Annotation of /branches/dev-api-3/xvidcore/src/motion/motion_est.c

Revision 658 - (view) (download)

1 :	Isibaar	3	/**************************************************************************
2 :			*
3 :	chl	259	* XVID MPEG-4 VIDEO CODEC
4 :			* motion estimation
5 :			*
6 :			* This program is an implementation of a part of one or more MPEG-4
7 :			* Video tools as specified in ISO/IEC 14496-2 standard. Those intending
8 :			* to use this software module in hardware or software products are
9 :			* advised that its use may infringe existing patents or copyrights, and
10 :			* any such use would be at such party's own risk. The original
11 :			* developer of this software module and his/her company, and subsequent
12 :			* editors and their companies, will have no liability for use of this
13 :			* software or modifications or derivatives thereof.
14 :			*
15 :			* This program is free software; you can redistribute it and/or modify
16 :			* it under the terms of the GNU General Public License as published by
17 :			* the Free Software Foundation; either version 2 of the License, or
18 :			* (at your option) any later version.
19 :			*
20 :			* This program is distributed in the hope that it will be useful,
21 :			* but WITHOUT ANY WARRANTY; without even the implied warranty of
22 :			* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
23 :			* GNU General Public License for more details.
24 :			*
25 :			* You should have received a copy of the GNU General Public License
26 :			* along with this program; if not, write to the Free Software
27 :			* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
28 :			*
29 :			*************************************************************************/
30 :
31 :	Isibaar	3	#include <assert.h>
32 :			#include <stdio.h>
33 :	chl	96	#include <stdlib.h>
34 :	Isibaar	3
35 :			#include "../encoder.h"
36 :			#include "../utils/mbfunctions.h"
37 :			#include "../prediction/mbprediction.h"
38 :			#include "../global.h"
39 :			#include "../utils/timer.h"
40 :	Isibaar	580	#include "../image/interpolate8x8.h"
41 :	chl	530	#include "motion_est.h"
42 :	suxen_drol	118	#include "motion.h"
43 :	Isibaar	3	#include "sad.h"
44 :	Isibaar	539	#include "../utils/emms.h"
45 :	Isibaar	3
46 :	chl	530	#define INITIAL_SKIP_THRESH (10)
47 :			#define FINAL_SKIP_THRESH (50)
48 :			#define MAX_SAD00_FOR_SKIP (20)
49 :	Isibaar	539	#define MAX_CHROMA_SAD_FOR_SKIP (22)
50 :			#define SKIP_THRESH_B (25)
51 :	Isibaar	3
52 :	chl	530	#define CHECK_CANDIDATE(X,Y,D) { \
53 :			(*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); }
54 :	Isibaar	3
55 :	Isibaar	580	#define GET_REFERENCE(X, Y, REF) { \
56 :	chl	596	switch ( (((X)&1)<<1) + ((Y)&1) ) \
57 :	Isibaar	580	{ \
58 :	syskin	601	case 0 : REF = (uint8_t )data->Ref + (X)/2 + ((Y)/2)(data->iEdgedWidth); break; \
59 :			case 1 : REF = (uint8_t )data->RefV + (X)/2 + (((Y)-1)/2)(data->iEdgedWidth); break; \
60 :			case 2 : REF = (uint8_t )data->RefH + ((X)-1)/2 + ((Y)/2)(data->iEdgedWidth); break; \
61 :			default : REF = (uint8_t )data->RefHV + ((X)-1)/2 + (((Y)-1)/2)(data->iEdgedWidth); break; \
62 :	Isibaar	580	} \
63 :			}
64 :	syskin	658	// I hate those macros :/
65 :			#define GET_REFERENCE2(X, Y, REF) { \
66 :			switch ( (((X)&1)<<1) + ((Y)&1) ) \
67 :			{ \
68 :			case 0 : REF = (uint8_t )data->bRef + (X)/2 + ((Y)/2)(data->iEdgedWidth); break; \
69 :			case 1 : REF = (uint8_t )data->bRefV + (X)/2 + (((Y)-1)/2)(data->iEdgedWidth); break; \
70 :			case 2 : REF = (uint8_t )data->bRefH + ((X)-1)/2 + ((Y)/2)(data->iEdgedWidth); break; \
71 :			default : REF = (uint8_t )data->bRefHV + ((X)-1)/2 + (((Y)-1)/2)(data->iEdgedWidth); break; \
72 :			} \
73 :			}
74 :	Isibaar	580
75 :	syskin	658
76 :	chl	530	#define iDiamondSize 2
77 :	chl	141
78 :	chl	530	static __inline int
79 :			d_mv_bits(int x, int y, const uint32_t iFcode)
80 :	Isibaar	3	{
81 :	chl	530	int xb, yb;
82 :
83 :			if (x == 0) xb = 1;
84 :			else {
85 :			if (x < 0) x = -x;
86 :			x += (1 << (iFcode - 1)) - 1;
87 :			x >>= (iFcode - 1);
88 :			if (x > 32) x = 32;
89 :			xb = mvtab[x] + iFcode;
90 :	edgomez	195	}
91 :	Isibaar	3
92 :	chl	530	if (y == 0) yb = 1;
93 :			else {
94 :			if (y < 0) y = -y;
95 :			y += (1 << (iFcode - 1)) - 1;
96 :			y >>= (iFcode - 1);
97 :			if (y > 32) y = 32;
98 :			yb = mvtab[y] + iFcode;
99 :			}
100 :			return xb + yb;
101 :	Isibaar	3	}
102 :
103 :	syskin	628	static int32_t
104 :			ChromaSAD(int dx, int dy, const SearchData * const data)
105 :	Isibaar	3	{
106 :	syskin	628	int sad;
107 :			dx = (dx >> 1) + roundtab_79[dx & 0x3];
108 :			dy = (dy >> 1) + roundtab_79[dy & 0x3];
109 :	Isibaar	3
110 :	syskin	628	switch (((dx & 1) << 1) + (dy & 1)) { // ((dx%2)?2:0)+((dy%2)?1:0)
111 :			case 0:
112 :			sad = sad8(data->CurU, data->RefCU + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2);
113 :			sad += sad8(data->CurV, data->RefCV + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2);
114 :			break;
115 :			case 1:
116 :			dx = dx / 2; dy = (dy - 1) / 2;
117 :			sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2);
118 :			sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2);
119 :			break;
120 :			case 2:
121 :			dx = (dx - 1) / 2; dy = dy / 2;
122 :			sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2);
123 :			sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2);
124 :			break;
125 :			default:
126 :			dx = (dx - 1) / 2; dy = (dy - 1) / 2;
127 :			interpolate8x8_halfpel_hv(data->RefQ,
128 :			data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2,
129 :			data->rounding);
130 :			sad = sad8(data->CurU, data->RefQ, data->iEdgedWidth/2);
131 :			interpolate8x8_halfpel_hv(data->RefQ,
132 :			data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2,
133 :			data->rounding);
134 :			sad += sad8(data->CurV, data->RefQ, data->iEdgedWidth/2);
135 :			break;
136 :	chl	530	}
137 :	syskin	628	return sad;
138 :			}
139 :	Isibaar	3
140 :	Isibaar	600
141 :	syskin	628	/* CHECK_CANDIATE FUNCTIONS START */
142 :	chl	184
143 :
144 :	chl	530	static void
145 :	syskin	628	CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
146 :	Isibaar	606	{
147 :			int t;
148 :			const uint8_t * Reference;
149 :
150 :			if (( x > data->max_dx) \|\| ( x < data->min_dx)
151 :			\|\| ( y > data->max_dy) \|\| (y < data->min_dy)) return;
152 :
153 :			switch ( ((x&1)<<1) + (y&1) ) {
154 :			case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break;
155 :			case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break;
156 :			case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break;
157 :			default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break;
158 :			}
159 :
160 :			data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
161 :
162 :	syskin	628	if (data->qpel) t = d_mv_bits(2x - data->predQMV.x, 2y - data->predQMV.y, data->iFcode);
163 :			else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode);
164 :	Isibaar	606
165 :	syskin	628	data->temp[0] += (data->lambda16 * t * data->temp[0])/1000;
166 :			data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100;
167 :
168 :			if (data->chroma) data->temp[0] += ChromaSAD(x, y, data);
169 :
170 :	Isibaar	606	if (data->temp[0] < data->iMinSAD[0]) {
171 :			data->iMinSAD[0] = data->temp[0];
172 :			data->currentMV[0].x = x; data->currentMV[0].y = y;
173 :			*dir = Direction; }
174 :
175 :			if (data->temp[1] < data->iMinSAD[1]) {
176 :			data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
177 :			if (data->temp[2] < data->iMinSAD[2]) {
178 :			data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
179 :			if (data->temp[3] < data->iMinSAD[3]) {
180 :			data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
181 :			if (data->temp[4] < data->iMinSAD[4]) {
182 :			data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
183 :
184 :			}
185 :
186 :			static void
187 :	chl	530	CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
188 :			{
189 :			int32_t sad;
190 :			const uint8_t * Reference;
191 :	chl	326
192 :	chl	530	if (( x > data->max_dx) \|\| ( x < data->min_dx)
193 :			\|\| ( y > data->max_dy) \|\| (y < data->min_dy)) return;
194 :	Isibaar	3
195 :	chl	530	switch ( ((x&1)<<1) + (y&1) )
196 :			{
197 :			case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break;
198 :			case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break;
199 :			case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break;
200 :			default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break;
201 :			}
202 :	Isibaar	3
203 :	syskin	628	sad = sad16(data->Cur, Reference, data->iEdgedWidth, MV_MAX_ERROR);
204 :	syskin	658	if (data->qpel) //only to be used in b-frames' ME
205 :			sad += (data->lambda16 * d_mv_bits(2x - data->predMV.x, 2y - data->predMV.y, data->iFcode) * sad)/1000;
206 :			else
207 :			sad += (data->lambda16 * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode) * sad)/1000;
208 :	Isibaar	3
209 :	chl	530	if (sad < *(data->iMinSAD)) {
210 :			*(data->iMinSAD) = sad;
211 :			data->currentMV[0].x = x; data->currentMV[0].y = y;
212 :			*dir = Direction; }
213 :	Isibaar	3	}
214 :
215 :	chl	530	static void
216 :	Isibaar	580	CheckCandidate16_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
217 :
218 :			// CheckCandidate16 variant which expects x and y in quarter pixel resolution
219 :			// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!)
220 :			// around currentMV!
221 :			{
222 :			int t;
223 :	syskin	601	uint8_t * Reference = (uint8_t *)data->RefQ;
224 :	Isibaar	580	const uint8_t ref1, ref2, ref3, ref4;
225 :			VECTOR halfpelMV = *(data->currentMV);
226 :
227 :			int32_t iEdgedWidth = data->iEdgedWidth;
228 :			uint32_t rounding = data->rounding;
229 :
230 :			if (( x > data->max_dx) \|\| ( x < data->min_dx)
231 :			\|\| ( y > data->max_dy) \|\| (y < data->min_dy)) return;
232 :
233 :	Isibaar	606	GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this refenrence is used in all cases
234 :	Isibaar	580	switch( ((x&1)<<1) + (y&1) )
235 :			{
236 :			case 0: // pure halfpel position - shouldn't happen during a refinement step
237 :	syskin	601	GET_REFERENCE(halfpelMV.x, halfpelMV.y, Reference);
238 :	Isibaar	580	break;
239 :
240 :			case 1: // x halfpel, y qpel - top or bottom during qpel refinement
241 :			GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2);
242 :			interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding);
243 :			interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding);
244 :			interpolate8x8_avg2(Reference+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding);
245 :			interpolate8x8_avg2(Reference+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding);
246 :			break;
247 :
248 :			case 2: // x qpel, y halfpel - left or right during qpel refinement
249 :			GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2);
250 :			interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding);
251 :			interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding);
252 :			interpolate8x8_avg2(Reference+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding);
253 :			interpolate8x8_avg2(Reference+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding);
254 :			break;
255 :
256 :			default: // x and y in qpel resolution - the "corners" (top left/right and
257 :			// bottom left/right) during qpel refinement
258 :			GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2);
259 :			GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3);
260 :			GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4);
261 :
262 :			interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
263 :			interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding);
264 :			interpolate8x8_avg4(Reference+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8iEdgedWidth, ref3+8iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding);
265 :			interpolate8x8_avg4(Reference+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8iEdgedWidth+8, ref3+8iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding);
266 :			break;
267 :			}
268 :
269 :	syskin	594	data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp+1);
270 :	Isibaar	580
271 :			t = d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode);
272 :	syskin	628	data->temp[0] += (data->lambda16 * t * data->temp[0])/1000;
273 :			data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100;
274 :	Isibaar	580
275 :	syskin	628	if (data->chroma)
276 :			data->temp[0] += ChromaSAD(x/2, y/2, data);
277 :
278 :	Isibaar	580	if (data->temp[0] < data->iMinSAD[0]) {
279 :			data->iMinSAD[0] = data->temp[0];
280 :			data->currentQMV[0].x = x; data->currentQMV[0].y = y;
281 :	syskin	594	/* dir = Direction;/ }
282 :	Isibaar	580
283 :			if (data->temp[1] < data->iMinSAD[1]) {
284 :			data->iMinSAD[1] = data->temp[1]; data->currentQMV[1].x = x; data->currentQMV[1].y = y; }
285 :			if (data->temp[2] < data->iMinSAD[2]) {
286 :			data->iMinSAD[2] = data->temp[2]; data->currentQMV[2].x = x; data->currentQMV[2].y = y; }
287 :			if (data->temp[3] < data->iMinSAD[3]) {
288 :			data->iMinSAD[3] = data->temp[3]; data->currentQMV[3].x = x; data->currentQMV[3].y = y; }
289 :			if (data->temp[4] < data->iMinSAD[4]) {
290 :			data->iMinSAD[4] = data->temp[4]; data->currentQMV[4].x = x; data->currentQMV[4].y = y; }
291 :			}
292 :
293 :			static void
294 :	syskin	658	CheckCandidate16no4v_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
295 :
296 :			// CheckCandidate16no4v variant which expects x and y in quarter pixel resolution
297 :			// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!)
298 :			// around currentMV!
