Annotation of /xvidcore/src/motion/motion_est.c

Revision 1.73 - (view) (download)

1 :	edgomez	1.55	/**************************************************************************
2 :	Isibaar	1.1	*
3 :	edgomez	1.55	* XVID MPEG-4 VIDEO CODEC
4 :			* motion estimation
5 :	chl	1.26	*
6 :	edgomez	1.55	* This program is an implementation of a part of one or more MPEG-4
7 :			* Video tools as specified in ISO/IEC 14496-2 standard. Those intending
8 :			* to use this software module in hardware or software products are
9 :			* advised that its use may infringe existing patents or copyrights, and
10 :			* any such use would be at such party's own risk. The original
11 :			* developer of this software module and his/her company, and subsequent
12 :			* editors and their companies, will have no liability for use of this
13 :			* software or modifications or derivatives thereof.
14 :	chl	1.45	*
15 :	edgomez	1.55	* This program is free software; you can redistribute it and/or modify
16 :			* it under the terms of the GNU General Public License as published by
17 :			* the Free Software Foundation; either version 2 of the License, or
18 :			* (at your option) any later version.
19 :	chl	1.45	*
20 :	edgomez	1.55	* This program is distributed in the hope that it will be useful,
21 :			* but WITHOUT ANY WARRANTY; without even the implied warranty of
22 :			* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
23 :			* GNU General Public License for more details.
24 :	chl	1.45	*
25 :	edgomez	1.55	* You should have received a copy of the GNU General Public License
26 :			* along with this program; if not, write to the Free Software
27 :			* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
28 :	chl	1.26	*
29 :			*************************************************************************/
30 :	Isibaar	1.1
31 :			#include <assert.h>
32 :			#include <stdio.h>
33 :	chl	1.3	#include <stdlib.h>
34 :	edgomez	1.55	#include <string.h> // memcpy
35 :			#include <math.h> // lrint
36 :	Isibaar	1.1
37 :			#include "../encoder.h"
38 :			#include "../utils/mbfunctions.h"
39 :			#include "../prediction/mbprediction.h"
40 :			#include "../global.h"
41 :			#include "../utils/timer.h"
42 :	edgomez	1.55	#include "../image/interpolate8x8.h"
43 :			#include "motion_est.h"
44 :	suxen_drol	1.6	#include "motion.h"
45 :	Isibaar	1.1	#include "sad.h"
46 :	edgomez	1.55	#include "../utils/emms.h"
47 :			#include "../dct/fdct.h"
48 :	Isibaar	1.1
49 :	edgomez	1.57	/*****************************************************************************
50 :			* Modified rounding tables -- declared in motion.h
51 :			* Original tables see ISO spec tables 7-6 -> 7-9
52 :			****************************************************************************/
53 :
54 :			const uint32_t roundtab[16] =
55 :			{0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2 };
56 :
57 :			/* K = 4 */
58 :			const uint32_t roundtab_76[16] =
59 :			{ 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1 };
60 :
61 :			/* K = 2 */
62 :			const uint32_t roundtab_78[8] =
63 :			{ 0, 0, 1, 1, 0, 0, 0, 1 };
64 :
65 :			/* K = 1 */
66 :			const uint32_t roundtab_79[4] =
67 :			{ 0, 1, 0, 0 };
68 :
69 :	edgomez	1.55	#define INITIAL_SKIP_THRESH (10)
70 :			#define FINAL_SKIP_THRESH (50)
71 :			#define MAX_SAD00_FOR_SKIP (20)
72 :			#define MAX_CHROMA_SAD_FOR_SKIP (22)
73 :	edgomez	1.20
74 :	edgomez	1.55	#define CHECK_CANDIDATE(X,Y,D) { \
75 :			CheckCandidate((X),(Y), (D), &iDirection, data ); }
76 :	edgomez	1.57
77 :			/*****************************************************************************
78 :			* Code
79 :			****************************************************************************/
80 :	edgomez	1.20
81 :	edgomez	1.55	static __inline uint32_t
82 :			d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv)
83 :			{
84 :	syskin	1.63	int bits;
85 :			const int q = (1 << (iFcode - 1)) - 1;
86 :
87 :			x <<= qpel;
88 :			y <<= qpel;
89 :	edgomez	1.55	if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); }
90 :
91 :			x -= pred.x;
92 :	syskin	1.63	bits = (x != 0 ? iFcode:0);
93 :			x = abs(x);
94 :			x += q;
95 :			x >>= (iFcode - 1);
96 :			bits += mvtab[x];
97 :
98 :	edgomez	1.55	y -= pred.y;
99 :	syskin	1.63	bits += (y != 0 ? iFcode:0);
100 :			y = abs(y);
101 :			y += q;
102 :			y >>= (iFcode - 1);
103 :			bits += mvtab[y];
104 :	edgomez	1.55
105 :	syskin	1.63	return bits;
106 :	edgomez	1.55	}
107 :	chl	1.10
108 :	syskin	1.64	static int32_t ChromaSAD2(const int fx, const int fy, const int bx, const int by,
109 :			const SearchData * const data)
110 :	edgomez	1.55	{
111 :			int sad;
112 :			const uint32_t stride = data->iEdgedWidth/2;
113 :			uint8_t * f_refu = data->RefQ,
114 :			* f_refv = data->RefQ + 8,
115 :			* b_refu = data->RefQ + 16,
116 :			* b_refv = data->RefQ + 24;
117 :	syskin	1.64	int offset = (fx>>1) + (fy>>1)*stride;
118 :	edgomez	1.55
119 :			switch (((fx & 1) << 1) \| (fy & 1)) {
120 :			case 0:
121 :	syskin	1.64	f_refu = (uint8_t*)data->RefP[4] + offset;
122 :			f_refv = (uint8_t*)data->RefP[5] + offset;
123 :	edgomez	1.55	break;
124 :			case 1:
125 :	syskin	1.64	interpolate8x8_halfpel_v(f_refu, data->RefP[4] + offset, stride, data->rounding);
126 :			interpolate8x8_halfpel_v(f_refv, data->RefP[5] + offset, stride, data->rounding);
127 :	edgomez	1.55	break;
128 :			case 2:
129 :	syskin	1.64	interpolate8x8_halfpel_h(f_refu, data->RefP[4] + offset, stride, data->rounding);
130 :			interpolate8x8_halfpel_h(f_refv, data->RefP[5] + offset, stride, data->rounding);
131 :	edgomez	1.55	break;
132 :			default:
133 :	syskin	1.64	interpolate8x8_halfpel_hv(f_refu, data->RefP[4] + offset, stride, data->rounding);
134 :			interpolate8x8_halfpel_hv(f_refv, data->RefP[5] + offset, stride, data->rounding);
135 :	edgomez	1.55	break;
136 :			}
137 :
138 :	syskin	1.64	offset = (bx>>1) + (by>>1)*stride;
139 :	edgomez	1.55	switch (((bx & 1) << 1) \| (by & 1)) {
140 :			case 0:
141 :	syskin	1.64	b_refu = (uint8_t*)data->b_RefP[4] + offset;
142 :			b_refv = (uint8_t*)data->b_RefP[5] + offset;
143 :	edgomez	1.55	break;
144 :			case 1:
145 :	syskin	1.64	interpolate8x8_halfpel_v(b_refu, data->b_RefP[4] + offset, stride, data->rounding);
146 :			interpolate8x8_halfpel_v(b_refv, data->b_RefP[5] + offset, stride, data->rounding);
147 :	edgomez	1.55	break;
148 :			case 2:
149 :	syskin	1.64	interpolate8x8_halfpel_h(b_refu, data->b_RefP[4] + offset, stride, data->rounding);
150 :			interpolate8x8_halfpel_h(b_refv, data->b_RefP[5] + offset, stride, data->rounding);
151 :	edgomez	1.55	break;
152 :			default:
153 :	syskin	1.64	interpolate8x8_halfpel_hv(b_refu, data->b_RefP[4] + offset, stride, data->rounding);
154 :			interpolate8x8_halfpel_hv(b_refv, data->b_RefP[5] + offset, stride, data->rounding);
155 :	edgomez	1.55	break;
156 :			}
157 :
158 :			sad = sad8bi(data->CurU, b_refu, f_refu, stride);
159 :			sad += sad8bi(data->CurV, b_refv, f_refv, stride);
160 :	chl	1.10
161 :	edgomez	1.55	return sad;
162 :			}
163 :	chl	1.10
164 :	edgomez	1.55	static int32_t
165 :	syskin	1.64	ChromaSAD(const int dx, const int dy, const SearchData * const data)
166 :	Isibaar	1.1	{
167 :	edgomez	1.55	int sad;
168 :			const uint32_t stride = data->iEdgedWidth/2;
169 :	syskin	1.64	int offset = (dx>>1) + (dy>>1)*stride;
170 :	Isibaar	1.1
171 :	edgomez	1.55	if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently
172 :			data->temp[5] = dx; data->temp[6] = dy; // backup
173 :	Isibaar	1.1
174 :	edgomez	1.55	switch (((dx & 1) << 1) \| (dy & 1)) {
175 :			case 0:
176 :	syskin	1.64	sad = sad8(data->CurU, data->RefP[4] + offset, stride);
177 :			sad += sad8(data->CurV, data->RefP[5] + offset, stride);
178 :	edgomez	1.55	break;
179 :			case 1:
180 :	syskin	1.64	sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + stride, stride);
181 :			sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + stride, stride);
182 :	edgomez	1.55	break;
183 :			case 2:
184 :	syskin	1.64	sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + 1, stride);
185 :			sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + 1, stride);
186 :	edgomez	1.55	break;
187 :			default:
188 :	syskin	1.64	interpolate8x8_halfpel_hv(data->RefQ, data->RefP[4] + offset, stride, data->rounding);
189 :	edgomez	1.55	sad = sad8(data->CurU, data->RefQ, stride);
190 :
191 :	syskin	1.64	interpolate8x8_halfpel_hv(data->RefQ, data->RefP[5] + offset, stride, data->rounding);
192 :	edgomez	1.55	sad += sad8(data->CurV, data->RefQ, stride);
193 :			break;
194 :			}
195 :			data->temp[7] = sad; //backup, part 2
196 :			return sad;
197 :			}
198 :	Isibaar	1.1
199 :	edgomez	1.55	static __inline const uint8_t *
200 :			GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data)
201 :			{
202 :			// dir : 0 = forward, 1 = backward
203 :	syskin	1.65	const uint8_t const const direction = ( dir == 0 ? data->RefP : data->b_RefP );
204 :	syskin	1.64	const int picture = ((x&1)<<1) \| (y&1);
205 :			const int offset = (x>>1) + (y>>1)*data->iEdgedWidth;
206 :			return direction[picture] + offset;
207 :	edgomez	1.55	}
208 :	Isibaar	1.1
209 :	edgomez	1.55	// this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate
210 :			static __inline const uint8_t *
211 :			GetReference(const int x, const int y, const SearchData * const data)
212 :			{
213 :	syskin	1.64	const int picture = ((x&1)<<1) \| (y&1);
214 :			const int offset = (x>>1) + (y>>1)*data->iEdgedWidth;
215 :			return data->RefP[picture] + offset;
216 :	Isibaar	1.1	}
217 :
218 :	edgomez	1.55	static uint8_t *
219 :			Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data)
220 :	Isibaar	1.1	{
221 :	edgomez	1.55	// create or find a qpel-precision reference picture; return pointer to it
222 :			uint8_t * Reference = data->RefQ + 16*dir;
223 :			const uint32_t iEdgedWidth = data->iEdgedWidth;
224 :			const uint32_t rounding = data->rounding;
225 :			const int halfpel_x = x/2;
226 :			const int halfpel_y = y/2;
227 :			const uint8_t ref1, ref2, ref3, ref4;
228 :
229 :			ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
230 :			ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
231 :			switch( ((x&1)<<1) + (y&1) ) {
232 :	edgomez	1.62	case 3: // x and y in qpel resolution - the "corners" (top left/right and
233 :			// bottom left/right) during qpel refinement
234 :			ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
235 :			ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
236 :			ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
237 :			ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
238 :			ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
239 :			ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
240 :			interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
241 :	edgomez	1.55	break;
242 :
243 :			case 1: // x halfpel, y qpel - top or bottom during qpel refinement
244 :			ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
245 :			ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
246 :			interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
247 :			break;
248 :
249 :			case 2: // x qpel, y halfpel - left or right during qpel refinement
250 :			ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
251 :			ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
252 :			interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
253 :			break;
254 :
255 :	edgomez	1.62	default: // pure halfpel position
256 :			return (uint8_t *) ref1;
257 :	syskin	1.64
258 :	edgomez	1.55	}
259 :			return Reference;
260 :	Isibaar	1.1	}
261 :
262 :	edgomez	1.55	static uint8_t *
263 :			Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data)
264 :	Isibaar	1.1	{
265 :	edgomez	1.55	// create or find a qpel-precision reference picture; return pointer to it
266 :			uint8_t * Reference = data->RefQ + 16*dir;
267 :			const uint32_t iEdgedWidth = data->iEdgedWidth;
268 :			const uint32_t rounding = data->rounding;
269 :			const int halfpel_x = x/2;
270 :			const int halfpel_y = y/2;
271 :			const uint8_t ref1, ref2, ref3, ref4;
272 :
273 :			ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
274 :			switch( ((x&1)<<1) + (y&1) ) {
275 :			case 3: // x and y in qpel resolution - the "corners" (top left/right and
276 :	syskin	1.64	// bottom left/right) during qpel refinement
277 :	edgomez	1.55	ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
278 :			ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
279 :			ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
280 :			interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
281 :			interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding);
282 :			interpolate8x8_avg4(Reference+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8iEdgedWidth, ref3+8iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding);
283 :			interpolate8x8_avg4(Reference+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8iEdgedWidth+8, ref3+8iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding);
