |
913 |
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
914 |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
915 |
} else { // what remains here is up or down |
} else { /* what remains here is up or down */ |
916 |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
917 |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); } |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
918 |
|
} |
919 |
bDirection += iDirection; |
bDirection += iDirection; |
920 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
921 |
} |
} |
925 |
|
|
926 |
/* MAINSEARCH FUNCTIONS END */ |
/* MAINSEARCH FUNCTIONS END */ |
927 |
|
|
|
/* HALFPELREFINE COULD BE A MAINSEARCH FUNCTION, BUT THERE IS NO NEED FOR IT */ |
|
|
|
|
928 |
static void |
static void |
929 |
HalfpelRefine(const SearchData * const data) |
SubpelRefine(const SearchData * const data) |
930 |
{ |
{ |
931 |
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
/* Do a half-pel or q-pel refinement */ |
932 |
|
const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV; |
933 |
VECTOR backupMV = *(data->currentMV); |
int iDirection; /* only needed because macro expects it */ |
934 |
int iDirection; //not needed |
|
935 |
|
CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0); |
936 |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); |
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0); |
937 |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); |
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0); |
938 |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); |
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0); |
939 |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0); |
CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0); |
940 |
|
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0); |
941 |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); |
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0); |
942 |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); |
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0); |
|
|
|
|
CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); |
|
|
CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); |
|
943 |
} |
} |
944 |
|
|
945 |
static __inline int |
static __inline int |
946 |
SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
947 |
const int x, const int y, |
const int x, const int y, |
948 |
const uint32_t iEdgedWidth, const uint32_t iQuant) |
const uint32_t stride, const uint32_t iQuant, int rrv) |
949 |
|
|
950 |
{ |
{ |
951 |
/* keep repeating checks for all b-frames before this P frame, |
int offset = (x + y*stride)*8; |
952 |
to make sure that SKIP is possible (todo) |
if(!rrv) { |
953 |
how: if skip is not possible set sad00 to a very high value */ |
uint32_t sadC = sad8(current->u + offset, |
954 |
|
reference->u + offset, stride); |
|
uint32_t sadC = sad8(current->u + x*8 + y*(iEdgedWidth/2)*8, |
|
|
reference->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); |
|
955 |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
956 |
sadC += sad8(current->v + (x + y*(iEdgedWidth/2))*8, |
sadC += sad8(current->v + offset, |
957 |
reference->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); |
reference->v + offset, stride); |
958 |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
959 |
|
return 1; |
960 |
|
|
961 |
|
} else { |
962 |
|
uint32_t sadC = sad16(current->u + 2*offset, |
963 |
|
reference->u + 2*offset, stride, 256*4096); |
964 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
965 |
|
sadC += sad16(current->v + 2*offset, |
966 |
|
reference->v + 2*offset, stride, 256*4096); |
967 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
968 |
return 1; |
return 1; |
969 |
} |
} |
970 |
|
} |
971 |
|
|
972 |
static __inline void |
static __inline void |
973 |
SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
ZeroMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
974 |
{ |
{ |
975 |
pMB->mode = MODE_NOT_CODED; |
pMB->mode = MODE_INTER; |
976 |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV; |
977 |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; |
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV; |
978 |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
979 |
} |
} |
980 |
|
|
981 |
bool |
static __inline void |
982 |
MotionEstimation(MBParam * const pParam, |
ModeDecision(SearchData * const Data, |
983 |
FRAMEINFO * const current, |
MACROBLOCK * const pMB, |
984 |
FRAMEINFO * const reference, |
const MACROBLOCK * const pMBs, |
985 |
const IMAGE * const pRefH, |
const int x, const int y, |
986 |
const IMAGE * const pRefV, |
const MBParam * const pParam, |
987 |
const IMAGE * const pRefHV, |
const uint32_t MotionFlags, |
988 |
const uint32_t iLimit) |
const uint32_t VopFlags, |
989 |
{ |
const uint32_t VolFlags, |
990 |
MACROBLOCK *const pMBs = current->mbs; |
const IMAGE * const pCurrent, |
991 |
const IMAGE *const pCurrent = ¤t->image; |
const IMAGE * const pRef, |
992 |
const IMAGE *const pRef = &reference->image; |
const IMAGE * const vGMC, |
993 |
|
const int coding_type) |
994 |
const VECTOR zeroMV = { 0, 0 }; |
{ |
995 |
|
int mode = MODE_INTER; |
996 |
uint32_t x, y; |
int mcsel = 0; |
997 |
uint32_t iIntra = 0; |
int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0); |
998 |
int32_t InterBias; |
const uint32_t iQuant = pMB->quant; |
999 |
|
|
1000 |
// some pre-initialized thingies for SearchP |
const int skip_possible = (coding_type == P_VOP) && (pMB->dquant == 0); |
1001 |
int32_t temp[5]; |
|
1002 |
VECTOR currentMV[5]; |
pMB->mcsel = 0; |
1003 |
int32_t iMinSAD[5]; |
|
1004 |
SearchData Data; |
if (!(VopFlags & XVID_VOP_MODEDECISION_RD)) { /* normal, fast, SAD-based mode decision */ |
1005 |
Data.iEdgedWidth = pParam->edged_width; |
int sad; |
1006 |
Data.currentMV = currentMV; |
int InterBias = MV16_INTER_BIAS; |
1007 |
Data.iMinSAD = iMinSAD; |
if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
1008 |
Data.temp = temp; |
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) { |
1009 |
Data.iFcode = current->fcode; |
mode = MODE_INTER; |
1010 |
|
sad = Data->iMinSAD[0]; |
1011 |
|
} else { |
1012 |
|
mode = MODE_INTER4V; |
1013 |
|
sad = Data->iMinSAD[1] + Data->iMinSAD[2] + |
1014 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant; |
1015 |
|
Data->iMinSAD[0] = sad; |
1016 |
|
} |
1017 |
|
|
1018 |
if (sadInit) (*sadInit) (); |
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
1019 |
|
if (skip_possible && (pMB->sad16 < (int)iQuant * MAX_SAD00_FOR_SKIP)) |
1020 |
|
if ( (100*sad)/(pMB->sad16+1) > FINAL_SKIP_THRESH) |
1021 |
|
if (Data->chroma || SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant, Data->rrv)) { |
1022 |
|
mode = MODE_NOT_CODED; |
1023 |
|
sad = 0; |
1024 |
|
} |
1025 |
|
|
1026 |
for (y = 0; y < pParam->mb_height; y++) { |
/* mcsel */ |
1027 |
for (x = 0; x < pParam->mb_width; x++) { |
if (coding_type == S_VOP) { |
1028 |
|
|
1029 |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
int32_t iSAD = sad16(Data->Cur, |
1030 |
int32_t sad00 = pMB->sad16 |
vGMC->y + 16*y*Data->iEdgedWidth + 16*x, Data->iEdgedWidth, 65536); |
|
= sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, |
|
|
pRef->y + (x + y * pParam->edged_width) * 16, |
|
|
pParam->edged_width, pMB->sad8 ); |
|
1031 |
|
|
1032 |
if (!(current->global_flags & XVID_LUMIMASKING)) { |
if (Data->chroma) { |
1033 |
pMB->dquant = NO_CHANGE; |
iSAD += sad8(Data->CurU, vGMC->u + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); |
1034 |
pMB->quant = current->quant; } |
iSAD += sad8(Data->CurV, vGMC->v + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); |
|
|
|
|
//initial skip decision |
|
|
|
|
|
if ((pMB->dquant == NO_CHANGE) && (sad00 <= MAX_SAD00_FOR_SKIP * pMB->quant) |
|
|
&& (SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) ) { |
|
|
if (pMB->sad16 < pMB->quant * INITIAL_SKIP_THRESH) { |
|
|
SkipMacroblockP(pMB, sad00); |
|
|
continue; |
|
1035 |
} |
} |
|
} else sad00 = 256*4096; // skip not allowed - for final skip decision |
|
1036 |
|
|
1037 |
SearchP(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
if (iSAD <= sad) { /* mode decision GMC */ |
1038 |
y, current->motion_flags, pMB->quant, |
mode = MODE_INTER; |
1039 |
&Data, pParam, pMBs, reference->mbs, |
mcsel = 1; |
1040 |
current->global_flags & XVID_INTER4V, pMB); |
sad = iSAD; |
1041 |
|
} |
1042 |
|
|
1043 |
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
} |
|
if (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
|
|
if ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) |
|
|
{ SkipMacroblockP(pMB, sad00); continue; } |
|
1044 |
|
|
1045 |
/* finally, intra decision */ |
/* intra decision */ |
1046 |
|
|
1047 |
InterBias = MV16_INTER_BIAS; |
if (iQuant > 8) InterBias += 100 * (iQuant - 8); /* to make high quants work */ |
|
if (pMB->quant > 8) InterBias += 50 * (pMB->quant - 8); // to make high quants work |
|
1048 |
if (y != 0) |
if (y != 0) |
1049 |
if ((pMB - pParam->mb_width)->mode == MODE_INTER ) InterBias -= 50; |
if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; |
1050 |
if (x != 0) |
if (x != 0) |
1051 |
if ((pMB - 1)->mode == MODE_INTER ) InterBias -= 50; |
if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; |
1052 |
|
|
1053 |
if (InterBias < pMB->sad16) { |
if (Data->chroma) InterBias += 50; /* dev8(chroma) ??? <-- yes, we need dev8 (no big difference though) */ |
1054 |
const int32_t deviation = |
if (Data->rrv) InterBias *= 4; |
|
dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
|
|
pParam->edged_width); |
|
1055 |
|
|
1056 |
if (deviation < (pMB->sad16 - InterBias)) { |
if (InterBias < sad) { |
1057 |
if (++iIntra >= iLimit) return 1; |
int32_t deviation; |
1058 |
pMB->mode = MODE_INTRA; |
if (!Data->rrv) |
1059 |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = |
deviation = dev16(Data->Cur, Data->iEdgedWidth); |
1060 |
pMB->mvs[3] = zeroMV; |
else |
1061 |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = |
deviation = dev16(Data->Cur, Data->iEdgedWidth) + /* dev32() */ |
1062 |
pMB->sad8[3] = 0; |
dev16(Data->Cur+16, Data->iEdgedWidth) + |
1063 |
} |
dev16(Data->Cur + 16*Data->iEdgedWidth, Data->iEdgedWidth) + |
1064 |
} |
dev16(Data->Cur+16+16*Data->iEdgedWidth, Data->iEdgedWidth); |
1065 |
} |
|
1066 |
} |
if (deviation < (sad - InterBias)) mode = MODE_INTRA; |
|
return 0; |
|
1067 |
} |
} |
1068 |
|
|
1069 |
|
pMB->cbp = 63; |
1070 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
1071 |
|
|
1072 |
|
} else { /* Rate-Distortion */ |
1073 |
|
|
1074 |
|
int min_rd, intra_rd, i, cbp, c[2] = {0, 0}; |
1075 |
|
VECTOR backup[5], *v; |
1076 |
|
Data->iQuant = iQuant; |
1077 |
|
Data->cbp = c; |
1078 |
|
|
1079 |
|
v = Data->qpel ? Data->currentQMV : Data->currentMV; |
1080 |
|
for (i = 0; i < 5; i++) { |
1081 |
|
Data->iMinSAD[i] = 256*4096; |
1082 |
|
backup[i] = v[i]; |
1083 |
|
} |
1084 |
|
|
1085 |
|
min_rd = findRDinter(Data, pMBs, x, y, pParam, MotionFlags); |
1086 |
|
cbp = *Data->cbp; |
1087 |
|
|
1088 |
|
if (coding_type == S_VOP) { |
1089 |
|
int gmc_rd; |
1090 |
|
*Data->iMinSAD = min_rd += BITS_MULT*1; /* mcsel */ |
1091 |
|
gmc_rd = findRDgmc(Data, vGMC, x, y); |
1092 |
|
if (gmc_rd < min_rd) { |
1093 |
|
mcsel = 1; |
1094 |
|
*Data->iMinSAD = min_rd = gmc_rd; |
1095 |
|
mode = MODE_INTER; |
1096 |
|
cbp = *Data->cbp; |
1097 |
|
} |
1098 |
|
} |
1099 |
|
|
1100 |
|
if (inter4v) { |
1101 |
|
int v4_rd; |
1102 |
|
v4_rd = findRDinter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup); |
1103 |
|
if (v4_rd < min_rd) { |
1104 |
|
Data->iMinSAD[0] = min_rd = v4_rd; |
1105 |
|
mode = MODE_INTER4V; |
1106 |
|
cbp = *Data->cbp; |
1107 |
|
} |
1108 |
|
} |
1109 |
|
|
1110 |
|
intra_rd = findRDintra(Data); |
1111 |
|
if (intra_rd < min_rd) { |
1112 |
|
*Data->iMinSAD = min_rd = intra_rd; |
1113 |
|
mode = MODE_INTRA; |
1114 |
|
} |
1115 |
|
|
1116 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = 0; |
1117 |
|
pMB->cbp = cbp; |
1118 |
|
} |
1119 |
|
|
1120 |
|
if (Data->rrv) { |
1121 |
|
Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); |
1122 |
|
Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); |
1123 |
|
} |
1124 |
|
|
1125 |
|
if (mode == MODE_INTER && mcsel == 0) { |
1126 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
1127 |
|
|
1128 |
|
if(Data->qpel) { |
1129 |
|
pMB->qmvs[0] = pMB->qmvs[1] |
1130 |
|
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
1131 |
|
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; |
1132 |
|
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; |
1133 |
|
} else { |
1134 |
|
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
1135 |
|
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
1136 |
|
} |
1137 |
|
|
1138 |
|
} else if (mode == MODE_INTER ) { // but mcsel == 1 |
1139 |
|
|
1140 |
|
pMB->mcsel = 1; |
1141 |
|
if (Data->qpel) { |
1142 |
|
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = pMB->amv; |
1143 |
|
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = pMB->amv.x/2; |
1144 |
|
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = pMB->amv.y/2; |
1145 |
|
} else |
1146 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->amv; |
1147 |
|
|
1148 |
|
} else |
1149 |
|
if (mode == MODE_INTER4V) ; /* anything here? */ |
1150 |
|
else /* INTRA, NOT_CODED */ |
1151 |
|
ZeroMacroblockP(pMB, 0); |
1152 |
|
|
1153 |
|
pMB->mode = mode; |
1154 |
|
} |
1155 |
|
|
1156 |
|
bool |
1157 |
|
MotionEstimation(MBParam * const pParam, |
1158 |
|
FRAMEINFO * const current, |
1159 |
|
FRAMEINFO * const reference, |
1160 |
|
const IMAGE * const pRefH, |
1161 |
|
const IMAGE * const pRefV, |
1162 |
|
const IMAGE * const pRefHV, |
1163 |
|
const IMAGE * const pGMC, |
1164 |
|
const uint32_t iLimit) |
1165 |
|
{ |
1166 |
|
MACROBLOCK *const pMBs = current->mbs; |
1167 |
|
const IMAGE *const pCurrent = ¤t->image; |
1168 |
|
const IMAGE *const pRef = &reference->image; |
1169 |
|
|
1170 |
|
uint32_t mb_width = pParam->mb_width; |
1171 |
|
uint32_t mb_height = pParam->mb_height; |
1172 |
|
const uint32_t iEdgedWidth = pParam->edged_width; |
1173 |
|
const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->vop_flags, current->vol_flags); |
1174 |
|
|
1175 |
|
uint32_t x, y; |
1176 |
|
uint32_t iIntra = 0; |
1177 |
|
int32_t sad00; |
1178 |
|
int skip_thresh = INITIAL_SKIP_THRESH * \ |
1179 |
|
(current->vop_flags & XVID_VOP_REDUCED ? 4:1) * \ |
1180 |
|
(current->vop_flags & XVID_VOP_MODEDECISION_RD ? 2:1); |
1181 |
|
|
1182 |
|
/* some pre-initialized thingies for SearchP */ |
1183 |
|
int32_t temp[8]; |
1184 |
|
VECTOR currentMV[5]; |
1185 |
|
VECTOR currentQMV[5]; |
1186 |
|
int32_t iMinSAD[5]; |
1187 |
|
DECLARE_ALIGNED_MATRIX(dct_space, 3, 64, int16_t, CACHE_LINE); |
1188 |
|
SearchData Data; |
1189 |
|
memset(&Data, 0, sizeof(SearchData)); |