299 :			// this function is for B-frames' search only
300 :			{
301 :			uint8_t * Reference = (uint8_t *)data->RefQ;
302 :			const uint8_t ref1, ref2, ref3, ref4;
303 :			VECTOR halfpelMV = *(data->currentMV);
304 :
305 :			int32_t iEdgedWidth = data->iEdgedWidth;
306 :			int32_t sad;
307 :
308 :			if (( x > data->max_dx) \|\| ( x < data->min_dx)
309 :			\|\| ( y > data->max_dy) \|\| (y < data->min_dy)) return;
310 :
311 :			GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this refenrence is used in all cases
312 :			switch( ((x&1)<<1) + (y&1) )
313 :			{
314 :			case 0: // pure halfpel position - shouldn't happen during a refinement step
315 :			GET_REFERENCE(halfpelMV.x, halfpelMV.y, Reference);
316 :			break;
317 :
318 :			case 1: // x halfpel, y qpel - top or bottom during qpel refinement
319 :			GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2);
320 :			interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, 0);
321 :			interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, 0);
322 :			interpolate8x8_avg2(Reference+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0);
323 :			interpolate8x8_avg2(Reference+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0);
324 :			break;
325 :
326 :			case 2: // x qpel, y halfpel - left or right during qpel refinement
327 :			GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2);
328 :			interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, 0);
329 :			interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, 0);
330 :			interpolate8x8_avg2(Reference+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0);
331 :			interpolate8x8_avg2(Reference+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0);
332 :			break;
333 :
334 :			default: // x and y in qpel resolution - the "corners" (top left/right and
335 :			// bottom left/right) during qpel refinement
336 :			GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2);
337 :			GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3);
338 :			GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4);
339 :
340 :			interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, 0);
341 :			interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, 0);
342 :			interpolate8x8_avg4(Reference+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8iEdgedWidth, ref3+8iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, 0);
343 :			interpolate8x8_avg4(Reference+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8iEdgedWidth+8, ref3+8iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, 0);
344 :			break;
345 :			}
346 :
347 :			sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096);
348 :			sad += (data->lambda16 * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode) * sad)/1000;
349 :
350 :			if (sad < data->iMinSAD[0]) {
351 :			data->iMinSAD[0] = sad;
352 :			data->currentQMV[0].x = x; data->currentQMV[0].y = y;
353 :			/* dir = Direction;/ }
354 :			}
355 :
356 :			static void
357 :	Isibaar	539	CheckCandidate16no4vI(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
358 :			{
359 :	syskin	658	// maximum speed - for P/B/I decision
360 :	Isibaar	539	int32_t sad;
361 :
362 :			if (( x > data->max_dx) \|\| ( x < data->min_dx)
363 :			\|\| ( y > data->max_dy) \|\| (y < data->min_dy)) return;
364 :
365 :	syskin	601	sad = sad16(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth),
366 :	Isibaar	539	data->iEdgedWidth, 256*4096);
367 :
368 :			if (sad < *(data->iMinSAD)) {
369 :			*(data->iMinSAD) = sad;
370 :			data->currentMV[0].x = x; data->currentMV[0].y = y;
371 :			*dir = Direction; }
372 :			}
373 :
374 :
375 :			static void
376 :	chl	530	CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data)
377 :			{
378 :			int32_t sad;
379 :			const int xb = data->currentMV[1].x;
380 :			const int yb = data->currentMV[1].y;
381 :			const uint8_t ReferenceF, ReferenceB;
382 :	Isibaar	3
383 :	chl	530	if (( xf > data->max_dx) \|\| ( xf < data->min_dx)
384 :			\|\| ( yf > data->max_dy) \|\| (yf < data->min_dy)) return;
385 :	Isibaar	3
386 :	chl	530	switch ( ((xf&1)<<1) + (yf&1) ) {
387 :			case 0 : ReferenceF = data->Ref + xf/2 + (yf/2)*(data->iEdgedWidth); break;
388 :			case 1 : ReferenceF = data->RefV + xf/2 + ((yf-1)/2)*(data->iEdgedWidth); break;
389 :			case 2 : ReferenceF = data->RefH + (xf-1)/2 + (yf/2)*(data->iEdgedWidth); break;
390 :			default : ReferenceF = data->RefHV + (xf-1)/2 + ((yf-1)/2)*(data->iEdgedWidth); break;
391 :			}
392 :	edgomez	195
393 :	chl	530	switch ( ((xb&1)<<1) + (yb&1) ) {
394 :			case 0 : ReferenceB = data->bRef + xb/2 + (yb/2)*(data->iEdgedWidth); break;
395 :			case 1 : ReferenceB = data->bRefV + xb/2 + ((yb-1)/2)*(data->iEdgedWidth); break;
396 :			case 2 : ReferenceB = data->bRefH + (xb-1)/2 + (yb/2)*(data->iEdgedWidth); break;
397 :			default : ReferenceB = data->bRefHV + (xb-1)/2 + ((yb-1)/2)*(data->iEdgedWidth); break;
398 :			}
399 :	Isibaar	3
400 :	syskin	628	sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
401 :
402 :	syskin	658	if (data->qpel)
403 :			sad += (data->lambda16 *
404 :			( d_mv_bits(2xf - data->predMV.x, 2yf - data->predMV.y, data->iFcode) +
405 :			d_mv_bits(2xb - data->bpredMV.x, 2yb - data->bpredMV.y, data->iFcode)) * sad)/1000;
406 :			else
407 :			sad += (data->lambda16 *
408 :	chl	530	( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) +
409 :	syskin	628	d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode)) * sad)/1000;
410 :	Isibaar	3
411 :	chl	530	if (sad < *(data->iMinSAD)) {
412 :			*(data->iMinSAD) = sad;
413 :			data->currentMV->x = xf; data->currentMV->y = yf;
414 :			*dir = Direction; }
415 :	Isibaar	3	}
416 :
417 :	syskin	658
418 :	chl	530	static void
419 :	syskin	658	CheckCandidateInt_qpel(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data)
420 :			{
421 :			// CheckCandidateInt variant which expects x and y in quarter pixel resolution
422 :
423 :			int32_t sad;
424 :			const int xb = data->currentQMV[1].x;
425 :			const int yb = data->currentQMV[1].y;
426 :			uint8_t * ReferenceF = (uint8_t *)data->RefQ;
427 :			uint8_t * ReferenceB = (uint8_t *)data->RefQ + 16;
428 :			const uint8_t ref1, ref2, ref3, ref4;
429 :			VECTOR halfpelMV;
430 :			const int32_t iEdgedWidth = data->iEdgedWidth;
431 :
432 :			if (( xf > data->max_dx) \|\| ( xf < data->min_dx)
433 :			\|\| ( yf > data->max_dy) \|\| (yf < data->min_dy)) return;
434 :
435 :			halfpelMV.x = xf/2; //forward first
436 :			halfpelMV.y = yf/2;
437 :			GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this reference is used in all cases
438 :			switch( ((xf&1)<<1) + (yf&1) )
439 :			{
440 :			case 0: // pure halfpel position - shouldn't happen during a refinement step
441 :			GET_REFERENCE(halfpelMV.x, halfpelMV.y, ReferenceF);
442 :			break;
443 :
444 :			case 1: // x halfpel, y qpel - top or bottom during qpel refinement
445 :			GET_REFERENCE(halfpelMV.x, yf - halfpelMV.y, ref2);
446 :			interpolate8x8_avg2(ReferenceF, ref1, ref2, iEdgedWidth, 0);
447 :			interpolate8x8_avg2(ReferenceF+8, ref1+8, ref2+8, iEdgedWidth, 0);
448 :			interpolate8x8_avg2(ReferenceF+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0);
449 :			interpolate8x8_avg2(ReferenceF+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0);
450 :			break;
451 :
452 :			case 2: // x qpel, y halfpel - left or right during qpel refinement
453 :			GET_REFERENCE(xf - halfpelMV.x, halfpelMV.y, ref2);
454 :			interpolate8x8_avg2(ReferenceF, ref1, ref2, iEdgedWidth, 0);
455 :			interpolate8x8_avg2(ReferenceF+8, ref1+8, ref2+8, iEdgedWidth, 0);
456 :			interpolate8x8_avg2(ReferenceF+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0);
457 :			interpolate8x8_avg2(ReferenceF+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0);
458 :			break;
459 :
460 :			default: // x and y in qpel resolution - the "corners" (top left/right and
461 :			// bottom left/right) during qpel refinement
462 :			GET_REFERENCE(halfpelMV.x, yf - halfpelMV.y, ref2);
463 :			GET_REFERENCE(xf - halfpelMV.x, halfpelMV.y, ref3);
464 :			GET_REFERENCE(xf - halfpelMV.x, yf - halfpelMV.y, ref4);
465 :
466 :			interpolate8x8_avg4(ReferenceF, ref1, ref2, ref3, ref4, iEdgedWidth, 0);
467 :			interpolate8x8_avg4(ReferenceF+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, 0);
468 :			interpolate8x8_avg4(ReferenceF+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8iEdgedWidth, ref3+8iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, 0);
469 :			interpolate8x8_avg4(ReferenceF+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8iEdgedWidth+8, ref3+8iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, 0);
470 :			break;
471 :			}
472 :
473 :			halfpelMV.x = xb/2; //backward
474 :			halfpelMV.y = yb/2;
475 :			GET_REFERENCE2(halfpelMV.x, halfpelMV.y, ref1); // this reference is used in all cases
476 :			switch( ((xb&1)<<1) + (yb&1) )
477 :			{
478 :			case 0: // pure halfpel position - shouldn't happen during a refinement step
479 :			GET_REFERENCE2(halfpelMV.x, halfpelMV.y, ReferenceB);
480 :			break;
481 :
482 :			case 1: // x halfpel, y qpel - top or bottom during qpel refinement
483 :			GET_REFERENCE2(halfpelMV.x, yb - halfpelMV.y, ref2);
484 :			interpolate8x8_avg2(ReferenceB, ref1, ref2, iEdgedWidth, 0);
485 :			interpolate8x8_avg2(ReferenceB+8, ref1+8, ref2+8, iEdgedWidth, 0);
486 :			interpolate8x8_avg2(ReferenceB+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0);
487 :			interpolate8x8_avg2(ReferenceB+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0);
488 :			break;
489 :
490 :			case 2: // x qpel, y halfpel - left or right during qpel refinement
491 :			GET_REFERENCE2(xb - halfpelMV.x, halfpelMV.y, ref2);
492 :			interpolate8x8_avg2(ReferenceB, ref1, ref2, iEdgedWidth, 0);
493 :			interpolate8x8_avg2(ReferenceB+8, ref1+8, ref2+8, iEdgedWidth, 0);
494 :			interpolate8x8_avg2(ReferenceB+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0);
495 :			interpolate8x8_avg2(ReferenceB+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0);
496 :			break;
497 :
498 :			default: // x and y in qpel resolution - the "corners" (top left/right and
499 :			// bottom left/right) during qpel refinement
500 :			GET_REFERENCE2(halfpelMV.x, yb - halfpelMV.y, ref2);
501 :			GET_REFERENCE2(xb - halfpelMV.x, halfpelMV.y, ref3);
502 :			GET_REFERENCE2(xb - halfpelMV.x, yb - halfpelMV.y, ref4);
503 :
504 :			interpolate8x8_avg4(ReferenceB, ref1, ref2, ref3, ref4, iEdgedWidth, 0);
505 :			interpolate8x8_avg4(ReferenceB+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, 0);
506 :			interpolate8x8_avg4(ReferenceB+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8iEdgedWidth, ref3+8iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, 0);
507 :			interpolate8x8_avg4(ReferenceB+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8iEdgedWidth+8, ref3+8iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, 0);
508 :			break;
509 :			}
510 :
511 :			sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
512 :
513 :			sad += (data->lambda16 *
514 :			( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) +
515 :			d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode)) * sad)/1000;
516 :
517 :			if (sad < *(data->iMinSAD)) {
518 :			*(data->iMinSAD) = sad;
519 :			data->currentQMV->x = xf; data->currentQMV->y = yf;
520 :			*dir = Direction; }
521 :			}
522 :
523 :			static void
524 :	chl	530	CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
525 :	Isibaar	3	{
526 :	syskin	628	int32_t sad = 0;
527 :	chl	530	int k;
528 :			const uint8_t *ReferenceF;
529 :			const uint8_t *ReferenceB;
530 :			VECTOR mvs, b_mvs;
531 :	Isibaar	3
532 :	chl	530	if (( x > 31) \|\| ( x < -32) \|\| ( y > 31) \|\| (y < -32)) return;
533 :	Isibaar	3
534 :	chl	530	for (k = 0; k < 4; k++) {
535 :			mvs.x = data->directmvF[k].x + x;
536 :			b_mvs.x = ((x == 0) ?
537 :			data->directmvB[k].x
538 :			: mvs.x - data->referencemv[k].x);
539 :	Isibaar	3
540 :	chl	530	mvs.y = data->directmvF[k].y + y;
541 :			b_mvs.y = ((y == 0) ?
542 :			data->directmvB[k].y
543 :			: mvs.y - data->referencemv[k].y);
544 :
545 :			if (( mvs.x > data->max_dx ) \|\| ( mvs.x < data->min_dx )
546 :			\|\| ( mvs.y > data->max_dy ) \|\| ( mvs.y < data->min_dy )
547 :			\|\| ( b_mvs.x > data->max_dx ) \|\| ( b_mvs.x < data->min_dx )
548 :			\|\| ( b_mvs.y > data->max_dy ) \|\| ( b_mvs.y < data->min_dy )) return;
549 :	Isibaar	3
550 :	chl	530	switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) {
551 :			case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break;
552 :			case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break;
553 :			case 2 : ReferenceF = data->RefH + (mvs.x-1)/2 + (mvs.y/2)*(data->iEdgedWidth); break;
554 :			default : ReferenceF = data->RefHV + (mvs.x-1)/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break;
555 :			}
556 :	Isibaar	3
557 :	chl	530	switch ( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) {
558 :			case 0 : ReferenceB = data->bRef + b_mvs.x/2 + (b_mvs.y/2)*(data->iEdgedWidth); break;
559 :			case 1 : ReferenceB = data->bRefV + b_mvs.x/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break;
560 :			case 2 : ReferenceB = data->bRefH + (b_mvs.x-1)/2 + (b_mvs.y/2)*(data->iEdgedWidth); break;
561 :			default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break;
562 :	chl	344	}
563 :	chl	530
564 :			sad += sad8bi(data->Cur + 8(k&1) + 8(k>>1)*(data->iEdgedWidth),
565 :			ReferenceF + 8(k&1) + 8(k>>1)*(data->iEdgedWidth),
566 :			ReferenceB + 8(k&1) + 8(k>>1)*(data->iEdgedWidth),
567 :			data->iEdgedWidth);
568 :			if (sad > *(data->iMinSAD)) return;
569 :	edgomez	78	}
570 :	chl	530
571 :	syskin	628	sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000;
572 :
573 :	chl	530	if (sad < *(data->iMinSAD)) {
574 :			*(data->iMinSAD) = sad;
575 :			data->currentMV->x = x; data->currentMV->y = y;
576 :			*dir = Direction; }
577 :	Isibaar	3	}
578 :
579 :	syskin	658
580 :	chl	530	static void
581 :	syskin	658	CheckCandidateDirect_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
582 :			{
583 :			int32_t sad = 0;
584 :			int k;
585 :			VECTOR mvs, b_mvs, halfpelMV;
586 :			const uint8_t ref1, ref2, ref3, ref4;
587 :			uint8_t ReferenceF, ReferenceB;
588 :			const uint32_t iEdgedWidth = data->iEdgedWidth;
589 :
590 :			if (( x > 31) \|\| ( x < -32) \|\| ( y > 31) \|\| (y < -32)) return;
591 :
592 :			for (k = 0; k < 4; k++) {
593 :			ReferenceF = (uint8_t *)data->RefQ;
594 :			ReferenceB = (uint8_t *)data->RefQ + 64;
595 :
596 :			mvs.x = data->directmvF[k].x + x;
597 :			b_mvs.x = ((x == 0) ?
598 :			data->directmvB[k].x
599 :			: mvs.x - data->referencemv[k].x);
600 :
601 :			mvs.y = data->directmvF[k].y + y;
602 :			b_mvs.y = ((y == 0) ?
603 :			data->directmvB[k].y
604 :			: mvs.y - data->referencemv[k].y);
605 :
606 :			if (( mvs.x > data->max_dx ) \|\| ( mvs.x < data->min_dx )
607 :			\|\| ( mvs.y > data->max_dy ) \|\| ( mvs.y < data->min_dy )
608 :			\|\| ( b_mvs.x > data->max_dx ) \|\| ( b_mvs.x < data->min_dx )
609 :			\|\| ( b_mvs.y > data->max_dy ) \|\| ( b_mvs.y < data->min_dy )) return;
610 :
611 :			halfpelMV.x = mvs.x/2; //forward first
612 :			halfpelMV.y = mvs.y/2;
613 :			GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this reference is used in all cases
614 :			switch( ((mvs.x&1)<<1) + (mvs.y&1) ) {
615 :			case 0: // pure halfpel position
616 :			GET_REFERENCE(halfpelMV.x + 16(k&1), halfpelMV.y + 16(k>>1), ReferenceF);
617 :			break;
618 :
619 :			case 1: // x halfpel, y qpel - top or bottom during qpel refinement
620 :			GET_REFERENCE(halfpelMV.x, mvs.y - halfpelMV.y, ref2);
621 :			interpolate8x8_avg2(ReferenceF, ref1+8(k&1) + 8(k>>1)*(data->iEdgedWidth),
622 :			ref2+ 8(k&1) + 8(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0);
623 :			break;
624 :
625 :			case 2: // x qpel, y halfpel - left or right during qpel refinement
626 :			GET_REFERENCE(mvs.x - halfpelMV.x, halfpelMV.y, ref2);
627 :			interpolate8x8_avg2(ReferenceF, ref1 + 8(k&1) + 8(k>>1)*(data->iEdgedWidth),
628 :			ref2 + 8(k&1) + 8(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0);
629 :			break;
630 :
631 :			default: // x and y in qpel resolution - the "corners" (top left/right and
632 :			// bottom left/right) during qpel refinement
633 :			GET_REFERENCE(halfpelMV.x, mvs.y - halfpelMV.y, ref2);
634 :			GET_REFERENCE(mvs.x - halfpelMV.x, halfpelMV.y, ref3);
635 :			GET_REFERENCE(mvs.x - halfpelMV.x, mvs.y - halfpelMV.y, ref4);
636 :			interpolate8x8_avg4(ReferenceF, ref1 + 8(k&1) + 8(k>>1)*(data->iEdgedWidth),
637 :			ref2 + 8(k&1) + 8(k>>1)*(data->iEdgedWidth),
638 :			ref3 + 8(k&1) + 8(k>>1)*(data->iEdgedWidth),
639 :			ref4 + 8(k&1) + 8(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0);
640 :			break;
641 :			}
642 :
643 :			halfpelMV.x = b_mvs.x/2;
644 :			halfpelMV.y = b_mvs.y/2;
645 :			GET_REFERENCE2(halfpelMV.x, halfpelMV.y, ref1); // this reference is used in most cases
646 :			switch( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) {
647 :			case 0: // pure halfpel position
648 :			GET_REFERENCE2(halfpelMV.x + 16(k&1), halfpelMV.y + 16(k>>1), ReferenceB);
649 :			break;
650 :
651 :			case 1: // x halfpel, y qpel - top or bottom during qpel refinement
652 :			GET_REFERENCE2(halfpelMV.x, b_mvs.y - halfpelMV.y, ref2);
653 :			interpolate8x8_avg2(ReferenceB, ref1+8(k&1) + 8(k>>1)*(data->iEdgedWidth),
654 :			ref2+ 8(k&1) + 8(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0);
655 :			break;
656 :
657 :			case 2: // x qpel, y halfpel - left or right during qpel refinement
658 :			GET_REFERENCE2(b_mvs.x - halfpelMV.x, halfpelMV.y, ref2);
659 :			interpolate8x8_avg2(ReferenceB, ref1 + 8(k&1) + 8(k>>1)*(data->iEdgedWidth),
660 :			ref2 + 8(k&1) + 8(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0);
661 :			break;
662 :
663 :			default: // x and y in qpel resolution - the "corners" (top left/right and
664 :			// bottom left/right) during qpel refinement
665 :			GET_REFERENCE2(halfpelMV.x, b_mvs.y - halfpelMV.y, ref2);
666 :			GET_REFERENCE2(b_mvs.x - halfpelMV.x, halfpelMV.y, ref3);
667 :			GET_REFERENCE2(b_mvs.x - halfpelMV.x, b_mvs.y - halfpelMV.y, ref4);
668 :			interpolate8x8_avg4(ReferenceB, ref1 + 8(k&1) + 8(k>>1)*(data->iEdgedWidth),
669 :			ref2 + 8(k&1) + 8(k>>1)*(data->iEdgedWidth),
670 :			ref3 + 8(k&1) + 8(k>>1)*(data->iEdgedWidth),
671 :			ref4 + 8(k&1) + 8(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0);
672 :			break;
673 :			}
674 :
675 :			sad += sad8bi(data->Cur + 8(k&1) + 8(k>>1)*(data->iEdgedWidth),
676 :			ReferenceF,
677 :			ReferenceB,
678 :			data->iEdgedWidth);
679 :			if (sad > *(data->iMinSAD)) return;
680 :			}
681 :
682 :			sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000;
683 :
684 :			if (sad < *(data->iMinSAD)) {
685 :			*(data->iMinSAD) = sad;
686 :			data->currentMV->x = x; data->currentMV->y = y;
687 :			*dir = Direction; }
688 :			}
689 :
690 :			static void
691 :			CheckCandidateDirectno4v_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
692 :			{
693 :			int32_t sad = 0;
694 :			VECTOR mvs, b_mvs, halfpelMV;
695 :			const uint8_t ref1, ref2, ref3, ref4;
696 :			const uint32_t iEdgedWidth = data->iEdgedWidth;
697 :			uint8_t * ReferenceF = (uint8_t *)data->RefQ;
698 :			uint8_t * ReferenceB = (uint8_t *)data->RefQ + 64;
699 :
700 :			if (( x > 31) \|\| ( x < -32) \|\| ( y > 31) \|\| (y < -32)) return;
701 :
702 :			mvs.x = data->directmvF[0].x + x;
703 :			b_mvs.x = ((x == 0) ?