284 :			break;
285 :
286 :			case 1: // x halfpel, y qpel - top or bottom during qpel refinement
287 :			ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
288 :			interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
289 :			interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
290 :			interpolate8x8_avg2(Reference+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
291 :			interpolate8x8_avg2(Reference+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
292 :			break;
293 :
294 :			case 2: // x qpel, y halfpel - left or right during qpel refinement
295 :			ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
296 :			interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
297 :			interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
298 :			interpolate8x8_avg2(Reference+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
299 :			interpolate8x8_avg2(Reference+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
300 :			break;
301 :
302 :	edgomez	1.62	default: // pure halfpel position
303 :	edgomez	1.55	return (uint8_t *) ref1;
304 :			}
305 :			return Reference;
306 :	Isibaar	1.1	}
307 :	suxen_drol	1.8
308 :	edgomez	1.55	/* CHECK_CANDIATE FUNCTIONS START */
309 :	suxen_drol	1.8
310 :	edgomez	1.55	static void
311 :			CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
312 :			{
313 :			int xc, yc;
314 :			const uint8_t * Reference;
315 :			VECTOR * current;
316 :			int32_t sad; uint32_t t;
317 :	chl	1.19
318 :	edgomez	1.55	if ( (x > data->max_dx) \|\| (x < data->min_dx)
319 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
320 :	edgomez	1.20
321 :	edgomez	1.55	if (!data->qpel_precision) {
322 :			Reference = GetReference(x, y, data);
323 :			current = data->currentMV;
324 :			xc = x; yc = y;
325 :			} else { // x and y are in 1/4 precision
326 :			Reference = Interpolate16x16qpel(x, y, 0, data);
327 :			xc = x/2; yc = y/2; //for chroma sad
328 :			current = data->currentQMV;
329 :			}
330 :	edgomez	1.20
331 :	edgomez	1.55	sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
332 :			t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
333 :	edgomez	1.20
334 :	edgomez	1.55	sad += (data->lambda16 * t * sad)>>10;
335 :			data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
336 :	edgomez	1.20
337 :	syskin	1.64	if (data->chroma) sad += ChromaSAD( (xc >> 1) + roundtab_79[xc & 0x3],
338 :			(yc >> 1) + roundtab_79[yc & 0x3], data);
339 :	edgomez	1.20
340 :	edgomez	1.55	if (sad < data->iMinSAD[0]) {
341 :			data->iMinSAD[0] = sad;
342 :			current[0].x = x; current[0].y = y;
343 :			*dir = Direction;
344 :			}
345 :	edgomez	1.20
346 :	edgomez	1.55	if (data->temp[1] < data->iMinSAD[1]) {
347 :			data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y = y; }
348 :			if (data->temp[2] < data->iMinSAD[2]) {
349 :			data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; }
350 :			if (data->temp[3] < data->iMinSAD[3]) {
351 :			data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; }
352 :			if (data->temp[4] < data->iMinSAD[4]) {
353 :			data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; }
354 :	Isibaar	1.1	}
355 :
356 :	edgomez	1.55	static void
357 :			CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
358 :			{
359 :			int32_t sad; uint32_t t;
360 :			const uint8_t * Reference;
361 :	syskin	1.63	VECTOR * current;
362 :	edgomez	1.55
363 :			if ( (x > data->max_dx) \|\| (x < data->min_dx)
364 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
365 :	chl	1.35
366 :	syskin	1.63	if (!data->qpel_precision) {
367 :			Reference = GetReference(x, y, data);
368 :			current = data->currentMV;
369 :			} else { // x and y are in 1/4 precision
370 :			Reference = Interpolate8x8qpel(x, y, 0, 0, data);
371 :			current = data->currentQMV;
372 :			}
373 :	Isibaar	1.1
374 :	edgomez	1.55	sad = sad8(data->Cur, Reference, data->iEdgedWidth);
375 :			t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
376 :	chl	1.3
377 :	edgomez	1.55	sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10;
378 :	edgomez	1.20
379 :	edgomez	1.55	if (sad < *(data->iMinSAD)) {
380 :			*(data->iMinSAD) = sad;
381 :	syskin	1.63	current->x = x; current->y = y;
382 :	edgomez	1.55	*dir = Direction;
383 :			}
384 :			}
385 :	edgomez	1.20
386 :	edgomez	1.55	static void
387 :			CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
388 :			{
389 :			uint32_t t;
390 :			const uint8_t * Reference;
391 :	edgomez	1.20
392 :	edgomez	1.62	if ( (!(x&1) && x !=0) \|\| (!(y&1) && y !=0) \|\| //non-zero even value
393 :	edgomez	1.55	(x > data->max_dx) \|\| (x < data->min_dx)
394 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
395 :
396 :			Reference = GetReference(x, y, data);
397 :			t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1);
398 :
399 :			data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
400 :
401 :			data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10;
402 :			data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
403 :
404 :			if (data->temp[0] < data->iMinSAD[0]) {
405 :			data->iMinSAD[0] = data->temp[0];
406 :			data->currentMV[0].x = x; data->currentMV[0].y = y;
407 :			*dir = Direction; }
408 :
409 :			if (data->temp[1] < data->iMinSAD[1]) {
410 :			data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
411 :			if (data->temp[2] < data->iMinSAD[2]) {
412 :			data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
413 :			if (data->temp[3] < data->iMinSAD[3]) {
414 :			data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
415 :			if (data->temp[4] < data->iMinSAD[4]) {
416 :			data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
417 :			}
418 :	edgomez	1.20
419 :	edgomez	1.55	static void
420 :			CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
421 :			{
422 :			int32_t sad, xc, yc;
423 :			const uint8_t * Reference;
424 :			uint32_t t;
425 :			VECTOR * current;
426 :
427 :	edgomez	1.62	if ( (x > data->max_dx) \|\| ( x < data->min_dx)
428 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
429 :	edgomez	1.55
430 :			if (data->rrv && (!(x&1) && x !=0) \| (!(y&1) && y !=0) ) return; //non-zero even value
431 :
432 :			if (data->qpel_precision) { // x and y are in 1/4 precision
433 :			Reference = Interpolate16x16qpel(x, y, 0, data);
434 :			current = data->currentQMV;
435 :			xc = x/2; yc = y/2;
436 :			} else {
437 :			Reference = GetReference(x, y, data);
438 :			current = data->currentMV;
439 :			xc = x; yc = y;
440 :			}
441 :			t = d_mv_bits(x, y, data->predMV, data->iFcode,
442 :			data->qpel^data->qpel_precision, data->rrv);
443 :	edgomez	1.20
444 :	edgomez	1.55	sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096);
445 :			sad += (data->lambda16 * t * sad)>>10;
446 :	edgomez	1.20
447 :	edgomez	1.55	if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
448 :			(yc >> 1) + roundtab_79[yc & 0x3], data);
449 :	edgomez	1.20
450 :	edgomez	1.55	if (sad < *(data->iMinSAD)) {
451 :			*(data->iMinSAD) = sad;
452 :			current->x = x; current->y = y;
453 :			*dir = Direction;
454 :	edgomez	1.20	}
455 :	chl	1.3	}
456 :
457 :	edgomez	1.55	static void
458 :			CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
459 :			{
460 :			// maximum speed - for P/B/I decision
461 :			int32_t sad;
462 :
463 :			if ( (x > data->max_dx) \|\| (x < data->min_dx)
464 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
465 :	chl	1.3
466 :	syskin	1.65	sad = sad32v_c(data->Cur, data->RefP[0] + (x>>1) + (y>>1)*(data->iEdgedWidth),
467 :			data->iEdgedWidth, data->temp+1);
468 :	chl	1.18
469 :	edgomez	1.55	if (sad < *(data->iMinSAD)) {
470 :			*(data->iMinSAD) = sad;
471 :			data->currentMV[0].x = x; data->currentMV[0].y = y;
472 :			*dir = Direction;
473 :			}
474 :			if (data->temp[1] < data->iMinSAD[1]) {
475 :			data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
476 :			if (data->temp[2] < data->iMinSAD[2]) {
477 :			data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
478 :			if (data->temp[3] < data->iMinSAD[3]) {
479 :			data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
480 :			if (data->temp[4] < data->iMinSAD[4]) {
481 :			data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
482 :	chl	1.18
483 :	edgomez	1.55	}
484 :	chl	1.18
485 :	edgomez	1.55	static void
486 :			CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data)
487 :			{
488 :			int32_t sad, xb, yb, xcf, ycf, xcb, ycb;
489 :			uint32_t t;
490 :			const uint8_t ReferenceF, ReferenceB;
491 :			VECTOR *current;
492 :
493 :	edgomez	1.62	if ((xf > data->max_dx) \|\| (xf < data->min_dx) \|\|
494 :			(yf > data->max_dy) \|\| (yf < data->min_dy))
495 :			return;
496 :	edgomez	1.55
497 :			if (!data->qpel_precision) {
498 :			ReferenceF = GetReference(xf, yf, data);
499 :			xb = data->currentMV[1].x; yb = data->currentMV[1].y;
500 :			ReferenceB = GetReferenceB(xb, yb, 1, data);
501 :			current = data->currentMV;
502 :			xcf = xf; ycf = yf;
503 :			xcb = xb; ycb = yb;
504 :	edgomez	1.20	} else {
505 :	edgomez	1.55	ReferenceF = Interpolate16x16qpel(xf, yf, 0, data);
506 :			xb = data->currentQMV[1].x; yb = data->currentQMV[1].y;
507 :			current = data->currentQMV;
508 :			ReferenceB = Interpolate16x16qpel(xb, yb, 1, data);
509 :			xcf = xf/2; ycf = yf/2;
510 :			xcb = xb/2; ycb = yb/2;
511 :			}
512 :	chl	1.18
513 :	edgomez	1.55	t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0)
514 :	syskin	1.64	+ d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0);
515 :	chl	1.18
516 :	edgomez	1.55	sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
517 :			sad += (data->lambda16 * t * sad)>>10;
518 :	chl	1.18
519 :	edgomez	1.55	if (data->chroma) sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3],
520 :			(ycf >> 1) + roundtab_79[ycf & 0x3],
521 :			(xcb >> 1) + roundtab_79[xcb & 0x3],
522 :			(ycb >> 1) + roundtab_79[ycb & 0x3], data);
523 :	chl	1.18
524 :	edgomez	1.55	if (sad < *(data->iMinSAD)) {
525 :			*(data->iMinSAD) = sad;
526 :			current->x = xf; current->y = yf;
527 :			*dir = Direction;
528 :	chl	1.18	}
529 :			}
530 :
531 :	edgomez	1.55	static void
532 :			CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
533 :	chl	1.18	{
534 :	edgomez	1.55	int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0;
535 :			uint32_t k;
536 :			const uint8_t *ReferenceF;
537 :			const uint8_t *ReferenceB;
538 :			VECTOR mvs, b_mvs;
539 :
540 :	syskin	1.60	if (( x > 31) \|\| ( x < -32) \|\| ( y > 31) \|\| (y < -32)) return;
541 :	edgomez	1.55
542 :			for (k = 0; k < 4; k++) {
543 :			mvs.x = data->directmvF[k].x + x;
544 :			b_mvs.x = ((x == 0) ?
545 :			data->directmvB[k].x
546 :			: mvs.x - data->referencemv[k].x);
547 :	chl	1.18
548 :	edgomez	1.55	mvs.y = data->directmvF[k].y + y;
549 :			b_mvs.y = ((y == 0) ?
550 :			data->directmvB[k].y
551 :			: mvs.y - data->referencemv[k].y);
552 :	chl	1.18
553 :	edgomez	1.62	if ((mvs.x > data->max_dx) \|\| (mvs.x < data->min_dx) \|\|
554 :			(mvs.y > data->max_dy) \|\| (mvs.y < data->min_dy) \|\|
555 :			(b_mvs.x > data->max_dx) \|\| (b_mvs.x < data->min_dx) \|\|
556 :			(b_mvs.y > data->max_dy) \|\| (b_mvs.y < data->min_dy) )
557 :			return;
558 :	chl	1.18
559 :	edgomez	1.55	if (data->qpel) {
560 :			xcf += mvs.x/2; ycf += mvs.y/2;
561 :			xcb += b_mvs.x/2; ycb += b_mvs.y/2;
562 :			} else {
563 :			xcf += mvs.x; ycf += mvs.y;
564 :			xcb += b_mvs.x; ycb += b_mvs.y;
565 :			mvs.x = 2; mvs.y = 2; //we move to qpel precision anyway
566 :			b_mvs.x = 2; b_mvs.y = 2;
567 :			}
568 :	edgomez	1.20
569 :	edgomez	1.55	ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data);
570 :			ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data);
571 :	chl	1.18
572 :	edgomez	1.55	sad += sad8bi(data->Cur + 8(k&1) + 8(k>>1)*(data->iEdgedWidth),
573 :			ReferenceF, ReferenceB, data->iEdgedWidth);
574 :			if (sad > *(data->iMinSAD)) return;
575 :			}
576 :	chl	1.18
577 :	edgomez	1.55	sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
578 :	chl	1.18
579 :	edgomez	1.55	if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
580 :			(ycf >> 3) + roundtab_76[ycf & 0xf],
581 :			(xcb >> 3) + roundtab_76[xcb & 0xf],
582 :			(ycb >> 3) + roundtab_76[ycb & 0xf], data);
583 :	chl	1.18
584 :	edgomez	1.55	if (sad < *(data->iMinSAD)) {
585 :			*(data->iMinSAD) = sad;
586 :			data->currentMV->x = x; data->currentMV->y = y;
587 :			*dir = Direction;
588 :	chl	1.18	}
589 :			}
590 :
591 :	edgomez	1.55	static void
592 :			CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
593 :	Isibaar	1.1	{
594 :	edgomez	1.55	int32_t sad, xcf, ycf, xcb, ycb;
595 :			const uint8_t *ReferenceF;
596 :			const uint8_t *ReferenceB;
597 :			VECTOR mvs, b_mvs;
598 :
599 :	edgomez	1.62	if (( x > 31) \|\| ( x < -32) \|\| ( y > 31) \|\| (y < -32)) return;
600 :	edgomez	1.55
601 :			mvs.x = data->directmvF[0].x + x;
602 :			b_mvs.x = ((x == 0) ?
603 :			data->directmvB[0].x
604 :			: mvs.x - data->referencemv[0].x);
605 :
606 :			mvs.y = data->directmvF[0].y + y;
607 :			b_mvs.y = ((y == 0) ?