1190 |
|
Data.iEdgedWidth = iEdgedWidth; |
1191 |
|
Data.currentMV = currentMV; |
1192 |
|
Data.currentQMV = currentQMV; |
1193 |
|
Data.iMinSAD = iMinSAD; |
1194 |
|
Data.temp = temp; |
1195 |
|
Data.iFcode = current->fcode; |
1196 |
|
Data.rounding = pParam->m_rounding_type; |
1197 |
|
Data.qpel = (current->vol_flags & XVID_VOL_QUARTERPEL ? 1:0); |
1198 |
|
Data.chroma = MotionFlags & XVID_ME_CHROMA_PVOP; |
1199 |
|
Data.rrv = (current->vop_flags & XVID_VOP_REDUCED) ? 1:0; |
1200 |
|
Data.dctSpace = dct_space; |
1201 |
|
Data.quant_type = !(pParam->vol_flags & XVID_VOL_MPEGQUANT); |
1202 |
|
|
1203 |
|
if ((current->vop_flags & XVID_VOP_REDUCED)) { |
1204 |
|
mb_width = (pParam->width + 31) / 32; |
1205 |
|
mb_height = (pParam->height + 31) / 32; |
1206 |
|
Data.qpel = 0; |
1207 |
|
} |
1208 |
|
|
1209 |
|
Data.RefQ = pRefV->u; /* a good place, also used in MC (for similar purpose) */ |
1210 |
|
if (sadInit) (*sadInit) (); |
1211 |
|
|
1212 |
|
for (y = 0; y < mb_height; y++) { |
1213 |
|
for (x = 0; x < mb_width; x++) { |
1214 |
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
1215 |
|
|
1216 |
|
if (!Data.rrv) pMB->sad16 = |
1217 |
|
sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16, |
1218 |
|
pRef->y + (x + y * iEdgedWidth) * 16, |
1219 |
|
pParam->edged_width, pMB->sad8 ); |
1220 |
|
|
1221 |
|
else pMB->sad16 = |
1222 |
|
sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32, |
1223 |
|
pRef->y + (x + y * iEdgedWidth) * 32, |
1224 |
|
pParam->edged_width, pMB->sad8 ); |
1225 |
|
|
1226 |
|
if (Data.chroma) { |
1227 |
|
Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8, |
1228 |
|
pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2) |
1229 |
|
+ sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8, |
1230 |
|
pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); |
1231 |
|
pMB->sad16 += Data.temp[7]; |
1232 |
|
} |
1233 |
|
|
1234 |
|
sad00 = pMB->sad16; |
1235 |
|
|
1236 |
|
/* initial skip decision */ |
1237 |
|
/* no early skip for GMC (global vector = skip vector is unknown!) */ |
1238 |
|
if (current->coding_type != S_VOP) { /* no fast SKIP for S(GMC)-VOPs */ |
1239 |
|
if (pMB->dquant == 0 && sad00 < pMB->quant * skip_thresh) |
1240 |
|
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) { |
1241 |
|
ZeroMacroblockP(pMB, sad00); |
1242 |
|
pMB->mode = MODE_NOT_CODED; |
1243 |
|
continue; |
1244 |
|
} |
1245 |
|
} |
1246 |
|
|
1247 |
|
if ((current->vop_flags & XVID_VOP_CARTOON) && |
1248 |
|
(sad00 < pMB->quant * 4 * skip_thresh)) { /* favorize (0,0) vector for cartoons */ |
1249 |
|
ZeroMacroblockP(pMB, sad00); |
1250 |
|
continue; |
1251 |
|
} |
1252 |
|
|
1253 |
|
SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
1254 |
|
y, MotionFlags, current->vop_flags, current->vol_flags, |
1255 |
|
&Data, pParam, pMBs, reference->mbs, pMB); |
1256 |
|
|
1257 |
|
ModeDecision(&Data, pMB, pMBs, x, y, pParam, |
1258 |
|
MotionFlags, current->vop_flags, current->vol_flags, |
1259 |
|
pCurrent, pRef, pGMC, current->coding_type); |
1260 |
|
|
1261 |
|
if (pMB->mode == MODE_INTRA) |
1262 |
|
if (++iIntra > iLimit) return 1; |
1263 |
|
} |
1264 |
|
} |
1265 |
|
|
1266 |
|
return 0; |
1267 |
|
} |
1268 |
|
|
|
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
|
1269 |
|
|
1270 |
static __inline int |
static __inline int |
1271 |
make_mask(const VECTOR * const pmv, const int i) |
make_mask(const VECTOR * const pmv, const int i) |
1272 |
{ |
{ |
1273 |
int mask = 255, j; |
int mask = 255, j; |
1274 |
for (j = 0; j < i; j++) { |
for (j = 0; j < i; j++) { |
1275 |
if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already |
if (MVequal(pmv[i], pmv[j])) return 0; /* same vector has been checked already */ |
1276 |
if (pmv[i].x == pmv[j].x) { |
if (pmv[i].x == pmv[j].x) { |
1277 |
if (pmv[i].y == pmv[j].y + iDiamondSize) { mask &= ~4; continue; } |
if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4; |
1278 |
if (pmv[i].y == pmv[j].y - iDiamondSize) { mask &= ~8; continue; } |
else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8; |
1279 |
} else |
} else |
1280 |
if (pmv[i].y == pmv[j].y) { |
if (pmv[i].y == pmv[j].y) { |
1281 |
if (pmv[i].x == pmv[j].x + iDiamondSize) { mask &= ~1; continue; } |
if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1; |
1282 |
if (pmv[i].x == pmv[j].x - iDiamondSize) { mask &= ~2; continue; } |
else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2; |
1283 |
} |
} |
1284 |
} |
} |
1285 |
return mask; |
return mask; |
1286 |
} |
} |
1287 |
|
|
1288 |
static __inline void |
static __inline void |
1289 |
PreparePredictionsP(VECTOR * const pmv, int x, int y, const int iWcount, |
PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, |
1290 |
const int iHcount, const MACROBLOCK * const prevMB) |
int iHcount, const MACROBLOCK * const prevMB, int rrv) |
1291 |
{ |
{ |
1292 |
|
/* this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself */ |
1293 |
|
if (rrv) { iWcount /= 2; iHcount /= 2; } |
1294 |
|
|
1295 |
//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself |
if ( (y != 0) && (x < (iWcount-1)) ) { /* [5] top-right neighbour */ |
|
|
|
|
if ( (y != 0) && (x != (iWcount-1)) ) { // [5] top-right neighbour |
|
1296 |
pmv[5].x = EVEN(pmv[3].x); |
pmv[5].x = EVEN(pmv[3].x); |
1297 |
pmv[5].y = EVEN(pmv[3].y); |
pmv[5].y = EVEN(pmv[3].y); |
1298 |
} else pmv[5].x = pmv[5].y = 0; |
} else pmv[5].x = pmv[5].y = 0; |
1299 |
|
|
1300 |
if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour |
if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }/* pmv[3] is left neighbour */ |
1301 |
else pmv[3].x = pmv[3].y = 0; |
else pmv[3].x = pmv[3].y = 0; |
1302 |
|
|
1303 |
if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour |
if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }/* [4] top neighbour */ |
1304 |
else pmv[4].x = pmv[4].y = 0; |
else pmv[4].x = pmv[4].y = 0; |
1305 |
|
|
1306 |
// [1] median prediction |
/* [1] median prediction */ |
1307 |
pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); |
pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); |
1308 |
|
|
1309 |
pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask |
pmv[0].x = pmv[0].y = 0; /* [0] is zero; not used in the loop (checked before) but needed here for make_mask */ |
1310 |
|
|
1311 |
pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
pmv[2].x = EVEN(prevMB->mvs[0].x); /* [2] is last frame */ |
1312 |
pmv[2].y = EVEN(prevMB->mvs[0].y); |
pmv[2].y = EVEN(prevMB->mvs[0].y); |
1313 |
|
|
1314 |
if ((x != iWcount-1) && (y != iHcount-1)) { |
if ((x < iWcount-1) && (y < iHcount-1)) { |
1315 |
pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame |
pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); /* [6] right-down neighbour in last frame */ |
1316 |
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
1317 |
} else pmv[6].x = pmv[6].y = 0; |
} else pmv[6].x = pmv[6].y = 0; |
1318 |
|
|
1319 |
|
if (rrv) { |
1320 |
|
int i; |
1321 |
|
for (i = 0; i < 7; i++) { |
1322 |
|
pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); |
1323 |
|
pmv[i].y = RRV_MV_SCALEUP(pmv[i].y); |
1324 |
|
} |
1325 |
|
} |
1326 |
} |
} |
1327 |
|
|
1328 |
static void |
static void |
1329 |
SearchP(const uint8_t * const pRef, |
SearchP(const IMAGE * const pRef, |
1330 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1331 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
1332 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
1334 |
const int x, |
const int x, |
1335 |
const int y, |
const int y, |
1336 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
1337 |
const uint32_t iQuant, |
const uint32_t VopFlags, |
1338 |
|
const uint32_t VolFlags, |
1339 |
SearchData * const Data, |
SearchData * const Data, |
1340 |
const MBParam * const pParam, |
const MBParam * const pParam, |
1341 |
const MACROBLOCK * const pMBs, |
const MACROBLOCK * const pMBs, |
1342 |
const MACROBLOCK * const prevMBs, |
const MACROBLOCK * const prevMBs, |
|
int inter4v, |
|
1343 |
MACROBLOCK * const pMB) |
MACROBLOCK * const pMB) |
1344 |
{ |
{ |
1345 |
|
|
1346 |
int i, iDirection = 255, mask, threshA; |
int i, iDirection = 255, mask, threshA; |
1347 |
VECTOR pmv[7]; |
VECTOR pmv[7]; |
1348 |
|
int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0); |
1349 |
|
|
|
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); //has to be changed to get_pmv(2)() |
|
1350 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1351 |
pParam->width, pParam->height, Data->iFcode); |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
|
|
|
|
Data->predMV = pmv[0]; |
|
|
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16; |
|
|
Data->Ref = pRef + (x + Data->iEdgedWidth*y)*16; |
|
|
Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16; |
|
|
|
|
|
Data->iQuant = iQuant; |
|
|
|
|
|
if (!(MotionFlags & PMV_HALFPEL16)) { |
|
|
Data->min_dx = EVEN(Data->min_dx); |
|
|
Data->max_dx = EVEN(Data->max_dx); |
|
|
Data->min_dy = EVEN(Data->min_dy); |
|
|
Data->max_dy = EVEN(Data->max_dy); } |
|
|
|
|
|
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
|
|
|
|
|
if (inter4v) CheckCandidate = CheckCandidate16; |
|
|
else CheckCandidate = CheckCandidate16no4v; |
|
1352 |
|
|
1353 |
for(i = 0; i < 5; i++) Data->currentMV[i].x = Data->currentMV[i].y = 0; |
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); |
1354 |
|
|
1355 |
i = d_mv_bits(Data->predMV.x, Data->predMV.y, Data->iFcode); |
Data->temp[5] = Data->temp[6] = 0; /* chroma-sad cache */ |
1356 |
Data->iMinSAD[0] = pMB->sad16 + lambda_vec16[iQuant] * i; |
i = Data->rrv ? 2 : 1; |
1357 |
Data->iMinSAD[1] = pMB->sad8[0] + lambda_vec8[iQuant] * i; |
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i; |
1358 |
|
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1359 |
|
Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1360 |
|
|
1361 |
|
Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16*i; |
1362 |
|
Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16*i; |
1363 |
|
Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16*i; |
1364 |
|
Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16*i; |
1365 |
|
Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1366 |
|
Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1367 |
|
|
1368 |
|
Data->lambda16 = lambda_vec16[pMB->quant]; |
1369 |
|
Data->lambda8 = lambda_vec8[pMB->quant]; |
1370 |
|
Data->qpel_precision = 0; |
1371 |
|
|
1372 |
|
memset(Data->currentMV, 0, 5*sizeof(VECTOR)); |
1373 |
|
|
1374 |
|
if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
1375 |
|
else Data->predMV = pmv[0]; |
1376 |
|
|
1377 |
|
i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0); |
1378 |
|
Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10); |
1379 |
|
Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10); |
1380 |
Data->iMinSAD[2] = pMB->sad8[1]; |
Data->iMinSAD[2] = pMB->sad8[1]; |
1381 |
Data->iMinSAD[3] = pMB->sad8[2]; |
Data->iMinSAD[3] = pMB->sad8[2]; |
1382 |
Data->iMinSAD[4] = pMB->sad8[3]; |
Data->iMinSAD[4] = pMB->sad8[3]; |
1383 |
|
|
1384 |
if ((x == 0) && (y == 0)) threshA = 512; |
if ((!(VopFlags & XVID_VOP_MODEDECISION_RD)) && (x | y)) { |
1385 |
else { |
threshA = Data->temp[0]; /* that's where we keep this SAD atm */ |
|
threshA = Data->temp[0]; // that's when we keep this SAD atm |
|
1386 |
if (threshA < 512) threshA = 512; |
if (threshA < 512) threshA = 512; |
1387 |
if (threshA > 1024) threshA = 1024; } |
else if (threshA > 1024) threshA = 1024; |
1388 |
|
} else |
1389 |
|
threshA = 512; |
1390 |
|
|
1391 |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
1392 |
prevMBs + x + y * pParam->mb_width); |
prevMBs + x + y * pParam->mb_width, Data->rrv); |
|
|
|
|
if (inter4v) CheckCandidate = CheckCandidate16; |
|
|
else CheckCandidate = CheckCandidate16no4v; |
|
1393 |
|
|
1394 |
|
if (!Data->rrv) { |
1395 |
|
if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16; |
1396 |
|
else CheckCandidate = CheckCandidate16no4v; /* for extra speed */ |
1397 |
|
} else CheckCandidate = CheckCandidate32; |
1398 |
|
|
1399 |
/* main loop. checking all predictions */ |
/* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/ |
1400 |
|
|
1401 |
for (i = 1; i < 7; i++) { |
for (i = 1; i < 7; i++) { |
1402 |
if (!(mask = make_mask(pmv, i)) ) continue; |
if (!(mask = make_mask(pmv, i)) ) continue; |
1403 |
(*CheckCandidate)(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1404 |
if (Data->iMinSAD[0] <= threshA) break; |
if (Data->iMinSAD[0] <= threshA) break; |
1405 |
} |
} |
1406 |
|
|
1407 |
if ((Data->iMinSAD[0] <= threshA) || |
if ((Data->iMinSAD[0] <= threshA) || |
1408 |
(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
1409 |
(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { |
(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) |
1410 |
inter4v = 0; |
inter4v = 0; |
1411 |
} else { |
else { |
1412 |
|
|
1413 |
MainSearchFunc * MainSearchPtr; |
MainSearchFunc * MainSearchPtr; |
1414 |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch; |
1415 |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1416 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1417 |
|
|
1418 |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
1419 |
|
|
1420 |
/* extended search, diamond starting in 0,0 and in prediction. |
/* extended search, diamond starting in 0,0 and in prediction. |
1421 |
note that this search is/might be done in halfpel positions, |
note that this search is/might be done in halfpel positions, |
1422 |
which makes it more different than the diamond above */ |
which makes it more different than the diamond above */ |
1423 |
|
|
1424 |
if (MotionFlags & PMV_EXTSEARCH16) { |
if (MotionFlags & XVID_ME_EXTSEARCH16) { |
1425 |
int32_t bSAD; |
int32_t bSAD; |
1426 |
VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; |
VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; |
1427 |
if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? |
if (Data->rrv) { |
1428 |
startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); |
startMV.x = RRV_MV_SCALEUP(startMV.x); |
1429 |
|
startMV.y = RRV_MV_SCALEUP(startMV.y); |
1430 |
|
} |
1431 |
if (!(MVequal(startMV, backupMV))) { |
if (!(MVequal(startMV, backupMV))) { |
1432 |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1433 |
|
|
1434 |
CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, Data); |
CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data); |
1435 |
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
MainSearchPtr(startMV.x, startMV.y, Data, 255); |
1436 |
if (bSAD < Data->iMinSAD[0]) { |
if (bSAD < Data->iMinSAD[0]) { |
1437 |
Data->currentMV[0] = backupMV; |
Data->currentMV[0] = backupMV; |
1438 |
Data->iMinSAD[0] = bSAD; } |
Data->iMinSAD[0] = bSAD; } |
1439 |
} |
} |
1440 |
|
|
1441 |
backupMV = Data->currentMV[0]; |