704 :			data->directmvB[0].x
705 :			: mvs.x - data->referencemv[0].x);
706 :
707 :			mvs.y = data->directmvF[0].y + y;
708 :			b_mvs.y = ((y == 0) ?
709 :			data->directmvB[0].y
710 :			: mvs.y - data->referencemv[0].y);
711 :
712 :			if (( mvs.x > data->max_dx ) \|\| ( mvs.x < data->min_dx )
713 :			\|\| ( mvs.y > data->max_dy ) \|\| ( mvs.y < data->min_dy )
714 :			\|\| ( b_mvs.x > data->max_dx ) \|\| ( b_mvs.x < data->min_dx )
715 :			\|\| ( b_mvs.y > data->max_dy ) \|\| ( b_mvs.y < data->min_dy )) return;
716 :
717 :			halfpelMV.x = mvs.x/2; //forward first
718 :			halfpelMV.y = mvs.y/2;
719 :			GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this reference is used in all cases
720 :			switch( ((mvs.x&1)<<1) + (mvs.y&1) ) {
721 :			case 0: // pure halfpel position
722 :			GET_REFERENCE(halfpelMV.x, halfpelMV.y, ReferenceF);
723 :			break;
724 :
725 :			case 1: // x halfpel, y qpel - top or bottom during qpel refinement
726 :			GET_REFERENCE(halfpelMV.x, mvs.y - halfpelMV.y, ref2);
727 :			interpolate8x8_avg2(ReferenceF, ref1, ref2, iEdgedWidth, 0);
728 :			interpolate8x8_avg2(ReferenceF+8, ref1+8, ref2+8, iEdgedWidth, 0);
729 :			interpolate8x8_avg2(ReferenceF+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0);
730 :			interpolate8x8_avg2(ReferenceF+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0);
731 :			break;
732 :
733 :			case 2: // x qpel, y halfpel - left or right during qpel refinement
734 :			GET_REFERENCE(mvs.x - halfpelMV.x, halfpelMV.y, ref2);
735 :			interpolate8x8_avg2(ReferenceF, ref1, ref2, iEdgedWidth, 0);
736 :			interpolate8x8_avg2(ReferenceF+8, ref1+8, ref2+8, iEdgedWidth, 0);
737 :			interpolate8x8_avg2(ReferenceF+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0);
738 :			interpolate8x8_avg2(ReferenceF+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0);
739 :			break;
740 :
741 :			default: // x and y in qpel resolution
742 :			GET_REFERENCE(halfpelMV.x, mvs.y - halfpelMV.y, ref2);
743 :			GET_REFERENCE(mvs.x - halfpelMV.x, halfpelMV.y, ref3);
744 :			GET_REFERENCE(mvs.x - halfpelMV.x, mvs.y - halfpelMV.y, ref4);
745 :
746 :			interpolate8x8_avg4(ReferenceF, ref1, ref2, ref3, ref4, iEdgedWidth, 0);
747 :			interpolate8x8_avg4(ReferenceF+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, 0);
748 :			interpolate8x8_avg4(ReferenceF+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8iEdgedWidth, ref3+8iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, 0);
749 :			interpolate8x8_avg4(ReferenceF+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8iEdgedWidth+8, ref3+8iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, 0);
750 :			break;
751 :			}
752 :
753 :			halfpelMV.x = b_mvs.x/2; //backward
754 :			halfpelMV.y = b_mvs.y/2;
755 :			GET_REFERENCE2(halfpelMV.x, halfpelMV.y, ref1);
756 :			switch( ((b_mvs.x&1)<<1) + (b_mvs.y&1) )
757 :			{
758 :			case 0: // pure halfpel position
759 :			GET_REFERENCE2(halfpelMV.x, halfpelMV.y, ReferenceB);
760 :			break;
761 :
762 :			case 1: // x halfpel, y qpel - top or bottom during qpel refinement
763 :			GET_REFERENCE2(halfpelMV.x, b_mvs.y - halfpelMV.y, ref2);
764 :			interpolate8x8_avg2(ReferenceB, ref1, ref2, iEdgedWidth, 0);
765 :			interpolate8x8_avg2(ReferenceB+8, ref1+8, ref2+8, iEdgedWidth, 0);
766 :			interpolate8x8_avg2(ReferenceB+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0);
767 :			interpolate8x8_avg2(ReferenceB+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0);
768 :			break;
769 :
770 :			case 2: // x qpel, y halfpel - left or right during qpel refinement
771 :			GET_REFERENCE2(b_mvs.x - halfpelMV.x, halfpelMV.y, ref2);
772 :			interpolate8x8_avg2(ReferenceB, ref1, ref2, iEdgedWidth, 0);
773 :			interpolate8x8_avg2(ReferenceB+8, ref1+8, ref2+8, iEdgedWidth, 0);
774 :			interpolate8x8_avg2(ReferenceB+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0);
775 :			interpolate8x8_avg2(ReferenceB+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0);
776 :			break;
777 :
778 :			default: // x and y in qpel resolution - the "corners" (top left/right and
779 :			// bottom left/right) during qpel refinement
780 :			GET_REFERENCE2(halfpelMV.x, b_mvs.y - halfpelMV.y, ref2);
781 :			GET_REFERENCE2(b_mvs.x - halfpelMV.x, halfpelMV.y, ref3);
782 :			GET_REFERENCE2(b_mvs.x - halfpelMV.x, b_mvs.y - halfpelMV.y, ref4);
783 :
784 :			interpolate8x8_avg4(ReferenceB, ref1, ref2, ref3, ref4, iEdgedWidth, 0);
785 :			interpolate8x8_avg4(ReferenceB+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, 0);
786 :			interpolate8x8_avg4(ReferenceB+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8iEdgedWidth, ref3+8iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, 0);
787 :			interpolate8x8_avg4(ReferenceB+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8iEdgedWidth+8, ref3+8iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, 0);
788 :			break;
789 :			}
790 :
791 :			sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
792 :			sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000;
793 :
794 :			if (sad < *(data->iMinSAD)) {
795 :			*(data->iMinSAD) = sad;
796 :			data->currentMV->x = x; data->currentMV->y = y;
797 :			*dir = Direction; }
798 :			}
799 :
800 :
801 :			static void
802 :	chl	530	CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
803 :	chl	96	{
804 :	chl	530	int32_t sad;
805 :			const uint8_t *ReferenceF;
806 :			const uint8_t *ReferenceB;
807 :			VECTOR mvs, b_mvs;
808 :	chl	96
809 :	syskin	576	if (( x > 31) \|\| ( x < -32) \|\| ( y > 31) \|\| (y < -32)) return;
810 :	edgomez	195
811 :	chl	530	mvs.x = data->directmvF[0].x + x;
812 :			b_mvs.x = ((x == 0) ?
813 :			data->directmvB[0].x
814 :			: mvs.x - data->referencemv[0].x);
815 :	edgomez	195
816 :	chl	530	mvs.y = data->directmvF[0].y + y;
817 :			b_mvs.y = ((y == 0) ?
818 :			data->directmvB[0].y
819 :			: mvs.y - data->referencemv[0].y);
820 :
821 :			if (( mvs.x > data->max_dx ) \|\| ( mvs.x < data->min_dx )
822 :			\|\| ( mvs.y > data->max_dy ) \|\| ( mvs.y < data->min_dy )
823 :			\|\| ( b_mvs.x > data->max_dx ) \|\| ( b_mvs.x < data->min_dx )
824 :			\|\| ( b_mvs.y > data->max_dy ) \|\| ( b_mvs.y < data->min_dy )) return;
825 :	chl	96
826 :	chl	530	switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) {
827 :			case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break;
828 :			case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break;
829 :			case 2 : ReferenceF = data->RefH + (mvs.x-1)/2 + (mvs.y/2)*(data->iEdgedWidth); break;
830 :			default : ReferenceF = data->RefHV + (mvs.x-1)/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break;
831 :			}
832 :	chl	96
833 :	chl	530	switch ( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) {
834 :			case 0 : ReferenceB = data->bRef + b_mvs.x/2 + (b_mvs.y/2)*(data->iEdgedWidth); break;
835 :			case 1 : ReferenceB = data->bRefV + b_mvs.x/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break;
836 :			case 2 : ReferenceB = data->bRefH + (b_mvs.x-1)/2 + (b_mvs.y/2)*(data->iEdgedWidth); break;
837 :			default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break;
838 :			}
839 :
840 :	syskin	628	sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
841 :			sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000;
842 :	chl	96
843 :	chl	530	if (sad < *(data->iMinSAD)) {
844 :			*(data->iMinSAD) = sad;
845 :			data->currentMV->x = x; data->currentMV->y = y;
846 :			*dir = Direction; }
847 :			}
848 :	chl	96
849 :	chl	530	static void
850 :			CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
851 :			{
852 :	syskin	628	int32_t sad; int t;
853 :	chl	530	const uint8_t * Reference;
854 :	edgomez	195
855 :	chl	530	if (( x > data->max_dx) \|\| ( x < data->min_dx)
856 :			\|\| ( y > data->max_dy) \|\| (y < data->min_dy)) return;
857 :	chl	96
858 :	chl	530	switch ( ((x&1)<<1) + (y&1) )
859 :			{
860 :			case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break;
861 :			case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break;
862 :			case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break;
863 :			default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break;
864 :			}
865 :	edgomez	195
866 :	chl	530	sad = sad8(data->Cur, Reference, data->iEdgedWidth);
867 :	syskin	628	if (data->qpel) t = d_mv_bits(2 * x - data->predQMV.x, 2 * y - data->predQMV.y, data->iFcode);
868 :			else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode);
869 :	edgomez	195
870 :	syskin	628	sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))/100;
871 :	chl	96
872 :	Isibaar	606	if (sad < *(data->iMinSAD)) {
873 :			*(data->iMinSAD) = sad;
874 :			data->currentMV->x = x; data->currentMV->y = y;
875 :			*dir = Direction; }
876 :			}
877 :
878 :			static void
879 :	Isibaar	580	CheckCandidate8_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
880 :	syskin	658	// CheckCandidate8 variant which expects x and y in quarter pixel resolution
881 :	Isibaar	580	// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!)
882 :			// around currentMV!
883 :
884 :			{
885 :			int32_t sad;
886 :			uint8_t Reference = (uint8_t ) data->RefQ;
887 :			const uint8_t ref1, ref2, ref3, ref4;
888 :			VECTOR halfpelMV = *(data->currentMV);
889 :
890 :			int32_t iEdgedWidth = data->iEdgedWidth;
891 :			uint32_t rounding = data->rounding;
892 :
893 :			if (( x > data->max_dx) \|\| ( x < data->min_dx)
894 :			\|\| ( y > data->max_dy) \|\| (y < data->min_dy)) return;
895 :
896 :	Isibaar	606	GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1);
897 :	Isibaar	580	switch( ((x&1)<<1) + (y&1) )
898 :			{
899 :			case 0: // pure halfpel position - shouldn't happen during a refinement step
900 :	syskin	601	GET_REFERENCE(halfpelMV.x, halfpelMV.y, Reference);
901 :	Isibaar	580	break;
902 :
903 :			case 1: // x halfpel, y qpel - top or bottom during qpel refinement
904 :			GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2);
905 :
906 :			interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding);
907 :			break;
908 :
909 :			case 2: // x qpel, y halfpel - left or right during qpel refinement
910 :			GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2);
911 :
912 :			interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding);
913 :			break;
914 :
915 :			default: // x and y in qpel resolution - the "corners" (top left/right and
916 :			// bottom left/right) during qpel refinement
917 :			GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2);
918 :			GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3);
919 :			GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4);
920 :
921 :			interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
922 :			break;
923 :			}
924 :
925 :			sad = sad8(data->Cur, Reference, data->iEdgedWidth);
926 :	syskin	628	sad += (data->lambda8 * d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode) * (sad+NEIGH_8X8_BIAS))/100;
927 :	Isibaar	580
928 :			if (sad < *(data->iMinSAD)) {
929 :			*(data->iMinSAD) = sad;
930 :			data->currentQMV->x = x; data->currentQMV->y = y;
931 :			*dir = Direction; }
932 :			}
933 :
934 :	Isibaar	539	/* CHECK_CANDIATE FUNCTIONS END */
935 :	chl	96
936 :	chl	530	/* MAINSEARCH FUNCTIONS START */
937 :	edgomez	195
938 :	chl	530	static void
939 :			AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection)
940 :	chl	181	{
941 :
942 :			/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
943 :
944 :	chl	530	int iDirection;
945 :	edgomez	195
946 :			do {
947 :	chl	181	iDirection = 0;
948 :	chl	530	if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
949 :			if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
950 :			if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
951 :			if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
952 :	chl	181
953 :			/* now we're doing diagonal checks near our candidate */
954 :
955 :	chl	530	if (iDirection) { //checking if anything found
956 :	chl	181	bDirection = iDirection;
957 :			iDirection = 0;
958 :	chl	530	x = data->currentMV->x; y = data->currentMV->y;
959 :			if (bDirection & 3) { //our candidate is left or right
960 :			CHECK_CANDIDATE(x, y + iDiamondSize, 8);
961 :			CHECK_CANDIDATE(x, y - iDiamondSize, 4);
962 :			} else { // what remains here is up or down
963 :			CHECK_CANDIDATE(x + iDiamondSize, y, 2);
964 :			CHECK_CANDIDATE(x - iDiamondSize, y, 1); }
965 :	chl	181
966 :	edgomez	195	if (iDirection) {
967 :			bDirection += iDirection;
968 :	chl	530	x = data->currentMV->x; y = data->currentMV->y; }
969 :			} else { //about to quit, eh? not so fast....