608 :			data->directmvB[0].y
609 :			: mvs.y - data->referencemv[0].y);
610 :
611 :	edgomez	1.62	if ( (mvs.x > data->max_dx) \|\| (mvs.x < data->min_dx)
612 :			\|\| (mvs.y > data->max_dy) \|\| (mvs.y < data->min_dy)
613 :			\|\| (b_mvs.x > data->max_dx) \|\| (b_mvs.x < data->min_dx)
614 :			\|\| (b_mvs.y > data->max_dy) \|\| (b_mvs.y < data->min_dy) ) return;
615 :	edgomez	1.55
616 :			if (data->qpel) {
617 :			xcf = 4(mvs.x/2); ycf = 4(mvs.y/2);
618 :			xcb = 4(b_mvs.x/2); ycb = 4(b_mvs.y/2);
619 :			ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data);
620 :			ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data);
621 :			} else {
622 :			xcf = 4mvs.x; ycf = 4mvs.y;
623 :			xcb = 4b_mvs.x; ycb = 4b_mvs.y;
624 :			ReferenceF = GetReference(mvs.x, mvs.y, data);
625 :			ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data);
626 :			}
627 :	Isibaar	1.1
628 :	edgomez	1.55	sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
629 :			sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
630 :	Isibaar	1.1
631 :	edgomez	1.55	if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
632 :			(ycf >> 3) + roundtab_76[ycf & 0xf],
633 :			(xcb >> 3) + roundtab_76[xcb & 0xf],
634 :			(ycb >> 3) + roundtab_76[ycb & 0xf], data);
635 :	edgomez	1.20
636 :	edgomez	1.55	if (sad < *(data->iMinSAD)) {
637 :			*(data->iMinSAD) = sad;
638 :			data->currentMV->x = x; data->currentMV->y = y;
639 :			*dir = Direction;
640 :			}
641 :			}
642 :	edgomez	1.20
643 :	Isibaar	1.1
644 :	edgomez	1.55	static void
645 :			CheckCandidateBits16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
646 :			{
647 :	edgomez	1.20
648 :	edgomez	1.61	int16_t in = data->dctSpace, coeff = data->dctSpace + 64;
649 :	syskin	1.69	int32_t bits = 0;
650 :	edgomez	1.55	VECTOR * current;
651 :			const uint8_t * ptr;
652 :			int i, cbp = 0, t, xc, yc;
653 :
654 :			if ( (x > data->max_dx) \|\| (x < data->min_dx)
655 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
656 :
657 :			if (!data->qpel_precision) {
658 :			ptr = GetReference(x, y, data);
659 :			current = data->currentMV;
660 :			xc = x; yc = y;
661 :			} else { // x and y are in 1/4 precision
662 :			ptr = Interpolate16x16qpel(x, y, 0, data);
663 :			current = data->currentQMV;
664 :			xc = x/2; yc = y/2;
665 :			}
666 :
667 :			for(i = 0; i < 4; i++) {
668 :			int s = 8((i&1) + (i>>1)data->iEdgedWidth);
669 :			transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth);
670 :	syskin	1.72	bits += data->temp[i] = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, i);
671 :	edgomez	1.55	}
672 :
673 :	syskin	1.70	bits += t = BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
674 :	edgomez	1.55
675 :	syskin	1.72	//8x8 blocks for inter4v mode
676 :	syskin	1.71	if (data->temp[0] + t < data->iMinSAD[1]) {
677 :			data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; }
678 :			if (data->temp[1] < data->iMinSAD[2]) {
679 :			data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; }
680 :			if (data->temp[2] < data->iMinSAD[3]) {
681 :			data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; }
682 :			if (data->temp[3] < data->iMinSAD[4]) {
683 :			data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; }
684 :
685 :	syskin	1.70	bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len;
686 :	syskin	1.69
687 :			if (bits >= data->iMinSAD[0]) return;
688 :
689 :			//chroma
690 :			xc = (xc >> 1) + roundtab_79[xc & 0x3];
691 :			yc = (yc >> 1) + roundtab_79[yc & 0x3];
692 :
693 :			//chroma U
694 :			ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefP[4], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
695 :			transfer_8to16subro(in, ptr, data->CurU, data->iEdgedWidth/2);
696 :	syskin	1.72	bits += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 4);
697 :	syskin	1.69	if (bits >= data->iMinSAD[0]) return;
698 :
699 :			//chroma V
700 :			ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefP[5], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
701 :			transfer_8to16subro(in, ptr, data->CurV, data->iEdgedWidth/2);
702 :	syskin	1.72	bits += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5);
703 :	chl	1.18
704 :	syskin	1.70	bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) \| ((cbp & 3) << 3)].len;
705 :	edgomez	1.20
706 :	edgomez	1.55	if (bits < data->iMinSAD[0]) {
707 :			data->iMinSAD[0] = bits;
708 :			current[0].x = x; current[0].y = y;
709 :			*dir = Direction;
710 :			}
711 :	syskin	1.71	}
712 :	edgomez	1.55	static void
713 :			CheckCandidateBits8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
714 :			{
715 :	Isibaar	1.1
716 :	edgomez	1.61	int16_t in = data->dctSpace, coeff = data->dctSpace + 64;
717 :	syskin	1.69	int32_t bits;
718 :	edgomez	1.55	VECTOR * current;
719 :			const uint8_t * ptr;
720 :	syskin	1.69	int cbp = 0;
721 :	Isibaar	1.1
722 :	edgomez	1.55	if ( (x > data->max_dx) \|\| (x < data->min_dx)
723 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
724 :	Isibaar	1.1
725 :	edgomez	1.55	if (!data->qpel_precision) {
726 :			ptr = GetReference(x, y, data);
727 :			current = data->currentMV;
728 :			} else { // x and y are in 1/4 precision
729 :			ptr = Interpolate8x8qpel(x, y, 0, 0, data);
730 :			current = data->currentQMV;
731 :	edgomez	1.20	}
732 :
733 :	edgomez	1.55	transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth);
734 :	syskin	1.72	bits = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5);
735 :	syskin	1.70	bits += BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
736 :	edgomez	1.20
737 :	edgomez	1.55	if (bits < data->iMinSAD[0]) {
738 :			data->temp[0] = cbp;
739 :			data->iMinSAD[0] = bits;
740 :			current[0].x = x; current[0].y = y;
741 :			*dir = Direction;
742 :	edgomez	1.20	}
743 :	edgomez	1.55	}
744 :	edgomez	1.20
745 :	edgomez	1.55	/* CHECK_CANDIATE FUNCTIONS END */
746 :	edgomez	1.20
747 :	edgomez	1.55	/* MAINSEARCH FUNCTIONS START */
748 :	chl	1.13
749 :	edgomez	1.55	static void
750 :			AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection)
751 :			{
752 :	Isibaar	1.1
753 :	edgomez	1.55	/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
754 :	chl	1.13
755 :	edgomez	1.55	int iDirection;
756 :	chl	1.13
757 :	edgomez	1.55	for(;;) { //forever
758 :			iDirection = 0;
759 :			if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
760 :			if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
761 :			if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
762 :			if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
763 :	chl	1.13
764 :	edgomez	1.55	/* now we're doing diagonal checks near our candidate */
765 :	chl	1.13
766 :	edgomez	1.55	if (iDirection) { //if anything found
767 :			bDirection = iDirection;
768 :			iDirection = 0;
769 :			x = data->currentMV->x; y = data->currentMV->y;
770 :			if (bDirection & 3) { //our candidate is left or right
771 :			CHECK_CANDIDATE(x, y + iDiamondSize, 8);
772 :			CHECK_CANDIDATE(x, y - iDiamondSize, 4);
773 :			} else { // what remains here is up or down
774 :			CHECK_CANDIDATE(x + iDiamondSize, y, 2);
775 :			CHECK_CANDIDATE(x - iDiamondSize, y, 1);
776 :			}
777 :	chl	1.13
778 :	edgomez	1.55	if (iDirection) {
779 :			bDirection += iDirection;
780 :			x = data->currentMV->x; y = data->currentMV->y;
781 :			}
782 :			} else { //about to quit, eh? not so fast....
783 :			switch (bDirection) {
784 :			case 2:
785 :			CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
786 :			CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
787 :			break;
788 :			case 1:
789 :			CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
790 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
791 :			break;
792 :			case 2 + 4:
793 :			CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
794 :			CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
795 :			CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
796 :			break;
797 :			case 4:
798 :			CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
799 :			CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
800 :			break;
801 :			case 8:
802 :			CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
803 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
804 :			break;
805 :			case 1 + 4:
806 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
807 :			CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
808 :			CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
809 :			break;
810 :			case 2 + 8:
811 :			CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
812 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
813 :			CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
814 :			break;
815 :			case 1 + 8:
816 :			CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
817 :			CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
818 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
819 :			break;
820 :			default: //1+2+4+8 == we didn't find anything at all
821 :			CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
822 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
823 :			CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
824 :			CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
825 :			break;
826 :			}
827 :			if (!iDirection) break; //ok, the end. really
828 :			bDirection = iDirection;
829 :			x = data->currentMV->x; y = data->currentMV->y;
830 :			}
831 :			}
832 :			}
833 :	chl	1.13
834 :	edgomez	1.55	static void
835 :			SquareSearch(int x, int y, const SearchData * const data, int bDirection)
836 :			{
837 :			int iDirection;
838 :	Isibaar	1.1
839 :	edgomez	1.55	do {
840 :			iDirection = 0;
841 :			if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64);
842 :			if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128);
843 :			if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32);
844 :			if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128);
845 :			if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64);
846 :			if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128);
847 :			if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128);
848 :			if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128);
849 :
850 :			bDirection = iDirection;
851 :			x = data->currentMV->x; y = data->currentMV->y;
852 :			} while (iDirection);
853 :			}
854 :	Isibaar	1.1
855 :	edgomez	1.55	static void
856 :			DiamondSearch(int x, int y, const SearchData * const data, int bDirection)
857 :			{
858 :	Isibaar	1.1
859 :	edgomez	1.55	/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
860 :	chl	1.13
861 :	edgomez	1.55	int iDirection;
862 :	edgomez	1.20
863 :	edgomez	1.55	do {
864 :			iDirection = 0;
865 :			if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
866 :			if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
867 :			if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
868 :			if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
869 :	chl	1.13
870 :	edgomez	1.55	/* now we're doing diagonal checks near our candidate */
871 :	Isibaar	1.1
872 :	edgomez	1.55	if (iDirection) { //checking if anything found
873 :			bDirection = iDirection;
874 :			iDirection = 0;
875 :			x = data->currentMV->x; y = data->currentMV->y;
876 :			if (bDirection & 3) { //our candidate is left or right
877 :			CHECK_CANDIDATE(x, y + iDiamondSize, 8);
878 :			CHECK_CANDIDATE(x, y - iDiamondSize, 4);
879 :			} else { // what remains here is up or down
880 :			CHECK_CANDIDATE(x + iDiamondSize, y, 2);
881 :			CHECK_CANDIDATE(x - iDiamondSize, y, 1);
882 :	edgomez	1.20	}
883 :	edgomez	1.55	bDirection += iDirection;
884 :			x = data->currentMV->x; y = data->currentMV->y;
885 :			}
886 :			}
887 :			while (iDirection);
888 :			}
889 :	edgomez	1.20
890 :	edgomez	1.55	/* MAINSEARCH FUNCTIONS END */
891 :	Isibaar	1.1
892 :	edgomez	1.55	static void
893 :			SubpelRefine(const SearchData * const data)
894 :			{
895 :			/* Do a half-pel or q-pel refinement */
896 :			const VECTOR centerMV = data->qpel_precision ? data->currentQMV : data->currentMV;
897 :			int iDirection; //only needed because macro expects it
898 :
899 :			CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0);
900 :			CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0);
901 :			CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0);
902 :			CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0);
903 :			CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0);
904 :			CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0);
905 :			CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0);
906 :			CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0);
907 :			}
908 :
909 :			static __inline int
910 :			SkipDecisionP(const IMAGE * current, const IMAGE * reference,
911 :	syskin	1.69	const int x, const int y,
912 :			const uint32_t stride, const uint32_t iQuant, int rrv)
913 :	edgomez	1.20
914 :	edgomez	1.55	{
915 :	edgomez	1.62	int offset = (x + ystride)8;
916 :	edgomez	1.55	if(!rrv) {
917 :	edgomez	1.62	uint32_t sadC = sad8(current->u + offset,
918 :			reference->u + offset, stride);
919 :	edgomez	1.55	if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
920 :	edgomez	1.62	sadC += sad8(current->v + offset,
921 :			reference->v + offset, stride);
922 :	edgomez	1.55	if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
923 :			return 1;
924 :	Isibaar	1.1
925 :	edgomez	1.55	} else {
926 :	edgomez	1.62	uint32_t sadC = sad16(current->u + 2*offset,
927 :			reference->u + 2offset, stride, 2564096);
928 :	edgomez	1.55	if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
929 :	edgomez	1.62	sadC += sad16(current->v + 2*offset,
930 :			reference->v + 2offset, stride, 2564096);
931 :	edgomez	1.55	if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
932 :			return 1;
933 :	edgomez	1.2	}
934 :	edgomez	1.55	}
935 :	Isibaar	1.1
936 :	edgomez	1.55	static __inline void
937 :			SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad)
938 :			{
939 :			pMB->mode = MODE_NOT_CODED;
940 :			pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV;
941 :			pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV;
942 :			pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
943 :			}
944 :	Isibaar	1.1
945 :	syskin	1.69	static __inline void
946 :			ModeDecision(SearchData * const Data,
947 :			MACROBLOCK * const pMB,
948 :			const MACROBLOCK * const pMBs,
949 :			const int x, const int y,
950 :			const MBParam * const pParam,
951 :			const uint32_t MotionFlags,
952 :			const uint32_t GlobalFlags,
953 :			const IMAGE * const pCurrent,
954 :			const IMAGE * const pRef)
955 :			{
956 :			int mode = MODE_INTER;
957 :			int inter4v = (GlobalFlags & XVID_INTER4V) && (pMB->dquant == NO_CHANGE);
958 :			const uint32_t iQuant = pMB->quant;
959 :
960 :			const int skip_possible = (!(GlobalFlags & XVID_GMC)) && (pMB->dquant == NO_CHANGE);
961 :
962 :			if (!(GlobalFlags & XVID_MODEDECISION_BITS)) { //normal, fast, SAD-based mode decision
963 :			int sad;
964 :			int InterBias = MV16_INTER_BIAS;
965 :			if (inter4v == 0 \|\| Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] +
966 :			Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) {
967 :			mode = MODE_INTER;
968 :			sad = Data->iMinSAD[0];
969 :			} else {
970 :			mode = MODE_INTER4V;
971 :			sad = Data->iMinSAD[1] + Data->iMinSAD[2] +
972 :			Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant;
973 :			Data->iMinSAD[0] = sad;
974 :			}
975 :
976 :	syskin	1.72	// final skip decision, a.k.a. "the vector you found, really that good?"
977 :	syskin	1.69	if (skip_possible && (pMB->sad16 < (int)iQuant * MAX_SAD00_FOR_SKIP))
978 :			if ( (100*sad)/(pMB->sad16+1) > FINAL_SKIP_THRESH)
979 :			if (Data->chroma \|\| SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant, Data->rrv)) {
980 :			mode = MODE_NOT_CODED;
981 :			sad = 0;
982 :			}
983 :
984 :	syskin	1.72	// intra decision
985 :	syskin	1.69
986 :			if (iQuant > 8) InterBias += 100 * (iQuant - 8); // to make high quants work
987 :			if (y != 0)
988 :			if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80;
989 :			if (x != 0)
990 :			if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80;
991 :
992 :			if (Data->chroma) InterBias += 50; // dev8(chroma) ???