backupMV = Data->currentMV[0]; |
1442 |
if (MotionFlags & PMV_HALFPELREFINE16) startMV.x = startMV.y = 1; |
startMV.x = startMV.y = 1; |
|
else startMV.x = startMV.y = 0; |
|
1443 |
if (!(MVequal(startMV, backupMV))) { |
if (!(MVequal(startMV, backupMV))) { |
1444 |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1445 |
|
|
1446 |
CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, Data); |
CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data); |
1447 |
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
MainSearchPtr(startMV.x, startMV.y, Data, 255); |
1448 |
if (bSAD < Data->iMinSAD[0]) { |
if (bSAD < Data->iMinSAD[0]) { |
1449 |
Data->currentMV[0] = backupMV; |
Data->currentMV[0] = backupMV; |
1450 |
Data->iMinSAD[0] = bSAD; } |
Data->iMinSAD[0] = bSAD; } |
1452 |
} |
} |
1453 |
} |
} |
1454 |
|
|
1455 |
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data); |
if (MotionFlags & XVID_ME_HALFPELREFINE16) |
1456 |
|
SubpelRefine(Data); |
1457 |
|
|
1458 |
|
for(i = 0; i < 5; i++) { |
1459 |
|
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* initialize qpel vectors */ |
1460 |
|
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
1461 |
|
} |
1462 |
|
|
1463 |
|
if (Data->qpel) { |
1464 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1465 |
|
pParam->width, pParam->height, Data->iFcode, 1, 0); |
1466 |
|
Data->qpel_precision = 1; |
1467 |
|
if (MotionFlags & XVID_ME_QUARTERPELREFINE16) |
1468 |
|
SubpelRefine(Data); |
1469 |
|
} |
1470 |
|
|
1471 |
|
if (Data->iMinSAD[0] < (int32_t)pMB->quant * 30) |
1472 |
|
inter4v = 0; |
1473 |
|
|
1474 |
if (inter4v) { |
if (inter4v) { |
1475 |
SearchData Data8; |
SearchData Data8; |
1476 |
Data8.iFcode = Data->iFcode; |
memcpy(&Data8, Data, sizeof(SearchData)); /* quick copy of common data */ |
1477 |
Data8.iQuant = Data->iQuant; |
|
|
Data8.iEdgedWidth = Data->iEdgedWidth; |
|
1478 |
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
1479 |
Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); |
Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); |
1480 |
Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); |
Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); |
1481 |
Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); |
Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); |
|
} |
|
|
|
|
|
if (!(inter4v) || |
|
|
(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
|
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
|
|
// INTER MODE |
|
|
pMB->mode = MODE_INTER; |
|
|
pMB->mvs[0] = pMB->mvs[1] |
|
|
= pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
|
1482 |
|
|
1483 |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
if ((Data->chroma) && (!(VopFlags & XVID_VOP_MODEDECISION_RD))) { |
1484 |
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
/* chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, it will not be used */ |
1485 |
|
int sumx = 0, sumy = 0; |
1486 |
|
|
1487 |
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
if (Data->qpel) |
1488 |
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
for (i = 1; i < 5; i++) { |
1489 |
} else { |
sumx += Data->currentQMV[i].x/2; |
1490 |
// INTER4V MODE; all other things are already set in Search8 |
sumy += Data->currentQMV[i].y/2; |
1491 |
pMB->mode = MODE_INTER4V; |
} |
1492 |
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + |
else |
1493 |
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant; |
for (i = 1; i < 5; i++) { |
1494 |
|
sumx += Data->currentMV[i].x; |
1495 |
|
sumy += Data->currentMV[i].y; |
1496 |
} |
} |
1497 |
|
|
1498 |
|
Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf], |
1499 |
|
(sumy >> 3) + roundtab_76[sumy & 0xf], Data); |
1500 |
|
} |
1501 |
|
} else Data->iMinSAD[1] = 4096*256; |
1502 |
} |
} |
1503 |
|
|
1504 |
static void |
static void |
1511 |
const int block, |
const int block, |
1512 |
SearchData * const Data) |
SearchData * const Data) |
1513 |
{ |
{ |
1514 |
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
int i = 0; |
1515 |
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
1516 |
Data->currentMV = OldData->currentMV + 1 + block; |
Data->currentMV = OldData->currentMV + 1 + block; |
1517 |
|
Data->currentQMV = OldData->currentQMV + 1 + block; |
1518 |
|
|
1519 |
|
if(Data->qpel) { |
1520 |
|
Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
1521 |
|
if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y, |
1522 |
|
Data->predMV, Data->iFcode, 0, 0); |
1523 |
|
} else { |
1524 |
|
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
1525 |
|
if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y, |
1526 |
|
Data->predMV, Data->iFcode, 0, Data->rrv); |
1527 |
|
} |
1528 |
|
|
1529 |
|
*(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10; |
1530 |
|
|
1531 |
|
if (MotionFlags & (XVID_ME_EXTSEARCH8|XVID_ME_HALFPELREFINE8|XVID_ME_QUARTERPELREFINE8)) { |
1532 |
|
|
1533 |
if (block != 0) |
if (Data->rrv) i = 16; else i = 8; |
|
*(Data->iMinSAD) += lambda_vec8[Data->iQuant] * |
|
|
d_mv_bits( Data->currentMV->x - Data->predMV.x, |
|
|
Data->currentMV->y - Data->predMV.y, |
|
|
Data->iFcode); |
|
|
|
|
|
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { |
|
|
|
|
|
Data->Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
|
Data->RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
|
Data->RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
|
Data->RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
1534 |
|
|
1535 |
Data->Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->RefP[0] = OldData->RefP[0] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1536 |
|
Data->RefP[1] = OldData->RefP[1] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1537 |
|
Data->RefP[2] = OldData->RefP[2] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1538 |
|
Data->RefP[3] = OldData->RefP[3] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1539 |
|
|
1540 |
|
Data->Cur = OldData->Cur + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1541 |
|
Data->qpel_precision = 0; |
1542 |
|
|
1543 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
1544 |
pParam->width, pParam->height, OldData->iFcode); |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
1545 |
|
|
1546 |
CheckCandidate = CheckCandidate8; |
if (!Data->rrv) CheckCandidate = CheckCandidate8; |
1547 |
|
else CheckCandidate = CheckCandidate16no4v; |
1548 |
|
|
1549 |
if (MotionFlags & PMV_EXTSEARCH8) { |
if (MotionFlags & XVID_ME_EXTSEARCH8 && (!(MotionFlags & XVID_ME_EXTSEARCH_RD))) { |
1550 |
|
int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */ |
1551 |
|
|
1552 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1553 |
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
if (MotionFlags & XVID_ME_USESQUARES8) MainSearchPtr = SquareSearch; |
1554 |
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
1555 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1556 |
|
|
1557 |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); } |
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1558 |
|
|
1559 |
|
if(*(Data->iMinSAD) < temp_sad) { |
1560 |
|
Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */ |
1561 |
|
Data->currentQMV->y = 2 * Data->currentMV->y; |
1562 |
|
} |
1563 |
|
} |
1564 |
|
|
1565 |
|
if (MotionFlags & XVID_ME_HALFPELREFINE8) { |
1566 |
|
int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */ |
1567 |
|
|
1568 |
|
SubpelRefine(Data); /* perform halfpel refine of current best vector */ |
1569 |
|
|
1570 |
|
if(*(Data->iMinSAD) < temp_sad) { /* we have found a better match */ |
1571 |
|
Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */ |
1572 |
|
Data->currentQMV->y = 2 * Data->currentMV->y; |
1573 |
|
} |
1574 |
|
} |
1575 |
|
|
1576 |
|
if (Data->qpel && MotionFlags & XVID_ME_QUARTERPELREFINE8) { |
1577 |
|
Data->qpel_precision = 1; |
1578 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
1579 |
|
pParam->width, pParam->height, Data->iFcode, 1, 0); |
1580 |
|
SubpelRefine(Data); |
1581 |
|
} |
1582 |
|
} |
1583 |
|
|
1584 |
if (MotionFlags & PMV_HALFPELREFINE8) HalfpelRefine(Data); |
if (Data->rrv) { |
1585 |
|
Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x); |
1586 |
|
Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y); |
1587 |
} |
} |
1588 |
|
|
1589 |
|
if(Data->qpel) { |
1590 |
|
pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x; |
1591 |
|
pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y; |
1592 |
|
pMB->qmvs[block] = *Data->currentQMV; |
1593 |
|
} else { |
1594 |
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
1595 |
pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; |
pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; |
1596 |
pMB->mvs[block] = *(Data->currentMV); |
} |
|
pMB->sad8[block] = 4 * (*Data->iMinSAD); |
|
1597 |
|
|
1598 |
|
pMB->mvs[block] = *Data->currentMV; |
1599 |
|
pMB->sad8[block] = 4 * *Data->iMinSAD; |
1600 |
} |
} |
1601 |
|
|
1602 |
/* B-frames code starts here */ |
/* motion estimation for B-frames */ |
1603 |
|
|
1604 |
static __inline VECTOR |
static __inline VECTOR |
1605 |
ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) |
ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) |
1606 |
{ |
{ |
1607 |
/* the stupidiest function ever */ |
/* the stupidiest function ever */ |
1608 |
if (mode == MODE_FORWARD) return pMB->mvs[0]; |
return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]); |
|
else return pMB->b_mvs[0]; |
|
1609 |
} |
} |
1610 |
|
|
1611 |
static void __inline |
static void __inline |
1615 |
const uint32_t mode_curr) |
const uint32_t mode_curr) |
1616 |
{ |
{ |
1617 |
|
|
1618 |
// [0] is prediction |
/* [0] is prediction */ |
1619 |
pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); |
pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); |
1620 |
|
|
1621 |
pmv[1].x = pmv[1].y = 0; // [1] is zero |
pmv[1].x = pmv[1].y = 0; /* [1] is zero */ |
1622 |
|
|
1623 |
pmv[2] = ChoosePred(pMB, mode_curr); |
pmv[2] = ChoosePred(pMB, mode_curr); |
1624 |
pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); |
pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); |
1625 |
|
|
1626 |
if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour |
if ((y != 0)&&(x != (int)(iWcount+1))) { /* [3] top-right neighbour */ |
1627 |
pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); |
pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); |
1628 |
pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); |
pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); |
1629 |
} else pmv[3].x = pmv[3].y = 0; |
} else pmv[3].x = pmv[3].y = 0; |
1638 |
pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
1639 |
} else pmv[5].x = pmv[5].y = 0; |
} else pmv[5].x = pmv[5].y = 0; |
1640 |
|
|
1641 |
if ((x != 0)&&(y != 0)) { |
if (x != 0 && y != 0) { |
1642 |
pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); |
pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); |
1643 |
pmv[6].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y); |
1644 |
} else pmv[6].x = pmv[6].y = 0; |
} else pmv[6].x = pmv[6].y = 0; |
|
|
|
|
// more? |
|
1645 |
} |
} |
1646 |
|
|
1647 |
|
|
1648 |
/* search backward or forward, for b-frames */ |
/* search backward or forward */ |
1649 |
static void |
static void |
1650 |
SearchBF( const uint8_t * const pRef, |
SearchBF( const IMAGE * const pRef, |
1651 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1652 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
1653 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
1663 |
SearchData * const Data) |
SearchData * const Data) |
1664 |
{ |
{ |
1665 |
|
|
1666 |
const int32_t iEdgedWidth = pParam->edged_width; |
int i, iDirection = 255, mask; |
|
|
|
|
int i, iDirection, mask; |
|
1667 |
VECTOR pmv[7]; |
VECTOR pmv[7]; |
1668 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1669 |
*Data->iMinSAD = MV_MAX_ERROR; |
*Data->iMinSAD = MV_MAX_ERROR; |
1670 |
Data->iFcode = iFcode; |
Data->iFcode = iFcode; |
1671 |
|
Data->qpel_precision = 0; |
1672 |
|
Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; /* reset chroma-sad cache */ |
1673 |
|
|
1674 |
Data->Ref = pRef + (x + y * iEdgedWidth) * 16; |
Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16; |
1675 |
Data->RefH = pRefH + (x + y * iEdgedWidth) * 16; |
Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16; |
1676 |
Data->RefV = pRefV + (x + y * iEdgedWidth) * 16; |
Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16; |
1677 |
Data->RefHV = pRefHV + (x + y * iEdgedWidth) * 16; |
Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16; |
1678 |
|
Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
1679 |
|
Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
1680 |
|
|
1681 |
Data->predMV = *predMV; |
Data->predMV = *predMV; |
1682 |
|
|
1683 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1684 |
pParam->width, pParam->height, iFcode); |
pParam->width, pParam->height, iFcode - Data->qpel, 0, 0); |
1685 |
|
|
1686 |
pmv[0] = Data->predMV; |
pmv[0] = Data->predMV; |
1687 |
|
if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } |
1688 |
|
|
1689 |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
1690 |
|
|
1691 |
Data->currentMV->x = Data->currentMV->y = 0; |
Data->currentMV->x = Data->currentMV->y = 0; |
|
|
|
1692 |
CheckCandidate = CheckCandidate16no4v; |
CheckCandidate = CheckCandidate16no4v; |
1693 |
|
|
1694 |
// main loop. checking all predictions |
/* main loop. checking all predictions */ |
1695 |
for (i = 0; i < 8; i++) { |
for (i = 0; i < 7; i++) { |
1696 |
if (!(mask = make_mask(pmv, i)) ) continue; |
if (!(mask = make_mask(pmv, i)) ) continue; |
1697 |
CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1698 |
} |
} |
1699 |
|
|
1700 |
if (MotionFlags & PMV_USESQUARES16) |
if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch; |
1701 |
MainSearchPtr = SquareSearch; |
else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
|
|
MainSearchPtr = AdvDiamondSearch; |
|
1702 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1703 |
|
|
1704 |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
1705 |
|
|
1706 |
|
SubpelRefine(Data); |
1707 |
|
|
1708 |
HalfpelRefine(Data); |
if (Data->qpel && *Data->iMinSAD < *best_sad + 300) { |
1709 |
|
Data->currentQMV->x = 2*Data->currentMV->x; |
1710 |
|
Data->currentQMV->y = 2*Data->currentMV->y; |
1711 |
|
Data->qpel_precision = 1; |
1712 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1713 |
|
pParam->width, pParam->height, iFcode, 1, 0); |
1714 |
|
SubpelRefine(Data); |
1715 |
|
} |
1716 |
|
|
1717 |
// three bits are needed to code backward mode. four for forward |
/* three bits are needed to code backward mode. four for forward */ |
|
// we treat the bits just like they were vector's |
|
|
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * lambda_vec16[Data->iQuant]; |