970 :	edgomez	195	switch (bDirection) {
971 :	chl	181	case 2:
972 :	chl	530	CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
973 :			CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
974 :	chl	181	break;
975 :			case 1:
976 :	chl	530	CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
977 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
978 :	chl	181	break;
979 :	edgomez	195	case 2 + 4:
980 :	chl	530	CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
981 :			CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
982 :			CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
983 :	chl	181	break;
984 :			case 4:
985 :	chl	530	CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
986 :			CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
987 :	chl	181	break;
988 :			case 8:
989 :	chl	530	CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
990 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
991 :	chl	181	break;
992 :	edgomez	195	case 1 + 4:
993 :	chl	530	CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
994 :			CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
995 :			CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
996 :	chl	181	break;
997 :	edgomez	195	case 2 + 8:
998 :	chl	530	CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
999 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
1000 :			CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
1001 :	chl	181	break;
1002 :	edgomez	195	case 1 + 8:
1003 :	chl	530	CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
1004 :			CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
1005 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
1006 :	chl	181	break;
1007 :	edgomez	195	default: //1+2+4+8 == we didn't find anything at all
1008 :	chl	530	CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
1009 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
1010 :			CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
1011 :			CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
1012 :	chl	181	break;
1013 :			}
1014 :	chl	530	if (!iDirection) break; //ok, the end. really
1015 :			bDirection = iDirection;
1016 :			x = data->currentMV->x; y = data->currentMV->y;
1017 :	chl	181	}
1018 :			}
1019 :	edgomez	195	while (1); //forever
1020 :	chl	181	}
1021 :
1022 :	chl	530	static void
1023 :			SquareSearch(int x, int y, const SearchData * const data, int bDirection)
1024 :	chl	326	{
1025 :	chl	530	int iDirection;
1026 :	chl	326
1027 :	chl	530	do {
1028 :			iDirection = 0;
1029 :			if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64);
1030 :			if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128);
1031 :			if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32);
1032 :			if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128);
1033 :			if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64);
1034 :			if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128);
1035 :			if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128);
1036 :			if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128);
1037 :	chl	326
1038 :	chl	530	bDirection = iDirection;
1039 :			x = data->currentMV->x; y = data->currentMV->y;
1040 :			} while (iDirection);
1041 :	chl	346	}
1042 :
1043 :	chl	530	static void
1044 :			DiamondSearch(int x, int y, const SearchData * const data, int bDirection)
1045 :	chl	346	{
1046 :
1047 :	chl	181	/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
1048 :
1049 :	chl	530	int iDirection;
1050 :	edgomez	195
1051 :			do {
1052 :	chl	181	iDirection = 0;
1053 :	chl	530	if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
1054 :			if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
1055 :			if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
1056 :			if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
1057 :	chl	181
1058 :			/* now we're doing diagonal checks near our candidate */
1059 :
1060 :	chl	530	if (iDirection) { //checking if anything found
1061 :	chl	181	bDirection = iDirection;
1062 :			iDirection = 0;
1063 :	chl	530	x = data->currentMV->x; y = data->currentMV->y;
1064 :			if (bDirection & 3) { //our candidate is left or right
1065 :			CHECK_CANDIDATE(x, y + iDiamondSize, 8);
1066 :			CHECK_CANDIDATE(x, y - iDiamondSize, 4);
1067 :			} else { // what remains here is up or down
1068 :			CHECK_CANDIDATE(x + iDiamondSize, y, 2);
1069 :			CHECK_CANDIDATE(x - iDiamondSize, y, 1); }
1070 :	chl	181
1071 :	chl	530	bDirection += iDirection;
1072 :			x = data->currentMV->x; y = data->currentMV->y;
1073 :	chl	181	}
1074 :			}
1075 :	chl	530	while (iDirection);
1076 :	chl	181	}
1077 :
1078 :	chl	530	/* MAINSEARCH FUNCTIONS END */
1079 :	chl	181
1080 :	chl	530	/* HALFPELREFINE COULD BE A MAINSEARCH FUNCTION, BUT THERE IS NO NEED FOR IT */
1081 :
1082 :			static void
1083 :			HalfpelRefine(const SearchData * const data)
1084 :	chl	96	{
1085 :	chl	530	/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */
1086 :	edgomez	195
1087 :	chl	530	VECTOR backupMV = *(data->currentMV);
1088 :			int iDirection; //not needed
1089 :	chl	96
1090 :	chl	530	CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0);
1091 :			CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0);
1092 :			CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0);
1093 :			CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0);
1094 :	edgomez	195
1095 :	chl	530	CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0);
1096 :			CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0);
1097 :
1098 :			CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0);
1099 :			CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0);
1100 :	chl	96	}
1101 :
1102 :	Isibaar	580
1103 :			static void
1104 :			QuarterpelRefine(const SearchData * const data)
1105 :			{
1106 :			/* Perform quarter pixel refinement*/
1107 :
1108 :			VECTOR backupMV = *(data->currentQMV);
1109 :			int iDirection; //not needed
1110 :
1111 :			CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0);
1112 :			CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0);
1113 :			CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0);
1114 :			CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0);
1115 :
1116 :			CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0);
1117 :			CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0);
1118 :
1119 :			CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0);
1120 :			CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0);
1121 :
1122 :			}
1123 :
1124 :	chl	530	static __inline int
1125 :			SkipDecisionP(const IMAGE * current, const IMAGE * reference,
1126 :			const int x, const int y,
1127 :			const uint32_t iEdgedWidth, const uint32_t iQuant)
1128 :	chl	96
1129 :	Isibaar	3	{
1130 :	chl	530	/* keep repeating checks for all b-frames before this P frame,
1131 :			to make sure that SKIP is possible (todo)
1132 :			how: if skip is not possible set sad00 to a very high value */
1133 :	Isibaar	3
1134 :	chl	530	uint32_t sadC = sad8(current->u + x8 + y(iEdgedWidth/2)*8,
1135 :			reference->u + x8 + y(iEdgedWidth/2)*8, iEdgedWidth/2);
1136 :			if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
1137 :	h	545	sadC += sad8(current->v + (x + y(iEdgedWidth/2))8,
1138 :			reference->v + (x + y(iEdgedWidth/2))8, iEdgedWidth/2);
1139 :	chl	530	if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
1140 :	edgomez	195
1141 :	chl	530	return 1;
1142 :			}
1143 :	edgomez	195
1144 :	chl	530	static __inline void
1145 :			SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad)
1146 :			{
1147 :			pMB->mode = MODE_NOT_CODED;
1148 :	Isibaar	539	pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0;
1149 :			pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0;
1150 :	syskin	628
1151 :	Isibaar	580	pMB->qmvs[0].x = pMB->qmvs[1].x = pMB->qmvs[2].x = pMB->qmvs[3].x = 0;
1152 :			pMB->qmvs[0].y = pMB->qmvs[1].y = pMB->qmvs[2].y = pMB->qmvs[3].y = 0;
1153 :
1154 :	chl	530	pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
1155 :	Isibaar	3	}
1156 :
1157 :	syskin	628	bool
1158 :	chl	530	MotionEstimation(MBParam * const pParam,
1159 :			FRAMEINFO * const current,
1160 :			FRAMEINFO * const reference,
1161 :			const IMAGE * const pRefH,
1162 :			const IMAGE * const pRefV,
1163 :			const IMAGE * const pRefHV,
1164 :			const uint32_t iLimit)
1165 :	Isibaar	3	{
1166 :	chl	530	MACROBLOCK *const pMBs = current->mbs;
1167 :			const IMAGE *const pCurrent = &current->image;
1168 :			const IMAGE *const pRef = &reference->image;
1169 :	Isibaar	3
1170 :	chl	530	const VECTOR zeroMV = { 0, 0 };
1171 :	Isibaar	3
1172 :	chl	530	uint32_t x, y;
1173 :			uint32_t iIntra = 0;
1174 :	syskin	628	int32_t InterBias, quant = current->quant, sad00;
1175 :	Isibaar	580	uint8_t *qimage;
1176 :	edgomez	195
1177 :	Isibaar	539	// some pre-initialized thingies for SearchP
1178 :			int32_t temp[5];
1179 :			VECTOR currentMV[5];
1180 :	Isibaar	580	VECTOR currentQMV[5];
1181 :	Isibaar	539	int32_t iMinSAD[5];
1182 :			SearchData Data;
1183 :			Data.iEdgedWidth = pParam->edged_width;
1184 :			Data.currentMV = currentMV;
1185 :	Isibaar	580	Data.currentQMV = currentQMV;
1186 :	Isibaar	539	Data.iMinSAD = iMinSAD;
1187 :			Data.temp = temp;
1188 :			Data.iFcode = current->fcode;
1189 :	Isibaar	580	Data.rounding = pParam->m_rounding_type;
1190 :	syskin	628	Data.qpel = pParam->m_quarterpel;
1191 :			Data.chroma = current->global_flags & XVID_ME_COLOUR;
1192 :	Isibaar	539
1193 :	Isibaar	580	if((qimage = (uint8_t ) malloc(32 pParam->edged_width)) == NULL)
1194 :	syskin	594	return 1; // allocate some mem for qpel interpolated blocks
1195 :	Isibaar	580	// somehow this is dirty since I think we shouldn't use malloc outside
1196 :			// encoder_create() - so please fix me!
1197 :	syskin	628	Data.RefQ = qimage;
1198 :	chl	530	if (sadInit) (*sadInit) ();
1199 :	edgomez	195
1200 :	chl	530	for (y = 0; y < pParam->mb_height; y++) {
1201 :			for (x = 0; x < pParam->mb_width; x++) {
1202 :			MACROBLOCK pMB = &pMBs[x + y pParam->mb_width];
1203 :	syskin	628
1204 :			pMB->sad16
1205 :	chl	530	= sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16,
1206 :			pRef->y + (x + y * pParam->edged_width) * 16,
1207 :			pParam->edged_width, pMB->sad8 );
1208 :	edgomez	195
1209 :	syskin	628	if (Data.chroma) {
1210 :			pMB->sad16 += sad8(pCurrent->u + x8 + y(pParam->edged_width/2)*8,
1211 :			pRef->u + x8 + y(pParam->edged_width/2)*8, pParam->edged_width/2);
1212 :
1213 :			pMB->sad16 += sad8(pCurrent->v + (x + y(pParam->edged_width/2))8,
1214 :			pRef->v + (x + y(pParam->edged_width/2))8, pParam->edged_width/2);
1215 :			}
1216 :
1217 :			sad00 = pMB->sad16; //if no gmc; else sad00 = (..)
1218 :
1219 :	chl	530	if (!(current->global_flags & XVID_LUMIMASKING)) {
1220 :			pMB->dquant = NO_CHANGE;
1221 :	syskin	601	pMB->quant = current->quant;
1222 :			} else {
1223 :	syskin	574	if (pMB->dquant != NO_CHANGE) {
1224 :			quant += DQtab[pMB->dquant];
1225 :			if (quant > 31) quant = 31;
1226 :			else if (quant < 1) quant = 1;
1227 :			}
1228 :	syskin	601	pMB->quant = quant;
1229 :			}
1230 :	chl	181
1231 :	syskin	628	//initial skip decision
1232 :	chl	619	/* no early skip for GMC (global vector = skip vector is unknown!) */
1233 :			if (current->coding_type == P_VOP) { /* no fast SKIP for S(GMC)-VOPs */
1234 :			if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH)
1235 :	syskin	628	if (Data.chroma \|\| SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) {
1236 :	chl	619	SkipMacroblockP(pMB, sad00);
1237 :			continue;
1238 :			}
1239 :			}
1240 :	syskin	628
1241 :			SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
1242 :	chl	530	y, current->motion_flags, pMB->quant,
1243 :	Isibaar	539	&Data, pParam, pMBs, reference->mbs,
1244 :	chl	530	current->global_flags & XVID_INTER4V, pMB);
1245 :	edgomez	195
1246 :	chl	530	/* final skip decision, a.k.a. "the vector you found, really that good?" */
1247 :	chl	619	if (current->coding_type == P_VOP) {
1248 :			if ( (pMB->dquant == NO_CHANGE) && (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP)
1249 :	syskin	601	&& ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) )
1250 :	syskin	628	if (Data.chroma \|\| SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) {
1251 :	chl	619	SkipMacroblockP(pMB, sad00);
1252 :			continue;
1253 :			}
1254 :			}
1255 :	syskin	628
1256 :	chl	530	/* finally, intra decision */
1257 :	Isibaar	3
1258 :	chl	530	InterBias = MV16_INTER_BIAS;
1259 :	syskin	639	if (pMB->quant > 8) InterBias += 100 * (pMB->quant - 8); // to make high quants work
1260 :	chl	530	if (y != 0)
1261 :	syskin	639	if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80;
1262 :	chl	530	if (x != 0)
1263 :	syskin	639	if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80;
1264 :	syskin	628
1265 :			if (Data.chroma) InterBias += 50; // to compensate bigger SAD
1266 :	Isibaar	3
1267 :	chl	530	if (InterBias < pMB->sad16) {
1268 :			const int32_t deviation =
1269 :			dev16(pCurrent->y + (x + y * pParam->edged_width) * 16,
1270 :			pParam->edged_width);
1271 :	Isibaar	3
1272 :	chl	530	if (deviation < (pMB->sad16 - InterBias)) {
1273 :	Isibaar	580	if (++iIntra >= iLimit) { free(qimage); return 1; }
1274 :	chl	530	pMB->mode = MODE_INTRA;
1275 :	Isibaar	539	pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] =
1276 :			pMB->mvs[3] = zeroMV;
1277 :	Isibaar	580	pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] =
1278 :			pMB->qmvs[3] = zeroMV;
1279 :	chl	530	pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] =
1280 :			pMB->sad8[3] = 0;
1281 :			}
1282 :			}
1283 :			}
1284 :	Isibaar	3	}
1285 :	Isibaar	580	free(qimage);
1286 :	chl	619
1287 :			if (current->coding_type == S_VOP) /* first GMC step only for S(GMC)-VOPs */
1288 :	syskin	628	current->GMC_MV = GlobalMotionEst( pMBs, pParam, current->fcode );
1289 :	chl	619	else
1290 :			current->GMC_MV = zeroMV;
1291 :
1292 :	syskin	628	return 0;
1293 :	chl	530	}
1294 :	Isibaar	3
1295 :	edgomez	195
1296 :	chl	530	#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16\|PMV_HALFPELREFINE16)
1297 :	Isibaar	3
1298 :	chl	530	static __inline int
1299 :			make_mask(const VECTOR * const pmv, const int i)
1300 :			{
1301 :	Isibaar	539	int mask = 255, j;
1302 :	chl	530	for (j = 0; j < i; j++) {
1303 :			if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already
1304 :			if (pmv[i].x == pmv[j].x) {
1305 :			if (pmv[i].y == pmv[j].y + iDiamondSize) { mask &= ~4; continue; }
1306 :			if (pmv[i].y == pmv[j].y - iDiamondSize) { mask &= ~8; continue; }
1307 :			} else
1308 :			if (pmv[i].y == pmv[j].y) {
1309 :			if (pmv[i].x == pmv[j].x + iDiamondSize) { mask &= ~1; continue; }
1310 :			if (pmv[i].x == pmv[j].x - iDiamondSize) { mask &= ~2; continue; }
1311 :			}
1312 :	Isibaar	3	}
1313 :	chl	530	return mask;
1314 :			}
1315 :	edgomez	195
1316 :	chl	530	static __inline void
1317 :			PreparePredictionsP(VECTOR * const pmv, int x, int y, const int iWcount,
1318 :			const int iHcount, const MACROBLOCK * const prevMB)
1319 :			{
1320 :	edgomez	195
1321 :	chl	530	//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself
1322 :	edgomez	195
1323 :	chl	530	if ( (y != 0) && (x != (iWcount-1)) ) { // [5] top-right neighbour
1324 :			pmv[5].x = EVEN(pmv[3].x);
1325 :	Isibaar	539	pmv[5].y = EVEN(pmv[3].y);
1326 :			} else pmv[5].x = pmv[5].y = 0;
1327 :	chl	169
1328 :	chl	530	if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour
1329 :			else pmv[3].x = pmv[3].y = 0;
1330 :	Isibaar	3
1331 :	chl	530	if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour
1332 :			else pmv[4].x = pmv[4].y = 0;
1333 :	chl	169
1334 :	chl	530	// [1] median prediction
1335 :			pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y);
1336 :	chl	169
1337 :	chl	530	pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask
1338 :	chl	169
1339 :	chl	530	pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame
1340 :			pmv[2].y = EVEN(prevMB->mvs[0].y);
1341 :	chl	169
1342 :	chl	530	if ((x != iWcount-1) && (y != iHcount-1)) {
1343 :			pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame
1344 :	Isibaar	539	pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y);
1345 :			} else pmv[6].x = pmv[6].y = 0;
1346 :	chl	530	}
1347 :	chl	169
1348 :	chl	530	static void
1349 :	syskin	628	SearchP(const IMAGE * const pRef,
1350 :	chl	530	const uint8_t * const pRefH,
1351 :			const uint8_t * const pRefV,
1352 :			const uint8_t * const pRefHV,
1353 :			const IMAGE * const pCur,
1354 :			const int x,
1355 :			const int y,
1356 :			const uint32_t MotionFlags,
1357 :			const uint32_t iQuant,
1358 :	Isibaar	539	SearchData * const Data,
1359 :	chl	530	const MBParam * const pParam,
1360 :			const MACROBLOCK * const pMBs,
1361 :			const MACROBLOCK * const prevMBs,
1362 :			int inter4v,
1363 :			MACROBLOCK * const pMB)
1364 :			{
1365 :	chl	169
1366 :	chl	530	int i, iDirection = 255, mask, threshA;
1367 :	Isibaar	539	VECTOR pmv[7];
1368 :	Isibaar	3
1369 :	Isibaar	539	get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); //has to be changed to get_pmv(2)()
1370 :			get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1371 :	Isibaar	580	pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel);
1372 :	Isibaar	3
1373 :	Isibaar	539	Data->predMV = pmv[0];
1374 :	Isibaar	580
1375 :	Isibaar	539	Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16;
1376 :	syskin	628	Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8;
1377 :			Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8;
1378 :
1379 :			Data->Ref = pRef->y + (x + Data->iEdgedWidthy) 16;
1380 :	Isibaar	539	Data->RefH = pRefH + (x + Data->iEdgedWidthy) 16;
1381 :			Data->RefV = pRefV + (x + Data->iEdgedWidthy) 16;
1382 :			Data->RefHV = pRefHV + (x + Data->iEdgedWidthy) 16;
1383 :	syskin	628	Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8;
1384 :			Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8;
1385 :	Isibaar	3
1386 :	syskin	601	Data->lambda16 = lambda_vec16[iQuant];
1387 :	Isibaar	606	Data->lambda8 = lambda_vec8[iQuant];
1388 :	chl	169
1389 :	chl	530	if (!(MotionFlags & PMV_HALFPEL16)) {
1390 :	Isibaar	539	Data->min_dx = EVEN(Data->min_dx);
1391 :			Data->max_dx = EVEN(Data->max_dx);
1392 :			Data->min_dy = EVEN(Data->min_dy);
1393 :			Data->max_dy = EVEN(Data->max_dy); }
1394 :	edgomez	568
1395 :			if (pMB->dquant != NO_CHANGE) inter4v = 0;
1396 :	edgomez	195
1397 :	Isibaar	580	for(i = 0; i < 5; i++)
1398 :			Data->currentMV[i].x = Data->currentMV[i].y = 0;
1399 :	Isibaar	539
1400 :	Isibaar	606	if (pParam->m_quarterpel) {
1401 :			Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0);
1402 :			i = d_mv_bits(Data->predQMV.x, Data->predQMV.y, Data->iFcode);
1403 :			} else i = d_mv_bits(Data->predMV.x, Data->predMV.y, Data->iFcode);
1404 :
1405 :	syskin	628	Data->iMinSAD[0] = pMB->sad16 + (Data->lambda16 * i * pMB->sad16)/1000;
1406 :			Data->iMinSAD[1] = pMB->sad8[0] + (Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS))/100;
1407 :	Isibaar	539	Data->iMinSAD[2] = pMB->sad8[1];
1408 :			Data->iMinSAD[3] = pMB->sad8[2];
1409 :			Data->iMinSAD[4] = pMB->sad8[3];
1410 :
1411 :	chl	530	if ((x == 0) && (y == 0)) threshA = 512;
1412 :			else {
1413 :	edgomez	568	threshA = Data->temp[0]; // that's when we keep this SAD atm
1414 :	chl	530	if (threshA < 512) threshA = 512;
1415 :			if (threshA > 1024) threshA = 1024; }
1416 :	Isibaar	3
1417 :	chl	530	PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height,
1418 :			prevMBs + x + y * pParam->mb_width);
1419 :	edgomez	195
1420 :	syskin	628	if (inter4v \|\| pParam->m_quarterpel \|\| Data->chroma) CheckCandidate = CheckCandidate16;
1421 :			else CheckCandidate = CheckCandidate16no4v;
1422 :	edgomez	568
1423 :	chl	530	/* main loop. checking all predictions */
1424 :
1425 :			for (i = 1; i < 7; i++) {
1426 :			if (!(mask = make_mask(pmv, i)) ) continue;
1427 :	edgomez	568	(*CheckCandidate)(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1428 :	Isibaar	539	if (Data->iMinSAD[0] <= threshA) break;
1429 :	chl	530	}
1430 :
1431 :	Isibaar	539	if ((Data->iMinSAD[0] <= threshA) \|\|
1432 :			(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) &&
1433 :			(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) {
1434 :	chl	530	inter4v = 0;
1435 :	Isibaar	539	} else {
1436 :	Isibaar	3
1437 :	Isibaar	539	MainSearchFunc * MainSearchPtr;
1438 :			if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1439 :			else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1440 :			else MainSearchPtr = DiamondSearch;
1441 :	chl	181
1442 :	Isibaar	539	(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1443 :	Isibaar	3
1444 :	chl	530	/* extended search, diamond starting in 0,0 and in prediction.