993 :			if (Data->rrv) InterBias *= 4;
994 :
995 :			if (InterBias < pMB->sad16) {
996 :			int32_t deviation;
997 :			if (!Data->rrv) deviation = dev16(Data->Cur, Data->iEdgedWidth);
998 :			else deviation = dev16(Data->Cur, Data->iEdgedWidth) +
999 :			dev16(Data->Cur+16, Data->iEdgedWidth) +
1000 :			dev16(Data->Cur + 16*Data->iEdgedWidth, Data->iEdgedWidth) +
1001 :			dev16(Data->Cur+16+16*Data->iEdgedWidth, Data->iEdgedWidth);
1002 :
1003 :			if (deviation < (sad - InterBias)) mode = MODE_INTRA;
1004 :			}
1005 :
1006 :			} else { // BITS
1007 :
1008 :			int bits, intra, i;
1009 :			VECTOR backup[5], *v;
1010 :			Data->iQuant = iQuant;
1011 :
1012 :			v = Data->qpel ? Data->currentQMV : Data->currentMV;
1013 :			for (i = 0; i < 5; i++) {
1014 :			Data->iMinSAD[i] = 256*4096;
1015 :			backup[i] = v[i];
1016 :			}
1017 :
1018 :			bits = CountMBBitsInter(Data, pMBs, x, y, pParam, MotionFlags);
1019 :			if (bits == 0)
1020 :			mode = MODE_INTER; // quick stop
1021 :			else {
1022 :			if (inter4v) {
1023 :			int bits_inter4v = CountMBBitsInter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup);
1024 :			if (bits_inter4v < bits) { Data->iMinSAD[0] = bits = bits_inter4v; mode = MODE_INTER4V; }
1025 :			}
1026 :
1027 :			intra = CountMBBitsIntra(Data);
1028 :
1029 :			if (intra < bits) { *Data->iMinSAD = bits = intra; mode = MODE_INTRA; }
1030 :			}
1031 :			}
1032 :
1033 :			if (Data->rrv) {
1034 :			Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x);
1035 :			Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y);
1036 :			}
1037 :
1038 :			if (mode == MODE_INTER) {
1039 :			pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
1040 :			pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0];
1041 :
1042 :			if(Data->qpel) {
1043 :			pMB->qmvs[0] = pMB->qmvs[1]
1044 :			= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
1045 :			pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x;
1046 :			pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y;
1047 :			} else {
1048 :			pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
1049 :			pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
1050 :			}
1051 :
1052 :			} else if (mode == MODE_INTER4V)
1053 :			pMB->sad16 = Data->iMinSAD[0];
1054 :			else // INTRA, NOT_CODED
1055 :			SkipMacroblockP(pMB, 0);
1056 :
1057 :			pMB->mode = mode;
1058 :			}
1059 :
1060 :	edgomez	1.55	bool
1061 :			MotionEstimation(MBParam * const pParam,
1062 :	syskin	1.64	FRAMEINFO * const current,
1063 :			FRAMEINFO * const reference,
1064 :			const IMAGE * const pRefH,
1065 :			const IMAGE * const pRefV,
1066 :			const IMAGE * const pRefHV,
1067 :			const uint32_t iLimit)
1068 :	edgomez	1.55	{
1069 :			MACROBLOCK *const pMBs = current->mbs;
1070 :			const IMAGE *const pCurrent = &current->image;
1071 :			const IMAGE *const pRef = &reference->image;
1072 :	Isibaar	1.1
1073 :	edgomez	1.55	uint32_t mb_width = pParam->mb_width;
1074 :			uint32_t mb_height = pParam->mb_height;
1075 :			const uint32_t iEdgedWidth = pParam->edged_width;
1076 :			const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->global_flags);
1077 :	chl	1.18
1078 :	edgomez	1.55	uint32_t x, y;
1079 :			uint32_t iIntra = 0;
1080 :			int32_t quant = current->quant, sad00;
1081 :	syskin	1.64	int skip_thresh = INITIAL_SKIP_THRESH *
1082 :			(current->global_flags & XVID_REDUCED ? 4:1) *
1083 :	syskin	1.60	(current->global_flags & XVID_MODEDECISION_BITS ? 2:1);
1084 :	Isibaar	1.1
1085 :	edgomez	1.55	// some pre-initialized thingies for SearchP
1086 :			int32_t temp[8];
1087 :			VECTOR currentMV[5];
1088 :			VECTOR currentQMV[5];
1089 :			int32_t iMinSAD[5];
1090 :	syskin	1.72	DECLARE_ALIGNED_MATRIX(dct_space, 3, 64, int16_t, CACHE_LINE);
1091 :	edgomez	1.55	SearchData Data;
1092 :			memset(&Data, 0, sizeof(SearchData));
1093 :			Data.iEdgedWidth = iEdgedWidth;
1094 :			Data.currentMV = currentMV;
1095 :			Data.currentQMV = currentQMV;
1096 :			Data.iMinSAD = iMinSAD;
1097 :			Data.temp = temp;
1098 :			Data.iFcode = current->fcode;
1099 :			Data.rounding = pParam->m_rounding_type;
1100 :			Data.qpel = pParam->m_quarterpel;
1101 :			Data.chroma = MotionFlags & PMV_CHROMA16;
1102 :			Data.rrv = current->global_flags & XVID_REDUCED;
1103 :	edgomez	1.61	Data.dctSpace = dct_space;
1104 :	syskin	1.69	Data.quant_type = pParam->m_quant_type;
1105 :	edgomez	1.55
1106 :			if ((current->global_flags & XVID_REDUCED)) {
1107 :			mb_width = (pParam->width + 31) / 32;
1108 :			mb_height = (pParam->height + 31) / 32;
1109 :			Data.qpel = 0;
1110 :			}
1111 :
1112 :			Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose)
1113 :			if (sadInit) (*sadInit) ();
1114 :
1115 :			for (y = 0; y < mb_height; y++) {
1116 :			for (x = 0; x < mb_width; x++) {
1117 :			MACROBLOCK pMB = &pMBs[x + y pParam->mb_width];
1118 :
1119 :			if (!Data.rrv) pMB->sad16 =
1120 :			sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16,
1121 :			pRef->y + (x + y * iEdgedWidth) * 16,
1122 :			pParam->edged_width, pMB->sad8 );
1123 :
1124 :			else pMB->sad16 =
1125 :			sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32,
1126 :			pRef->y + (x + y * iEdgedWidth) * 32,
1127 :			pParam->edged_width, pMB->sad8 );
1128 :
1129 :			if (Data.chroma) {
1130 :			Data.temp[7] = sad8(pCurrent->u + x8 + y(iEdgedWidth/2)*8,
1131 :			pRef->u + x8 + y(iEdgedWidth/2)*8, iEdgedWidth/2)
1132 :			+ sad8(pCurrent->v + (x + y(iEdgedWidth/2))8,
1133 :			pRef->v + (x + y(iEdgedWidth/2))8, iEdgedWidth/2);
1134 :			pMB->sad16 += Data.temp[7];
1135 :			}
1136 :	chl	1.26
1137 :	edgomez	1.55	sad00 = pMB->sad16;
1138 :	edgomez	1.20
1139 :	syskin	1.69	if (!(current->global_flags & XVID_LUMIMASKING))
1140 :	edgomez	1.55	pMB->dquant = NO_CHANGE;
1141 :	syskin	1.69	else {
1142 :	edgomez	1.55	if (pMB->dquant != NO_CHANGE) {
1143 :			quant += DQtab[pMB->dquant];
1144 :			if (quant > 31) quant = 31;
1145 :			else if (quant < 1) quant = 1;
1146 :			}
1147 :			}
1148 :	syskin	1.69	pMB->quant = quant;
1149 :	Isibaar	1.1
1150 :	edgomez	1.55	//initial skip decision
1151 :			/* no early skip for GMC (global vector = skip vector is unknown!) */
1152 :			if (!(current->global_flags & XVID_GMC)) { /* no fast SKIP for S(GMC)-VOPs */
1153 :	syskin	1.70	if (pMB->dquant == NO_CHANGE && sad00 < quant * skip_thresh)
1154 :	edgomez	1.55	if (Data.chroma \|\| SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) {
1155 :			SkipMacroblockP(pMB, sad00);
1156 :			continue;
1157 :			}
1158 :			}
1159 :	Isibaar	1.1
1160 :	edgomez	1.55	SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
1161 :	syskin	1.69	y, MotionFlags, current->global_flags,
1162 :			&Data, pParam, pMBs, reference->mbs, pMB);
1163 :
1164 :			ModeDecision(&Data, pMB, pMBs, x, y, pParam,
1165 :			MotionFlags, current->global_flags,
1166 :			pCurrent, pRef);
1167 :
1168 :	edgomez	1.55	if (pMB->mode == MODE_INTRA)
1169 :			if (++iIntra > iLimit) return 1;
1170 :	Isibaar	1.1	}
1171 :	edgomez	1.55	}
1172 :
1173 :			if (current->global_flags & XVID_GMC ) /* GMC only for S(GMC)-VOPs */
1174 :			{
1175 :			current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV);
1176 :			}
1177 :			return 0;
1178 :			}
1179 :	Isibaar	1.1
1180 :	edgomez	1.20
1181 :	edgomez	1.55	static __inline int
1182 :			make_mask(const VECTOR * const pmv, const int i)
1183 :			{
1184 :			int mask = 255, j;
1185 :			for (j = 0; j < i; j++) {
1186 :			if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already
1187 :			if (pmv[i].x == pmv[j].x) {
1188 :			if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4;
1189 :			else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8;
1190 :			} else
1191 :			if (pmv[i].y == pmv[j].y) {
1192 :			if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1;
1193 :			else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2;
1194 :	edgomez	1.20	}
1195 :	Isibaar	1.1	}
1196 :	edgomez	1.55	return mask;
1197 :			}
1198 :	Isibaar	1.1
1199 :	edgomez	1.55	static __inline void
1200 :			PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount,
1201 :			int iHcount, const MACROBLOCK * const prevMB, int rrv)
1202 :			{
1203 :
1204 :			//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself
1205 :			if (rrv) { iWcount /= 2; iHcount /= 2; }
1206 :
1207 :			if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour
1208 :			pmv[5].x = EVEN(pmv[3].x);
1209 :			pmv[5].y = EVEN(pmv[3].y);
1210 :			} else pmv[5].x = pmv[5].y = 0;
1211 :	Isibaar	1.1
1212 :	edgomez	1.55	if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour
1213 :			else pmv[3].x = pmv[3].y = 0;
1214 :	Isibaar	1.1
1215 :	edgomez	1.55	if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour
1216 :			else pmv[4].x = pmv[4].y = 0;
1217 :	Isibaar	1.1
1218 :	edgomez	1.55	// [1] median prediction
1219 :			pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y);
1220 :	Isibaar	1.1
1221 :	edgomez	1.55	pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask
1222 :	Isibaar	1.1
1223 :	edgomez	1.55	pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame
1224 :			pmv[2].y = EVEN(prevMB->mvs[0].y);
1225 :	Isibaar	1.1
1226 :	edgomez	1.55	if ((x < iWcount-1) && (y < iHcount-1)) {
1227 :			pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame
1228 :			pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y);
1229 :			} else pmv[6].x = pmv[6].y = 0;
1230 :	Isibaar	1.1
1231 :	edgomez	1.55	if (rrv) {
1232 :			int i;
1233 :			for (i = 0; i < 7; i++) {
1234 :			pmv[i].x = RRV_MV_SCALEUP(pmv[i].x);
1235 :			pmv[i].y = RRV_MV_SCALEUP(pmv[i].y);
1236 :	chl	1.38	}
1237 :	edgomez	1.2	}
1238 :	Isibaar	1.1	}
1239 :
1240 :	edgomez	1.55	static void
1241 :			SearchP(const IMAGE * const pRef,
1242 :			const uint8_t * const pRefH,
1243 :			const uint8_t * const pRefV,
1244 :			const uint8_t * const pRefHV,
1245 :			const IMAGE * const pCur,
1246 :			const int x,
1247 :			const int y,
1248 :			const uint32_t MotionFlags,
1249 :			const uint32_t GlobalFlags,
1250 :			SearchData * const Data,
1251 :			const MBParam * const pParam,
1252 :			const MACROBLOCK * const pMBs,
1253 :			const MACROBLOCK * const prevMBs,
1254 :			MACROBLOCK * const pMB)
1255 :			{
1256 :
1257 :			int i, iDirection = 255, mask, threshA;
1258 :			VECTOR pmv[7];
1259 :	syskin	1.69	int inter4v = (GlobalFlags & XVID_INTER4V) && (pMB->dquant == NO_CHANGE);
1260 :	chl	1.39
1261 :	edgomez	1.55	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1262 :	syskin	1.69	pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1263 :	chl	1.39
1264 :	edgomez	1.55	get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp);
1265 :
1266 :			Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache
1267 :			i = Data->rrv ? 2 : 1;
1268 :			Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i;
1269 :			Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1270 :			Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1271 :
1272 :	syskin	1.64	Data->RefP[0] = pRef->y + (x + Data->iEdgedWidthy) 16*i;
1273 :			Data->RefP[2] = pRefH + (x + Data->iEdgedWidthy) 16*i;
1274 :			Data->RefP[1] = pRefV + (x + Data->iEdgedWidthy) 16*i;
1275 :			Data->RefP[3] = pRefHV + (x + Data->iEdgedWidthy) 16*i;
1276 :			Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1277 :			Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1278 :	edgomez	1.55
1279 :	syskin	1.69	Data->lambda16 = lambda_vec16[pMB->quant];
1280 :			Data->lambda8 = lambda_vec8[pMB->quant];
1281 :	edgomez	1.55	Data->qpel_precision = 0;
1282 :
1283 :	edgomez	1.62	memset(Data->currentMV, 0, 5*sizeof(VECTOR));
1284 :	edgomez	1.55
1285 :			if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0);
1286 :			else Data->predMV = pmv[0];
1287 :
1288 :			i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0);
1289 :			Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10);
1290 :			Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10);
1291 :			Data->iMinSAD[2] = pMB->sad8[1];
1292 :			Data->iMinSAD[3] = pMB->sad8[2];
1293 :			Data->iMinSAD[4] = pMB->sad8[3];
1294 :
1295 :	syskin	1.71	if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) && (x \| y)) {
1296 :	edgomez	1.55	threshA = Data->temp[0]; // that's where we keep this SAD atm
1297 :			if (threshA < 512) threshA = 512;
1298 :			else if (threshA > 1024) threshA = 1024;
1299 :			} else
1300 :			threshA = 512;
1301 :	chl	1.39
1302 :	edgomez	1.55	PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height,
1303 :			prevMBs + x + y * pParam->mb_width, Data->rrv);
1304 :	chl	1.39
1305 :	edgomez	1.55	if (!Data->rrv) {
1306 :			if (inter4v \| Data->chroma) CheckCandidate = CheckCandidate16;
1307 :			else CheckCandidate = CheckCandidate16no4v; //for extra speed
1308 :			} else CheckCandidate = CheckCandidate32;
1309 :
1310 :			/* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/
1311 :
1312 :			for (i = 1; i < 7; i++) {
1313 :			if (!(mask = make_mask(pmv, i)) ) continue;
1314 :			CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1315 :			if (Data->iMinSAD[0] <= threshA) break;
1316 :			}
1317 :
1318 :			if ((Data->iMinSAD[0] <= threshA) \|\|
1319 :			(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) &&
1320 :	syskin	1.69	(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16)))
1321 :			inter4v = 0;
1322 :	edgomez	1.55	else {
1323 :
1324 :			MainSearchFunc * MainSearchPtr;
1325 :			if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1326 :			else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1327 :			else MainSearchPtr = DiamondSearch;
1328 :
1329 :			MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1330 :
1331 :			/* extended search, diamond starting in 0,0 and in prediction.