|
|
else *Data->iMinSAD += 3 * lambda_vec16[Data->iQuant]; |
|
1718 |
|
|
1719 |
|
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
1720 |
|
else *Data->iMinSAD += 3 * Data->lambda16; |
1721 |
|
|
1722 |
if (*Data->iMinSAD < *best_sad) { |
if (*Data->iMinSAD < *best_sad) { |
1723 |
*best_sad = *Data->iMinSAD; |
*best_sad = *Data->iMinSAD; |
1724 |
pMB->mode = mode_current; |
pMB->mode = mode_current; |
1725 |
|
if (Data->qpel) { |
1726 |
|
pMB->pmvs[0].x = Data->currentQMV->x - predMV->x; |
1727 |
|
pMB->pmvs[0].y = Data->currentQMV->y - predMV->y; |
1728 |
|
if (mode_current == MODE_FORWARD) |
1729 |
|
pMB->qmvs[0] = *Data->currentQMV; |
1730 |
|
else |
1731 |
|
pMB->b_qmvs[0] = *Data->currentQMV; |
1732 |
|
} else { |
1733 |
pMB->pmvs[0].x = Data->currentMV->x - predMV->x; |
pMB->pmvs[0].x = Data->currentMV->x - predMV->x; |
1734 |
pMB->pmvs[0].y = Data->currentMV->y - predMV->y; |
pMB->pmvs[0].y = Data->currentMV->y - predMV->y; |
1735 |
|
} |
1736 |
if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; |
if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; |
1737 |
else pMB->b_mvs[0] = *Data->currentMV; |
else pMB->b_mvs[0] = *Data->currentMV; |
1738 |
} |
} |
1739 |
|
|
1740 |
|
if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; |
1741 |
|
else *(Data->currentMV+1) = *Data->currentMV; /* we store currmv for interpolate search */ |
1742 |
} |
} |
1743 |
|
|
1744 |
static int32_t |
static void |
1745 |
|
SkipDecisionB(const IMAGE * const pCur, |
1746 |
|
const IMAGE * const f_Ref, |
1747 |
|
const IMAGE * const b_Ref, |
1748 |
|
MACROBLOCK * const pMB, |
1749 |
|
const uint32_t x, const uint32_t y, |
1750 |
|
const SearchData * const Data) |
1751 |
|
{ |
1752 |
|
int dx = 0, dy = 0, b_dx = 0, b_dy = 0; |
1753 |
|
int32_t sum; |
1754 |
|
const int div = 1 + Data->qpel; |
1755 |
|
int k; |
1756 |
|
const uint32_t stride = Data->iEdgedWidth/2; |
1757 |
|
/* this is not full chroma compensation, only it's fullpel approximation. should work though */ |
1758 |
|
|
1759 |
|
for (k = 0; k < 4; k++) { |
1760 |
|
dy += Data->directmvF[k].y / div; |
1761 |
|
dx += Data->directmvF[k].x / div; |
1762 |
|
b_dy += Data->directmvB[k].y / div; |
1763 |
|
b_dx += Data->directmvB[k].x / div; |
1764 |
|
} |
1765 |
|
|
1766 |
|
dy = (dy >> 3) + roundtab_76[dy & 0xf]; |
1767 |
|
dx = (dx >> 3) + roundtab_76[dx & 0xf]; |
1768 |
|
b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf]; |
1769 |
|
b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf]; |
1770 |
|
|
1771 |
|
sum = sad8bi(pCur->u + 8 * x + 8 * y * stride, |
1772 |
|
f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2, |
1773 |
|
b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, |
1774 |
|
stride); |
1775 |
|
|
1776 |
|
if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; /* no skip */ |
1777 |
|
|
1778 |
|
sum += sad8bi(pCur->v + 8*x + 8 * y * stride, |
1779 |
|
f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2, |
1780 |
|
b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, |
1781 |
|
stride); |
1782 |
|
|
1783 |
|
if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) { |
1784 |
|
pMB->mode = MODE_DIRECT_NONE_MV; /* skipped */ |
1785 |
|
for (k = 0; k < 4; k++) { |
1786 |
|
pMB->qmvs[k] = pMB->mvs[k]; |
1787 |
|
pMB->b_qmvs[k] = pMB->b_mvs[k]; |
1788 |
|
} |
1789 |
|
} |
1790 |
|
} |
1791 |
|
|
1792 |
|
static __inline uint32_t |
1793 |
SearchDirect(const IMAGE * const f_Ref, |
SearchDirect(const IMAGE * const f_Ref, |
1794 |
const uint8_t * const f_RefH, |
const uint8_t * const f_RefH, |
1795 |
const uint8_t * const f_RefV, |
const uint8_t * const f_RefV, |
1810 |
|
|
1811 |
{ |
{ |
1812 |
int32_t skip_sad; |
int32_t skip_sad; |
1813 |
int k; |
int k = (x + Data->iEdgedWidth*y) * 16; |
|
|
|
1814 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1815 |
|
|
1816 |
*Data->iMinSAD = 256*4096; |
*Data->iMinSAD = 256*4096; |
1817 |
Data->referencemv = b_mb->mvs; |
Data->RefP[0] = f_Ref->y + k; |
1818 |
|
Data->RefP[2] = f_RefH + k; |
1819 |
|
Data->RefP[1] = f_RefV + k; |
1820 |
|
Data->RefP[3] = f_RefHV + k; |
1821 |
|
Data->b_RefP[0] = b_Ref->y + k; |
1822 |
|
Data->b_RefP[2] = b_RefH + k; |
1823 |
|
Data->b_RefP[1] = b_RefV + k; |
1824 |
|
Data->b_RefP[3] = b_RefHV + k; |
1825 |
|
Data->RefP[4] = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; |
1826 |
|
Data->RefP[5] = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; |
1827 |
|
Data->b_RefP[4] = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; |
1828 |
|
Data->b_RefP[5] = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; |
1829 |
|
|
1830 |
|
k = Data->qpel ? 4 : 2; |
1831 |
|
Data->max_dx = k * (pParam->width - x * 16); |
1832 |
|
Data->max_dy = k * (pParam->height - y * 16); |
1833 |
|
Data->min_dx = -k * (16 + x * 16); |
1834 |
|
Data->min_dy = -k * (16 + y * 16); |
1835 |
|
|
1836 |
Data->Ref = f_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs; |
1837 |
Data->RefH = f_RefH + (x + Data->iEdgedWidth*y) * 16; |
Data->qpel_precision = 0; |
|
Data->RefV = f_RefV + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefHV = f_RefHV + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->bRef = b_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->bRefH = b_RefH + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->bRefV = b_RefV + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->bRefHV = b_RefHV + (x + Data->iEdgedWidth*y) * 16; |
|
|
|
|
|
Data->max_dx = 2 * pParam->width - 2 * (x) * 16; |
|
|
Data->max_dy = 2 * pParam->height - 2 * (y) * 16; |
|
|
Data->min_dx = -(2 * 16 + 2 * (x) * 16); |
|
|
Data->min_dy = -(2 * 16 + 2 * (y) * 16); |
|
1838 |
|
|
1839 |
for (k = 0; k < 4; k++) { |
for (k = 0; k < 4; k++) { |
1840 |
pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); |
pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); |
1842 |
pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); |
pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); |
1843 |
pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; |
pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; |
1844 |
|
|
1845 |
if ( ( pMB->b_mvs[k].x > Data->max_dx ) || ( pMB->b_mvs[k].x < Data->min_dx ) |
if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx) |
1846 |
|| ( pMB->b_mvs[k].y > Data->max_dy ) || ( pMB->b_mvs[k].y < Data->min_dy )) { |
| (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) { |
1847 |
|
|
1848 |
*best_sad = 256*4096; // in that case, we won't use direct mode |
*best_sad = 256*4096; /* in that case, we won't use direct mode */ |
1849 |
pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV" |
pMB->mode = MODE_DIRECT; /* just to make sure it doesn't say "MODE_DIRECT_NONE_MV" */ |
1850 |
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
1851 |
return 0; |
return 256*4096; |
1852 |
} |
} |
1853 |
if (b_mb->mode != MODE_INTER4V) { |
if (b_mb->mode != MODE_INTER4V) { |
1854 |
pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; |
pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; |
1859 |
} |
} |
1860 |
} |
} |
1861 |
|
|
1862 |
if (b_mb->mode == MODE_INTER4V) |
CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v; |
1863 |
CheckCandidate = CheckCandidateDirect; |
|
1864 |
else CheckCandidate = CheckCandidateDirectno4v; |
CheckCandidate(0, 0, 255, &k, Data); |
|
|
|
|
(*CheckCandidate)(0, 0, 255, &k, Data); |
|
|
|
|
|
// skip decision |
|
|
if (*Data->iMinSAD - 2 * lambda_vec16[Data->iQuant] < (int32_t)Data->iQuant * SKIP_THRESH_B) { |
|
|
//checking chroma. everything copied from MC |
|
|
//this is not full chroma compensation, only it's fullpel approximation. should work though |
|
|
int sum, dx, dy, b_dx, b_dy; |
|
|
|
|
|
sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
|
|
dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
|
|
|
|
sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
|
|
dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
|
|
|
|
sum = pMB->b_mvs[0].x + pMB->b_mvs[1].x + pMB->b_mvs[2].x + pMB->b_mvs[3].x; |
|
|
b_dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
|
|
|
|
sum = pMB->b_mvs[0].y + pMB->b_mvs[1].y + pMB->b_mvs[2].y + pMB->b_mvs[3].y; |
|
|
b_dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
|
|
|
|
sum = sad8bi(pCur->u + 8*x + 8*y*(Data->iEdgedWidth/2), |
|
|
f_Ref->u + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
|
|
b_Ref->u + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
|
|
Data->iEdgedWidth/2); |
|
|
sum += sad8bi(pCur->v + 8*x + 8*y*(Data->iEdgedWidth/2), |
|
|
f_Ref->v + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
|
|
b_Ref->v + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
|
|
Data->iEdgedWidth/2); |
|
1865 |
|
|
1866 |
if ((uint32_t) sum < MAX_CHROMA_SAD_FOR_SKIP * Data->iQuant) { |
/* initial (fast) skip decision */ |
1867 |
|
if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (Data->chroma?3:2)) { |
1868 |
|
/* possible skip */ |
1869 |
|
if (Data->chroma) { |
1870 |
pMB->mode = MODE_DIRECT_NONE_MV; |
pMB->mode = MODE_DIRECT_NONE_MV; |
1871 |
return *Data->iMinSAD; |
return *Data->iMinSAD; /* skip. */ |
1872 |
|
} else { |
1873 |
|
SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data); |
1874 |
|
if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; /* skip. */ |
1875 |
} |
} |
1876 |
} |
} |
1877 |
|
|
1878 |
|
*Data->iMinSAD += Data->lambda16; |
1879 |
skip_sad = *Data->iMinSAD; |
skip_sad = *Data->iMinSAD; |
1880 |
|
|
1881 |
// DIRECT MODE DELTA VECTOR SEARCH. |
/* |
1882 |
// This has to be made more effective, but at the moment I'm happy it's running at all |
* DIRECT MODE DELTA VECTOR SEARCH. |
1883 |
|
* This has to be made more effective, but at the moment I'm happy it's running at all |
1884 |
|
*/ |
1885 |
|
|
1886 |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch; |
1887 |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1888 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1889 |
|
|
1890 |
(*MainSearchPtr)(0, 0, Data, 255); |
MainSearchPtr(0, 0, Data, 255); |
1891 |
|
|
1892 |
HalfpelRefine(Data); |
SubpelRefine(Data); |
1893 |
|
|
|
*Data->iMinSAD += 1 * lambda_vec16[Data->iQuant]; // one bit is needed to code direct mode. we treat this bit just like it was vector's |
|
1894 |
*best_sad = *Data->iMinSAD; |
*best_sad = *Data->iMinSAD; |
1895 |
|
|
1896 |
if (b_mb->mode == MODE_INTER4V) |
if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT; |
1897 |
pMB->mode = MODE_DIRECT; |
else pMB->mode = MODE_DIRECT_NO4V; /* for faster compensation */ |
|
else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
|
1898 |
|
|
1899 |
pMB->pmvs[3] = *Data->currentMV; |
pMB->pmvs[3] = *Data->currentMV; |
1900 |
|
|
1907 |
pMB->b_mvs[k].y = ((Data->currentMV->y == 0) |
pMB->b_mvs[k].y = ((Data->currentMV->y == 0) |
1908 |
? Data->directmvB[k].y |
? Data->directmvB[k].y |
1909 |
: pMB->mvs[k].y - Data->referencemv[k].y); |
: pMB->mvs[k].y - Data->referencemv[k].y); |
1910 |
|
if (Data->qpel) { |
1911 |
|
pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2; |
1912 |
|
pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2; |
1913 |
|
pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2; |
1914 |
|
pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2; |
1915 |
|
} |
1916 |
|
|
1917 |
if (b_mb->mode != MODE_INTER4V) { |
if (b_mb->mode != MODE_INTER4V) { |
1918 |
pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; |
pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; |
1919 |
pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; |
pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; |
1920 |
|
pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0]; |
1921 |
|
pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0]; |
1922 |
break; |
break; |
1923 |
} |
} |
1924 |
} |
} |
1925 |
return skip_sad; |
return skip_sad; |
1926 |
} |
} |
1927 |
|
|
1928 |
|
static void |
1929 |
static __inline void |
SearchInterpolate(const IMAGE * const f_Ref, |
|
SearchInterpolate(const uint8_t * const f_Ref, |
|
1930 |
const uint8_t * const f_RefH, |
const uint8_t * const f_RefH, |
1931 |
const uint8_t * const f_RefV, |
const uint8_t * const f_RefV, |
1932 |
const uint8_t * const f_RefHV, |
const uint8_t * const f_RefHV, |
1933 |
const uint8_t * const b_Ref, |
const IMAGE * const b_Ref, |
1934 |
const uint8_t * const b_RefH, |
const uint8_t * const b_RefH, |
1935 |
const uint8_t * const b_RefV, |
const uint8_t * const b_RefV, |
1936 |
const uint8_t * const b_RefHV, |
const uint8_t * const b_RefHV, |
1948 |
|
|
1949 |
{ |
{ |
1950 |
|
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
|
|
|
1951 |
int iDirection, i, j; |
int iDirection, i, j; |
1952 |
SearchData bData; |
SearchData bData; |
1953 |
|
|
1954 |
bData.iMinSAD = fData->iMinSAD; |
fData->qpel_precision = 0; |
1955 |
*bData.iMinSAD = 4096*256; |
memcpy(&bData, fData, sizeof(SearchData)); /* quick copy of common data */ |
1956 |
bData.Cur = fData->Cur; |
*fData->iMinSAD = 4096*256; |
1957 |
fData->iEdgedWidth = bData.iEdgedWidth = iEdgedWidth; |
bData.currentMV++; bData.currentQMV++; |
|
bData.currentMV = fData->currentMV + 1; |
|
|
bData.iQuant = fData->iQuant; |
|
1958 |
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
1959 |
|
|
1960 |
bData.bRef = fData->Ref = f_Ref + (x + y * iEdgedWidth) * 16; |
i = (x + y * fData->iEdgedWidth) * 16; |
1961 |
bData.bRefH = fData->RefH = f_RefH + (x + y * iEdgedWidth) * 16; |
|
1962 |
bData.bRefV = fData->RefV = f_RefV + (x + y * iEdgedWidth) * 16; |
bData.b_RefP[0] = fData->RefP[0] = f_Ref->y + i; |
1963 |
bData.bRefHV = fData->RefHV = f_RefHV + (x + y * iEdgedWidth) * 16; |
bData.b_RefP[2] = fData->RefP[2] = f_RefH + i; |
1964 |
bData.Ref = fData->bRef = b_Ref + (x + y * iEdgedWidth) * 16; |
bData.b_RefP[1] = fData->RefP[1] = f_RefV + i; |
1965 |
bData.RefH = fData->bRefH = b_RefH + (x + y * iEdgedWidth) * 16; |
bData.b_RefP[3] = fData->RefP[3] = f_RefHV + i; |
1966 |
bData.RefV = fData->bRefV = b_RefV + (x + y * iEdgedWidth) * 16; |
bData.RefP[0] = fData->b_RefP[0] = b_Ref->y + i; |
1967 |
bData.RefHV = fData->bRefHV = b_RefHV + (x + y * iEdgedWidth) * 16; |
bData.RefP[2] = fData->b_RefP[2] = b_RefH + i; |
1968 |
|
bData.RefP[1] = fData->b_RefP[1] = b_RefV + i; |
1969 |
|
bData.RefP[3] = fData->b_RefP[3] = b_RefHV + i; |
1970 |
|
bData.b_RefP[4] = fData->RefP[4] = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; |
1971 |
|
bData.b_RefP[5] = fData->RefP[5] = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; |
1972 |
|
bData.RefP[4] = fData->b_RefP[4] = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; |
1973 |
|
bData.RefP[5] = fData->b_RefP[5] = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; |
1974 |
|
|
1975 |
bData.bpredMV = fData->predMV = *f_predMV; |
bData.bpredMV = fData->predMV = *f_predMV; |
1976 |
fData->bpredMV = bData.predMV = *b_predMV; |