1445 :			note that this search is/might be done in halfpel positions,
1446 :			which makes it more different than the diamond above */
1447 :	chl	259
1448 :	Isibaar	539	if (MotionFlags & PMV_EXTSEARCH16) {
1449 :			int32_t bSAD;
1450 :			VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0];
1451 :			if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel?
1452 :			startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y);
1453 :			if (!(MVequal(startMV, backupMV))) {
1454 :			bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1455 :	Isibaar	3
1456 :	Isibaar	606	(*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data);
1457 :	Isibaar	539	(*MainSearchPtr)(startMV.x, startMV.y, Data, 255);
1458 :			if (bSAD < Data->iMinSAD[0]) {
1459 :			Data->currentMV[0] = backupMV;
1460 :			Data->iMinSAD[0] = bSAD; }
1461 :			}
1462 :	Isibaar	3
1463 :	Isibaar	539	backupMV = Data->currentMV[0];
1464 :			if (MotionFlags & PMV_HALFPELREFINE16) startMV.x = startMV.y = 1;
1465 :			else startMV.x = startMV.y = 0;
1466 :			if (!(MVequal(startMV, backupMV))) {
1467 :			bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1468 :	edgomez	195
1469 :	Isibaar	606	(*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data);
1470 :	Isibaar	539	(*MainSearchPtr)(startMV.x, startMV.y, Data, 255);
1471 :			if (bSAD < Data->iMinSAD[0]) {
1472 :			Data->currentMV[0] = backupMV;
1473 :			Data->iMinSAD[0] = bSAD; }
1474 :			}
1475 :	Isibaar	3	}
1476 :			}
1477 :
1478 :	Isibaar	539	if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data);
1479 :	Isibaar	3
1480 :	Isibaar	580	for(i = 0; i < 5; i++) {
1481 :			Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors
1482 :			Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
1483 :			}
1484 :
1485 :			if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) {
1486 :
1487 :	Isibaar	606	CheckCandidate = CheckCandidate16_qpel;
1488 :			get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1489 :	syskin	601	pParam->width, pParam->height, Data->iFcode, 0);
1490 :
1491 :	Isibaar	580	QuarterpelRefine(Data);
1492 :			}
1493 :
1494 :	syskin	628	if (Data->iMinSAD[0] < (int32_t)iQuant * 30 ) inter4v = 0;
1495 :	Isibaar	539	if (inter4v) {
1496 :			SearchData Data8;
1497 :			Data8.iFcode = Data->iFcode;
1498 :	Isibaar	606	Data8.lambda8 = Data->lambda8;
1499 :	Isibaar	539	Data8.iEdgedWidth = Data->iEdgedWidth;
1500 :	Isibaar	606	Data8.RefQ = Data->RefQ;
1501 :	syskin	628	Data8.qpel = Data->qpel;
1502 :	Isibaar	539	Search8(Data, 2x, 2y, MotionFlags, pParam, pMB, pMBs, 0, &Data8);
1503 :			Search8(Data, 2x + 1, 2y, MotionFlags, pParam, pMB, pMBs, 1, &Data8);
1504 :			Search8(Data, 2x, 2y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8);
1505 :			Search8(Data, 2x + 1, 2y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8);
1506 :	syskin	628
1507 :			if (Data->chroma) {
1508 :			int sum, dx, dy;
1509 :
1510 :			if(pParam->m_quarterpel) {
1511 :			sum = pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2;
1512 :			} else sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y;
1513 :			dy = (sum >> 3) + roundtab_76[sum & 0xf];
1514 :
1515 :			if(pParam->m_quarterpel) {
1516 :			sum = pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2;
1517 :			} else sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x;
1518 :			dx = (sum >> 3) + roundtab_76[sum & 0xf];
1519 :
1520 :			Data->iMinSAD[1] += ChromaSAD(dx, dy, Data);
1521 :			}
1522 :	Isibaar	539	}
1523 :	Isibaar	3
1524 :	chl	530	if (!(inter4v) \|\|
1525 :	Isibaar	539	(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] +
1526 :			Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) {
1527 :	chl	530	// INTER MODE
1528 :			pMB->mode = MODE_INTER;
1529 :	Isibaar	539	pMB->mvs[0] = pMB->mvs[1]
1530 :			= pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
1531 :	Isibaar	3
1532 :	Isibaar	580	pMB->qmvs[0] = pMB->qmvs[1]
1533 :			= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
1534 :
1535 :	chl	530	pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] =
1536 :	Isibaar	539	pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0];
1537 :	Isibaar	3
1538 :	Isibaar	580	if(pParam->m_quarterpel) {
1539 :			pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predQMV.x;
1540 :			pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predQMV.y;
1541 :	syskin	628	} else {
1542 :	Isibaar	580	pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
1543 :			pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
1544 :			}
1545 :	edgomez	195	} else {
1546 :	chl	530	// INTER4V MODE; all other things are already set in Search8
1547 :			pMB->mode = MODE_INTER4V;
1548 :	Isibaar	539	pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] +
1549 :			Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant;
1550 :	edgomez	78	}
1551 :	Isibaar	3	}
1552 :
1553 :	chl	530	static void
1554 :			Search8(const SearchData * const OldData,
1555 :			const int x, const int y,
1556 :			const uint32_t MotionFlags,
1557 :			const MBParam * const pParam,
1558 :			MACROBLOCK * const pMB,
1559 :			const MACROBLOCK * const pMBs,
1560 :	Isibaar	539	const int block,
1561 :			SearchData * const Data)
1562 :	chl	345	{
1563 :	Isibaar	539	Data->iMinSAD = OldData->iMinSAD + 1 + block;
1564 :			Data->currentMV = OldData->currentMV + 1 + block;
1565 :	Isibaar	580	Data->currentQMV = OldData->currentQMV + 1 + block;
1566 :	chl	345
1567 :	syskin	601	if(pParam->m_quarterpel) {
1568 :	Isibaar	606	Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block);
1569 :	syskin	628	if (block != 0) (Data->iMinSAD) += (Data->lambda8
1570 :	syskin	601	d_mv_bits( Data->currentQMV->x - Data->predQMV.x,
1571 :			Data->currentQMV->y - Data->predQMV.y,
1572 :	syskin	628	Data->iFcode) * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100;
1573 :	Isibaar	606	} else {
1574 :			Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block);
1575 :	syskin	628	if (block != 0) (Data->iMinSAD) += (Data->lambda8
1576 :	syskin	601	d_mv_bits( Data->currentMV->x - Data->predMV.x,
1577 :			Data->currentMV->y - Data->predMV.y,
1578 :	syskin	628	Data->iFcode) * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100;
1579 :	Isibaar	606	}
1580 :	chl	345
1581 :	chl	530	if (MotionFlags & (PMV_EXTSEARCH8\|PMV_HALFPELREFINE8)) {
1582 :	chl	345
1583 :	Isibaar	539	Data->Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1));
1584 :			Data->RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1));
1585 :			Data->RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1));
1586 :			Data->RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1));
1587 :	chl	345
1588 :	Isibaar	539	Data->Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1));
1589 :	chl	530
1590 :	Isibaar	539	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1591 :	Isibaar	580	pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel);
1592 :	syskin	628	CheckCandidate = CheckCandidate8;
1593 :	chl	345
1594 :	chl	530	if (MotionFlags & PMV_EXTSEARCH8) {
1595 :	Isibaar	580	int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1596 :
1597 :	chl	530	MainSearchFunc *MainSearchPtr;
1598 :			if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch;
1599 :			else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch;
1600 :			else MainSearchPtr = DiamondSearch;
1601 :	chl	345
1602 :	Isibaar	580	(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255);
1603 :	chl	345
1604 :	syskin	601	if(*(Data->iMinSAD) < temp_sad) {
1605 :	Isibaar	580	Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1606 :			Data->currentQMV->y = 2 * Data->currentMV->y;
1607 :			}
1608 :			}
1609 :
1610 :			if (MotionFlags & PMV_HALFPELREFINE8) {
1611 :			int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1612 :
1613 :			HalfpelRefine(Data); // perform halfpel refine of current best vector
1614 :
1615 :			if(*(Data->iMinSAD) < temp_sad) { // we have found a better match
1616 :			Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1617 :			Data->currentQMV->y = 2 * Data->currentMV->y;
1618 :			}
1619 :			}
1620 :
1621 :	syskin	601	if(pParam->m_quarterpel) {
1622 :			if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) &&
1623 :			(MotionFlags & PMV_QUARTERPELREFINE8)) {
1624 :	Isibaar	600	CheckCandidate = CheckCandidate8_qpel;
1625 :	Isibaar	606	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1626 :			pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel);
1627 :			QuarterpelRefine(Data);
1628 :	syskin	601	}
1629 :	Isibaar	580	}
1630 :	chl	345	}
1631 :	chl	530
1632 :	Isibaar	580	if(pParam->m_quarterpel) {
1633 :			pMB->pmvs[block].x = Data->currentQMV->x - Data->predQMV.x;
1634 :			pMB->pmvs[block].y = Data->currentQMV->y - Data->predQMV.y;
1635 :			}
1636 :			else {
1637 :			pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x;
1638 :			pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y;
1639 :			}
1640 :
1641 :	Isibaar	539	pMB->mvs[block] = *(Data->currentMV);
1642 :	Isibaar	580	pMB->qmvs[block] = *(Data->currentQMV);
1643 :
1644 :	syskin	601	pMB->sad8[block] = 4 * (*Data->iMinSAD);
1645 :	chl	345	}
1646 :
1647 :	chl	530	/* B-frames code starts here */
1648 :	chl	345
1649 :	chl	530	static __inline VECTOR
1650 :			ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode)
1651 :	Isibaar	3	{
1652 :	chl	530	/* the stupidiest function ever */
1653 :			if (mode == MODE_FORWARD) return pMB->mvs[0];
1654 :			else return pMB->b_mvs[0];
1655 :	Isibaar	3	}
1656 :
1657 :	chl	530	static void __inline
1658 :			PreparePredictionsBF(VECTOR * const pmv, const int x, const int y,
1659 :			const uint32_t iWcount,
1660 :			const MACROBLOCK * const pMB,
1661 :			const uint32_t mode_curr)
1662 :	Isibaar	3	{
1663 :
1664 :	chl	530	// [0] is prediction
1665 :			pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y);
1666 :	Isibaar	3
1667 :	chl	530	pmv[1].x = pmv[1].y = 0; // [1] is zero
1668 :	Isibaar	3
1669 :	chl	530	pmv[2] = ChoosePred(pMB, mode_curr);
1670 :			pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y);
1671 :	edgomez	195
1672 :	chl	530	if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour
1673 :			pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr);
1674 :	Isibaar	539	pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y);
1675 :			} else pmv[3].x = pmv[3].y = 0;
1676 :	Isibaar	3
1677 :	chl	530	if (y != 0) {
1678 :			pmv[4] = ChoosePred(pMB-iWcount, mode_curr);
1679 :			pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y);
1680 :			} else pmv[4].x = pmv[4].y = 0;
1681 :	Isibaar	3
1682 :	chl	530	if (x != 0) {
1683 :			pmv[5] = ChoosePred(pMB-1, mode_curr);
1684 :			pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y);
1685 :			} else pmv[5].x = pmv[5].y = 0;
1686 :	Isibaar	3
1687 :	chl	530	if ((x != 0)&&(y != 0)) {
1688 :			pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr);
1689 :			pmv[6].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y);
1690 :			} else pmv[6].x = pmv[6].y = 0;
1691 :	edgomez	195
1692 :	chl	530	// more?
