Annotation of /branches/dev-api-4/xvidcore/src/motion/motion_est.c

Revision 1129 - (view) (download)

1 :	edgomez	1054	/*****************************************************************************
2 :	Isibaar	3	*
3 :	edgomez	1054	* XVID MPEG-4 VIDEO CODEC
4 :			* - Motion Estimation related code -
5 :	chl	259	*
6 :	edgomez	1054	* Copyright(C) 2002 Christoph Lampert <gruel@web.de>
7 :			* 2002 Michael Militzer <michael@xvid.org>
8 :			* 2002-2003 Radoslaw Czyz <xvid@syskin.cjb.net>
9 :	chl	259	*
10 :	edgomez	1054	* This program is free software ; you can redistribute it and/or modify
11 :			* it under the terms of the GNU General Public License as published by
12 :			* the Free Software Foundation ; either version 2 of the License, or
13 :			* (at your option) any later version.
14 :	chl	259	*
15 :	edgomez	1054	* This program is distributed in the hope that it will be useful,
16 :			* but WITHOUT ANY WARRANTY ; without even the implied warranty of
17 :			* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 :			* GNU General Public License for more details.
19 :	chl	259	*
20 :	edgomez	1054	* You should have received a copy of the GNU General Public License
21 :			* along with this program ; if not, write to the Free Software
22 :			* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
23 :	chl	259	*
24 :	syskin	1129	* $Id: motion_est.c,v 1.58.2.28 2003-08-25 15:10:13 syskin Exp $
25 :	edgomez	1054	*
26 :			****************************************************************************/
27 :	chl	259
28 :	Isibaar	3	#include <assert.h>
29 :			#include <stdio.h>
30 :	chl	96	#include <stdlib.h>
31 :	edgomez	1053	#include <string.h> /* memcpy */
32 :			#include <math.h> /* lrint */
33 :	Isibaar	3
34 :			#include "../encoder.h"
35 :			#include "../utils/mbfunctions.h"
36 :			#include "../prediction/mbprediction.h"
37 :	syskin	1084	#include "../global.h"
38 :	Isibaar	3	#include "../utils/timer.h"
39 :	edgomez	851	#include "../image/interpolate8x8.h"
40 :			#include "motion_est.h"
41 :	suxen_drol	118	#include "motion.h"
42 :	Isibaar	3	#include "sad.h"
43 :	chl	1077	#include "gmc.h"
44 :	edgomez	851	#include "../utils/emms.h"
45 :			#include "../dct/fdct.h"
46 :	Isibaar	3
47 :	edgomez	872	/*****************************************************************************
48 :			* Modified rounding tables -- declared in motion.h
49 :			* Original tables see ISO spec tables 7-6 -> 7-9
50 :			****************************************************************************/
51 :
52 :			const uint32_t roundtab[16] =
53 :			{0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2 };
54 :
55 :			/* K = 4 */
56 :			const uint32_t roundtab_76[16] =
57 :			{ 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1 };
58 :
59 :			/* K = 2 */
60 :			const uint32_t roundtab_78[8] =
61 :			{ 0, 0, 1, 1, 0, 0, 0, 1 };
62 :
63 :			/* K = 1 */
64 :			const uint32_t roundtab_79[4] =
65 :			{ 0, 1, 0, 0 };
66 :
67 :	edgomez	851	#define INITIAL_SKIP_THRESH (10)
68 :			#define FINAL_SKIP_THRESH (50)
69 :			#define MAX_SAD00_FOR_SKIP (20)
70 :			#define MAX_CHROMA_SAD_FOR_SKIP (22)
71 :	Isibaar	3
72 :	edgomez	851	#define CHECK_CANDIDATE(X,Y,D) { \
73 :			CheckCandidate((X),(Y), (D), &iDirection, data ); }
74 :	Isibaar	3
75 :	syskin	1084
76 :	edgomez	872	/*****************************************************************************
77 :			* Code
78 :			****************************************************************************/
79 :
80 :	edgomez	851	static __inline uint32_t
81 :			d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv)
82 :			{
83 :	edgomez	974	int bits;
84 :			const int q = (1 << (iFcode - 1)) - 1;
85 :
86 :			x <<= qpel;
87 :			y <<= qpel;
88 :	edgomez	851	if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); }
89 :	chl	141
90 :	edgomez	851	x -= pred.x;
91 :	edgomez	974	bits = (x != 0 ? iFcode:0);
92 :			x = abs(x);
93 :			x += q;
94 :			x >>= (iFcode - 1);
95 :			bits += mvtab[x];
96 :
97 :	edgomez	851	y -= pred.y;
98 :	edgomez	974	bits += (y != 0 ? iFcode:0);
99 :			y = abs(y);
100 :			y += q;
101 :			y >>= (iFcode - 1);
102 :			bits += mvtab[y];
103 :	chl	141
104 :	edgomez	974	return bits;
105 :	edgomez	851	}
106 :	chl	141
107 :	edgomez	978	static int32_t ChromaSAD2(const int fx, const int fy, const int bx, const int by,
108 :			const SearchData * const data)
109 :	edgomez	851	{
110 :			int sad;
111 :			const uint32_t stride = data->iEdgedWidth/2;