fData->bpredMV = bData.predMV = *b_predMV; |
1977 |
|
fData->currentMV[0] = fData->currentMV[2]; |
1978 |
|
|
1979 |
fData->currentMV[0] = pMB->mvs[0]; |
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); |
1980 |
fData->currentMV[1] = pMB->b_mvs[0]; |
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); |
|
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode); |
|
|
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode); |
|
1981 |
|
|
1982 |
if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; |
if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; |
1983 |
if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dy; |
if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx; |
1984 |
if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dx; |
if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy; |
1985 |
if (fData->currentMV[0].y > fData->min_dy) fData->currentMV[0].y = fData->min_dy; |
if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy; |
1986 |
|
|
1987 |
if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx; |
if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx; |
1988 |
if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dy; |
if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx; |
1989 |
if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dx; |
if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy; |
1990 |
if (fData->currentMV[1].y > bData.min_dy) fData->currentMV[1].y = bData.min_dy; |
if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy; |
1991 |
|
|
1992 |
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
1993 |
|
|
1994 |
//diamond. I wish we could use normal mainsearch functions (square, advdiamond) |
/* diamond */ |
|
|
|
1995 |
do { |
do { |
1996 |
iDirection = 255; |
iDirection = 255; |
1997 |
// forward MV moves |
/* forward MV moves */ |
1998 |
i = fData->currentMV[0].x; j = fData->currentMV[0].y; |
i = fData->currentMV[0].x; j = fData->currentMV[0].y; |
1999 |
|
|
2000 |
CheckCandidateInt(i + 1, j, 0, &iDirection, fData); |
CheckCandidateInt(i + 1, j, 0, &iDirection, fData); |
2002 |
CheckCandidateInt(i - 1, j, 0, &iDirection, fData); |
CheckCandidateInt(i - 1, j, 0, &iDirection, fData); |
2003 |
CheckCandidateInt(i, j - 1, 0, &iDirection, fData); |
CheckCandidateInt(i, j - 1, 0, &iDirection, fData); |
2004 |
|
|
2005 |
// backward MV moves |
/* backward MV moves */ |
2006 |
i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
2007 |
fData->currentMV[2] = fData->currentMV[0]; |
fData->currentMV[2] = fData->currentMV[0]; |
|
|
|
2008 |
CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); |
CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); |
2009 |
CheckCandidateInt(i, j + 1, 0, &iDirection, &bData); |
CheckCandidateInt(i, j + 1, 0, &iDirection, &bData); |
2010 |
CheckCandidateInt(i - 1, j, 0, &iDirection, &bData); |
CheckCandidateInt(i - 1, j, 0, &iDirection, &bData); |
2012 |
|
|
2013 |
} while (!(iDirection)); |
} while (!(iDirection)); |
2014 |
|
|
2015 |
// two bits are needed to code interpolate mode. we treat the bits just like they were vector's |
/* qpel refinement */ |
2016 |
*fData->iMinSAD += 2 * lambda_vec16[fData->iQuant]; |
if (fData->qpel) { |
2017 |
|
if (*fData->iMinSAD > *best_sad + 500) return; |
2018 |
|
CheckCandidate = CheckCandidateInt; |
2019 |
|
fData->qpel_precision = bData.qpel_precision = 1; |
2020 |
|
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0); |
2021 |
|
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0); |
2022 |
|
fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x; |
2023 |
|
fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y; |
2024 |
|
fData->currentQMV[1].x = 2 * fData->currentMV[1].x; |
2025 |
|
fData->currentQMV[1].y = 2 * fData->currentMV[1].y; |
2026 |
|
SubpelRefine(fData); |
2027 |
|
if (*fData->iMinSAD > *best_sad + 300) return; |
2028 |
|
fData->currentQMV[2] = fData->currentQMV[0]; |
2029 |
|
SubpelRefine(&bData); |
2030 |
|
} |
2031 |
|
|
2032 |
|
*fData->iMinSAD += (2+3) * fData->lambda16; /* two bits are needed to code interpolate mode. */ |
2033 |
|
|
2034 |
if (*fData->iMinSAD < *best_sad) { |
if (*fData->iMinSAD < *best_sad) { |
2035 |
*best_sad = *fData->iMinSAD; |
*best_sad = *fData->iMinSAD; |
2036 |
pMB->mvs[0] = fData->currentMV[0]; |
pMB->mvs[0] = fData->currentMV[0]; |
2037 |
pMB->b_mvs[0] = fData->currentMV[1]; |
pMB->b_mvs[0] = fData->currentMV[1]; |
2038 |
pMB->mode = MODE_INTERPOLATE; |
pMB->mode = MODE_INTERPOLATE; |
2039 |
|
if (fData->qpel) { |
2040 |
|
pMB->qmvs[0] = fData->currentQMV[0]; |
2041 |
|
pMB->b_qmvs[0] = fData->currentQMV[1]; |
2042 |
|
pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x; |
2043 |
|
pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y; |
2044 |
|
pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x; |
2045 |
|
pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y; |
2046 |
|
} else { |
2047 |
pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; |
pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; |
2048 |
pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; |
pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; |
2049 |
pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; |
pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; |
2050 |
pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; |
pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; |
2051 |
} |
} |
2052 |
} |
} |
2053 |
|
} |
2054 |
|
|
2055 |
void |
void |
2056 |
MotionEstimationBVOP(MBParam * const pParam, |
MotionEstimationBVOP(MBParam * const pParam, |
2057 |
FRAMEINFO * const frame, |
FRAMEINFO * const frame, |
2058 |
const int32_t time_bp, |
const int32_t time_bp, |
2059 |
const int32_t time_pp, |
const int32_t time_pp, |
2060 |
// forward (past) reference |
/* forward (past) reference */ |
2061 |
const MACROBLOCK * const f_mbs, |
const MACROBLOCK * const f_mbs, |
2062 |
const IMAGE * const f_ref, |
const IMAGE * const f_ref, |
2063 |
const IMAGE * const f_refH, |
const IMAGE * const f_refH, |
2064 |
const IMAGE * const f_refV, |
const IMAGE * const f_refV, |
2065 |
const IMAGE * const f_refHV, |
const IMAGE * const f_refHV, |
2066 |
// backward (future) reference |
/* backward (future) reference */ |
2067 |
const MACROBLOCK * const b_mbs, |
const FRAMEINFO * const b_reference, |
2068 |
const IMAGE * const b_ref, |
const IMAGE * const b_ref, |
2069 |
const IMAGE * const b_refH, |
const IMAGE * const b_refH, |
2070 |
const IMAGE * const b_refV, |
const IMAGE * const b_refV, |
2071 |
const IMAGE * const b_refHV) |
const IMAGE * const b_refHV) |
2072 |
{ |
{ |
2073 |
uint32_t i, j; |
uint32_t i, j; |
2074 |
int32_t best_sad, skip_sad; |
int32_t best_sad; |
2075 |
|
uint32_t skip_sad; |
2076 |
int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; |
int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; |
2077 |
static const VECTOR zeroMV={0,0}; |
const MACROBLOCK * const b_mbs = b_reference->mbs; |
2078 |
|
|
2079 |
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
2080 |
|
|
2081 |
const int32_t TRB = time_pp - time_bp; |
const int32_t TRB = time_pp - time_bp; |
2082 |
const int32_t TRD = time_pp; |
const int32_t TRD = time_pp; |
2083 |
|
|
2084 |
// some pre-inintialized data for the rest of the search |
/* some pre-inintialized data for the rest of the search */ |
2085 |
|
|
2086 |
SearchData Data; |
SearchData Data; |
2087 |
int32_t iMinSAD; |
int32_t iMinSAD; |
2088 |
VECTOR currentMV[3]; |
VECTOR currentMV[3]; |
2089 |
|
VECTOR currentQMV[3]; |
2090 |
|
int32_t temp[8]; |
2091 |
|
memset(&Data, 0, sizeof(SearchData)); |
2092 |
Data.iEdgedWidth = pParam->edged_width; |
Data.iEdgedWidth = pParam->edged_width; |
2093 |
Data.currentMV = currentMV; |
Data.currentMV = currentMV; Data.currentQMV = currentQMV; |
2094 |
Data.iMinSAD = &iMinSAD; |
Data.iMinSAD = &iMinSAD; |
2095 |
Data.iQuant = frame->quant; |
Data.lambda16 = lambda_vec16[frame->quant]; |
2096 |
|
Data.qpel = pParam->vol_flags & XVID_VOL_QUARTERPEL; |
2097 |
|
Data.rounding = 0; |
2098 |
|
Data.chroma = frame->motion_flags & XVID_ME_CHROMA_BVOP; |
2099 |
|
Data.temp = temp; |
2100 |
|
|
2101 |
// note: i==horizontal, j==vertical |
Data.RefQ = f_refV->u; /* a good place, also used in MC (for similar purpose) */ |
2102 |
|
|
2103 |
|
/* note: i==horizontal, j==vertical */ |
2104 |
for (j = 0; j < pParam->mb_height; j++) { |
for (j = 0; j < pParam->mb_height; j++) { |
2105 |
|
|
2106 |
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
2109 |
MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; |
MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; |
2110 |
const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; |
const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; |
2111 |
|
|
2112 |
/* special case, if collocated block is SKIPed: encoding is forward (0,0), cpb=0 without further ado */ |
/* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */ |
2113 |
|
if (b_reference->coding_type != S_VOP) |
2114 |
if (b_mb->mode == MODE_NOT_CODED) { |
if (b_mb->mode == MODE_NOT_CODED) { |
2115 |
pMB->mode = MODE_NOT_CODED; |
pMB->mode = MODE_NOT_CODED; |
2116 |
continue; |
continue; |
2117 |
} |
} |
2118 |
|
|
2119 |
Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
2120 |
|
Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8; |
2121 |
|
Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8; |
2122 |
|
pMB->quant = frame->quant; |
2123 |
|
|
2124 |
/* direct search comes first, because it (1) checks for SKIP-mode |
/* direct search comes first, because it (1) checks for SKIP-mode |
2125 |
and (2) sets very good predictions for forward and backward search */ |
and (2) sets very good predictions for forward and backward search */ |
|
|
|
2126 |
skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
2127 |
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
2128 |
&frame->image, |
&frame->image, |
2136 |
|
|
2137 |
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
2138 |
|
|
2139 |
// best_sad = 256*4096; //uncomment to disable Directsearch. |
/* forward search */ |
2140 |
// To disable any other mode, just comment the function call |
SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
|
|
|
|
// forward search |
|
|
SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
2141 |
&frame->image, i, j, |
&frame->image, i, j, |
2142 |
frame->motion_flags, |
frame->motion_flags, |
2143 |
frame->fcode, pParam, |
frame->fcode, pParam, |
2144 |
pMB, &f_predMV, &best_sad, |
pMB, &f_predMV, &best_sad, |
2145 |
MODE_FORWARD, &Data); |
MODE_FORWARD, &Data); |
2146 |
|
|
2147 |
// backward search |
/* backward search */ |
2148 |
SearchBF(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y, |
2149 |
&frame->image, i, j, |
&frame->image, i, j, |
2150 |
frame->motion_flags, |
frame->motion_flags, |
2151 |
frame->bcode, pParam, |
frame->bcode, pParam, |
2152 |
pMB, &b_predMV, &best_sad, |
pMB, &b_predMV, &best_sad, |
2153 |
MODE_BACKWARD, &Data); |
MODE_BACKWARD, &Data); |
2154 |
|
|
2155 |
// interpolate search comes last, because it uses data from forward and backward as prediction |
/* interpolate search comes last, because it uses data from forward and backward as prediction */ |
2156 |
|
SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
2157 |
SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
|
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
2158 |
&frame->image, |
&frame->image, |
2159 |
i, j, |
i, j, |
2160 |
frame->fcode, frame->bcode, |
frame->fcode, frame->bcode, |
2164 |
pMB, &best_sad, |
pMB, &best_sad, |
2165 |
&Data); |
&Data); |
2166 |
|
|
2167 |
|
/* final skip decision */ |
2168 |
|
if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2) |
2169 |
|
&& ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) |
2170 |
|
SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data); |
2171 |
|
|
2172 |
switch (pMB->mode) { |
switch (pMB->mode) { |
2173 |
case MODE_FORWARD: |
case MODE_FORWARD: |
2174 |
f_count++; |
f_count++; |
2175 |
f_predMV = pMB->mvs[0]; |
f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; |
2176 |
break; |
break; |
2177 |
case MODE_BACKWARD: |
case MODE_BACKWARD: |
2178 |
b_count++; |
b_count++; |
2179 |
b_predMV = pMB->b_mvs[0]; |
b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; |
2180 |
break; |
break; |
2181 |
case MODE_INTERPOLATE: |
case MODE_INTERPOLATE: |
2182 |
i_count++; |
i_count++; |
2183 |
f_predMV = pMB->mvs[0]; |
f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; |
2184 |
b_predMV = pMB->b_mvs[0]; |
b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; |
2185 |
break; |
break; |
2186 |
case MODE_DIRECT: |
case MODE_DIRECT: |
2187 |
case MODE_DIRECT_NO4V: |
case MODE_DIRECT_NO4V: |
2188 |
d_count++; |
d_count++; |
|
break; |
|
2189 |
default: |
default: |
2190 |
break; |
break; |
2191 |
} |
} |
2192 |
} |
} |
2193 |
} |
} |
|
|
|
|
// fprintf(debug,"B-Stat: F: %04d B: %04d I: %04d D: %04d, N: %04d\n", |
|
|
// f_count,b_count,i_count,d_count,n_count); |
|
|
|
|
2194 |
} |
} |
2195 |
|
|
|
/* Hinted ME starts here */ |
|
|
|
|
2196 |
static __inline void |
static __inline void |
2197 |
Search8hinted( const SearchData * const OldData, |
MEanalyzeMB ( const uint8_t * const pRef, |
2198 |
const int x, const int y, |
const uint8_t * const pCur, |
2199 |
const uint32_t MotionFlags, |
const int x, |
2200 |
|
const int y, |
2201 |
const MBParam * const pParam, |
const MBParam * const pParam, |
2202 |
MACROBLOCK * const pMB, |
MACROBLOCK * const pMBs, |
2203 |
const MACROBLOCK * const pMBs, |
SearchData * const Data) |
|
const int block) |
|
2204 |
{ |
{ |
|
SearchData Data; |
|
|
MainSearchFunc *MainSearchPtr; |
|
2205 |
|
|
2206 |
Data.predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
int i, mask; |
2207 |
Data.iMinSAD = OldData->iMinSAD + 1 + block; |
int quarterpel = (pParam->vol_flags & XVID_VOL_QUARTERPEL)? 1: 0; |
2208 |
Data.currentMV = OldData->currentMV+1+block; |
VECTOR pmv[3]; |
2209 |
Data.iFcode = OldData->iFcode; |
MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width]; |
|
Data.iQuant = OldData->iQuant; |
|
|
|
|
|
Data.Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
|
Data.RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
|
Data.RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
|
Data.RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
|
Data.iEdgedWidth = pParam->edged_width; |
|
|
Data.Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
2210 |
|
|
2211 |
CheckCandidate = CheckCandidate8; |
for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
2212 |
|
|
2213 |
if (block != 0) |
/* median is only used as prediction. it doesn't have to be real */ |
2214 |
*(Data.iMinSAD) += lambda_vec8[Data.iQuant] * |
if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; |
2215 |
d_mv_bits( Data.currentMV->x - Data.predMV.x, |