1693 :			}
1694 :	chl	181
1695 :	edgomez	170
1696 :	chl	530	/* search backward or forward, for b-frames */
1697 :			static void
1698 :			SearchBF( const uint8_t * const pRef,
1699 :			const uint8_t * const pRefH,
1700 :			const uint8_t * const pRefV,
1701 :			const uint8_t * const pRefHV,
1702 :			const IMAGE * const pCur,
1703 :			const int x, const int y,
1704 :			const uint32_t MotionFlags,
1705 :			const uint32_t iFcode,
1706 :			const MBParam * const pParam,
1707 :			MACROBLOCK * const pMB,
1708 :			const VECTOR * const predMV,
1709 :			int32_t * const best_sad,
1710 :	h	545	const int32_t mode_current,
1711 :			SearchData * const Data)
1712 :	chl	530	{
1713 :	Isibaar	3
1714 :	chl	530	const int32_t iEdgedWidth = pParam->edged_width;
1715 :
1716 :			int i, iDirection, mask;
1717 :	h	545	VECTOR pmv[7];
1718 :	chl	530	MainSearchFunc *MainSearchPtr;
1719 :	h	545	*Data->iMinSAD = MV_MAX_ERROR;
1720 :			Data->iFcode = iFcode;
1721 :	Isibaar	3
1722 :	h	545	Data->Ref = pRef + (x + y * iEdgedWidth) * 16;
1723 :			Data->RefH = pRefH + (x + y * iEdgedWidth) * 16;
1724 :			Data->RefV = pRefV + (x + y * iEdgedWidth) * 16;
1725 :			Data->RefHV = pRefHV + (x + y * iEdgedWidth) * 16;
1726 :	Isibaar	3
1727 :	h	545	Data->predMV = *predMV;
1728 :	edgomez	195
1729 :	h	545	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1730 :	Isibaar	580	pParam->width, pParam->height, iFcode, pParam->m_quarterpel);
1731 :	Isibaar	3
1732 :	h	545	pmv[0] = Data->predMV;
1733 :	syskin	658	if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; }
1734 :	h	545	PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current);
1735 :	edgomez	195
1736 :	h	545	Data->currentMV->x = Data->currentMV->y = 0;
1737 :	Isibaar	3
1738 :	chl	530	CheckCandidate = CheckCandidate16no4v;
1739 :	chl	181
1740 :	chl	530	// main loop. checking all predictions
1741 :			for (i = 0; i < 8; i++) {
1742 :			if (!(mask = make_mask(pmv, i)) ) continue;
1743 :	h	545	CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1744 :	edgomez	78	}
1745 :	Isibaar	3
1746 :	chl	530	if (MotionFlags & PMV_USESQUARES16)
1747 :			MainSearchPtr = SquareSearch;
1748 :			else if (MotionFlags & PMV_ADVANCEDDIAMOND16)
1749 :			MainSearchPtr = AdvDiamondSearch;
1750 :			else MainSearchPtr = DiamondSearch;
1751 :	chl	96
1752 :	h	545	(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255);
1753 :	chl	169
1754 :	h	545	HalfpelRefine(Data);
1755 :	syskin	658
1756 :			if (Data->qpel) {
1757 :			Data->currentQMV->x = 2*Data->currentMV->x;
1758 :			Data->currentQMV->y = 2*Data->currentMV->y;
1759 :			CheckCandidate = CheckCandidate16no4v_qpel;
1760 :			get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1761 :			pParam->width, pParam->height, iFcode, pParam->m_quarterpel);
1762 :			QuarterpelRefine(Data);
1763 :			}
1764 :	chl	169
1765 :	chl	530	// three bits are needed to code backward mode. four for forward
1766 :			// we treat the bits just like they were vector's
1767 :	syskin	628	if (mode_current == MODE_FORWARD) Data->iMinSAD += 4 Data->lambda16;
1768 :			else Data->iMinSAD += 3 Data->lambda16;
1769 :	chl	169
1770 :	h	545	if (Data->iMinSAD < best_sad) {
1771 :			best_sad = Data->iMinSAD;
1772 :	chl	530	pMB->mode = mode_current;
1773 :	syskin	658	if (Data->qpel) {
1774 :			pMB->pmvs[0].x = Data->currentQMV->x - predMV->x;
1775 :			pMB->pmvs[0].y = Data->currentQMV->y - predMV->y;
1776 :			if (mode_current == MODE_FORWARD)
1777 :			pMB->qmvs[0] = *Data->currentQMV;
1778 :			else
1779 :			pMB->b_qmvs[0] = *Data->currentQMV;
1780 :			} else {
1781 :			pMB->pmvs[0].x = Data->currentMV->x - predMV->x;
1782 :			pMB->pmvs[0].y = Data->currentMV->y - predMV->y;
1783 :			}
1784 :			if (mode_current == MODE_FORWARD)
1785 :			pMB->mvs[0] = (Data->currentMV+2) = Data->currentMV;
1786 :			else
1787 :			pMB->b_mvs[0] = (Data->currentMV+1) = Data->currentMV; //we store currmv for interpolate search
1788 :
1789 :	chl	169	}
1790 :	chl	530
1791 :			}
1792 :	chl	169
1793 :	chl	530	static int32_t
1794 :	Isibaar	539	SearchDirect(const IMAGE * const f_Ref,
1795 :	chl	530	const uint8_t * const f_RefH,
1796 :			const uint8_t * const f_RefV,
1797 :			const uint8_t * const f_RefHV,
1798 :	Isibaar	539	const IMAGE * const b_Ref,
1799 :	chl	530	const uint8_t * const b_RefH,
1800 :			const uint8_t * const b_RefV,
1801 :			const uint8_t * const b_RefHV,
1802 :			const IMAGE * const pCur,
1803 :			const int x, const int y,
1804 :			const uint32_t MotionFlags,
1805 :			const int32_t TRB, const int32_t TRD,
1806 :			const MBParam * const pParam,
1807 :			MACROBLOCK * const pMB,
1808 :			const MACROBLOCK * const b_mb,
1809 :	h	545	int32_t * const best_sad,
1810 :			SearchData * const Data)
1811 :	edgomez	195
1812 :	chl	530	{
1813 :	h	545	int32_t skip_sad;
1814 :	chl	530	int k;
1815 :	h	545
1816 :	chl	530	MainSearchFunc *MainSearchPtr;
1817 :	Isibaar	3
1818 :	h	545	Data->iMinSAD = 2564096;
1819 :	chl	140
1820 :	h	545	Data->Ref = f_Ref->y + (x + Data->iEdgedWidthy) 16;
1821 :			Data->RefH = f_RefH + (x + Data->iEdgedWidthy) 16;
1822 :			Data->RefV = f_RefV + (x + Data->iEdgedWidthy) 16;
1823 :			Data->RefHV = f_RefHV + (x + Data->iEdgedWidthy) 16;
1824 :			Data->bRef = b_Ref->y + (x + Data->iEdgedWidthy) 16;
1825 :			Data->bRefH = b_RefH + (x + Data->iEdgedWidthy) 16;
1826 :			Data->bRefV = b_RefV + (x + Data->iEdgedWidthy) 16;
1827 :			Data->bRefHV = b_RefHV + (x + Data->iEdgedWidthy) 16;
1828 :	chl	140
1829 :	h	545	Data->max_dx = 2 * pParam->width - 2 * (x) * 16;
1830 :			Data->max_dy = 2 * pParam->height - 2 * (y) * 16;
1831 :			Data->min_dx = -(2 * 16 + 2 * (x) * 16);
1832 :			Data->min_dy = -(2 * 16 + 2 * (y) * 16);
1833 :	syskin	658	if (Data->qpel) { //we measure in qpixels
1834 :			Data->max_dx *= 2;
1835 :			Data->max_dy *= 2;
1836 :			Data->min_dx *= 2;
1837 :			Data->min_dy *= 2;
1838 :			Data->referencemv = b_mb->qmvs;
1839 :			} else Data->referencemv = b_mb->mvs;
1840 :	Isibaar	3
1841 :	chl	530	for (k = 0; k < 4; k++) {
1842 :	h	545	pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD);
1843 :			pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD;
1844 :			pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD);
1845 :			pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD;
1846 :	Isibaar	3
1847 :	h	545	if ( ( pMB->b_mvs[k].x > Data->max_dx ) \|\| ( pMB->b_mvs[k].x < Data->min_dx )
1848 :			\|\| ( pMB->b_mvs[k].y > Data->max_dy ) \|\| ( pMB->b_mvs[k].y < Data->min_dy )) {
1849 :	Isibaar	3
1850 :	Isibaar	539	best_sad = 2564096; // in that case, we won't use direct mode
1851 :			pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV"
1852 :			pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0;
1853 :			return 0;
1854 :			}
1855 :			if (b_mb->mode != MODE_INTER4V) {
1856 :			pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0];
1857 :			pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0];
1858 :	h	545	Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0];
1859 :			Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0];
1860 :	Isibaar	539	break;
1861 :			}
1862 :			}
1863 :	Isibaar	3
1864 :	syskin	658	if (Data->qpel) {
1865 :			if (b_mb->mode == MODE_INTER4V)
1866 :			CheckCandidate = CheckCandidateDirect_qpel;
1867 :			else CheckCandidate = CheckCandidateDirectno4v_qpel;
1868 :			} else {
1869 :			if (b_mb->mode == MODE_INTER4V) CheckCandidate = CheckCandidateDirect;
1870 :			else CheckCandidate = CheckCandidateDirectno4v;
1871 :			}
1872 :
1873 :	h	545	(*CheckCandidate)(0, 0, 255, &k, Data);
1874 :	Isibaar	3
1875 :	chl	530	// skip decision
1876 :	syskin	628	if (Data->iMinSAD < pMB->quant SKIP_THRESH_B) {
1877 :	syskin	594	//possible skip - checking chroma. everything copied from MC
1878 :	Isibaar	539	//this is not full chroma compensation, only it's fullpel approximation. should work though
1879 :			int sum, dx, dy, b_dx, b_dy;
1880 :	Isibaar	3
1881 :	syskin	658	if (Data->qpel) {
1882 :			sum = pMB->mvs[0].y/2 + pMB->mvs[1].y/2 + pMB->mvs[2].y/2 + pMB->mvs[3].y/2;
1883 :			dy = (sum >> 3) + roundtab_76[sum & 0xf];
1884 :			sum = pMB->mvs[0].x/2 + pMB->mvs[1].x/2 + pMB->mvs[2].x/2 + pMB->mvs[3].x/2;
1885 :			dx = (sum >> 3) + roundtab_76[sum & 0xf];
1886 :	Isibaar	539
1887 :	syskin	658	sum = pMB->b_mvs[0].y/2 + pMB->b_mvs[1].y/2 + pMB->b_mvs[2].y/2 + pMB->b_mvs[3].y/2;
1888 :			b_dy = (sum >> 3) + roundtab_76[sum & 0xf];
1889 :			sum = pMB->b_mvs[0].x/2 + pMB->b_mvs[1].x/2 + pMB->b_mvs[2].x/2 + pMB->b_mvs[3].x/2;
1890 :			b_dx = (sum >> 3) + roundtab_76[sum & 0xf];
1891 :	Isibaar	539
1892 :	syskin	658	} else {
1893 :			sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x;
1894 :			dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2));
1895 :			sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y;
1896 :			dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2));
1897 :	Isibaar	539
1898 :	syskin	658	sum = pMB->b_mvs[0].x + pMB->b_mvs[1].x + pMB->b_mvs[2].x + pMB->b_mvs[3].x;
1899 :			b_dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2));
1900 :			sum = pMB->b_mvs[0].y + pMB->b_mvs[1].y + pMB->b_mvs[2].y + pMB->b_mvs[3].y;
1901 :			b_dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2));
1902 :			}
1903 :	h	545	sum = sad8bi(pCur->u + 8x + 8y*(Data->iEdgedWidth/2),
1904 :			f_Ref->u + (y8 + dy/2) (Data->iEdgedWidth/2) + x*8 + dx/2,
1905 :			b_Ref->u + (y8 + b_dy/2) (Data->iEdgedWidth/2) + x*8 + b_dx/2,
1906 :			Data->iEdgedWidth/2);
1907 :			sum += sad8bi(pCur->v + 8x + 8y*(Data->iEdgedWidth/2),
1908 :			f_Ref->v + (y8 + dy/2) (Data->iEdgedWidth/2) + x*8 + dx/2,
1909 :			b_Ref->v + (y8 + b_dy/2) (Data->iEdgedWidth/2) + x*8 + b_dx/2,
1910 :			Data->iEdgedWidth/2);
1911 :	Isibaar	539
1912 :	syskin	601	if (sum < MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) {
1913 :	Isibaar	539	pMB->mode = MODE_DIRECT_NONE_MV;
1914 :	h	545	return *Data->iMinSAD;
1915 :	Isibaar	539	}
1916 :			}
1917 :
1918 :	h	545	skip_sad = *Data->iMinSAD;
1919 :	edgomez	195
1920 :	chl	530	// DIRECT MODE DELTA VECTOR SEARCH.
1921 :			// This has to be made more effective, but at the moment I'm happy it's running at all
1922 :	Isibaar	3
1923 :	chl	530	if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1924 :			else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1925 :			else MainSearchPtr = DiamondSearch;
1926 :	edgomez	195
1927 :	h	545	(*MainSearchPtr)(0, 0, Data, 255);
1928 :	Isibaar	3
1929 :	syskin	658	HalfpelRefine(Data); //or qpel refine, if we're in qpel mode
1930 :	Isibaar	3
1931 :	syskin	628	Data->iMinSAD += 1 Data->lambda16; // one bit is needed to code direct mode
1932 :	h	545	best_sad = Data->iMinSAD;
1933 :	Isibaar	3
1934 :	chl	530	if (b_mb->mode == MODE_INTER4V)
1935 :			pMB->mode = MODE_DIRECT;
1936 :			else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation
1937 :	edgomez	195
1938 :	h	545	pMB->pmvs[3] = *Data->currentMV;
1939 :	edgomez	195
1940 :	chl	530	for (k = 0; k < 4; k++) {
1941 :	h	545	pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x;
1942 :	syskin	658	pMB->b_mvs[k].x = ( (Data->currentMV->x == 0)
1943 :	h	545	? Data->directmvB[k].x
1944 :	syskin	658	:pMB->mvs[k].x - Data->referencemv[k].x);
1945 :	h	545	pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y);
1946 :			pMB->b_mvs[k].y = ((Data->currentMV->y == 0)
1947 :			? Data->directmvB[k].y
1948 :			: pMB->mvs[k].y - Data->referencemv[k].y);
1949 :	syskin	658	if (Data->qpel) {
1950 :			pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2;
1951 :			pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2;
1952 :			pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2;
1953 :			pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2;
1954 :			}
1955 :
1956 :	chl	530	if (b_mb->mode != MODE_INTER4V) {
1957 :			pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0];
1958 :			pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0];
1959 :	syskin	658	pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0];
1960 :			pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0];
1961 :	chl	530	break;
1962 :			}
1963 :			}
1964 :	h	545	return skip_sad;
1965 :	Isibaar	3	}
1966 :	chl	96
1967 :	h	545
1968 :	chl	530	static __inline void
1969 :			SearchInterpolate(const uint8_t * const f_Ref,
1970 :			const uint8_t * const f_RefH,
1971 :			const uint8_t * const f_RefV,
1972 :			const uint8_t * const f_RefHV,
1973 :			const uint8_t * const b_Ref,
1974 :			const uint8_t * const b_RefH,
1975 :			const uint8_t * const b_RefV,
1976 :			const uint8_t * const b_RefHV,
1977 :			const IMAGE * const pCur,
1978 :			const int x, const int y,
1979 :			const uint32_t fcode,
1980 :			const uint32_t bcode,
1981 :			const uint32_t MotionFlags,
1982 :			const MBParam * const pParam,
1983 :			const VECTOR * const f_predMV,
1984 :			const VECTOR * const b_predMV,
1985 :			MACROBLOCK * const pMB,
1986 :	h	545	int32_t * const best_sad,
1987 :			SearchData * const fData)
1988 :	chl	530
1989 :	chl	96	{
1990 :
1991 :	edgomez	195	const int32_t iEdgedWidth = pParam->edged_width;
1992 :	chl	530	int iDirection, i, j;
1993 :	h	545	SearchData bData;
1994 :	chl	96
1995 :	syskin	601	(bData.iMinSAD = fData->iMinSAD) = 4096256;
1996 :	h	545	bData.Cur = fData->Cur;
1997 :			fData->iEdgedWidth = bData.iEdgedWidth = iEdgedWidth;
1998 :	syskin	658	bData.currentMV = fData->currentMV + 1; bData.currentQMV = fData->currentQMV + 1;
1999 :	syskin	601	bData.lambda16 = fData->lambda16;
2000 :	h	545	fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode;
2001 :	edgomez	195
2002 :	h	545	bData.bRef = fData->Ref = f_Ref + (x + y * iEdgedWidth) * 16;
2003 :			bData.bRefH = fData->RefH = f_RefH + (x + y * iEdgedWidth) * 16;
2004 :			bData.bRefV = fData->RefV = f_RefV + (x + y * iEdgedWidth) * 16;
2005 :			bData.bRefHV = fData->RefHV = f_RefHV + (x + y * iEdgedWidth) * 16;
2006 :			bData.Ref = fData->bRef = b_Ref + (x + y * iEdgedWidth) * 16;
2007 :			bData.RefH = fData->bRefH = b_RefH + (x + y * iEdgedWidth) * 16;
2008 :			bData.RefV = fData->bRefV = b_RefV + (x + y * iEdgedWidth) * 16;
2009 :			bData.RefHV = fData->bRefHV = b_RefHV + (x + y * iEdgedWidth) * 16;
2010 :	syskin	658	bData.RefQ = fData->RefQ;
2011 :	chl	96
2012 :	h	545	bData.bpredMV = fData->predMV = *f_predMV;
2013 :			fData->bpredMV = bData.predMV = *b_predMV;
2014 :	edgomez	195
2015 :	syskin	658	fData->currentMV[0] = fData->currentMV[3];
2016 :	Isibaar	580	get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, pParam->m_quarterpel);
2017 :			get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, pParam->m_quarterpel);
2018 :	chl	96
2019 :	h	545	if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx;
2020 :			if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dy;
2021 :			if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dx;
2022 :			if (fData->currentMV[0].y > fData->min_dy) fData->currentMV[0].y = fData->min_dy;
2023 :	Isibaar	539
2024 :	h	545	if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx;
2025 :			if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dy;
2026 :			if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dx;
2027 :			if (fData->currentMV[1].y > bData.min_dy) fData->currentMV[1].y = bData.min_dy;
2028 :	Isibaar	539
2029 :	h	545	CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData);
2030 :	chl	96
2031 :	chl	530	//diamond. I wish we could use normal mainsearch functions (square, advdiamond)
2032 :	chl	96
2033 :	chl	530	do {
2034 :			iDirection = 255;
2035 :			// forward MV moves
2036 :	h	545	i = fData->currentMV[0].x; j = fData->currentMV[0].y;
2037 :	chl	96
2038 :	h	545	CheckCandidateInt(i + 1, j, 0, &iDirection, fData);
2039 :			CheckCandidateInt(i, j + 1, 0, &iDirection, fData);
2040 :			CheckCandidateInt(i - 1, j, 0, &iDirection, fData);
2041 :			CheckCandidateInt(i, j - 1, 0, &iDirection, fData);
2042 :	chl	96
2043 :	chl	530	// backward MV moves
2044 :	h	545	i = fData->currentMV[1].x; j = fData->currentMV[1].y;
2045 :			fData->currentMV[2] = fData->currentMV[0];
2046 :	Isibaar	539	CheckCandidateInt(i + 1, j, 0, &iDirection, &bData);
2047 :			CheckCandidateInt(i, j + 1, 0, &iDirection, &bData);
2048 :			CheckCandidateInt(i - 1, j, 0, &iDirection, &bData);
2049 :			CheckCandidateInt(i, j - 1, 0, &iDirection, &bData);
2050 :	chl	96
2051 :	chl	530	} while (!(iDirection));
2052 :	edgomez	195
2053 :	syskin	628	fData->iMinSAD += 2 fData->lambda16; // two bits are needed to code interpolate mode.