1332 :			note that this search is/might be done in halfpel positions,
1333 :			which makes it more different than the diamond above */
1334 :
1335 :			if (MotionFlags & PMV_EXTSEARCH16) {
1336 :			int32_t bSAD;
1337 :			VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0];
1338 :			if (Data->rrv) {
1339 :			startMV.x = RRV_MV_SCALEUP(startMV.x);
1340 :			startMV.y = RRV_MV_SCALEUP(startMV.y);
1341 :			}
1342 :			if (!(MVequal(startMV, backupMV))) {
1343 :			bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1344 :
1345 :			CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1346 :			MainSearchPtr(startMV.x, startMV.y, Data, 255);
1347 :			if (bSAD < Data->iMinSAD[0]) {
1348 :			Data->currentMV[0] = backupMV;
1349 :			Data->iMinSAD[0] = bSAD; }
1350 :			}
1351 :	edgomez	1.20
1352 :	edgomez	1.55	backupMV = Data->currentMV[0];
1353 :			startMV.x = startMV.y = 1;
1354 :			if (!(MVequal(startMV, backupMV))) {
1355 :			bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1356 :
1357 :			CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1358 :			MainSearchPtr(startMV.x, startMV.y, Data, 255);
1359 :			if (bSAD < Data->iMinSAD[0]) {
1360 :			Data->currentMV[0] = backupMV;
1361 :			Data->iMinSAD[0] = bSAD; }
1362 :			}
1363 :			}
1364 :			}
1365 :
1366 :			if (MotionFlags & PMV_HALFPELREFINE16)
1367 :	syskin	1.65	SubpelRefine(Data);
1368 :	edgomez	1.55
1369 :			for(i = 0; i < 5; i++) {
1370 :			Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors
1371 :			Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
1372 :			}
1373 :	Isibaar	1.1
1374 :	syskin	1.70	if (Data->qpel) {
1375 :	syskin	1.60	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1376 :			pParam->width, pParam->height, Data->iFcode, 1, 0);
1377 :	syskin	1.65	Data->qpel_precision = 1;
1378 :	syskin	1.70	if (MotionFlags & PMV_QUARTERPELREFINE16)
1379 :			SubpelRefine(Data);
1380 :	syskin	1.60	}
1381 :	Isibaar	1.1
1382 :	syskin	1.70	if (Data->iMinSAD[0] < (int32_t)pMB->quant * 30)
1383 :	syskin	1.69	inter4v = 0;
1384 :	chl	1.18
1385 :	syskin	1.65	if (inter4v) {
1386 :	edgomez	1.55	SearchData Data8;
1387 :			memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data
1388 :	chl	1.18
1389 :	edgomez	1.55	Search8(Data, 2x, 2y, MotionFlags, pParam, pMB, pMBs, 0, &Data8);
1390 :			Search8(Data, 2x + 1, 2y, MotionFlags, pParam, pMB, pMBs, 1, &Data8);
1391 :			Search8(Data, 2x, 2y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8);
1392 :			Search8(Data, 2x + 1, 2y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8);
1393 :	edgomez	1.20
1394 :	edgomez	1.55	if ((Data->chroma) && (!(GlobalFlags & XVID_MODEDECISION_BITS))) {
1395 :	syskin	1.69	// chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, it will not be used
1396 :	edgomez	1.55	int sumx = 0, sumy = 0;
1397 :	chl	1.13
1398 :	syskin	1.69	if (Data->qpel)
1399 :			for (i = 1; i < 5; i++) {
1400 :			sumx += Data->currentQMV[i].x/2;
1401 :			sumy += Data->currentQMV[i].y/2;
1402 :			}
1403 :			else
1404 :			for (i = 1; i < 5; i++) {
1405 :			sumx += Data->currentMV[i].x;
1406 :			sumy += Data->currentMV[i].y;
1407 :			}
1408 :	chl	1.13
1409 :	edgomez	1.55	Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf],
1410 :			(sumy >> 3) + roundtab_76[sumy & 0xf], Data);
1411 :			}
1412 :	syskin	1.69	} else Data->iMinSAD[1] = 4096*256;
1413 :	edgomez	1.55	}
1414 :	Isibaar	1.1
1415 :	edgomez	1.55	static void
1416 :			Search8(const SearchData * const OldData,
1417 :			const int x, const int y,
1418 :			const uint32_t MotionFlags,
1419 :			const MBParam * const pParam,
1420 :			MACROBLOCK * const pMB,
1421 :			const MACROBLOCK * const pMBs,
1422 :			const int block,
1423 :			SearchData * const Data)
1424 :			{
1425 :			int i = 0;
1426 :			Data->iMinSAD = OldData->iMinSAD + 1 + block;
1427 :			Data->currentMV = OldData->currentMV + 1 + block;
1428 :			Data->currentQMV = OldData->currentQMV + 1 + block;
1429 :
1430 :			if(Data->qpel) {
1431 :			Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1432 :			if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y,
1433 :			Data->predMV, Data->iFcode, 0, 0);
1434 :			} else {
1435 :			Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1436 :			if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y,
1437 :			Data->predMV, Data->iFcode, 0, Data->rrv);
1438 :			}
1439 :	Isibaar	1.1
1440 :	edgomez	1.55	(Data->iMinSAD) += (Data->lambda8 i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10;
1441 :	chl	1.9
1442 :	edgomez	1.55	if (MotionFlags & (PMV_EXTSEARCH8\|PMV_HALFPELREFINE8\|PMV_QUARTERPELREFINE8)) {
1443 :	chl	1.9
1444 :	syskin	1.64	if (Data->rrv) i = 16; else i = 8;
1445 :
1446 :			Data->RefP[0] = OldData->RefP[0] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1447 :			Data->RefP[1] = OldData->RefP[1] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1448 :			Data->RefP[2] = OldData->RefP[2] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1449 :			Data->RefP[3] = OldData->RefP[3] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1450 :	Isibaar	1.1
1451 :	syskin	1.64	Data->Cur = OldData->Cur + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1452 :	edgomez	1.55	Data->qpel_precision = 0;
1453 :	Isibaar	1.1
1454 :	edgomez	1.55	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1455 :			pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1456 :	Isibaar	1.1
1457 :	edgomez	1.55	if (!Data->rrv) CheckCandidate = CheckCandidate8;
1458 :			else CheckCandidate = CheckCandidate16no4v;
1459 :	edgomez	1.20
1460 :	edgomez	1.55	if (MotionFlags & PMV_EXTSEARCH8 && (!(MotionFlags & EXTSEARCH_BITS))) {
1461 :			int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1462 :	Isibaar	1.1
1463 :	edgomez	1.55	MainSearchFunc *MainSearchPtr;
1464 :			if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch;
1465 :			else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch;
1466 :			else MainSearchPtr = DiamondSearch;
1467 :	Isibaar	1.1
1468 :	edgomez	1.55	MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255);
1469 :	edgomez	1.20
1470 :	edgomez	1.55	if(*(Data->iMinSAD) < temp_sad) {
1471 :			Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1472 :			Data->currentQMV->y = 2 * Data->currentMV->y;
1473 :	edgomez	1.20	}
1474 :	Isibaar	1.1	}
1475 :
1476 :	edgomez	1.55	if (MotionFlags & PMV_HALFPELREFINE8) {
1477 :			int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1478 :	edgomez	1.20
1479 :	edgomez	1.55	SubpelRefine(Data); // perform halfpel refine of current best vector
1480 :
1481 :			if(*(Data->iMinSAD) < temp_sad) { // we have found a better match
1482 :			Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1483 :			Data->currentQMV->y = 2 * Data->currentMV->y;
1484 :	edgomez	1.20	}
1485 :	Isibaar	1.1	}
1486 :	edgomez	1.55
1487 :			if (Data->qpel && MotionFlags & PMV_QUARTERPELREFINE8) {
1488 :			Data->qpel_precision = 1;
1489 :			get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1490 :			pParam->width, pParam->height, Data->iFcode, 1, 0);
1491 :			SubpelRefine(Data);
1492 :			}
1493 :			}
1494 :
1495 :			if (Data->rrv) {
1496 :			Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x);
1497 :			Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y);
1498 :	Isibaar	1.1	}
1499 :
1500 :	edgomez	1.55	if(Data->qpel) {
1501 :			pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x;
1502 :			pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y;
1503 :			pMB->qmvs[block] = *Data->currentQMV;
1504 :			} else {
1505 :			pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x;
1506 :			pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y;
1507 :			}
1508 :
1509 :			pMB->mvs[block] = *Data->currentMV;
1510 :			pMB->sad8[block] = 4 * *Data->iMinSAD;
1511 :			}
1512 :
1513 :			/* motion estimation for B-frames */
1514 :
1515 :			static __inline VECTOR
1516 :			ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode)
1517 :			{
1518 :			/* the stupidiest function ever */
1519 :			return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]);
1520 :			}
1521 :
1522 :			static void __inline
1523 :			PreparePredictionsBF(VECTOR * const pmv, const int x, const int y,
1524 :			const uint32_t iWcount,
1525 :			const MACROBLOCK * const pMB,
1526 :			const uint32_t mode_curr)
1527 :			{
1528 :
1529 :			// [0] is prediction
1530 :			pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y);
1531 :
1532 :			pmv[1].x = pmv[1].y = 0; // [1] is zero
1533 :	edgomez	1.20
1534 :	edgomez	1.55	pmv[2] = ChoosePred(pMB, mode_curr);
1535 :			pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y);
1536 :	chl	1.3
1537 :	edgomez	1.55	if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour
1538 :			pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr);
1539 :			pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y);
1540 :			} else pmv[3].x = pmv[3].y = 0;
1541 :	chl	1.3
1542 :	edgomez	1.55	if (y != 0) {
1543 :			pmv[4] = ChoosePred(pMB-iWcount, mode_curr);
1544 :			pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y);
1545 :			} else pmv[4].x = pmv[4].y = 0;
1546 :	chl	1.35
1547 :	edgomez	1.55	if (x != 0) {
1548 :			pmv[5] = ChoosePred(pMB-1, mode_curr);
1549 :			pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y);
1550 :			} else pmv[5].x = pmv[5].y = 0;
1551 :	chl	1.3
1552 :	edgomez	1.55	if (x != 0 && y != 0) {
1553 :			pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr);
1554 :			pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y);
1555 :			} else pmv[6].x = pmv[6].y = 0;
1556 :			}
1557 :	chl	1.3
1558 :
1559 :	edgomez	1.55	/* search backward or forward */
1560 :			static void
1561 :			SearchBF( const IMAGE * const pRef,
1562 :			const uint8_t * const pRefH,
1563 :			const uint8_t * const pRefV,
1564 :			const uint8_t * const pRefHV,
1565 :			const IMAGE * const pCur,
1566 :			const int x, const int y,
1567 :			const uint32_t MotionFlags,
1568 :			const uint32_t iFcode,
1569 :			const MBParam * const pParam,
1570 :			MACROBLOCK * const pMB,
1571 :			const VECTOR * const predMV,
1572 :			int32_t * const best_sad,
1573 :			const int32_t mode_current,
1574 :			SearchData * const Data)
1575 :			{
1576 :
1577 :			int i, iDirection = 255, mask;
1578 :			VECTOR pmv[7];
1579 :			MainSearchFunc *MainSearchPtr;
1580 :			*Data->iMinSAD = MV_MAX_ERROR;
1581 :			Data->iFcode = iFcode;
1582 :			Data->qpel_precision = 0;
1583 :			Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; // reset chroma-sad cache
1584 :
1585 :	syskin	1.64	Data->RefP[0] = pRef->y + (x + Data->iEdgedWidthy) 16;
1586 :			Data->RefP[2] = pRefH + (x + Data->iEdgedWidthy) 16;
1587 :			Data->RefP[1] = pRefV + (x + Data->iEdgedWidthy) 16;
1588 :			Data->RefP[3] = pRefHV + (x + Data->iEdgedWidthy) 16;
1589 :			Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8;
1590 :			Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8;
1591 :	edgomez	1.55
1592 :			Data->predMV = *predMV;
1593 :
1594 :			get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1595 :			pParam->width, pParam->height, iFcode - Data->qpel, 0, 0);
1596 :
1597 :			pmv[0] = Data->predMV;
1598 :			if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; }
1599 :
1600 :			PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current);
1601 :
1602 :			Data->currentMV->x = Data->currentMV->y = 0;
1603 :			CheckCandidate = CheckCandidate16no4v;
1604 :
1605 :			// main loop. checking all predictions
1606 :			for (i = 0; i < 7; i++) {
1607 :			if (!(mask = make_mask(pmv, i)) ) continue;
1608 :			CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1609 :			}
1610 :
1611 :			if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1612 :			else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1613 :			else MainSearchPtr = DiamondSearch;
1614 :
1615 :			MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1616 :
1617 :			SubpelRefine(Data);
1618 :
1619 :			if (Data->qpel && Data->iMinSAD < best_sad + 300) {
1620 :			Data->currentQMV->x = 2*Data->currentMV->x;
1621 :			Data->currentQMV->y = 2*Data->currentMV->y;
1622 :			Data->qpel_precision = 1;
1623 :			get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1624 :			pParam->width, pParam->height, iFcode, 1, 0);
1625 :			SubpelRefine(Data);
1626 :			}
1627 :
1628 :			// three bits are needed to code backward mode. four for forward
1629 :
1630 :			if (mode_current == MODE_FORWARD) Data->iMinSAD += 4 Data->lambda16;
1631 :			else Data->iMinSAD += 3 Data->lambda16;
1632 :
1633 :			if (Data->iMinSAD < best_sad) {
1634 :			best_sad = Data->iMinSAD;
1635 :			pMB->mode = mode_current;
1636 :			if (Data->qpel) {
1637 :			pMB->pmvs[0].x = Data->currentQMV->x - predMV->x;
1638 :			pMB->pmvs[0].y = Data->currentQMV->y - predMV->y;
1639 :			if (mode_current == MODE_FORWARD)
1640 :			pMB->qmvs[0] = *Data->currentQMV;
1641 :			else
1642 :			pMB->b_qmvs[0] = *Data->currentQMV;
1643 :			} else {
1644 :			pMB->pmvs[0].x = Data->currentMV->x - predMV->x;
1645 :			pMB->pmvs[0].y = Data->currentMV->y - predMV->y;
1646 :			}
1647 :			if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV;
1648 :			else pMB->b_mvs[0] = *Data->currentMV;
1649 :			}
1650 :
1651 :			if (mode_current == MODE_FORWARD) (Data->currentMV+2) = Data->currentMV;
1652 :			else (Data->currentMV+1) = Data->currentMV; //we store currmv for interpolate search
1653 :			}
1654 :
1655 :			static void
1656 :			SkipDecisionB(const IMAGE * const pCur,
1657 :			const IMAGE * const f_Ref,
1658 :			const IMAGE * const b_Ref,
1659 :			MACROBLOCK * const pMB,
1660 :			const uint32_t x, const uint32_t y,
1661 :			const SearchData * const Data)
1662 :			{
1663 :			int dx = 0, dy = 0, b_dx = 0, b_dy = 0;
1664 :			int32_t sum;
1665 :			const int div = 1 + Data->qpel;
1666 :			int k;
1667 :			const uint32_t stride = Data->iEdgedWidth/2;
1668 :			//this is not full chroma compensation, only it's fullpel approximation. should work though
1669 :
1670 :			for (k = 0; k < 4; k++) {
1671 :			dy += Data->directmvF[k].y / div;
1672 :	edgomez	1.62	dx += Data->directmvF[k].x / div;
1673 :			b_dy += Data->directmvB[k].y / div;
1674 :			b_dx += Data->directmvB[k].x / div;
1675 :	edgomez	1.55	}
1676 :
1677 :			dy = (dy >> 3) + roundtab_76[dy & 0xf];
1678 :			dx = (dx >> 3) + roundtab_76[dx & 0xf];
1679 :			b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf];
1680 :			b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf];
1681 :
1682 :			sum = sad8bi(pCur->u + 8 * x + 8 * y * stride,
1683 :			f_Ref->u + (y8 + dy/2) stride + x*8 + dx/2,
1684 :			b_Ref->u + (y8 + b_dy/2) stride + x*8 + b_dx/2,
1685 :			stride);
1686 :
1687 :			if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; //no skip
1688 :
1689 :			sum += sad8bi(pCur->v + 8x + 8 y * stride,
1690 :			f_Ref->v + (y8 + dy/2) stride + x*8 + dx/2,
1691 :			b_Ref->v + (y8 + b_dy/2) stride + x*8 + b_dx/2,
1692 :			stride);
1693 :
1694 :	edgomez	1.62	if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) {
1695 :			pMB->mode = MODE_DIRECT_NONE_MV; //skipped
1696 :	syskin	1.64	for (k = 0; k < 4; k++) {
1697 :	edgomez	1.62	pMB->qmvs[k] = pMB->mvs[k];
1698 :			pMB->b_qmvs[k] = pMB->b_mvs[k];
1699 :			}
1700 :			}
1701 :	edgomez	1.55	}
1702 :	chl	1.3
1703 :	edgomez	1.55	static __inline uint32_t
1704 :			SearchDirect(const IMAGE * const f_Ref,
1705 :			const uint8_t * const f_RefH,
1706 :			const uint8_t * const f_RefV,
1707 :			const uint8_t * const f_RefHV,
1708 :			const IMAGE * const b_Ref,
1709 :			const uint8_t * const b_RefH,
1710 :			const uint8_t * const b_RefV,
1711 :			const uint8_t * const b_RefHV,
1712 :			const IMAGE * const pCur,
1713 :			const int x, const int y,
1714 :			const uint32_t MotionFlags,
1715 :			const int32_t TRB, const int32_t TRD,
1716 :			const MBParam * const pParam,
1717 :			MACROBLOCK * const pMB,
1718 :			const MACROBLOCK * const b_mb,
1719 :			int32_t * const best_sad,
1720 :			SearchData * const Data)
1721 :	chl	1.3
1722 :	edgomez	1.55	{
1723 :			int32_t skip_sad;
1724 :			int k = (x + Data->iEdgedWidthy) 16;
1725 :			MainSearchFunc *MainSearchPtr;
1726 :
1727 :			Data->iMinSAD = 2564096;
1728 :	syskin	1.64	Data->RefP[0] = f_Ref->y + k;
1729 :			Data->RefP[2] = f_RefH + k;
1730 :			Data->RefP[1] = f_RefV + k;
1731 :			Data->RefP[3] = f_RefHV + k;
1732 :			Data->b_RefP[0] = b_Ref->y + k;
1733 :			Data->b_RefP[2] = b_RefH + k;
1734 :			Data->b_RefP[1] = b_RefV + k;
1735 :			Data->b_RefP[3] = b_RefHV + k;
1736 :			Data->RefP[4] = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1737 :			Data->RefP[5] = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1738 :			Data->b_RefP[4] = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1739 :			Data->b_RefP[5] = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1740 :	edgomez	1.55
1741 :			k = Data->qpel ? 4 : 2;
1742 :			Data->max_dx = k * (pParam->width - x * 16);
1743 :			Data->max_dy = k * (pParam->height - y * 16);
1744 :			Data->min_dx = -k * (16 + x * 16);
1745 :			Data->min_dy = -k * (16 + y * 16);
1746 :
1747 :			Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs;
1748 :			Data->qpel_precision = 0;
1749 :
1750 :			for (k = 0; k < 4; k++) {
1751 :			pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD);
1752 :			pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD;
1753 :			pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD);
1754 :			pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD;
1755 :
1756 :			if ( (pMB->b_mvs[k].x > Data->max_dx) \| (pMB->b_mvs[k].x < Data->min_dx)
1757 :			\| (pMB->b_mvs[k].y > Data->max_dy) \| (pMB->b_mvs[k].y < Data->min_dy) ) {
1758 :
1759 :			best_sad = 2564096; // in that case, we won't use direct mode
1760 :			pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV"
1761 :			pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0;
1762 :			return 256*4096;
1763 :			}
1764 :			if (b_mb->mode != MODE_INTER4V) {
1765 :			pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0];
1766 :			pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0];
1767 :			Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0];
1768 :			Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0];
1769 :			break;
1770 :			}
1771 :			}
1772 :
1773 :			CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v;
1774 :
1775 :			CheckCandidate(0, 0, 255, &k, Data);
1776 :
1777 :			// initial (fast) skip decision
1778 :	syskin	1.70	if (Data->iMinSAD < pMB->quant INITIAL_SKIP_THRESH * (Data->chroma?3:2)) {
1779 :	edgomez	1.55	//possible skip
1780 :			if (Data->chroma) {
1781 :			pMB->mode = MODE_DIRECT_NONE_MV;
1782 :			return *Data->iMinSAD; // skip.