else |
2216 |
Data.currentMV->y - Data.predMV.y, |
if (x == 1) /* left macroblock does not have any vector now */ |
2217 |
Data.iFcode); |
Data->predMV = (pMB - pParam->mb_width)->mvs[0]; /* top instead of median */ |
2218 |
|
else if (y == 1) /* top macroblock doesn't have it's vector */ |
2219 |
|
Data->predMV = (pMB - 1)->mvs[0]; /* left instead of median */ |
2220 |
|
else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); /* else median */ |
2221 |
|
|
2222 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2223 |
|
pParam->width, pParam->height, Data->iFcode - quarterpel, 0, 0); |
2224 |
|
|
2225 |
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 8, |
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
2226 |
pParam->width, pParam->height, OldData->iFcode); |
Data->RefP[0] = pRef + (x + y * pParam->edged_width) * 16; |
2227 |
|
|
2228 |
if (pMB->mode == MODE_INTER4V) { |
pmv[1].x = EVEN(pMB->mvs[0].x); |
2229 |
int dummy; |
pmv[1].y = EVEN(pMB->mvs[0].y); |
2230 |
CheckCandidate8(pMB->mvs[block].x, pMB->mvs[block].y, 0, &dummy, &Data); } |
pmv[2].x = EVEN(Data->predMV.x); |
2231 |
|
pmv[2].y = EVEN(Data->predMV.y); |
2232 |
|
pmv[0].x = pmv[0].y = 0; |
2233 |
|
|
2234 |
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
CheckCandidate32I(0, 0, 255, &i, Data); |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
|
|
else MainSearchPtr = DiamondSearch; |
|
2235 |
|
|
2236 |
(*MainSearchPtr)(Data.currentMV->x, Data.currentMV->y, &Data, 255); |
if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) { |
2237 |
|
|
2238 |
if (MotionFlags & PMV_HALFPELREFINE8) HalfpelRefine(&Data); |
if (!(mask = make_mask(pmv, 1))) |
2239 |
|
CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data); |
2240 |
|
if (!(mask = make_mask(pmv, 2))) |
2241 |
|
CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data); |
2242 |
|
|
2243 |
pMB->pmvs[block].x = Data.currentMV->x - Data.predMV.x; |
if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) /* diamond only if needed */ |
2244 |
pMB->pmvs[block].y = Data.currentMV->y - Data.predMV.y; |
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
|
pMB->mvs[block] = *(Data.currentMV); |
|
|
pMB->sad8[block] = 4 * (*(Data.iMinSAD)); |
|
2245 |
} |
} |
2246 |
|
|
2247 |
|
for (i = 0; i < 4; i++) { |
2248 |
|
MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width]; |
2249 |
|
MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i]; |
2250 |
|
MB->mode = MODE_INTER; |
2251 |
|
MB->sad16 = Data->iMinSAD[i+1]; |
2252 |
|
} |
2253 |
|
} |
2254 |
|
|
2255 |
static void |
#define INTRA_THRESH 2200 |
2256 |
SearchPhinted ( const uint8_t * const pRef, |
#define INTER_THRESH 50 |
2257 |
const uint8_t * const pRefH, |
#define INTRA_THRESH2 95 |
2258 |
const uint8_t * const pRefV, |
|
2259 |
const uint8_t * const pRefHV, |
int |
2260 |
const IMAGE * const pCur, |
MEanalysis( const IMAGE * const pRef, |
2261 |
const int x, |
const FRAMEINFO * const Current, |
|
const int y, |
|
|
const uint32_t MotionFlags, |
|
|
const uint32_t iQuant, |
|
2262 |
const MBParam * const pParam, |
const MBParam * const pParam, |
2263 |
const MACROBLOCK * const pMBs, |
const int maxIntra, //maximum number if non-I frames |
2264 |
int inter4v, |
const int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame |
2265 |
MACROBLOCK * const pMB, |
const int bCount, // number of B frames in a row |
2266 |
SearchData * const Data) |
const int b_thresh) |
2267 |
{ |
{ |
2268 |
|
uint32_t x, y, intra = 0; |
2269 |
|
int sSAD = 0; |
2270 |
|
MACROBLOCK * const pMBs = Current->mbs; |
2271 |
|
const IMAGE * const pCurrent = &Current->image; |
2272 |
|
int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + b_thresh; |
2273 |
|
int blocks = 0; |
2274 |
|
int complexity = 0; |
2275 |
|
|
2276 |
|
int32_t iMinSAD[5], temp[5]; |
2277 |
|
VECTOR currentMV[5]; |
2278 |
|
SearchData Data; |
2279 |
|
Data.iEdgedWidth = pParam->edged_width; |
2280 |
|
Data.currentMV = currentMV; |
2281 |
|
Data.iMinSAD = iMinSAD; |
2282 |
|
Data.iFcode = Current->fcode; |
2283 |
|
Data.temp = temp; |
2284 |
|
CheckCandidate = CheckCandidate32I; |
2285 |
|
|
2286 |
const int32_t iEdgedWidth = pParam->edged_width; |
if (intraCount != 0) { |
2287 |
|
if (intraCount < 10) // we're right after an I frame |
2288 |
|
IntraThresh += 15* (intraCount - 10) * (intraCount - 10); |
2289 |
|
else |
2290 |
|
if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec |
2291 |
|
IntraThresh -= (IntraThresh * (maxIntra - 8*(maxIntra - intraCount)))/maxIntra; |
2292 |
|
} |
2293 |
|
|
2294 |
int i, t; |
InterThresh -= 12 * bCount; |
2295 |
MainSearchFunc * MainSearchPtr; |
if (InterThresh < 15 + b_thresh) InterThresh = 15 + b_thresh; |
2296 |
|
|
2297 |
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
if (sadInit) (*sadInit) (); |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
|
|
pParam->width, pParam->height, Data->iFcode); |
|
2298 |
|
|
2299 |
Data->Cur = pCur->y + (x + y * iEdgedWidth) * 16; |
for (y = 1; y < pParam->mb_height-1; y += 2) { |
2300 |
Data->Ref = pRef + (x + iEdgedWidth*y)*16; |
for (x = 1; x < pParam->mb_width-1; x += 2) { |
2301 |
Data->RefH = pRefH + (x + iEdgedWidth*y) * 16; |
int i; |
2302 |
Data->RefV = pRefV + (x + iEdgedWidth*y) * 16; |
blocks += 10; |
2303 |
Data->RefHV = pRefHV + (x + iEdgedWidth*y) * 16; |
|
2304 |
Data->iQuant = iQuant; |
if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; |
2305 |
|
else { //extrapolation of the vector found for last frame |
2306 |
|
pMBs[x + y * pParam->mb_width].mvs[0].x = |
2307 |
|
(pMBs[x + y * pParam->mb_width].mvs[0].x * (bCount+1) ) / bCount; |
2308 |
|
pMBs[x + y * pParam->mb_width].mvs[0].y = |
2309 |
|
(pMBs[x + y * pParam->mb_width].mvs[0].y * (bCount+1) ) / bCount; |
2310 |
|
} |
2311 |
|
|
2312 |
|
MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data); |
2313 |
|
|
2314 |
|
for (i = 0; i < 4; i++) { |
2315 |
|
int dev; |
2316 |
|
MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width]; |
2317 |
|
dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16, |
2318 |
|
pParam->edged_width); |
2319 |
|
|
2320 |
if (!(MotionFlags & PMV_HALFPEL16)) { |
complexity += MAX(dev, 300); |
2321 |
Data->min_dx = EVEN(Data->min_dx); |
if (dev + IntraThresh < pMB->sad16) { |
2322 |
Data->max_dx = EVEN(Data->max_dx); |
pMB->mode = MODE_INTRA; |
2323 |
Data->min_dy = EVEN(Data->min_dy); |
if (++intra > ((pParam->mb_height-2)*(pParam->mb_width-2))/2) return I_VOP; |
|
Data->max_dy = EVEN(Data->max_dy); |
|
2324 |
} |
} |
2325 |
|
|
2326 |
for(i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0) |
2327 |
|
if (dev > 500 && pMB->sad16 < 1000) |
2328 |
|
sSAD += 1000; |
2329 |
|
|
2330 |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
sSAD += (dev < 3000) ? pMB->sad16 : pMB->sad16/2; /* blocks with big contrast differences usually have large SAD - while they look very good in b-frames */ |
2331 |
|
} |
2332 |
|
} |
2333 |
|
} |
2334 |
|
complexity >>= 7; |
2335 |
|
|
2336 |
if (inter4v) |
sSAD /= complexity + 4*blocks; |
|
CheckCandidate = CheckCandidate16; |
|
|
else CheckCandidate = CheckCandidate16no4v; |
|
2337 |
|
|
2338 |
|
if (intraCount > 80 && sSAD > INTRA_THRESH2 ) return I_VOP; |
2339 |
|
if (sSAD > InterThresh ) return P_VOP; |
2340 |
|
emms(); |
2341 |
|
return B_VOP; |
2342 |
|
} |
2343 |
|
|
|
pMB->mvs[0].x = EVEN(pMB->mvs[0].x); |
|
|
pMB->mvs[0].y = EVEN(pMB->mvs[0].y); |
|
|
if (pMB->mvs[0].x > Data->max_dx) pMB->mvs[0].x = Data->max_dx; // this is in case iFcode changed |
|
|
if (pMB->mvs[0].x < Data->min_dx) pMB->mvs[0].x = Data->min_dx; |
|
|
if (pMB->mvs[0].y > Data->max_dy) pMB->mvs[0].y = Data->max_dy; |
|
|
if (pMB->mvs[0].y < Data->min_dy) pMB->mvs[0].y = Data->min_dy; |
|
|
|
|
|
(*CheckCandidate)(pMB->mvs[0].x, pMB->mvs[0].y, 0, &t, Data); |
|
|
|
|
|
if (pMB->mode == MODE_INTER4V) |
|
|
for (i = 1; i < 4; i++) { // all four vectors will be used as four predictions for 16x16 search |
|
|
pMB->mvs[i].x = EVEN(pMB->mvs[i].x); |
|
|
pMB->mvs[i].y = EVEN(pMB->mvs[i].y); |
|
|
if (!(make_mask(pMB->mvs, i))) |
|
|
(*CheckCandidate)(pMB->mvs[i].x, pMB->mvs[i].y, 0, &t, Data); |
|
|
} |
|
|
|
|
|
if (MotionFlags & PMV_USESQUARES16) |
|
|
MainSearchPtr = SquareSearch; |
|
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
|
|
MainSearchPtr = AdvDiamondSearch; |
|
|
else MainSearchPtr = DiamondSearch; |
|
2344 |
|
|
2345 |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
/* functions which perform BITS-based search/bitcount */ |
2346 |
|
|
2347 |
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data); |
static int |
2348 |
|
findRDinter(SearchData * const Data, |
2349 |
|
const MACROBLOCK * const pMBs, const int x, const int y, |
2350 |
|
const MBParam * const pParam, |
2351 |
|
const uint32_t MotionFlags) |
2352 |
|
{ |
2353 |
|
int i, iDirection; |
2354 |
|
int32_t bsad[5]; |
2355 |
|
|
2356 |
if (inter4v) |
CheckCandidate = CheckCandidateRD16; |
|
for(i = 0; i < 4; i++) |
|
|
Search8hinted(Data, 2*x+(i&1), 2*y+(i>>1), MotionFlags, pParam, pMB, pMBs, i); |
|
|
|
|
|
if (!(inter4v) || |
|
|
(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] + |
|
|
Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
|
|
// INTER MODE |
|
2357 |
|
|
2358 |
pMB->mode = MODE_INTER; |
if (Data->qpel) { |
2359 |
pMB->mvs[0] = pMB->mvs[1] |
for(i = 0; i < 5; i++) { |
2360 |
= pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
Data->currentMV[i].x = Data->currentQMV[i].x/2; |
2361 |
|
Data->currentMV[i].y = Data->currentQMV[i].y/2; |
2362 |
|
} |
2363 |
|
Data->qpel_precision = 1; |
2364 |
|
CheckCandidateRD16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data); |
2365 |
|
|
2366 |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
if (MotionFlags & (XVID_ME_HALFPELREFINE16_RD | XVID_ME_EXTSEARCH_RD)) { /* we have to prepare for halfpixel-precision search */ |
2367 |
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i]; |
2368 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2369 |
|
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
2370 |
|
Data->qpel_precision = 0; |
2371 |
|
if (Data->currentQMV->x & 1 || Data->currentQMV->y & 1) |
2372 |
|
CheckCandidateRD16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); |
2373 |
|
} |
2374 |
|
|
2375 |
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
} else { /* not qpel */ |
2376 |
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
|
2377 |
|
CheckCandidateRD16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); |
2378 |
|
} |
2379 |
|
|
2380 |
|
if (MotionFlags&XVID_ME_EXTSEARCH_RD) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
2381 |
|
|
2382 |
|
if (MotionFlags&XVID_ME_HALFPELREFINE16_RD) SubpelRefine(Data); |
2383 |
|
|
2384 |
|
if (Data->qpel) { |
2385 |
|
if (MotionFlags&(XVID_ME_EXTSEARCH_RD | XVID_ME_HALFPELREFINE16_RD)) { /* there was halfpel-precision search */ |
2386 |
|
for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) { |
2387 |
|
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* we have found a better match */ |
2388 |
|
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
2389 |
|
} |
2390 |
|
|
2391 |
|
/* preparing for qpel-precision search */ |
2392 |
|
Data->qpel_precision = 1; |
2393 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2394 |
|
pParam->width, pParam->height, Data->iFcode, 1, 0); |
2395 |
|
} |
2396 |
|
if (MotionFlags&XVID_ME_QUARTERPELREFINE16_RD) SubpelRefine(Data); |
2397 |
|
} |
2398 |
|
|
2399 |
|
if (MotionFlags&XVID_ME_CHECKPREDICTION_RD) { /* let's check vector equal to prediction */ |
2400 |
|
VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV; |
2401 |
|
if (!(Data->predMV.x == v->x && Data->predMV.y == v->y)) |
2402 |
|
CheckCandidateRD16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data); |
2403 |
|
} |
2404 |
|
return Data->iMinSAD[0]; |
2405 |
|
} |
2406 |
|
|
2407 |
|
static int |
2408 |
|
findRDinter4v(const SearchData * const Data, |
2409 |
|
MACROBLOCK * const pMB, const MACROBLOCK * const pMBs, |
2410 |
|
const int x, const int y, |
2411 |
|
const MBParam * const pParam, const uint32_t MotionFlags, |
2412 |
|
const VECTOR * const backup) |
2413 |
|
{ |
2414 |
|
|
2415 |
|
int cbp = 0, bits = 0, t = 0, i, iDirection; |
2416 |
|
SearchData Data2, *Data8 = &Data2; |
2417 |
|
int sumx = 0, sumy = 0; |
2418 |
|
int16_t *in = Data->dctSpace, *coeff = Data->dctSpace + 64; |
2419 |
|
uint8_t * ptr; |
2420 |
|
|
2421 |
|
memcpy(Data8, Data, sizeof(SearchData)); |
2422 |
|
CheckCandidate = CheckCandidateRD8; |
2423 |
|
|
2424 |
|
for (i = 0; i < 4; i++) { /* for all luma blocks */ |
2425 |
|
|
2426 |
|
Data8->iMinSAD = Data->iMinSAD + i + 1; |
2427 |
|
Data8->currentMV = Data->currentMV + i + 1; |
2428 |
|
Data8->currentQMV = Data->currentQMV + i + 1; |
2429 |
|
Data8->Cur = Data->Cur + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2430 |
|
Data8->RefP[0] = Data->RefP[0] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2431 |
|
Data8->RefP[2] = Data->RefP[2] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2432 |
|
Data8->RefP[1] = Data->RefP[1] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2433 |
|
Data8->RefP[3] = Data->RefP[3] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2434 |
|
*Data8->cbp = (Data->cbp[1] & (1<<(5-i))) ? 1:0; // copy corresponding cbp bit |
2435 |
|
|
2436 |
|
if(Data->qpel) { |
2437 |
|
Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i); |
2438 |
|
if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y, |
2439 |
|
Data8->predMV, Data8->iFcode, 0, 0); |
2440 |
} else { |
} else { |
2441 |
// INTER4V MODE; all other things are already set in Search8hinted |
Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i); |
2442 |
pMB->mode = MODE_INTER4V; |
if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y, |
2443 |
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] |
Data8->predMV, Data8->iFcode, 0, 0); |
2444 |
+ Data->iMinSAD[4] + IMV16X16 * iQuant; |
} |
2445 |
|
|
2446 |
|
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, |
2447 |
|
pParam->width, pParam->height, Data8->iFcode, Data8->qpel, 0); |
2448 |
|
|
2449 |
|
*Data8->iMinSAD += BITS_MULT*t; |
2450 |
|
|
2451 |
|
Data8->qpel_precision = Data8->qpel; |
2452 |
|
/* checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far) */ |
2453 |
|
{ |
2454 |
|
VECTOR *v = Data8->qpel ? Data8->currentQMV : Data8->currentMV; |
2455 |
|
if (!MVequal (*v, backup[i+1]) ) |
2456 |
|
CheckCandidateRD8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8); |
2457 |
|
} |
2458 |
|
|
2459 |
|
if (Data8->qpel) { |
2460 |
|
if (MotionFlags&XVID_ME_HALFPELREFINE8_RD || (MotionFlags&XVID_ME_EXTSEARCH8 && MotionFlags&XVID_ME_EXTSEARCH_RD)) { /* halfpixel motion search follows */ |
2461 |
|
int32_t s = *Data8->iMinSAD; |
2462 |
|
Data8->currentMV->x = Data8->currentQMV->x/2; |