2054 :	Isibaar	606
2055 :	syskin	658	if (fData->qpel) {
2056 :			CheckCandidate = CheckCandidateInt_qpel;
2057 :			get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 0);
2058 :			get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 0);
2059 :			fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x;
2060 :			fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y;
2061 :			fData->currentQMV[1].x = 2 * fData->currentMV[1].x;
2062 :			fData->currentQMV[1].y = 2 * fData->currentMV[1].y;
2063 :			QuarterpelRefine(fData);
2064 :			fData->currentQMV[2] = fData->currentQMV[0];
2065 :			QuarterpelRefine(&bData);
2066 :			}
2067 :
2068 :	h	545	if (fData->iMinSAD < best_sad) {
2069 :			best_sad = fData->iMinSAD;
2070 :			pMB->mvs[0] = fData->currentMV[0];
2071 :			pMB->b_mvs[0] = fData->currentMV[1];
2072 :	chl	530	pMB->mode = MODE_INTERPOLATE;
2073 :	syskin	658	if (fData->qpel) {
2074 :			pMB->qmvs[0] = fData->currentQMV[0];
2075 :			pMB->b_qmvs[0] = fData->currentQMV[1];
2076 :			pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x;
2077 :			pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y;
2078 :			pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x;
2079 :			pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y;
2080 :			} else {
2081 :			pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x;
2082 :			pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y;
2083 :			pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x;
2084 :			pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y;
2085 :			}
2086 :	chl	530	}
2087 :			}
2088 :	chl	96
2089 :	chl	530	void
2090 :			MotionEstimationBVOP(MBParam * const pParam,
2091 :			FRAMEINFO * const frame,
2092 :			const int32_t time_bp,
2093 :			const int32_t time_pp,
2094 :			// forward (past) reference
2095 :			const MACROBLOCK * const f_mbs,
2096 :			const IMAGE * const f_ref,
2097 :			const IMAGE * const f_refH,
2098 :			const IMAGE * const f_refV,
2099 :			const IMAGE * const f_refHV,
2100 :			// backward (future) reference
2101 :	syskin	644	const FRAMEINFO * const b_reference,
2102 :	chl	530	const IMAGE * const b_ref,
2103 :			const IMAGE * const b_refH,
2104 :			const IMAGE * const b_refV,
2105 :			const IMAGE * const b_refHV)
2106 :			{
2107 :			uint32_t i, j;
2108 :			int32_t best_sad, skip_sad;
2109 :			int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0;
2110 :			static const VECTOR zeroMV={0,0};
2111 :	syskin	644	const MACROBLOCK * const b_mbs = b_reference->mbs;
2112 :	chl	96
2113 :	chl	530	VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/
2114 :	chl	96
2115 :	chl	530	const int32_t TRB = time_pp - time_bp;
2116 :			const int32_t TRD = time_pp;
2117 :	syskin	658	uint8_t * qimage;
2118 :	chl	96
2119 :	h	545	// some pre-inintialized data for the rest of the search
2120 :
2121 :			SearchData Data;
2122 :			int32_t iMinSAD;
2123 :			VECTOR currentMV[3];
2124 :	syskin	658	VECTOR currentQMV[3];
2125 :	h	545	Data.iEdgedWidth = pParam->edged_width;
2126 :	syskin	658	Data.currentMV = currentMV; Data.currentQMV = currentQMV;
2127 :	h	545	Data.iMinSAD = &iMinSAD;
2128 :	syskin	601	Data.lambda16 = lambda_vec16[frame->quant];
2129 :	syskin	658	Data.qpel = pParam->m_quarterpel;
2130 :	h	545
2131 :	syskin	658	if((qimage = (uint8_t ) malloc(32 pParam->edged_width)) == NULL)
2132 :			return; // allocate some mem for qpel interpolated blocks
2133 :			// somehow this is dirty since I think we shouldn't use malloc outside
2134 :			// encoder_create() - so please fix me!
2135 :			Data.RefQ = qimage;
2136 :
2137 :	chl	530	// note: i==horizontal, j==vertical
2138 :			for (j = 0; j < pParam->mb_height; j++) {
2139 :	chl	96
2140 :	chl	530	f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */
2141 :	edgomez	195
2142 :	chl	530	for (i = 0; i < pParam->mb_width; i++) {
2143 :			MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width;
2144 :			const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width;
2145 :	edgomez	195
2146 :	chl	619	/* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */
2147 :	syskin	644	if (b_reference->coding_type != S_VOP)
2148 :			if (b_mb->mode == MODE_NOT_CODED) {
2149 :			pMB->mode = MODE_NOT_CODED;
2150 :			continue;
2151 :			}
2152 :	chl	96
2153 :	h	545	Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16;
2154 :	syskin	601	pMB->quant = frame->quant;
2155 :	syskin	628
2156 :	chl	530	/* direct search comes first, because it (1) checks for SKIP-mode
2157 :			and (2) sets very good predictions for forward and backward search */
2158 :	Isibaar	539	skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2159 :			b_ref, b_refH->y, b_refV->y, b_refHV->y,
2160 :	chl	530	&frame->image,
2161 :			i, j,
2162 :			frame->motion_flags,
2163 :			TRB, TRD,
2164 :			pParam,
2165 :			pMB, b_mb,
2166 :	h	545	&best_sad,
2167 :			&Data);
2168 :	chl	96
2169 :	Isibaar	539	if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; }
2170 :	syskin	658
2171 :	chl	530	// forward search
2172 :			SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y,
2173 :			&frame->image, i, j,
2174 :			frame->motion_flags,
2175 :	h	545	frame->fcode, pParam,
2176 :	chl	530	pMB, &f_predMV, &best_sad,
2177 :	h	545	MODE_FORWARD, &Data);
2178 :	chl	96
2179 :	chl	530	// backward search
2180 :			SearchBF(b_ref->y, b_refH->y, b_refV->y, b_refHV->y,
2181 :			&frame->image, i, j,
2182 :			frame->motion_flags,
2183 :	h	545	frame->bcode, pParam,
2184 :	chl	530	pMB, &b_predMV, &best_sad,
2185 :	h	545	MODE_BACKWARD, &Data);
2186 :	chl	96
2187 :	chl	530	// interpolate search comes last, because it uses data from forward and backward as prediction
2188 :	chl	96
2189 :	chl	530	SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y,
2190 :			b_ref->y, b_refH->y, b_refV->y, b_refHV->y,
2191 :			&frame->image,
2192 :			i, j,
2193 :			frame->fcode, frame->bcode,
2194 :			frame->motion_flags,
2195 :	h	545	pParam,
2196 :	chl	530	&f_predMV, &b_predMV,
2197 :	h	545	pMB, &best_sad,
2198 :			&Data);
2199 :	chl	96
2200 :	chl	530	switch (pMB->mode) {
2201 :			case MODE_FORWARD:
2202 :			f_count++;
2203 :	syskin	658	if (pParam->m_quarterpel) f_predMV = pMB->qmvs[0];
2204 :			else f_predMV = pMB->mvs[0];
2205 :	chl	530	break;
2206 :			case MODE_BACKWARD:
2207 :			b_count++;
2208 :	syskin	658	if (pParam->m_quarterpel) b_predMV = pMB->b_qmvs[0];
2209 :			else b_predMV = pMB->b_mvs[0];
2210 :	chl	530	break;
2211 :			case MODE_INTERPOLATE:
2212 :			i_count++;
2213 :	syskin	658	if (pParam->m_quarterpel) {
2214 :			f_predMV = pMB->qmvs[0];
2215 :			b_predMV = pMB->b_qmvs[0];
2216 :			} else {
2217 :			f_predMV = pMB->mvs[0];
2218 :			b_predMV = pMB->b_mvs[0];
2219 :			}
2220 :	chl	530	break;
2221 :			case MODE_DIRECT:
2222 :			case MODE_DIRECT_NO4V:
2223 :			d_count++;
2224 :			break;
2225 :			default:
2226 :			break;
2227 :	chl	96	}
2228 :			}
2229 :			}
2230 :	syskin	658	free(qimage);
2231 :	chl	96	}
2232 :
2233 :	chl	530	/* Hinted ME starts here */
2234 :	chl	96
2235 :	syskin	594	static void
2236 :	syskin	628	SearchPhinted ( const IMAGE * const pRef,
2237 :	chl	289	const uint8_t * const pRefH,
2238 :			const uint8_t * const pRefV,
2239 :			const uint8_t * const pRefHV,
2240 :			const IMAGE * const pCur,
2241 :			const int x,
2242 :			const int y,
2243 :			const uint32_t MotionFlags,
2244 :			const uint32_t iQuant,
2245 :			const MBParam * const pParam,
2246 :			const MACROBLOCK * const pMBs,
2247 :	chl	530	int inter4v,
2248 :	h	545	MACROBLOCK * const pMB,
2249 :	edgomez	568	SearchData * const Data)
2250 :	chl	289	{
2251 :	chl	530
2252 :	edgomez	568	int i, t;
2253 :	chl	530	MainSearchFunc * MainSearchPtr;
2254 :	suxen_drol	118
2255 :	h	545	Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0);
2256 :	syskin	628	Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0);
2257 :	h	545	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2258 :	Isibaar	580	pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel);
2259 :	suxen_drol	118
2260 :	syskin	628	Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16;
2261 :			Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8;
2262 :			Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8;
2263 :	Isibaar	600
2264 :	syskin	628	Data->Ref = pRef->y + (x + Data->iEdgedWidthy) 16;
2265 :			Data->RefH = pRefH + (x + Data->iEdgedWidthy) 16;
2266 :			Data->RefV = pRefV + (x + Data->iEdgedWidthy) 16;
2267 :			Data->RefHV = pRefHV + (x + Data->iEdgedWidthy) 16;
2268 :			Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8;
2269 :			Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8;
2270 :
2271 :	chl	530	if (!(MotionFlags & PMV_HALFPEL16)) {
2272 :	h	545	Data->min_dx = EVEN(Data->min_dx);
2273 :			Data->max_dx = EVEN(Data->max_dx);
2274 :			Data->min_dy = EVEN(Data->min_dy);
2275 :			Data->max_dy = EVEN(Data->max_dy);
2276 :	chl	530	}
2277 :	chl	289
2278 :	h	545	for(i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR;
2279 :	chl	289
2280 :	chl	530	if (pMB->dquant != NO_CHANGE) inter4v = 0;
2281 :	chl	289
2282 :	syskin	628	if (inter4v \|\| pParam->m_quarterpel \|\| Data->chroma) CheckCandidate = CheckCandidate16;
2283 :			else CheckCandidate = CheckCandidate16no4v;
2284 :	chl	289
2285 :	chl	530	pMB->mvs[0].x = EVEN(pMB->mvs[0].x);
2286 :			pMB->mvs[0].y = EVEN(pMB->mvs[0].y);
2287 :	h	545	if (pMB->mvs[0].x > Data->max_dx) pMB->mvs[0].x = Data->max_dx; // this is in case iFcode changed
2288 :			if (pMB->mvs[0].x < Data->min_dx) pMB->mvs[0].x = Data->min_dx;
2289 :			if (pMB->mvs[0].y > Data->max_dy) pMB->mvs[0].y = Data->max_dy;
2290 :			if (pMB->mvs[0].y < Data->min_dy) pMB->mvs[0].y = Data->min_dy;
2291 :	chl	370
2292 :	edgomez	568	(*CheckCandidate)(pMB->mvs[0].x, pMB->mvs[0].y, 0, &t, Data);
2293 :	chl	289
2294 :	chl	530	if (pMB->mode == MODE_INTER4V)
2295 :			for (i = 1; i < 4; i++) { // all four vectors will be used as four predictions for 16x16 search
2296 :			pMB->mvs[i].x = EVEN(pMB->mvs[i].x);
2297 :			pMB->mvs[i].y = EVEN(pMB->mvs[i].y);
2298 :			if (!(make_mask(pMB->mvs, i)))
2299 :	edgomez	568	(*CheckCandidate)(pMB->mvs[i].x, pMB->mvs[i].y, 0, &t, Data);
2300 :	chl	289	}
2301 :
2302 :			if (MotionFlags & PMV_USESQUARES16)
2303 :	chl	530	MainSearchPtr = SquareSearch;
2304 :	chl	289	else if (MotionFlags & PMV_ADVANCEDDIAMOND16)
2305 :	chl	530	MainSearchPtr = AdvDiamondSearch;
2306 :			else MainSearchPtr = DiamondSearch;
2307 :	chl	289
2308 :	h	545	(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255);
2309 :	chl	289
2310 :	h	545	if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data);
2311 :	chl	289
2312 :	syskin	594	for(i = 0; i < 5; i++) {
2313 :			Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors
2314 :			Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
2315 :			}
2316 :	chl	289
2317 :	syskin	594	if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) {
2318 :	Isibaar	606	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2319 :			pParam->width, pParam->height, Data->iFcode, 0);
2320 :			CheckCandidate = CheckCandidate16_qpel;
2321 :	syskin	594	QuarterpelRefine(Data);
2322 :			}
2323 :
2324 :			if (inter4v) {
2325 :			SearchData Data8;
2326 :			Data8.iFcode = Data->iFcode;
2327 :	Isibaar	606	Data8.lambda8 = Data->lambda8;
2328 :	syskin	594	Data8.iEdgedWidth = Data->iEdgedWidth;
2329 :	Isibaar	606	Data8.RefQ = Data->RefQ;
2330 :	syskin	628	Data8.qpel = Data->qpel;
2331 :			Search8(Data, 2x, 2y, MotionFlags, pParam, pMB, pMBs, 0, &Data8);
2332 :			Search8(Data, 2x + 1, 2y, MotionFlags, pParam, pMB, pMBs, 1, &Data8);
2333 :			Search8(Data, 2x, 2y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8);
2334 :			Search8(Data, 2x + 1, 2y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8);
2335 :
2336 :			if (Data->chroma) {
2337 :			int sum, dx, dy;
2338 :
2339 :			if(pParam->m_quarterpel)
2340 :	syskin	658	sum = (pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2);
2341 :	syskin	628	else sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y;
2342 :			dy = (sum ? SIGN(sum) *
2343 :			(roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2) : 0);
2344 :
2345 :			if(pParam->m_quarterpel)
2346 :	syskin	658	sum = (pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2);
2347 :	syskin	628	else sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x;
2348 :			dx = (sum ? SIGN(sum) *
2349 :			(roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2) : 0);
2350 :			Data->iMinSAD[1] += ChromaSAD(dx, dy, Data);
2351 :			}
2352 :	syskin	594	}
2353 :
2354 :	chl	530	if (!(inter4v) \|\|
2355 :	h	545	(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] +
2356 :			Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) {
2357 :	chl	530	// INTER MODE
2358 :			pMB->mode = MODE_INTER;
2359 :	Isibaar	539	pMB->mvs[0] = pMB->mvs[1]
2360 :	h	545	= pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
2361 :	chl	289
2362 :	Isibaar	606	pMB->qmvs[0] = pMB->qmvs[1]
2363 :			= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
2364 :
2365 :	chl	530	pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] =
2366 :	h	545	pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0];
2367 :	chl	289
2368 :	Isibaar	606	if(pParam->m_quarterpel) {
2369 :			pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predQMV.x;
2370 :			pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predQMV.y;
2371 :	syskin	628	} else {
2372 :	Isibaar	606	pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
2373 :			pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
2374 :			}
2375 :	chl	530	} else {
2376 :	syskin	628	// INTER4V MODE; all other things are already set in Search8
2377 :	chl	530	pMB->mode = MODE_INTER4V;
2378 :	h	545	pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3]
2379 :			+ Data->iMinSAD[4] + IMV16X16 * iQuant;
2380 :	chl	289	}
2381 :
2382 :			}
2383 :
2384 :	edgomez	195	void
2385 :	chl	530	MotionEstimationHinted( MBParam * const pParam,
2386 :			FRAMEINFO * const current,
2387 :			FRAMEINFO * const reference,
2388 :			const IMAGE * const pRefH,
2389 :			const IMAGE * const pRefV,
2390 :			const IMAGE * const pRefHV)
2391 :	suxen_drol	118	{
2392 :	chl	530	MACROBLOCK *const pMBs = current->mbs;
2393 :			const IMAGE *const pCurrent = &current->image;
2394 :			const IMAGE *const pRef = &reference->image;
2395 :	suxen_drol	118
2396 :	chl	530	uint32_t x, y;
2397 :	syskin	601	uint8_t * qimage;
2398 :	syskin	574	int32_t temp[5], quant = current->quant;
2399 :	h	545	int32_t iMinSAD[5];
2400 :	Isibaar	606	VECTOR currentMV[5], currentQMV[5];
2401 :	h	545	SearchData Data;
2402 :			Data.iEdgedWidth = pParam->edged_width;
2403 :			Data.currentMV = currentMV;
2404 :	Isibaar	606	Data.currentQMV = currentQMV;
2405 :	h	545	Data.iMinSAD = iMinSAD;
2406 :			Data.temp = temp;
2407 :			Data.iFcode = current->fcode;
2408 :	syskin	594	Data.rounding = pParam->m_rounding_type;
2409 :	syskin	628	Data.qpel = pParam->m_quarterpel;
2410 :			Data.chroma = current->global_flags & XVID_ME_COLOUR;
2411 :	h	545
2412 :	syskin	594	if((qimage = (uint8_t ) malloc(32 pParam->edged_width)) == NULL)
2413 :			return; // allocate some mem for qpel interpolated blocks
2414 :			// somehow this is dirty since I think we shouldn't use malloc outside
2415 :			// encoder_create() - so please fix me!