1783 :			} else {
1784 :			SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data);
1785 :			if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip.
1786 :			}
1787 :			}
1788 :
1789 :	syskin	1.60	*Data->iMinSAD += Data->lambda16;
1790 :	edgomez	1.55	skip_sad = *Data->iMinSAD;
1791 :
1792 :			// DIRECT MODE DELTA VECTOR SEARCH.
1793 :			// This has to be made more effective, but at the moment I'm happy it's running at all
1794 :
1795 :			if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1796 :			else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1797 :			else MainSearchPtr = DiamondSearch;
1798 :
1799 :			MainSearchPtr(0, 0, Data, 255);
1800 :
1801 :			SubpelRefine(Data);
1802 :
1803 :			best_sad = Data->iMinSAD;
1804 :
1805 :			if (Data->qpel \|\| b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT;
1806 :			else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation
1807 :
1808 :			pMB->pmvs[3] = *Data->currentMV;
1809 :
1810 :			for (k = 0; k < 4; k++) {
1811 :			pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x;
1812 :			pMB->b_mvs[k].x = ( (Data->currentMV->x == 0)
1813 :			? Data->directmvB[k].x
1814 :			:pMB->mvs[k].x - Data->referencemv[k].x);
1815 :			pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y);
1816 :			pMB->b_mvs[k].y = ((Data->currentMV->y == 0)
1817 :			? Data->directmvB[k].y
1818 :			: pMB->mvs[k].y - Data->referencemv[k].y);
1819 :			if (Data->qpel) {
1820 :			pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2;
1821 :			pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2;
1822 :			pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2;
1823 :			pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2;
1824 :			}
1825 :
1826 :			if (b_mb->mode != MODE_INTER4V) {
1827 :			pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0];
1828 :			pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0];
1829 :			pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0];
1830 :			pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0];
1831 :			break;
1832 :			}
1833 :			}
1834 :			return skip_sad;
1835 :			}
1836 :
1837 :			static void
1838 :			SearchInterpolate(const IMAGE * const f_Ref,
1839 :			const uint8_t * const f_RefH,
1840 :			const uint8_t * const f_RefV,
1841 :			const uint8_t * const f_RefHV,
1842 :			const IMAGE * const b_Ref,
1843 :			const uint8_t * const b_RefH,
1844 :			const uint8_t * const b_RefV,
1845 :			const uint8_t * const b_RefHV,
1846 :			const IMAGE * const pCur,
1847 :			const int x, const int y,
1848 :			const uint32_t fcode,
1849 :			const uint32_t bcode,
1850 :			const uint32_t MotionFlags,
1851 :			const MBParam * const pParam,
1852 :			const VECTOR * const f_predMV,
1853 :			const VECTOR * const b_predMV,
1854 :			MACROBLOCK * const pMB,
1855 :			int32_t * const best_sad,
1856 :			SearchData * const fData)
1857 :	chl	1.3
1858 :	edgomez	1.55	{
1859 :	chl	1.3
1860 :	edgomez	1.55	int iDirection, i, j;
1861 :			SearchData bData;
1862 :	edgomez	1.20
1863 :	edgomez	1.55	fData->qpel_precision = 0;
1864 :			memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data
1865 :			fData->iMinSAD = 4096256;
1866 :			bData.currentMV++; bData.currentQMV++;
1867 :			fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode;
1868 :
1869 :			i = (x + y * fData->iEdgedWidth) * 16;
1870 :
1871 :	syskin	1.64	bData.b_RefP[0] = fData->RefP[0] = f_Ref->y + i;
1872 :			bData.b_RefP[2] = fData->RefP[2] = f_RefH + i;
1873 :			bData.b_RefP[1] = fData->RefP[1] = f_RefV + i;
1874 :			bData.b_RefP[3] = fData->RefP[3] = f_RefHV + i;
1875 :			bData.RefP[0] = fData->b_RefP[0] = b_Ref->y + i;
1876 :			bData.RefP[2] = fData->b_RefP[2] = b_RefH + i;
1877 :			bData.RefP[1] = fData->b_RefP[1] = b_RefV + i;
1878 :			bData.RefP[3] = fData->b_RefP[3] = b_RefHV + i;
1879 :			bData.b_RefP[4] = fData->RefP[4] = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1880 :			bData.b_RefP[5] = fData->RefP[5] = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1881 :			bData.RefP[4] = fData->b_RefP[4] = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1882 :			bData.RefP[5] = fData->b_RefP[5] = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1883 :	edgomez	1.55
1884 :			bData.bpredMV = fData->predMV = *f_predMV;
1885 :			fData->bpredMV = bData.predMV = *b_predMV;
1886 :			fData->currentMV[0] = fData->currentMV[2];
1887 :
1888 :			get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode - fData->qpel, 0, 0);
1889 :			get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode - fData->qpel, 0, 0);
1890 :
1891 :			if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx;
1892 :			if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx;
1893 :			if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy;
1894 :			if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy;
1895 :
1896 :			if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx;
1897 :			if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx;
1898 :			if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy;
1899 :			if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy;
1900 :
1901 :			CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData);
1902 :
1903 :			//diamond
1904 :			do {
1905 :			iDirection = 255;
1906 :			// forward MV moves
1907 :			i = fData->currentMV[0].x; j = fData->currentMV[0].y;
1908 :
1909 :			CheckCandidateInt(i + 1, j, 0, &iDirection, fData);
1910 :			CheckCandidateInt(i, j + 1, 0, &iDirection, fData);
1911 :			CheckCandidateInt(i - 1, j, 0, &iDirection, fData);
1912 :			CheckCandidateInt(i, j - 1, 0, &iDirection, fData);
1913 :
1914 :			// backward MV moves
1915 :			i = fData->currentMV[1].x; j = fData->currentMV[1].y;
1916 :			fData->currentMV[2] = fData->currentMV[0];
1917 :			CheckCandidateInt(i + 1, j, 0, &iDirection, &bData);
1918 :			CheckCandidateInt(i, j + 1, 0, &iDirection, &bData);
1919 :			CheckCandidateInt(i - 1, j, 0, &iDirection, &bData);
1920 :			CheckCandidateInt(i, j - 1, 0, &iDirection, &bData);
1921 :
1922 :			} while (!(iDirection));
1923 :
1924 :			//qpel refinement
1925 :			if (fData->qpel) {
1926 :			if (fData->iMinSAD > best_sad + 500) return;
1927 :			CheckCandidate = CheckCandidateInt;
1928 :			fData->qpel_precision = bData.qpel_precision = 1;
1929 :			get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0);
1930 :			get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0);
1931 :			fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x;
1932 :			fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y;
1933 :			fData->currentQMV[1].x = 2 * fData->currentMV[1].x;
1934 :			fData->currentQMV[1].y = 2 * fData->currentMV[1].y;
1935 :			SubpelRefine(fData);
1936 :			if (fData->iMinSAD > best_sad + 300) return;
1937 :			fData->currentQMV[2] = fData->currentQMV[0];
1938 :			SubpelRefine(&bData);
1939 :			}
1940 :
1941 :			fData->iMinSAD += (2+3) fData->lambda16; // two bits are needed to code interpolate mode.
1942 :
1943 :			if (fData->iMinSAD < best_sad) {
1944 :			best_sad = fData->iMinSAD;
1945 :			pMB->mvs[0] = fData->currentMV[0];
1946 :			pMB->b_mvs[0] = fData->currentMV[1];
1947 :			pMB->mode = MODE_INTERPOLATE;
1948 :			if (fData->qpel) {
1949 :			pMB->qmvs[0] = fData->currentQMV[0];
1950 :			pMB->b_qmvs[0] = fData->currentQMV[1];
1951 :			pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x;
1952 :			pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y;
1953 :			pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x;
1954 :			pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y;
1955 :			} else {
1956 :			pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x;
1957 :			pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y;
1958 :			pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x;
1959 :			pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y;
1960 :			}
1961 :			}
1962 :			}
1963 :
1964 :			void
1965 :			MotionEstimationBVOP(MBParam * const pParam,
1966 :	syskin	1.64	FRAMEINFO * const frame,
1967 :			const int32_t time_bp,
1968 :			const int32_t time_pp,
1969 :			// forward (past) reference
1970 :			const MACROBLOCK * const f_mbs,
1971 :			const IMAGE * const f_ref,
1972 :			const IMAGE * const f_refH,
1973 :			const IMAGE * const f_refV,
1974 :			const IMAGE * const f_refHV,
1975 :			// backward (future) reference
1976 :			const FRAMEINFO * const b_reference,
1977 :			const IMAGE * const b_ref,
1978 :			const IMAGE * const b_refH,
1979 :			const IMAGE * const b_refV,
1980 :			const IMAGE * const b_refHV)
1981 :	edgomez	1.55	{
1982 :			uint32_t i, j;
1983 :			int32_t best_sad;
1984 :			uint32_t skip_sad;
1985 :			int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0;
1986 :			const MACROBLOCK * const b_mbs = b_reference->mbs;
1987 :
1988 :			VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/
1989 :
1990 :			const int32_t TRB = time_pp - time_bp;
1991 :			const int32_t TRD = time_pp;
1992 :
1993 :			// some pre-inintialized data for the rest of the search
1994 :
1995 :			SearchData Data;
1996 :			int32_t iMinSAD;
1997 :			VECTOR currentMV[3];
1998 :			VECTOR currentQMV[3];
1999 :			int32_t temp[8];
2000 :			memset(&Data, 0, sizeof(SearchData));
2001 :			Data.iEdgedWidth = pParam->edged_width;
2002 :			Data.currentMV = currentMV; Data.currentQMV = currentQMV;
2003 :			Data.iMinSAD = &iMinSAD;
2004 :			Data.lambda16 = lambda_vec16[frame->quant];
2005 :			Data.qpel = pParam->m_quarterpel;
2006 :			Data.rounding = 0;
2007 :			Data.chroma = frame->motion_flags & PMV_CHROMA8;
2008 :			Data.temp = temp;
2009 :
2010 :			Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose)
2011 :			// note: i==horizontal, j==vertical
2012 :			for (j = 0; j < pParam->mb_height; j++) {
2013 :
2014 :			f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */
2015 :
2016 :			for (i = 0; i < pParam->mb_width; i++) {
2017 :			MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width;
2018 :			const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width;
2019 :
2020 :			/* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */
2021 :			if (b_reference->coding_type != S_VOP)
2022 :			if (b_mb->mode == MODE_NOT_CODED) {
2023 :			pMB->mode = MODE_NOT_CODED;
2024 :			continue;
2025 :			}
2026 :	edgomez	1.20
2027 :	edgomez	1.55	Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16;
2028 :			Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8;
2029 :			Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8;
2030 :			pMB->quant = frame->quant;
2031 :
2032 :			/* direct search comes first, because it (1) checks for SKIP-mode
2033 :			and (2) sets very good predictions for forward and backward search */
2034 :			skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2035 :			b_ref, b_refH->y, b_refV->y, b_refHV->y,
2036 :			&frame->image,
2037 :			i, j,
2038 :			frame->motion_flags,
2039 :			TRB, TRD,
2040 :			pParam,
2041 :			pMB, b_mb,
2042 :			&best_sad,
2043 :			&Data);
2044 :
2045 :			if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; }
2046 :
2047 :			// forward search
2048 :			SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2049 :			&frame->image, i, j,
2050 :			frame->motion_flags,
2051 :			frame->fcode, pParam,
2052 :			pMB, &f_predMV, &best_sad,
2053 :			MODE_FORWARD, &Data);
2054 :
2055 :			// backward search
2056 :			SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y,
2057 :			&frame->image, i, j,
2058 :			frame->motion_flags,
2059 :			frame->bcode, pParam,
2060 :			pMB, &b_predMV, &best_sad,
2061 :			MODE_BACKWARD, &Data);
2062 :
2063 :			// interpolate search comes last, because it uses data from forward and backward as prediction
2064 :			SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2065 :			b_ref, b_refH->y, b_refV->y, b_refHV->y,
2066 :			&frame->image,
2067 :			i, j,
2068 :			frame->fcode, frame->bcode,
2069 :			frame->motion_flags,
2070 :			pParam,
2071 :			&f_predMV, &b_predMV,
2072 :			pMB, &best_sad,
2073 :			&Data);
2074 :
2075 :			// final skip decision
2076 :			if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2)
2077 :			&& ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) )
2078 :			SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data);
2079 :
2080 :			switch (pMB->mode) {
2081 :			case MODE_FORWARD:
2082 :			f_count++;
2083 :			f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2084 :			break;