2463 |
|
Data8->currentMV->y = Data8->currentQMV->y/2; |
2464 |
|
Data8->qpel_precision = 0; |
2465 |
|
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, |
2466 |
|
pParam->width, pParam->height, Data8->iFcode - 1, 0, 0); |
2467 |
|
|
2468 |
|
if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1) |
2469 |
|
CheckCandidateRD8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8); |
2470 |
|
|
2471 |
|
if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_RD) |
2472 |
|
SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255); |
2473 |
|
|
2474 |
|
if (MotionFlags & XVID_ME_HALFPELREFINE8_RD) |
2475 |
|
SubpelRefine(Data8); |
2476 |
|
|
2477 |
|
if(s > *Data8->iMinSAD) { /* we have found a better match */ |
2478 |
|
Data8->currentQMV->x = 2*Data8->currentMV->x; |
2479 |
|
Data8->currentQMV->y = 2*Data8->currentMV->y; |
2480 |
|
} |
2481 |
|
|
2482 |
|
Data8->qpel_precision = 1; |
2483 |
|
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, |
2484 |
|
pParam->width, pParam->height, Data8->iFcode, 1, 0); |
2485 |
|
|
2486 |
|
} |
2487 |
|
if (MotionFlags & XVID_ME_QUARTERPELREFINE8_RD) SubpelRefine(Data8); |
2488 |
|
|
2489 |
|
} else { /* not qpel */ |
2490 |
|
|
2491 |
|
if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_RD) /* extsearch */ |
2492 |
|
SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255); |
2493 |
|
|
2494 |
|
if (MotionFlags & XVID_ME_HALFPELREFINE8_RD) |
2495 |
|
SubpelRefine(Data8); /* halfpel refinement */ |
2496 |
|
} |
2497 |
|
|
2498 |
|
/* checking vector equal to predicion */ |
2499 |
|
if (i != 0 && MotionFlags & XVID_ME_CHECKPREDICTION_RD) { |
2500 |
|
const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV; |
2501 |
|
if (!MVequal(*v, Data8->predMV)) |
2502 |
|
CheckCandidateRD8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8); |
2503 |
|
} |
2504 |
|
|
2505 |
|
bits += *Data8->iMinSAD; |
2506 |
|
if (bits >= Data->iMinSAD[0]) return bits; /* no chances for INTER4V */ |
2507 |
|
|
2508 |
|
/* MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else */ |
2509 |
|
if(Data->qpel) { |
2510 |
|
pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x; |
2511 |
|
pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y; |
2512 |
|
pMB->qmvs[i] = *Data8->currentQMV; |
2513 |
|
sumx += Data8->currentQMV->x/2; |
2514 |
|
sumy += Data8->currentQMV->y/2; |
2515 |
|
} else { |
2516 |
|
pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x; |
2517 |
|
pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y; |
2518 |
|
sumx += Data8->currentMV->x; |
2519 |
|
sumy += Data8->currentMV->y; |
2520 |
|
} |
2521 |
|
pMB->mvs[i] = *Data8->currentMV; |
2522 |
|
pMB->sad8[i] = 4 * *Data8->iMinSAD; |
2523 |
|
if (Data8->cbp[0]) cbp |= 1 << (5 - i); |
2524 |
|
|
2525 |
|
} /* end - for all luma blocks */ |
2526 |
|
|
2527 |
|
bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; |
2528 |
|
|
2529 |
|
/* let's check chroma */ |
2530 |
|
sumx = (sumx >> 3) + roundtab_76[sumx & 0xf]; |
2531 |
|
sumy = (sumy >> 3) + roundtab_76[sumy & 0xf]; |
2532 |
|
|
2533 |
|
/* chroma U */ |
2534 |
|
ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[4], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
2535 |
|
transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2); |
2536 |
|
bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4); |
2537 |
|
|
2538 |
|
if (bits >= *Data->iMinSAD) return bits; |
2539 |
|
|
2540 |
|
/* chroma V */ |
2541 |
|
ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[5], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
2542 |
|
transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2); |
2543 |
|
bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5); |
2544 |
|
|
2545 |
|
bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len; |
2546 |
|
|
2547 |
|
*Data->cbp = cbp; |
2548 |
|
return bits; |
2549 |
|
} |
2550 |
|
|
2551 |
|
static int |
2552 |
|
findRDintra(const SearchData * const Data) |
2553 |
|
{ |
2554 |
|
int bits = BITS_MULT*1; /* this one is ac/dc prediction flag bit */ |
2555 |
|
int cbp = 0, i, dc = 0; |
2556 |
|
int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64; |
2557 |
|
|
2558 |
|
for(i = 0; i < 4; i++) { |
2559 |
|
int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2560 |
|
transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth); |
2561 |
|
bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i, &dc); |
2562 |
|
|
2563 |
|
if (bits >= Data->iMinSAD[0]) return bits; |
2564 |
|
} |
2565 |
|
|
2566 |
|
bits += BITS_MULT*xvid_cbpy_tab[cbp>>2].len; |
2567 |
|
|
2568 |
|
/*chroma U */ |
2569 |
|
transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2); |
2570 |
|
bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4, &dc); |
2571 |
|
|
2572 |
|
if (bits >= Data->iMinSAD[0]) return bits; |
2573 |
|
|
2574 |
|
/* chroma V */ |
2575 |
|
transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2); |
2576 |
|
bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5, &dc); |
2577 |
|
|
2578 |
|
bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len; |
2579 |
|
|
2580 |
|
return bits; |
2581 |
} |
} |
2582 |
|
|
2583 |
|
static int |
2584 |
|
findRDgmc(const SearchData * const Data, const IMAGE * const vGMC, const int x, const int y) |
2585 |
|
{ |
2586 |
|
int bits = BITS_MULT*1; /* this one is mcsel */ |
2587 |
|
int cbp = 0, i; |
2588 |
|
int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64; |
2589 |
|
|
2590 |
|
for(i = 0; i < 4; i++) { |
2591 |
|
int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2592 |
|
transfer_8to16subro(in, Data->Cur + s, vGMC->y + s + 16*(x+y*Data->iEdgedWidth), Data->iEdgedWidth); |
2593 |
|
bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i); |
2594 |
|
if (bits >= Data->iMinSAD[0]) return bits; |
2595 |
|
} |
2596 |
|
|
2597 |
|
bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; |
2598 |
|
|
2599 |
|
/*chroma U */ |
2600 |
|
transfer_8to16subro(in, Data->CurU, vGMC->u + 8*(x+y*(Data->iEdgedWidth/2)), Data->iEdgedWidth/2); |
2601 |
|
bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4); |
2602 |
|
|
2603 |
|
if (bits >= Data->iMinSAD[0]) return bits; |
2604 |
|
|
2605 |
|
/* chroma V */ |
2606 |
|
transfer_8to16subro(in, Data->CurV , vGMC->v + 8*(x+y*(Data->iEdgedWidth/2)), Data->iEdgedWidth/2); |
2607 |
|
bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5); |
2608 |
|
|
2609 |
|
bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; |
2610 |
|
|
2611 |
|
*Data->cbp = cbp; |
2612 |
|
|
2613 |
|
return bits; |
2614 |
|
} |
2615 |
|
|
2616 |
|
|
2617 |
|
|
2618 |
|
|
2619 |
|
static __inline void |
2620 |
|
GMEanalyzeMB ( const uint8_t * const pCur, |
2621 |
|
const uint8_t * const pRef, |
2622 |
|
const uint8_t * const pRefH, |
2623 |
|
const uint8_t * const pRefV, |
2624 |
|
const uint8_t * const pRefHV, |
2625 |
|
const int x, |
2626 |
|
const int y, |
2627 |
|
const MBParam * const pParam, |
2628 |
|
MACROBLOCK * const pMBs, |
2629 |
|
SearchData * const Data) |
2630 |
|
{ |
2631 |
|
|
2632 |
|
int i=0; |
2633 |
|
MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width]; |
2634 |
|
|
2635 |
|
Data->iMinSAD[0] = MV_MAX_ERROR; |
2636 |
|
|
2637 |
|
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
2638 |
|
|
2639 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2640 |
|
pParam->width, pParam->height, 16, 0, 0); |
2641 |
|
|
2642 |
|
Data->Cur = pCur + 16*(x + y * pParam->edged_width); |
2643 |
|
Data->RefP[0] = pRef + 16*(x + y * pParam->edged_width); |
2644 |
|
Data->RefP[1] = pRefV + 16*(x + y * pParam->edged_width); |
2645 |
|
Data->RefP[2] = pRefH + 16*(x + y * pParam->edged_width); |
2646 |
|
Data->RefP[3] = pRefHV + 16*(x + y * pParam->edged_width); |
2647 |
|
|
2648 |
|
Data->currentMV[0].x = Data->currentMV[0].y = 0; |
2649 |
|
CheckCandidate16I(0, 0, 255, &i, Data); |
2650 |
|
|
2651 |
|
if ( (Data->predMV.x !=0) || (Data->predMV.y != 0) ) |
2652 |
|
CheckCandidate16I(Data->predMV.x, Data->predMV.y, 255, &i, Data); |
2653 |
|
|
2654 |
|
AdvDiamondSearch(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255); |
2655 |
|
|
2656 |
|
SubpelRefine(Data); |
2657 |
|
|
2658 |
|
|
2659 |
|
/* for QPel halfpel positions are worse than in halfpel mode :( */ |
2660 |
|
/* if (Data->qpel) { |
2661 |
|
Data->currentQMV->x = 2*Data->currentMV->x; |
2662 |
|
Data->currentQMV->y = 2*Data->currentMV->y; |
2663 |
|
Data->qpel_precision = 1; |
2664 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2665 |
|
pParam->width, pParam->height, iFcode, 1, 0); |
2666 |
|
SubpelRefine(Data); |
2667 |
|
} |
2668 |
|
*/ |
2669 |
|
|
2670 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
2671 |
|
pMB->sad16 = Data->iMinSAD[0]; |
2672 |
|
pMB->mode = MODE_INTER; |
2673 |
|
pMB->sad16 += 10*d_mv_bits(pMB->mvs[0].x, pMB->mvs[0].y, Data->predMV, Data->iFcode, 0, 0); |
2674 |
|
return; |
2675 |
} |
} |
2676 |
|
|
2677 |
void |
void |
2678 |
MotionEstimationHinted( MBParam * const pParam, |
GMEanalysis(const MBParam * const pParam, |
2679 |
FRAMEINFO * const current, |
const FRAMEINFO * const current, |
2680 |
FRAMEINFO * const reference, |
const FRAMEINFO * const reference, |
2681 |
const IMAGE * const pRefH, |
const IMAGE * const pRefH, |
2682 |
const IMAGE * const pRefV, |
const IMAGE * const pRefV, |
2683 |
const IMAGE * const pRefHV) |
const IMAGE * const pRefHV) |
2684 |
{ |
{ |
2685 |
|
uint32_t x, y; |
2686 |
MACROBLOCK *const pMBs = current->mbs; |
MACROBLOCK *const pMBs = current->mbs; |
2687 |
const IMAGE *const pCurrent = ¤t->image; |
const IMAGE *const pCurrent = ¤t->image; |
2688 |
const IMAGE *const pRef = &reference->image; |
const IMAGE * const pReference = &reference->image; |
2689 |
|
|
2690 |
uint32_t x, y; |
int32_t iMinSAD[5], temp[5]; |
|
int32_t temp[5]; |
|
|
int32_t iMinSAD[5]; |
|
2691 |
VECTOR currentMV[5]; |
VECTOR currentMV[5]; |
2692 |
SearchData Data; |
SearchData Data; |
2693 |
|
memset(&Data, 0, sizeof(SearchData)); |
2694 |
|
|
2695 |
Data.iEdgedWidth = pParam->edged_width; |
Data.iEdgedWidth = pParam->edged_width; |
2696 |
Data.currentMV = currentMV; |
Data.rounding = pParam->m_rounding_type; |
2697 |
Data.iMinSAD = iMinSAD; |
|
2698 |
Data.temp = temp; |
Data.currentMV = ¤tMV[0]; |
2699 |
|
Data.iMinSAD = &iMinSAD[0]; |
2700 |
Data.iFcode = current->fcode; |
Data.iFcode = current->fcode; |
2701 |
|
Data.temp = temp; |
2702 |
|
|
2703 |
|
CheckCandidate = CheckCandidate16I; |
2704 |
|
|
2705 |
if (sadInit) (*sadInit) (); |
if (sadInit) (*sadInit) (); |
2706 |
|
|
2707 |
for (y = 0; y < pParam->mb_height; y++) { |
for (y = 0; y < pParam->mb_height; y++) { |
2708 |
for (x = 0; x < pParam->mb_width; x++) { |
for (x = 0; x < pParam->mb_width; x++) { |
2709 |
|
|
2710 |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
GMEanalyzeMB(pCurrent->y, pReference->y, pRefH->y, pRefV->y, pRefHV->y, x, y, pParam, pMBs, &Data); |
2711 |
|
} |
2712 |
|
} |
2713 |
|
return; |
2714 |
|
} |
2715 |
|
|
|
//intra mode is copied from the first pass. At least for the time being |
|
|
if ((pMB->mode == MODE_INTRA) || (pMB->mode == MODE_NOT_CODED) ) continue; |
|
2716 |
|
|
2717 |
if (!(current->global_flags & XVID_LUMIMASKING)) { |
WARPPOINTS |
2718 |
pMB->dquant = NO_CHANGE; |
GlobalMotionEst(MACROBLOCK * const pMBs, |
2719 |
pMB->quant = current->quant; } |
const MBParam * const pParam, |
2720 |
|
const FRAMEINFO * const current, |
2721 |
SearchPhinted(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
const FRAMEINFO * const reference, |
2722 |
y, current->motion_flags, pMB->quant, |
const IMAGE * const pRefH, |
2723 |
pParam, pMBs, current->global_flags & XVID_INTER4V, pMB, |
const IMAGE * const pRefV, |
2724 |
&Data); |
const IMAGE * const pRefHV) |
2725 |
|
{ |
2726 |
|
|
2727 |
|
const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs |
2728 |
|
const int deltay=8; |
2729 |
|
const unsigned int gradx=512; // lower bound for gradient in MB (ignore "flat" blocks) |
2730 |
|
const unsigned int grady=512; |
2731 |
|
|
2732 |
|
double sol[4] = { 0., 0., 0., 0. }; |
2733 |
|
|
2734 |
|
WARPPOINTS gmc; |
2735 |
|
|
2736 |
|
uint32_t mx, my; |
2737 |
|
|
2738 |
|
int MBh = pParam->mb_height; |
2739 |
|
int MBw = pParam->mb_width; |
2740 |
|
const int minblocks = 9; //MBh*MBw/32+3; /* just some reasonable number 3% + 3 */ |
2741 |
|
const int maxblocks = MBh*MBw/4; /* just some reasonable number 3% + 3 */ |
2742 |
|
|
2743 |
|
int num=0; |
2744 |
|
int oldnum; |
2745 |
|
|
2746 |
|
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; |
2747 |
|
|
2748 |
|
GMEanalysis(pParam,current, reference, pRefH, pRefV, pRefHV); |
2749 |
|
|
2750 |
|
/* block based ME isn't done, yet, so do a quick presearch */ |
2751 |
|
|
2752 |
|
// filter mask of all blocks |
2753 |
|
|
2754 |
|
for (my = 0; my < (uint32_t)MBh; my++) |
2755 |
|
for (mx = 0; mx < (uint32_t)MBw; mx++) |
2756 |
|
{ |
2757 |
|
const int mbnum = mx + my * MBw; |
2758 |
|
pMBs[mbnum].mcsel = 0; |
2759 |
|
} |
2760 |
|
|
2761 |
|
|
2762 |
|
for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */ |
2763 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */ |
2764 |
|
{ |
2765 |
|
const int mbnum = mx + my * MBw; |
2766 |
|
MACROBLOCK *const pMB = &pMBs[mbnum]; |
2767 |
|
const VECTOR mv = pMB->mvs[0]; |
2768 |
|
|
2769 |
|
/* don't use object boundaries */ |
2770 |
|
if ( (abs(mv.x - (pMB-1)->mvs[0].x) < deltax) |
2771 |
|
&& (abs(mv.y - (pMB-1)->mvs[0].y) < deltay) |
2772 |
|
&& (abs(mv.x - (pMB+1)->mvs[0].x) < deltax) |
2773 |
|
&& (abs(mv.y - (pMB+1)->mvs[0].y) < deltay) |
2774 |
|
&& (abs(mv.x - (pMB-MBw)->mvs[0].x) < deltax) |
2775 |
|
&& (abs(mv.y - (pMB-MBw)->mvs[0].y) < deltay) |
2776 |
|
&& (abs(mv.x - (pMB+MBw)->mvs[0].x) < deltax) |
2777 |
|
&& (abs(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) |
2778 |
|
{ const int iEdgedWidth = pParam->edged_width; |
2779 |
|
const uint8_t *const pCur = current->image.y + 16*(my*iEdgedWidth + mx); |
2780 |
|
if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx ) |
2781 |
|
&& (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) ) |
2782 |
|
{ pMB->mcsel = 1; |
2783 |
|
num++; |
2784 |
} |
} |
2785 |
|
|
2786 |
|
/* only use "structured" blocks */ |
2787 |
} |
} |
2788 |
} |
} |
2789 |
|
emms(); |
2790 |
|
|
2791 |
static __inline int |
/* further filtering would be possible, but during iteration, remaining |
2792 |
MEanalyzeMB ( const uint8_t * const pRef, |
outliers usually are removed, too */ |
2793 |
const uint8_t * const pCur, |
|
2794 |
const int x, |
if (num>= minblocks) |
2795 |
const int y, |
do { /* until convergence */ |
2796 |
const MBParam * const pParam, |
double DtimesF[4]; |
2797 |
const MACROBLOCK * const pMBs, |
double a,b,c,n,invdenom; |
2798 |
MACROBLOCK * const pMB, |
double meanx,meany; |
2799 |
SearchData * const Data) |
|
2800 |
|
a = b = c = n = 0; |
2801 |
|
DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.; |
2802 |
|
for (my = 1; my < (uint32_t)MBh-1; my++) |
2803 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
2804 |
{ |
{ |
2805 |
|
const int mbnum = mx + my * MBw; |
2806 |
|
const VECTOR mv = pMBs[mbnum].mvs[0]; |
2807 |
|
|
2808 |
int i, mask; |
if (!pMBs[mbnum].mcsel) |
2809 |
VECTOR pmv[3]; |
continue; |
2810 |
|
|
2811 |
*(Data->iMinSAD) = MV_MAX_ERROR; |
n++; |
2812 |
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
a += 16*mx+8; |
2813 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
b += 16*my+8; |
2814 |
pParam->width, pParam->height, Data->iFcode); |
c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8); |
2815 |
|
|
2816 |
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
DtimesF[0] += (double)mv.x; |
2817 |
Data->Ref = pRef + (x + y * pParam->edged_width) * 16; |
DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8); |
2818 |
|
DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8); |
2819 |
|
DtimesF[3] += (double)mv.y; |
2820 |
|
} |
2821 |
|
|
2822 |
CheckCandidate = CheckCandidate16no4vI; |
invdenom = a*a+b*b-c*n; |
2823 |
|
|
2824 |
pmv[1].x = EVEN(pMB->mvs[0].x); |
/* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */ |
2825 |
pmv[1].y = EVEN(pMB->mvs[0].y); |
/* D'*E*F has been calculated in the same loop as matrix */ |
|
pmv[0].x = EVEN(Data->predMV.x); |
|
|
pmv[0].y = EVEN(Data->predMV.y); |
|
|
pmv[2].x = pmv[2].y = 0; |
|
2826 |
|
|
2827 |
CheckCandidate16no4vI(pmv[0].x, pmv[0].y, 255, &i, Data); |
sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2]; |
2828 |
if (!(mask = make_mask(pmv, 1))) |
sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3]; |
2829 |
CheckCandidate16no4vI(pmv[1].x, pmv[1].y, mask, &i, Data); |
sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3]; |
2830 |
if (!(mask = make_mask(pmv, 2))) |
sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3]; |
|
CheckCandidate16no4vI(0, 0, mask, &i, Data); |
|
2831 |
|
|
2832 |
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
sol[0] /= invdenom; |
2833 |
|
sol[1] /= invdenom; |
2834 |
|
sol[2] /= invdenom; |
2835 |
|
sol[3] /= invdenom; |
2836 |
|
|
2837 |
|
meanx = meany = 0.; |
2838 |
|
oldnum = 0; |
2839 |
|
for (my = 1; my < (uint32_t)MBh-1; my++) |
2840 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
2841 |
|
{ |
2842 |
|
const int mbnum = mx + my * MBw; |
2843 |
|
const VECTOR mv = pMBs[mbnum].mvs[0]; |
2844 |
|
|
2845 |
pMB->mvs[0] = pMB->mvs[1] |
if (!pMBs[mbnum].mcsel) |
2846 |
= pMB->mvs[2] = pMB->mvs[3] = *Data->currentMV; // all, for future get_pmv() |
continue; |
2847 |
|
|
2848 |
return *(Data->iMinSAD); |
oldnum++; |
2849 |
|
meanx += fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - (double)mv.x ); |
2850 |
|
meany += fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - (double)mv.y ); |
2851 |
} |
} |
2852 |
|
|
2853 |
#define INTRA_THRESH 1350 |
if (4*meanx > oldnum) /* better fit than 0.25 (=1/4pel) is useless */ |
2854 |
#define INTER_THRESH 900 |
meanx /= oldnum; |
2855 |
|
else |
2856 |
|
meanx = 0.25; |
2857 |
|
|
2858 |
|
if (4*meany > oldnum) |
2859 |
|
meany /= oldnum; |
2860 |
|
else |
2861 |
|
meany = 0.25; |
2862 |
|
|
2863 |
|
num = 0; |
2864 |
|
for (my = 0; my < (uint32_t)MBh; my++) |
2865 |
|
for (mx = 0; mx < (uint32_t)MBw; mx++) |
2866 |
|
{ |
2867 |
|
const int mbnum = mx + my * MBw; |
2868 |
|
const VECTOR mv = pMBs[mbnum].mvs[0]; |
2869 |
|
|
2870 |
|
if (!pMBs[mbnum].mcsel) |
2871 |
|
continue; |
2872 |
|
|
2873 |
|
if ( ( fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - (double)mv.x ) > meanx ) |
2874 |
|
|| ( fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - (double)mv.y ) > meany ) ) |
2875 |
|
pMBs[mbnum].mcsel=0; |
2876 |
|
else |
2877 |
|
num++; |
2878 |
|
} |
2879 |
|
|
2880 |
|
} while ( (oldnum != num) && (num>= minblocks) ); |
2881 |
|
|
2882 |
|
if (num < minblocks) |
2883 |
|
{ |
2884 |
|
const int iEdgedWidth = pParam->edged_width; |
2885 |
|
num = 0; |
2886 |
|
|
2887 |
|
/* fprintf(stderr,"Warning! Unreliable GME (%d/%d blocks), falling back to translation.\n",num,MBh*MBw); |
2888 |
|
*/ |
2889 |
|
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; |
2890 |
|
|
2891 |
|
if (!(current->motion_flags & XVID_ME_GME_REFINE)) |
2892 |
|
return gmc; |
2893 |
|
|
2894 |
|
for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */ |
2895 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */ |
2896 |
|
{ |
2897 |
|
const int mbnum = mx + my * MBw; |
2898 |
|
MACROBLOCK *const pMB = &pMBs[mbnum]; |
2899 |
|
const uint8_t *const pCur = current->image.y + 16*(my*iEdgedWidth + mx); |
2900 |
|
if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx ) |
2901 |
|
&& (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) ) |
2902 |
|
{ pMB->mcsel = 1; |
2903 |
|
gmc.duv[0].x += pMB->mvs[0].x; |
2904 |
|
gmc.duv[0].y += pMB->mvs[0].y; |
2905 |
|
num++; |
2906 |
|
} |
2907 |
|
} |
2908 |
|
|
2909 |
|
if (gmc.duv[0].x) |
2910 |
|
gmc.duv[0].x /= num; |
2911 |
|
if (gmc.duv[0].y) |
2912 |
|
gmc.duv[0].y /= num; |
2913 |
|
} else { |
2914 |
|
|
2915 |
|
gmc.duv[0].x=(int)(sol[0]+0.5); |
2916 |
|
gmc.duv[0].y=(int)(sol[3]+0.5); |
2917 |
|
|
2918 |
|
gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5); |
2919 |
|
gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5); |
2920 |
|
|
2921 |
|
gmc.duv[2].x=-gmc.duv[1].y; /* two warp points only */ |
2922 |
|
gmc.duv[2].y=gmc.duv[1].x; |
2923 |
|
} |
2924 |
|
if (num>maxblocks) |
2925 |
|
{ for (my = 1; my < (uint32_t)MBh-1; my++) |
2926 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
2927 |
|
{ |
2928 |
|
const int mbnum = mx + my * MBw; |
2929 |
|
if (pMBs[mbnum-1].mcsel) |
2930 |
|
pMBs[mbnum].mcsel=0; |
2931 |
|
else |
2932 |
|
if (pMBs[mbnum-MBw].mcsel) |
2933 |
|
pMBs[mbnum].mcsel=0; |
2934 |
|
} |
2935 |
|
} |
2936 |
|
return gmc; |
2937 |
|
} |
2938 |
|
|
2939 |
int |
int |
2940 |
MEanalysis( const IMAGE * const pRef, |
GlobalMotionEstRefine( |
2941 |
const IMAGE * const pCurrent, |
WARPPOINTS *const startwp, |
|
MBParam * const pParam, |
|
2942 |
MACROBLOCK * const pMBs, |
MACROBLOCK * const pMBs, |
2943 |
const uint32_t iFcode) |
const MBParam * const pParam, |
2944 |
|
const FRAMEINFO * const current, |
2945 |
|
const FRAMEINFO * const reference, |
2946 |
|
const IMAGE * const pCurr, |
2947 |
|
const IMAGE * const pRef, |
2948 |
|
const IMAGE * const pRefH, |
2949 |
|
const IMAGE * const pRefV, |
2950 |
|
const IMAGE * const pRefHV) |
2951 |
{ |
{ |
2952 |
uint32_t x, y, intra = 0; |
uint8_t* GMCblock = (uint8_t*)malloc(16*pParam->edged_width); |
2953 |
int sSAD = 0; |
WARPPOINTS bestwp=*startwp; |
2954 |
|
WARPPOINTS centerwp,currwp; |
2955 |
|
int gmcminSAD=0; |
2956 |
|
int gmcSAD=0; |
2957 |
|
int direction; |
2958 |
|
// int mx,my; |
2959 |
|
|
2960 |
|
/* use many blocks... */ |
2961 |
|
/* for (my = 0; my < (uint32_t)pParam->mb_height; my++) |
2962 |
|
for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++) |
2963 |
|
{ |
2964 |
|
const int mbnum = mx + my * pParam->mb_width; |
2965 |
|
pMBs[mbnum].mcsel=1; |
2966 |
|
} |
2967 |
|
*/ |
2968 |
|
|
2969 |
|
/* or rather don't use too many blocks... */ |
2970 |
|
/* |
2971 |
|
for (my = 1; my < (uint32_t)MBh-1; my++) |
2972 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
2973 |
|
{ |
2974 |
|
const int mbnum = mx + my * MBw; |
2975 |
|
if (MBmask[mbnum-1]) |
2976 |
|
MBmask[mbnum-1]=0; |
2977 |
|
else |
2978 |
|
if (MBmask[mbnum-MBw]) |
2979 |
|
MBmask[mbnum-1]=0; |
2980 |
|
|
2981 |
|
} |
2982 |
|
*/ |
2983 |
|
gmcminSAD = globalSAD(&bestwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
2984 |
|
|
2985 |
|
if ( (reference->coding_type == S_VOP) |
2986 |
|
&& ( (reference->warp.duv[1].x != bestwp.duv[1].x) |
2987 |
|
|| (reference->warp.duv[1].y != bestwp.duv[1].y) |
2988 |
|
|| (reference->warp.duv[0].x != bestwp.duv[0].x) |
2989 |
|
|| (reference->warp.duv[0].y != bestwp.duv[0].y) |
2990 |
|
|| (reference->warp.duv[2].x != bestwp.duv[2].x) |
2991 |
|
|| (reference->warp.duv[2].y != bestwp.duv[2].y) ) ) |
2992 |
|
{ |
2993 |
|
gmcSAD = globalSAD(&reference->warp, pParam, pMBs, |
2994 |
|
current, pRef, pCurr, GMCblock); |
2995 |
|
|
2996 |
|
if (gmcSAD < gmcminSAD) |
2997 |
|
{ bestwp = reference->warp; |
2998 |
|
gmcminSAD = gmcSAD; |
2999 |
|
} |
3000 |
|
} |
3001 |
|
|
3002 |
VECTOR currentMV; |
do { |
3003 |
int32_t iMinSAD; |
direction = 0; |
3004 |
SearchData Data; |
centerwp = bestwp; |
|
Data.iEdgedWidth = pParam->edged_width; |
|
|
Data.currentMV = ¤tMV; |
|
|
Data.iMinSAD = &iMinSAD; |
|
|
Data.iFcode = iFcode; |
|
|
Data.iQuant = 2; |
|
3005 |
|
|
3006 |
if (sadInit) (*sadInit) (); |
currwp = centerwp; |
3007 |
|
|
3008 |
for (y = 0; y < pParam->mb_height-1; y++) { |
currwp.duv[0].x--; |
3009 |
for (x = 0; x < pParam->mb_width; x++) { |
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3010 |
int sad, dev; |
if (gmcSAD < gmcminSAD) |
3011 |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
{ bestwp = currwp; |
3012 |
|
gmcminSAD = gmcSAD; |
3013 |
|
direction = 1; |
3014 |
|
} |
3015 |
|
else |
3016 |
|
{ |
3017 |
|
currwp = centerwp; currwp.duv[0].x++; |
3018 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3019 |
|
if (gmcSAD < gmcminSAD) |
3020 |
|
{ bestwp = currwp; |
3021 |
|
gmcminSAD = gmcSAD; |
3022 |
|
direction = 2; |
3023 |
|
} |
3024 |
|
} |
3025 |
|
if (direction) continue; |
3026 |
|
|
3027 |
sad = MEanalyzeMB(pRef->y, pCurrent->y, x, y, |
currwp = centerwp; currwp.duv[0].y--; |
3028 |
pParam, pMBs, pMB, &Data); |
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3029 |
|
if (gmcSAD < gmcminSAD) |
3030 |
|
{ bestwp = currwp; |
3031 |
|
gmcminSAD = gmcSAD; |
3032 |
|
direction = 4; |
3033 |
|
} |
3034 |
|
else |
3035 |
|
{ |
3036 |
|
currwp = centerwp; currwp.duv[0].y++; |
3037 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3038 |
|
if (gmcSAD < gmcminSAD) |
3039 |
|
{ bestwp = currwp; |
3040 |
|
gmcminSAD = gmcSAD; |
3041 |
|
direction = 8; |
3042 |
|
} |
3043 |
|
} |
3044 |
|
if (direction) continue; |
3045 |
|
|
3046 |
if ( x != 0 && y != 0 && x != pParam->mb_width-1 ) { //no edge macroblocks, they just don't work |
currwp = centerwp; currwp.duv[1].x++; |
3047 |
if (sad > INTRA_THRESH) { |
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3048 |
dev = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
if (gmcSAD < gmcminSAD) |
3049 |
pParam->edged_width); |
{ bestwp = currwp; |
3050 |
if (dev + INTRA_THRESH < sad) intra++; |
gmcminSAD = gmcSAD; |
3051 |
if (intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return 2; // I frame |
direction = 32; |
3052 |
} |
} |
3053 |
sSAD += sad; |
currwp.duv[2].y++; |
3054 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3055 |
|
if (gmcSAD < gmcminSAD) |
3056 |
|
{ bestwp = currwp; |
3057 |
|
gmcminSAD = gmcSAD; |
3058 |
|
direction = 1024; |
3059 |
} |
} |
3060 |
|
|
3061 |
|
currwp = centerwp; currwp.duv[1].x--; |
3062 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3063 |
|
if (gmcSAD < gmcminSAD) |
3064 |
|
{ bestwp = currwp; |
3065 |
|
gmcminSAD = gmcSAD; |
3066 |
|
direction = 16; |
3067 |
} |
} |
3068 |
|
else |
3069 |
|
{ |
3070 |
|
currwp = centerwp; currwp.duv[1].x++; |
3071 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3072 |
|
if (gmcSAD < gmcminSAD) |
3073 |
|
{ bestwp = currwp; |
3074 |
|
gmcminSAD = gmcSAD; |
3075 |
|
direction = 32; |
3076 |
} |
} |
3077 |
sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); |
} |
3078 |
if (sSAD > INTER_THRESH ) return 1; //P frame |
if (direction) continue; |
3079 |
emms(); |
|
|
return 0; // B frame |
|
3080 |
|
|
3081 |
|
currwp = centerwp; currwp.duv[1].y--; |
3082 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3083 |
|
if (gmcSAD < gmcminSAD) |
3084 |
|
{ bestwp = currwp; |
3085 |
|
gmcminSAD = gmcSAD; |
3086 |
|
direction = 64; |
3087 |
|
} |
3088 |
|
else |
3089 |
|
{ |
3090 |
|
currwp = centerwp; currwp.duv[1].y++; |
3091 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3092 |
|
if (gmcSAD < gmcminSAD) |
3093 |
|
{ bestwp = currwp; |
3094 |
|
gmcminSAD = gmcSAD; |
3095 |
|
direction = 128; |
3096 |
} |
} |
3097 |
|
} |
3098 |
|
if (direction) continue; |
3099 |
|
|
3100 |
|
currwp = centerwp; currwp.duv[2].x--; |
3101 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3102 |
|
if (gmcSAD < gmcminSAD) |
3103 |
|
{ bestwp = currwp; |
3104 |
|
gmcminSAD = gmcSAD; |
3105 |
|
direction = 256; |
3106 |
|
} |
3107 |
|
else |
3108 |
|
{ |
3109 |
|
currwp = centerwp; currwp.duv[2].x++; |
3110 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3111 |
|
if (gmcSAD < gmcminSAD) |
3112 |
|
{ bestwp = currwp; |
3113 |
|
gmcminSAD = gmcSAD; |
3114 |
|
direction = 512; |
3115 |
|
} |
3116 |
|
} |
3117 |
|
if (direction) continue; |
3118 |
|
|
3119 |
|
currwp = centerwp; currwp.duv[2].y--; |
3120 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3121 |
|
if (gmcSAD < gmcminSAD) |
3122 |
|
{ bestwp = currwp; |
3123 |
|
gmcminSAD = gmcSAD; |
3124 |
|
direction = 1024; |
3125 |
|
} |
3126 |
|
else |
3127 |
|
{ |
3128 |
|
currwp = centerwp; currwp.duv[2].y++; |
3129 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3130 |
|
if (gmcSAD < gmcminSAD) |
3131 |
|
{ bestwp = currwp; |
3132 |
|
gmcminSAD = gmcSAD; |
3133 |
|
direction = 2048; |
3134 |
|
} |
3135 |
|
} |
3136 |
|
} while (direction); |
3137 |
|
free(GMCblock); |
3138 |
|
|
3139 |
|
*startwp = bestwp; |
3140 |
|
|
3141 |
|
return gmcminSAD; |
3142 |
|
} |
3143 |
|
|
3144 |
|
int |
3145 |
|
globalSAD(const WARPPOINTS *const wp, |
3146 |
|
const MBParam * const pParam, |
3147 |
|
const MACROBLOCK * const pMBs, |
3148 |
|
const FRAMEINFO * const current, |
3149 |
|
const IMAGE * const pRef, |
3150 |
|
const IMAGE * const pCurr, |
3151 |
|
uint8_t *const GMCblock) |
3152 |
|
{ |
3153 |
|
NEW_GMC_DATA gmc_data; |
3154 |
|
int iSAD, gmcSAD=0; |
3155 |
|
int num=0; |
3156 |
|
unsigned int mx, my; |
3157 |
|
|
3158 |
|
generate_GMCparameters( 3, 3, wp, pParam->width, pParam->height, &gmc_data); |
3159 |
|
|
3160 |
|
for (my = 0; my < (uint32_t)pParam->mb_height; my++) |
3161 |
|
for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++) { |
3162 |
|
|
3163 |
|
const int mbnum = mx + my * pParam->mb_width; |
3164 |
|
const int iEdgedWidth = pParam->edged_width; |
3165 |
|
|
3166 |
|
if (!pMBs[mbnum].mcsel) |
3167 |
|
continue; |
3168 |
|
|
3169 |
|
gmc_data.predict_16x16(&gmc_data, GMCblock, |
3170 |
|
pRef->y, |
3171 |
|
iEdgedWidth, |
3172 |
|
iEdgedWidth, |
3173 |
|
mx, my, |
3174 |
|
pParam->m_rounding_type); |
3175 |
|
|
3176 |
|
iSAD = sad16 ( pCurr->y + 16*(my*iEdgedWidth + mx), |
3177 |
|
GMCblock , iEdgedWidth, 65536); |
3178 |
|
iSAD -= pMBs[mbnum].sad16; |
3179 |
|
|
3180 |
|
if (iSAD<0) |
3181 |
|
gmcSAD += iSAD; |
3182 |
|
num++; |
3183 |
|
} |
3184 |
|
return gmcSAD; |
3185 |
|
} |
3186 |
|
|