2416 :
2417 :	syskin	601	Data.RefQ = qimage;
2418 :
2419 :	chl	530	if (sadInit) (*sadInit) ();
2420 :	suxen_drol	118
2421 :	chl	530	for (y = 0; y < pParam->mb_height; y++) {
2422 :			for (x = 0; x < pParam->mb_width; x++) {
2423 :	chl	312
2424 :	chl	530	MACROBLOCK pMB = &pMBs[x + y pParam->mb_width];
2425 :	suxen_drol	118
2426 :	chl	530	//intra mode is copied from the first pass. At least for the time being
2427 :			if ((pMB->mode == MODE_INTRA) \|\| (pMB->mode == MODE_NOT_CODED) ) continue;
2428 :	chl	346
2429 :	chl	530	if (!(current->global_flags & XVID_LUMIMASKING)) {
2430 :			pMB->dquant = NO_CHANGE;
2431 :			pMB->quant = current->quant; }
2432 :	Isibaar	606	else {
2433 :	syskin	574	if (pMB->dquant != NO_CHANGE) {
2434 :			quant += DQtab[pMB->dquant];
2435 :			if (quant > 31) quant = 31;
2436 :			else if (quant < 1) quant = 1;
2437 :			}
2438 :	Isibaar	606	pMB->quant = quant;
2439 :			}
2440 :	chl	341
2441 :	syskin	628	SearchPhinted(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
2442 :	Isibaar	539	y, current->motion_flags, pMB->quant,
2443 :	h	545	pParam, pMBs, current->global_flags & XVID_INTER4V, pMB,
2444 :			&Data);
2445 :	chl	317
2446 :	Isibaar	539	}
2447 :			}
2448 :	syskin	594	free(qimage);
2449 :	Isibaar	539	}
2450 :	suxen_drol	118
2451 :	Isibaar	539	static __inline int
2452 :			MEanalyzeMB ( const uint8_t * const pRef,
2453 :			const uint8_t * const pCur,
2454 :			const int x,
2455 :			const int y,
2456 :			const MBParam * const pParam,
2457 :			const MACROBLOCK * const pMBs,
2458 :	h	545	MACROBLOCK * const pMB,
2459 :			SearchData * const Data)
2460 :	Isibaar	539	{
2461 :	chl	317
2462 :	Isibaar	606	int i = 255, mask;
2463 :	h	545	VECTOR pmv[3];
2464 :	syskin	640	*(Data->iMinSAD) = MV_MAX_ERROR;
2465 :	suxen_drol	118
2466 :	syskin	640	//median is only used as prediction. it doesn't have to be real
2467 :			if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0;
2468 :			else
2469 :			if (x == 1) //left macroblock does not have any vector now
2470 :			Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median
2471 :			else if (y == 1) // top macroblock don't have it's vector
2472 :			Data->predMV = (pMB - 1)->mvs[0]; // left instead of median
2473 :			else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median
2474 :
2475 :	h	545	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2476 :	Isibaar	580	pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel);
2477 :	chl	326
2478 :	h	545	Data->Cur = pCur + (x + y * pParam->edged_width) * 16;
2479 :			Data->Ref = pRef + (x + y * pParam->edged_width) * 16;
2480 :	Isibaar	539
2481 :			pmv[1].x = EVEN(pMB->mvs[0].x);
2482 :			pmv[1].y = EVEN(pMB->mvs[0].y);
2483 :	syskin	640	pmv[2].x = EVEN(Data->predMV.x);
2484 :			pmv[2].y = EVEN(Data->predMV.y);
2485 :			pmv[0].x = pmv[0].y = 0;
2486 :	chl	326
2487 :	syskin	640	(*CheckCandidate)(0, 0, 255, &i, Data);
2488 :
2489 :			//early skip for 0,0
2490 :			if (Data->iMinSAD < MAX_SAD00_FOR_SKIP 4) {
2491 :			pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
2492 :			pMB->mode = MODE_NOT_CODED;
2493 :			return 0;
2494 :			}
2495 :
2496 :	Isibaar	539	if (!(mask = make_mask(pmv, 1)))
2497 :	syskin	640	(*CheckCandidate)(pmv[1].x, pmv[1].y, mask, &i, Data);
2498 :	Isibaar	539	if (!(mask = make_mask(pmv, 2)))
2499 :	syskin	640	(*CheckCandidate)(pmv[2].x, pmv[2].y, mask, &i, Data);
2500 :	Isibaar	539
2501 :	syskin	640	if (Data->iMinSAD > MAX_SAD00_FOR_SKIP 4) // diamond only if needed
2502 :			DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i);
2503 :	Isibaar	539
2504 :	syskin	640	pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
2505 :	syskin	628	pMB->mode = MODE_INTER;
2506 :	h	545	return *(Data->iMinSAD);
2507 :	Isibaar	539	}
2508 :
2509 :			#define INTRA_THRESH 1350
2510 :			#define INTER_THRESH 900
2511 :
2512 :	syskin	644
2513 :	Isibaar	539	int
2514 :			MEanalysis( const IMAGE * const pRef,
2515 :	syskin	644	FRAMEINFO * const Current,
2516 :	Isibaar	539	MBParam * const pParam,
2517 :	syskin	644	int maxIntra, //maximum number if non-I frames
2518 :			int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame
2519 :			int bCount) // number if B frames in a row
2520 :	Isibaar	539	{
2521 :			uint32_t x, y, intra = 0;
2522 :			int sSAD = 0;
2523 :	syskin	644	MACROBLOCK * const pMBs = Current->mbs;
2524 :			const IMAGE * const pCurrent = &Current->image;
2525 :			int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH;
2526 :	h	545
2527 :			VECTOR currentMV;
2528 :			int32_t iMinSAD;
2529 :			SearchData Data;
2530 :			Data.iEdgedWidth = pParam->edged_width;
2531 :			Data.currentMV = &currentMV;
2532 :			Data.iMinSAD = &iMinSAD;
2533 :	syskin	644	Data.iFcode = Current->fcode;
2534 :	syskin	640	CheckCandidate = CheckCandidate16no4vI;
2535 :	h	545
2536 :	syskin	658	if (intraCount < 10) // we're right after an I frame
2537 :			IntraThresh += 4 * (intraCount - 10) * (intraCount - 10);
2538 :	syskin	644	else
2539 :			if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec
2540 :			IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra;
2541 :
2542 :
2543 :			InterThresh += 300 * (1 - bCount);
2544 :			if (InterThresh < 200) InterThresh = 200;
2545 :
2546 :	Isibaar	539	if (sadInit) (*sadInit) ();
2547 :
2548 :	syskin	601	for (y = 1; y < pParam->mb_height-1; y++) {
2549 :			for (x = 1; x < pParam->mb_width-1; x++) {
2550 :	Isibaar	539	int sad, dev;
2551 :			MACROBLOCK pMB = &pMBs[x + y pParam->mb_width];
2552 :
2553 :			sad = MEanalyzeMB(pRef->y, pCurrent->y, x, y,
2554 :	h	545	pParam, pMBs, pMB, &Data);
2555 :	Isibaar	539
2556 :	syskin	644	if (sad > IntraThresh) {
2557 :	syskin	601	dev = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16,
2558 :			pParam->edged_width);
2559 :	syskin	644	if (dev + IntraThresh < sad) {
2560 :			pMB->mode = MODE_INTRA;
2561 :			if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return 2; // I frame
2562 :	Isibaar	539	}
2563 :	syskin	644	}
2564 :	syskin	601	sSAD += sad;
2565 :	suxen_drol	118	}
2566 :			}
2567 :	Isibaar	539	sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2);
2568 :	syskin	644	if (sSAD > InterThresh ) return 1; //P frame
2569 :	Isibaar	539	emms();
2570 :			return 0; // B frame
2571 :
2572 :	suxen_drol	118	}
2573 :	syskin	576
2574 :			int
2575 :			FindFcode( const MBParam * const pParam,
2576 :			const FRAMEINFO * const current)
2577 :			{
2578 :			uint32_t x, y;
2579 :			int max = 0, min = 0, i;
2580 :
2581 :			for (y = 0; y < pParam->mb_height; y++) {
2582 :			for (x = 0; x < pParam->mb_width; x++) {
2583 :
2584 :			MACROBLOCK pMB = &current->mbs[x + y pParam->mb_width];
2585 :			for(i = 0; i < (pMB->mode == MODE_INTER4V ? 4:1); i++) {
2586 :			if (pMB->mvs[i].x > max) max = pMB->mvs[i].x;
2587 :			if (pMB->mvs[i].y > max) max = pMB->mvs[i].y;
2588 :
2589 :			if (pMB->mvs[i].x < min) min = pMB->mvs[i].x;
2590 :			if (pMB->mvs[i].y < min) min = pMB->mvs[i].y;
2591 :			}
2592 :			}
2593 :			}
2594 :
2595 :			min = -min;
2596 :			max += 1;
2597 :			if (min > max) max = min;
2598 :	Isibaar	606	if (pParam->m_quarterpel) max *= 2;
2599 :	syskin	576
2600 :			for (i = 1; (max > 32 << (i - 1)); i++);
2601 :			return i;
2602 :			}
2603 :	chl	619
2604 :	syskin	628	static void
2605 :			CheckGMC(int x, int y, const int dir, int * iDirection,
2606 :			const MACROBLOCK * const pMBs, uint32_t * bestcount, VECTOR * GMC,
2607 :			const MBParam * const pParam)
2608 :			{
2609 :			uint32_t mx, my, a, count = 0;
2610 :	chl	619
2611 :	syskin	628	for (my = 1; my < pParam->mb_height-1; my++)
2612 :			for (mx = 1; mx < pParam->mb_width-1; mx++) {
2613 :			VECTOR mv;
2614 :			const MACROBLOCK pMB = &pMBs[mx + my pParam->mb_width];
2615 :			if (pMB->mode == MODE_INTRA \|\| pMB->mode == MODE_NOT_CODED) continue;
2616 :			mv = pMB->mvs[0];
2617 :			a = ABS(mv.x - x) + ABS(mv.y - y);
2618 :			if (a < 6) count += 6 - a;
2619 :			}
2620 :	chl	619
2621 :	syskin	628	if (count > *bestcount) {
2622 :			*bestcount = count;
2623 :			*iDirection = dir;
2624 :			GMC->x = x; GMC->y = y;
2625 :			}
2626 :			}
2627 :
2628 :
2629 :	chl	619	static VECTOR
2630 :			GlobalMotionEst(const MACROBLOCK * const pMBs, const MBParam * const pParam, const uint32_t iFcode)
2631 :			{
2632 :
2633 :	syskin	628	uint32_t count, bestcount = 0;
2634 :	chl	619	int x, y;
2635 :	syskin	628	VECTOR gmc = {0,0};
2636 :	chl	619	int step, min_x, max_x, min_y, max_y;
2637 :			uint32_t mx, my;
2638 :	syskin	628	int iDirection, bDirection;
2639 :	chl	619
2640 :			min_x = min_y = -32<<iFcode;
2641 :			max_x = max_y = 32<<iFcode;
2642 :
2643 :	syskin	628	//step1: let's find a rough camera panning
2644 :	chl	619	for (step = 32; step >= 2; step /= 2) {
2645 :			bestcount = 0;
2646 :			for (y = min_y; y <= max_y; y += step)
2647 :			for (x = min_x ; x <= max_x; x += step) {
2648 :			count = 0;
2649 :			//for all macroblocks
2650 :			for (my = 1; my < pParam->mb_height-1; my++)
2651 :			for (mx = 1; mx < pParam->mb_width-1; mx++) {
2652 :			const MACROBLOCK pMB = &pMBs[mx + my pParam->mb_width];
2653 :			VECTOR mv;
2654 :
2655 :			if (pMB->mode == MODE_INTRA \|\| pMB->mode == MODE_NOT_CODED)
2656 :			continue;
2657 :
2658 :			mv = pMB->mvs[0];
2659 :			if ( ABS(mv.x - x) <= step && ABS(mv.y - y) <= step ) /* GMC translation is always halfpel-res */
2660 :			count++;
2661 :			}
2662 :			if (count >= bestcount) { bestcount = count; gmc.x = x; gmc.y = y; }
2663 :			}
2664 :			min_x = gmc.x - step;
2665 :			max_x = gmc.x + step;
2666 :			min_y = gmc.y - step;
2667 :			max_y = gmc.y + step;
2668 :
2669 :			}
2670 :	syskin	628
2671 :			if (bestcount < (pParam->mb_height-2)*(pParam->mb_width-2)/10)
2672 :			gmc.x = gmc.y = 0; //no camara pan, no GMC
2673 :
2674 :			// step2: let's refine camera panning using gradiend-descent approach.
2675 :			// TODO: more warping points may be evaluated here (like in interpolate mode search - two vectors in one diamond)
2676 :			bestcount = 0;
2677 :			CheckGMC(gmc.x, gmc.y, 255, &iDirection, pMBs, &bestcount, &gmc, pParam);
2678 :			do {
2679 :			x = gmc.x; y = gmc.y;
2680 :			bDirection = iDirection; iDirection = 0;
2681 :			if (bDirection & 1) CheckGMC(x - 1, y, 1+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam);
2682 :			if (bDirection & 2) CheckGMC(x + 1, y, 2+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam);
2683 :			if (bDirection & 4) CheckGMC(x, y - 1, 1+2+4, &iDirection, pMBs, &bestcount, &gmc, pParam);
2684 :			if (bDirection & 8) CheckGMC(x, y + 1, 1+2+8, &iDirection, pMBs, &bestcount, &gmc, pParam);
2685 :
2686 :			} while (iDirection);
2687 :
2688 :	chl	619	if (pParam->m_quarterpel) {
2689 :			gmc.x *= 2;
2690 :			gmc.y = 2; / we store the halfpel value as pseudo-qpel to make comparison easier */
2691 :			}
2692 :
2693 :			return gmc;
2694 :			}

No admin address has been configured	ViewVC Help
Powered by ViewVC 1.0.4