2085 :			case MODE_BACKWARD:
2086 :			b_count++;
2087 :			b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2088 :			break;
2089 :			case MODE_INTERPOLATE:
2090 :			i_count++;
2091 :			f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2092 :			b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2093 :			break;
2094 :			case MODE_DIRECT:
2095 :			case MODE_DIRECT_NO4V:
2096 :			d_count++;
2097 :			default:
2098 :			break;
2099 :	chl	1.3	}
2100 :			}
2101 :			}
2102 :	edgomez	1.55	}
2103 :
2104 :			static __inline void
2105 :			MEanalyzeMB ( const uint8_t * const pRef,
2106 :			const uint8_t * const pCur,
2107 :			const int x,
2108 :			const int y,
2109 :			const MBParam * const pParam,
2110 :			MACROBLOCK * const pMBs,
2111 :			SearchData * const Data)
2112 :			{
2113 :	chl	1.3
2114 :	edgomez	1.55	int i, mask;
2115 :			VECTOR pmv[3];
2116 :	syskin	1.60	MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width];
2117 :	chl	1.3
2118 :	edgomez	1.55	for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR;
2119 :	chl	1.3
2120 :	edgomez	1.55	//median is only used as prediction. it doesn't have to be real
2121 :			if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0;
2122 :			else
2123 :			if (x == 1) //left macroblock does not have any vector now
2124 :			Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median
2125 :			else if (y == 1) // top macroblock doesn't have it's vector
2126 :			Data->predMV = (pMB - 1)->mvs[0]; // left instead of median
2127 :			else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median
2128 :	chl	1.3
2129 :	edgomez	1.55	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2130 :	syskin	1.60	pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, 0);
2131 :	chl	1.3
2132 :	edgomez	1.55	Data->Cur = pCur + (x + y * pParam->edged_width) * 16;
2133 :	syskin	1.64	Data->RefP[0] = pRef + (x + y * pParam->edged_width) * 16;
2134 :	chl	1.3
2135 :	edgomez	1.55	pmv[1].x = EVEN(pMB->mvs[0].x);
2136 :			pmv[1].y = EVEN(pMB->mvs[0].y);
2137 :			pmv[2].x = EVEN(Data->predMV.x);
2138 :			pmv[2].y = EVEN(Data->predMV.y);
2139 :			pmv[0].x = pmv[0].y = 0;
2140 :	chl	1.3
2141 :	edgomez	1.55	CheckCandidate32I(0, 0, 255, &i, Data);
2142 :	chl	1.3
2143 :	syskin	1.60	if (Data->iMinSAD > 4 MAX_SAD00_FOR_SKIP) {
2144 :	chl	1.3
2145 :	edgomez	1.55	if (!(mask = make_mask(pmv, 1)))
2146 :			CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data);
2147 :			if (!(mask = make_mask(pmv, 2)))
2148 :			CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data);
2149 :	chl	1.3
2150 :	syskin	1.60	if (Data->iMinSAD > 4 MAX_SAD00_FOR_SKIP) // diamond only if needed
2151 :	edgomez	1.55	DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i);
2152 :	syskin	1.60	}
2153 :	chl	1.3
2154 :	syskin	1.60	for (i = 0; i < 4; i++) {
2155 :			MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width];
2156 :			MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i];
2157 :			MB->mode = MODE_INTER;
2158 :			MB->sad16 = Data->iMinSAD[i+1];
2159 :	edgomez	1.20	}
2160 :	edgomez	1.55	}
2161 :	chl	1.3
2162 :	syskin	1.72	#define INTRA_THRESH 2200
2163 :			#define INTER_THRESH 50
2164 :			#define INTRA_THRESH2 95
2165 :	chl	1.3
2166 :	edgomez	1.55	int
2167 :			MEanalysis( const IMAGE * const pRef,
2168 :	syskin	1.60	const FRAMEINFO * const Current,
2169 :			const MBParam * const pParam,
2170 :			const int maxIntra, //maximum number if non-I frames
2171 :			const int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame
2172 :			const int bCount, // number of B frames in a row
2173 :			const int b_thresh)
2174 :	edgomez	1.55	{
2175 :			uint32_t x, y, intra = 0;
2176 :			int sSAD = 0;
2177 :			MACROBLOCK * const pMBs = Current->mbs;
2178 :			const IMAGE * const pCurrent = &Current->image;
2179 :	syskin	1.72	int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + b_thresh;
2180 :	syskin	1.60	int s = 0, blocks = 0;
2181 :	syskin	1.72	int complexity = 0;
2182 :	edgomez	1.55
2183 :			int32_t iMinSAD[5], temp[5];
2184 :			VECTOR currentMV[5];
2185 :			SearchData Data;
2186 :			Data.iEdgedWidth = pParam->edged_width;
2187 :			Data.currentMV = currentMV;
2188 :			Data.iMinSAD = iMinSAD;
2189 :			Data.iFcode = Current->fcode;
2190 :			Data.temp = temp;
2191 :			CheckCandidate = CheckCandidate32I;
2192 :	chl	1.3
2193 :	syskin	1.72
2194 :	syskin	1.70	if (intraCount != 0) {
2195 :			if (intraCount < 10) // we're right after an I frame
2196 :			IntraThresh += 15* (intraCount - 10) * (intraCount - 10);
2197 :			else
2198 :			if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec
2199 :			IntraThresh -= (IntraThresh * (maxIntra - 8*(maxIntra - intraCount)))/maxIntra;
2200 :			}
2201 :	chl	1.19
2202 :	syskin	1.72	InterThresh -= 12 * bCount;
2203 :			if (InterThresh < 15 + b_thresh) InterThresh = 15 + b_thresh;
2204 :	chl	1.3
2205 :	edgomez	1.55	if (sadInit) (*sadInit) ();
2206 :	edgomez	1.20
2207 :	edgomez	1.55	for (y = 1; y < pParam->mb_height-1; y += 2) {
2208 :			for (x = 1; x < pParam->mb_width-1; x += 2) {
2209 :			int i;
2210 :	syskin	1.72	blocks += 10;
2211 :	edgomez	1.55
2212 :			if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV;
2213 :	syskin	1.60	else { //extrapolation of the vector found for last frame
2214 :	syskin	1.64	pMBs[x + y * pParam->mb_width].mvs[0].x =
2215 :	syskin	1.60	(pMBs[x + y * pParam->mb_width].mvs[0].x * (bCount+1) ) / bCount;
2216 :	syskin	1.64	pMBs[x + y * pParam->mb_width].mvs[0].y =
2217 :	syskin	1.60	(pMBs[x + y * pParam->mb_width].mvs[0].y * (bCount+1) ) / bCount;
2218 :			}
2219 :	edgomez	1.55
2220 :			MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data);
2221 :
2222 :			for (i = 0; i < 4; i++) {
2223 :			int dev;
2224 :			MACROBLOCK pMB = &pMBs[x+(i&1) + (y+(i>>1)) pParam->mb_width];
2225 :	syskin	1.72	dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16,
2226 :			pParam->edged_width);
2227 :
2228 :			complexity += dev;
2229 :			if (dev + IntraThresh < pMB->sad16) {
2230 :			pMB->mode = MODE_INTRA;
2231 :			if (++intra > ((pParam->mb_height-2)*(pParam->mb_width-2))/2) return I_VOP;
2232 :	edgomez	1.55	}
2233 :	syskin	1.72
2234 :			if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0)
2235 :			if (dev > 500 && pMB->sad16 < 1000)
2236 :			sSAD += 1000;
2237 :	syskin	1.60
2238 :	edgomez	1.55	sSAD += pMB->sad16;
2239 :			}
2240 :			}
2241 :	chl	1.3	}
2242 :	syskin	1.72	complexity >>= 7;
2243 :	syskin	1.60
2244 :	syskin	1.72	sSAD /= complexity + 4*blocks;
2245 :	syskin	1.60
2246 :	syskin	1.72	if (intraCount > 12 && sSAD > INTRA_THRESH2 ) return I_VOP;
2247 :	edgomez	1.55	if (sSAD > InterThresh ) return P_VOP;
2248 :			emms();
2249 :			return B_VOP;
2250 :			}
2251 :
2252 :	chl	1.3
2253 :	edgomez	1.55	static WARPPOINTS
2254 :			GlobalMotionEst(const MACROBLOCK * const pMBs,
2255 :			const MBParam * const pParam,
2256 :			const FRAMEINFO * const current,
2257 :			const FRAMEINFO * const reference,
2258 :			const IMAGE * const pRefH,
2259 :			const IMAGE * const pRefV,
2260 :			const IMAGE * const pRefHV )
2261 :			{
2262 :	chl	1.3
2263 :	edgomez	1.55	const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs
2264 :			const int deltay=8;
2265 :			const int grad=512; // lower bound for deviation in MB
2266 :	chl	1.3
2267 :	edgomez	1.55	WARPPOINTS gmc;
2268 :	edgomez	1.20
2269 :	edgomez	1.55	uint32_t mx, my;
2270 :	edgomez	1.20
2271 :	edgomez	1.55	int MBh = pParam->mb_height;
2272 :			int MBw = pParam->mb_width;
2273 :	edgomez	1.20
2274 :	edgomez	1.55	int MBmask= calloc(MBhMBw,sizeof(int));
2275 :			double DtimesF[4] = { 0.,0., 0., 0. };
2276 :			double sol[4] = { 0., 0., 0., 0. };
2277 :			double a,b,c,n,denom;
2278 :			double meanx,meany;
2279 :			int num,oldnum;
2280 :	chl	1.3
2281 :	syskin	1.64	if (!MBmask) { fprintf(stderr,"Mem error\n");
2282 :			gmc.duv[0].x= gmc.duv[0].y =
2283 :			gmc.duv[1].x= gmc.duv[1].y =
2284 :	chl	1.56	gmc.duv[2].x= gmc.duv[2].y = 0;
2285 :			return gmc; }
2286 :	chl	1.3
2287 :	edgomez	1.55	// filter mask of all blocks
2288 :	chl	1.3
2289 :	edgomez	1.58	for (my = 1; my < (uint32_t)MBh-1; my++)
2290 :			for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2291 :	edgomez	1.55	{
2292 :			const int mbnum = mx + my * MBw;
2293 :			const MACROBLOCK *pMB = &pMBs[mbnum];
2294 :			const VECTOR mv = pMB->mvs[0];
2295 :	chl	1.3
2296 :	edgomez	1.55	if (pMB->mode == MODE_INTRA \|\| pMB->mode == MODE_NOT_CODED)
2297 :			continue;
2298 :	edgomez	1.20
2299 :	syskin	1.65	if ( ( (abs(mv.x - (pMB-1)->mvs[0].x) < deltax) && (abs(mv.y - (pMB-1)->mvs[0].y) < deltay) )
2300 :			&& ( (abs(mv.x - (pMB+1)->mvs[0].x) < deltax) && (abs(mv.y - (pMB+1)->mvs[0].y) < deltay) )
2301 :			&& ( (abs(mv.x - (pMB-MBw)->mvs[0].x) < deltax) && (abs(mv.y - (pMB-MBw)->mvs[0].y) < deltay) )
2302 :			&& ( (abs(mv.x - (pMB+MBw)->mvs[0].x) < deltax) && (abs(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) )
2303 :	edgomez	1.55	MBmask[mbnum]=1;
2304 :			}
2305 :	chl	1.3
2306 :	edgomez	1.58	for (my = 1; my < (uint32_t)MBh-1; my++)
2307 :			for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2308 :	edgomez	1.55	{
2309 :			const uint8_t const pCur = current->image.y + 16mypParam->edged_width + 16mx;
2310 :	chl	1.3
2311 :	edgomez	1.55	const int mbnum = mx + my * MBw;
2312 :			if (!MBmask[mbnum])
2313 :			continue;
2314 :	chl	1.10
2315 :	edgomez	1.58	if (sad16 ( pCur, pCur+1 , pParam->edged_width, 65536) <= (uint32_t)grad )
2316 :	edgomez	1.55	MBmask[mbnum] = 0;
2317 :	edgomez	1.58	if (sad16 ( pCur, pCur+pParam->edged_width, pParam->edged_width, 65536) <= (uint32_t)grad )
2318 :	edgomez	1.55	MBmask[mbnum] = 0;
2319 :	edgomez	1.20
2320 :	edgomez	1.55	}
2321 :	chl	1.3
2322 :	edgomez	1.55	emms();
2323 :	chl	1.3
2324 :	edgomez	1.55	do { /* until convergence */
2325 :	edgomez	1.20
2326 :	edgomez	1.55	a = b = c = n = 0;
2327 :			DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.;
2328 :	edgomez	1.58	for (my = 0; my < (uint32_t)MBh; my++)
2329 :			for (mx = 0; mx < (uint32_t)MBw; mx++)
2330 :	edgomez	1.55	{
2331 :			const int mbnum = mx + my * MBw;
2332 :			const MACROBLOCK *pMB = &pMBs[mbnum];
2333 :			const VECTOR mv = pMB->mvs[0];
2334 :
2335 :			if (!MBmask[mbnum])
2336 :			continue;
2337 :
2338 :			n++;
2339 :			a += 16*mx+8;
2340 :			b += 16*my+8;
2341 :			c += (16mx+8)(16mx+8)+(16my+8)(16my+8);
2342 :
2343 :			DtimesF[0] += (double)mv.x;
2344 :			DtimesF[1] += (double)mv.x(16mx+8) + (double)mv.y(16my+8);
2345 :			DtimesF[2] += (double)mv.x(16my+8) - (double)mv.y(16mx+8);
2346 :			DtimesF[3] += (double)mv.y;
2347 :			}
2348 :
2349 :			denom = aa+bb-c*n;
2350 :
2351 :	syskin	1.64	/* Solve the system: sol = (D'ED)^{-1} D'EF */
2352 :	edgomez	1.55	/* D'EF has been calculated in the same loop as matrix */
2353 :
2354 :			sol[0] = -cDtimesF[0] + aDtimesF[1] + b*DtimesF[2];
2355 :	syskin	1.64	sol[1] = aDtimesF[0] - nDtimesF[1] + b*DtimesF[3];
2356 :			sol[2] = bDtimesF[0] - nDtimesF[2] - a*DtimesF[3];
2357 :			sol[3] = bDtimesF[1] - aDtimesF[2] - c*DtimesF[3];
2358 :	edgomez	1.55
2359 :			sol[0] /= denom;
2360 :			sol[1] /= denom;
2361 :			sol[2] /= denom;
2362 :			sol[3] /= denom;
2363 :
2364 :			meanx = meany = 0.;
2365 :			oldnum = 0;
2366 :	edgomez	1.58	for (my = 0; my < (uint32_t)MBh; my++)
2367 :			for (mx = 0; mx < (uint32_t)MBw; mx++)
2368 :	edgomez	1.55	{
2369 :			const int mbnum = mx + my * MBw;
2370 :			const MACROBLOCK *pMB = &pMBs[mbnum];
2371 :			const VECTOR mv = pMB->mvs[0];
2372 :	edgomez	1.20
2373 :	edgomez	1.55	if (!MBmask[mbnum])
2374 :			continue;
2375 :	chl	1.3
2376 :	edgomez	1.55	oldnum++;
2377 :	syskin	1.65	meanx += fabs(( sol[0] + (16mx+8)sol[1] + (16my+8)sol[2] ) - mv.x );
2378 :			meany += fabs(( sol[3] - (16mx+8)sol[2] + (16my+8)sol[1] ) - mv.y );
2379 :	edgomez	1.55	}
2380 :	chl	1.3
2381 :	edgomez	1.55	if (4meanx > oldnum) / better fit than 0.25 is useless */
2382 :			meanx /= oldnum;
2383 :			else
2384 :			meanx = 0.25;
2385 :	edgomez	1.20
2386 :	edgomez	1.55	if (4*meany > oldnum)
2387 :			meany /= oldnum;
2388 :			else
2389 :			meany = 0.25;
2390 :	chl	1.3
2391 :	edgomez	1.55	/* fprintf(stderr,"sol = (%8.5f, %8.5f, %8.5f, %8.5f)\n",sol[0],sol[1],sol[2],sol[3]);
2392 :			fprintf(stderr,"meanx = %8.5f meany = %8.5f %d\n",meanx,meany, oldnum);
2393 :			*/
2394 :			num = 0;
2395 :	edgomez	1.58	for (my = 0; my < (uint32_t)MBh; my++)
2396 :			for (mx = 0; mx < (uint32_t)MBw; mx++)
2397 :	edgomez	1.55	{
2398 :			const int mbnum = mx + my * MBw;
2399 :			const MACROBLOCK *pMB = &pMBs[mbnum];
2400 :			const VECTOR mv = pMB->mvs[0];
2401 :	chl	1.3
2402 :	edgomez	1.55	if (!MBmask[mbnum])
2403 :			continue;
2404 :	chl	1.3
2405 :	syskin	1.65	if ( ( fabs(( sol[0] + (16mx+8)sol[1] + (16my+8)sol[2] ) - mv.x ) > meanx )
2406 :			\|\| ( fabs(( sol[3] - (16mx+8)sol[2] + (16my+8)sol[1] ) - mv.y ) > meany ) )
2407 :	edgomez	1.55	MBmask[mbnum]=0;
2408 :			else
2409 :			num++;
2410 :			}
2411 :	chl	1.3
2412 :	edgomez	1.55	} while ( (oldnum != num) && (num>=4) );
2413 :	chl	1.3
2414 :	edgomez	1.55	if (num < 4)
2415 :			{
2416 :			gmc.duv[0].x= gmc.duv[0].y= gmc.duv[1].x= gmc.duv[1].y= gmc.duv[2].x= gmc.duv[2].y=0;
2417 :			} else {
2418 :	chl	1.3
2419 :	edgomez	1.55	gmc.duv[0].x=(int)(sol[0]+0.5);
2420 :			gmc.duv[0].y=(int)(sol[3]+0.5);
2421 :	chl	1.3
2422 :	edgomez	1.55	gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5);
2423 :			gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5);
2424 :	edgomez	1.20
2425 :	edgomez	1.55	gmc.duv[2].x=0;
2426 :			gmc.duv[2].y=0;
2427 :	chl	1.3	}
2428 :	edgomez	1.55	// fprintf(stderr,"wp1 = ( %4d, %4d) wp2 = ( %4d, %4d) \n", gmc.duv[0].x, gmc.duv[0].y, gmc.duv[1].x, gmc.duv[1].y);
2429 :	chl	1.3
2430 :	edgomez	1.55	free(MBmask);
2431 :	edgomez	1.20
2432 :	edgomez	1.55	return gmc;
2433 :			}
2434 :	edgomez	1.20
2435 :	edgomez	1.55	// functions which perform BITS-based search/bitcount
2436 :	edgomez	1.20
2437 :	edgomez	1.55	static int
2438 :			CountMBBitsInter(SearchData * const Data,
2439 :			const MACROBLOCK * const pMBs, const int x, const int y,
2440 :			const MBParam * const pParam,
2441 :			const uint32_t MotionFlags)
2442 :			{
2443 :			int i, iDirection;
2444 :			int32_t bsad[5];
2445 :	edgomez	1.20
2446 :	edgomez	1.55	CheckCandidate = CheckCandidateBits16;
2447 :	edgomez	1.20
2448 :	edgomez	1.55	if (Data->qpel) {
2449 :			for(i = 0; i < 5; i++) {
2450 :			Data->currentMV[i].x = Data->currentQMV[i].x/2;
2451 :			Data->currentMV[i].y = Data->currentQMV[i].y/2;
2452 :			}
2453 :			Data->qpel_precision = 1;
2454 :			CheckCandidateBits16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data);
2455 :	chl	1.3
2456 :	edgomez	1.55	if (MotionFlags & (HALFPELREFINE16_BITS \| EXTSEARCH_BITS)) { //we have to prepare for halfpixel-precision search
2457 :			for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i];
2458 :			get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2459 :			pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
2460 :			Data->qpel_precision = 0;
2461 :			if (Data->currentQMV->x & 1 \|\| Data->currentQMV->y & 1)
2462 :			CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2463 :			}
2464 :	chl	1.10
2465 :	edgomez	1.55	} else { // not qpel
2466 :	chl	1.10
2467 :	edgomez	1.55	CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2468 :	chl	1.10	}
2469 :	edgomez	1.20
2470 :	edgomez	1.55	if (MotionFlags&EXTSEARCH_BITS) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
2471 :	syskin	1.69
2472 :	edgomez	1.55	if (MotionFlags&HALFPELREFINE16_BITS) SubpelRefine(Data);
2473 :
2474 :			if (Data->qpel) {
2475 :			if (MotionFlags&(EXTSEARCH_BITS \| HALFPELREFINE16_BITS)) { // there was halfpel-precision search
2476 :			for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) {
2477 :			Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // we have found a better match
2478 :			Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
2479 :	chl	1.10	}
2480 :	edgomez	1.55
2481 :			// preparing for qpel-precision search
2482 :			Data->qpel_precision = 1;
2483 :			get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2484 :			pParam->width, pParam->height, Data->iFcode, 1, 0);
2485 :	chl	1.10	}
2486 :	edgomez	1.55	if (MotionFlags&QUARTERPELREFINE16_BITS) SubpelRefine(Data);
2487 :	chl	1.10	}
2488 :
2489 :	edgomez	1.55	if (MotionFlags&CHECKPREDICTION_BITS) { //let's check vector equal to prediction
2490 :			VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV;
2491 :			if (!(Data->predMV.x == v->x && Data->predMV.y == v->y))
2492 :			CheckCandidateBits16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data);
2493 :			}
2494 :			return Data->iMinSAD[0];
2495 :			}
2496 :	chl	1.10
2497 :	edgomez	1.55	static int
2498 :			CountMBBitsInter4v(const SearchData * const Data,
2499 :			MACROBLOCK * const pMB, const MACROBLOCK * const pMBs,
2500 :			const int x, const int y,
2501 :			const MBParam * const pParam, const uint32_t MotionFlags,
2502 :			const VECTOR * const backup)
2503 :			{
2504 :	chl	1.3
2505 :	edgomez	1.55	int cbp = 0, bits = 0, t = 0, i, iDirection;
2506 :			SearchData Data2, *Data8 = &Data2;
2507 :			int sumx = 0, sumy = 0;
2508 :	edgomez	1.61	int16_t in = Data->dctSpace, coeff = Data->dctSpace + 64;
2509 :	syskin	1.69	uint8_t * ptr;
2510 :	edgomez	1.55
2511 :			memcpy(Data8, Data, sizeof(SearchData));
2512 :			CheckCandidate = CheckCandidateBits8;
2513 :
2514 :	syskin	1.69	for (i = 0; i < 4; i++) { //for all luma blocks
2515 :
2516 :	edgomez	1.55	Data8->iMinSAD = Data->iMinSAD + i + 1;
2517 :			Data8->currentMV = Data->currentMV + i + 1;
2518 :			Data8->currentQMV = Data->currentQMV + i + 1;
2519 :			Data8->Cur = Data->Cur + 8((i&1) + (i>>1)Data->iEdgedWidth);
2520 :	syskin	1.64	Data8->RefP[0] = Data->RefP[0] + 8((i&1) + (i>>1)Data->iEdgedWidth);
2521 :			Data8->RefP[2] = Data->RefP[2] + 8((i&1) + (i>>1)Data->iEdgedWidth);
2522 :			Data8->RefP[1] = Data->RefP[1] + 8((i&1) + (i>>1)Data->iEdgedWidth);
2523 :			Data8->RefP[3] = Data->RefP[3] + 8((i&1) + (i>>1)Data->iEdgedWidth);
2524 :	edgomez	1.55
2525 :			if(Data->qpel) {
2526 :			Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i);
2527 :			if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y,
2528 :			Data8->predMV, Data8->iFcode, 0, 0);
2529 :			} else {
2530 :			Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i);
2531 :			if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y,
2532 :			Data8->predMV, Data8->iFcode, 0, 0);
2533 :			}
2534 :
2535 :			get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2x + (i&1), 2y + (i>>1), 8,
2536 :			pParam->width, pParam->height, Data8->iFcode, Data8->qpel, 0);
2537 :
2538 :	syskin	1.70	Data8->iMinSAD += BITS_MULTt;
2539 :	edgomez	1.55
2540 :			Data8->qpel_precision = Data8->qpel;
2541 :			// checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far)
2542 :	syskin	1.65	{
2543 :			VECTOR *v = Data8->qpel ? Data8->currentQMV : Data8->currentMV;
2544 :	syskin	1.69	if (!MVequal (*v, backup[i+1]) )
2545 :	edgomez	1.55	CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8);
2546 :			}
2547 :
2548 :			if (Data8->qpel) {
2549 :			if (MotionFlags&HALFPELREFINE8_BITS \|\| (MotionFlags&PMV_EXTSEARCH8 && MotionFlags&EXTSEARCH_BITS)) { // halfpixel motion search follows
2550 :			int32_t s = *Data8->iMinSAD;
2551 :			Data8->currentMV->x = Data8->currentQMV->x/2;
2552 :			Data8->currentMV->y = Data8->currentQMV->y/2;
2553 :			Data8->qpel_precision = 0;
2554 :			get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2x + (i&1), 2y + (i>>1), 8,
2555 :			pParam->width, pParam->height, Data8->iFcode - 1, 0, 0);
2556 :
2557 :			if (Data8->currentQMV->x & 1 \|\| Data8->currentQMV->y & 1)
2558 :			CheckCandidateBits8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8);
2559 :
2560 :			if (MotionFlags & PMV_EXTSEARCH8 && MotionFlags & EXTSEARCH_BITS)
2561 :			SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255);
2562 :
2563 :	syskin	1.69	if (MotionFlags & HALFPELREFINE8_BITS)
2564 :			SubpelRefine(Data8);
2565 :	edgomez	1.55
2566 :	syskin	1.69	if (s > *Data8->iMinSAD) { //we have found a better match
2567 :	edgomez	1.55	Data8->currentQMV->x = 2*Data8->currentMV->x;
2568 :			Data8->currentQMV->y = 2*Data8->currentMV->y;
2569 :			}
2570 :	chl	1.3
2571 :	edgomez	1.55	Data8->qpel_precision = 1;
2572 :			get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2x + (i&1), 2y + (i>>1), 8,
2573 :			pParam->width, pParam->height, Data8->iFcode, 1, 0);
2574 :	chl	1.3
2575 :	edgomez	1.55	}
2576 :			if (MotionFlags & QUARTERPELREFINE8_BITS) SubpelRefine(Data8);
2577 :	chl	1.3
2578 :	syskin	1.69	} else { // not qpel
2579 :
2580 :			if (MotionFlags & PMV_EXTSEARCH8 && MotionFlags & EXTSEARCH_BITS) //extsearch
2581 :			SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255);
2582 :
2583 :			if (MotionFlags & HALFPELREFINE8_BITS)
2584 :			SubpelRefine(Data8); //halfpel refinement
2585 :			}
2586 :	chl	1.3
2587 :	edgomez	1.55	//checking vector equal to predicion
2588 :			if (i != 0 && MotionFlags & CHECKPREDICTION_BITS) {
2589 :			const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV;
2590 :	syskin	1.69	if (!MVequal(*v, Data8->predMV))
2591 :	edgomez	1.55	CheckCandidateBits8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8);
2592 :			}
2593 :
2594 :			bits += *Data8->iMinSAD;
2595 :	syskin	1.69	if (bits >= Data->iMinSAD[0]) return bits; // no chances for INTER4V
2596 :	edgomez	1.55
2597 :			// MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else
2598 :			if(Data->qpel) {
2599 :			pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x;
2600 :			pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y;
2601 :			pMB->qmvs[i] = *Data8->currentQMV;
2602 :			sumx += Data8->currentQMV->x/2;
2603 :			sumy += Data8->currentQMV->y/2;
2604 :			} else {
2605 :			pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x;
2606 :			pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y;
2607 :			sumx += Data8->currentMV->x;
2608 :			sumy += Data8->currentMV->y;
2609 :			}
2610 :			pMB->mvs[i] = *Data8->currentMV;
2611 :			pMB->sad8[i] = 4 * *Data8->iMinSAD;
2612 :			if (Data8->temp[0]) cbp \|= 1 << (5 - i);
2613 :
2614 :	syskin	1.69	} // /for all luma blocks
2615 :
2616 :	syskin	1.70	bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len;
2617 :	syskin	1.69
2618 :			// let's check chroma
2619 :			sumx = (sumx >> 3) + roundtab_76[sumx & 0xf];
2620 :			sumy = (sumy >> 3) + roundtab_76[sumy & 0xf];
2621 :
2622 :			//chroma U
2623 :			ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[4], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2624 :			transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2);
2625 :	syskin	1.72	bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4);
2626 :	syskin	1.69
2627 :			if (bits >= *Data->iMinSAD) return bits;
2628 :
2629 :			//chroma V
2630 :			ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[5], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2631 :			transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2);
2632 :	syskin	1.72	bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5);
2633 :	syskin	1.69
2634 :	syskin	1.70	bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER4V & 7) \| ((cbp & 3) << 3)].len;
2635 :	edgomez	1.20
2636 :	edgomez	1.55	return bits;
2637 :			}
2638 :	chl	1.10
2639 :	edgomez	1.55	static int
2640 :			CountMBBitsIntra(const SearchData * const Data)
2641 :			{
2642 :	syskin	1.70	int bits = BITS_MULT*1; //this one is ac/dc prediction flag bit
2643 :			int cbp = 0, i, dc = 0;
2644 :	edgomez	1.61	int16_t in = Data->dctSpace, coeff = Data->dctSpace + 64;
2645 :	chl	1.18
2646 :	edgomez	1.55	for(i = 0; i < 4; i++) {
2647 :			int s = 8((i&1) + (i>>1)Data->iEdgedWidth);
2648 :			transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth);
2649 :	syskin	1.72	bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i, &dc);
2650 :	syskin	1.70
2651 :	syskin	1.69	if (bits >= Data->iMinSAD[0]) return bits;
2652 :	chl	1.3	}
2653 :
2654 :	syskin	1.70	bits += BITS_MULT*xvid_cbpy_tab[cbp>>2].len;
2655 :	syskin	1.69
2656 :			//chroma U
2657 :			transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2);
2658 :	syskin	1.72	bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4, &dc);
2659 :	syskin	1.70
2660 :	syskin	1.69	if (bits >= Data->iMinSAD[0]) return bits;
2661 :
2662 :			//chroma V
2663 :			transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2);
2664 :	syskin	1.72	bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5, &dc);
2665 :	chl	1.3
2666 :	syskin	1.70	bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTRA & 7) \| ((cbp & 3) << 3)].len;
2667 :	edgomez	1.20
2668 :	edgomez	1.55	return bits;
2669 :	chl	1.3	}

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