112 :	syskin	1129	uint8_t f_refu, f_refv, b_refu, b_refv;
113 :
114 :			const INTERPOLATE8X8_PTR interpolate8x8_halfpel[] = {
115 :			NULL,
116 :			interpolate8x8_halfpel_v,
117 :			interpolate8x8_halfpel_h,
118 :			interpolate8x8_halfpel_hv
119 :			};
120 :
121 :	edgomez	978	int offset = (fx>>1) + (fy>>1)*stride;
122 :	syskin	1129	int filter = ((fx & 1) << 1) \| (fy & 1);
123 :	Isibaar	3
124 :	syskin	1129	if (filter != 0) {
125 :			f_refu = data->RefQ;
126 :			f_refv = data->RefQ + 8;
127 :			interpolate8x8_halfpel[filter](f_refu, data->RefP[4] + offset, stride, data->rounding);
128 :			interpolate8x8_halfpel[filter](f_refv, data->RefP[5] + offset, stride, data->rounding);
129 :			} else {
130 :			f_refu = (uint8_t*)data->RefP[4] + offset;
131 :			f_refv = (uint8_t*)data->RefP[5] + offset;
132 :	edgomez	851	}
133 :	Isibaar	3
134 :	edgomez	978	offset = (bx>>1) + (by>>1)*stride;
135 :	syskin	1129	filter = ((bx & 1) << 1) \| (by & 1);
136 :
137 :			if (filter != 0) {
138 :			b_refu = data->RefQ + 16;
139 :			b_refv = data->RefQ + 24;
140 :			interpolate8x8_halfpel[filter](b_refu, data->b_RefP[4] + offset, stride, data->rounding);
141 :			interpolate8x8_halfpel[filter](b_refv, data->b_RefP[5] + offset, stride, data->rounding);
142 :			} else {
143 :			b_refu = (uint8_t*)data->b_RefP[4] + offset;
144 :			b_refv = (uint8_t*)data->b_RefP[5] + offset;
145 :	edgomez	851	}
146 :	Isibaar	3
147 :	edgomez	851	sad = sad8bi(data->CurU, b_refu, f_refu, stride);
148 :			sad += sad8bi(data->CurV, b_refv, f_refv, stride);
149 :	Isibaar	3
150 :	edgomez	851	return sad;
151 :			}
152 :	Isibaar	3
153 :	edgomez	851	static int32_t
154 :	edgomez	978	ChromaSAD(const int dx, const int dy, const SearchData * const data)
155 :	edgomez	851	{
156 :			int sad;
157 :			const uint32_t stride = data->iEdgedWidth/2;
158 :	edgomez	978	int offset = (dx>>1) + (dy>>1)*stride;
159 :	Isibaar	3
160 :	edgomez	1053	if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; /* it has been checked recently */
161 :			data->temp[5] = dx; data->temp[6] = dy; /* backup */
162 :	Isibaar	3
163 :	edgomez	851	switch (((dx & 1) << 1) \| (dy & 1)) {
164 :			case 0:
165 :	edgomez	978	sad = sad8(data->CurU, data->RefP[4] + offset, stride);
166 :			sad += sad8(data->CurV, data->RefP[5] + offset, stride);
167 :	edgomez	851	break;
168 :			case 1:
169 :	edgomez	978	sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + stride, stride);
170 :			sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + stride, stride);
171 :	edgomez	851	break;
172 :			case 2:
173 :	edgomez	978	sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + 1, stride);
174 :			sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + 1, stride);
175 :	edgomez	851	break;
176 :			default:
177 :	edgomez	978	interpolate8x8_halfpel_hv(data->RefQ, data->RefP[4] + offset, stride, data->rounding);
178 :	edgomez	851	sad = sad8(data->CurU, data->RefQ, stride);
179 :
180 :	edgomez	978	interpolate8x8_halfpel_hv(data->RefQ, data->RefP[5] + offset, stride, data->rounding);
181 :	edgomez	851	sad += sad8(data->CurV, data->RefQ, stride);
182 :			break;
183 :			}
184 :	edgomez	1053	data->temp[7] = sad; /* backup, part 2 */
185 :	edgomez	851	return sad;
186 :	Isibaar	3	}
187 :
188 :	edgomez	851	static __inline const uint8_t *
189 :			GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data)
190 :	Isibaar	3	{
191 :	edgomez	1053	/* dir : 0 = forward, 1 = backward */
192 :	edgomez	987	const uint8_t const const direction = ( dir == 0 ? data->RefP : data->b_RefP );
193 :	edgomez	978	const int picture = ((x&1)<<1) \| (y&1);
194 :			const int offset = (x>>1) + (y>>1)*data->iEdgedWidth;
195 :			return direction[picture] + offset;
196 :	Isibaar	3	}
197 :
198 :	edgomez	1053	/* this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate */
199 :	edgomez	851	static __inline const uint8_t *
200 :			GetReference(const int x, const int y, const SearchData * const data)
201 :	Isibaar	3	{
202 :	edgomez	978	const int picture = ((x&1)<<1) \| (y&1);
203 :			const int offset = (x>>1) + (y>>1)*data->iEdgedWidth;
204 :			return data->RefP[picture] + offset;
205 :	Isibaar	3	}
206 :	edgomez	851
207 :			static uint8_t *
208 :			Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data)
209 :	Isibaar	3	{
210 :	edgomez	1053	/* create or find a qpel-precision reference picture; return pointer to it */
211 :	edgomez	851	uint8_t * Reference = data->RefQ + 16*dir;
212 :			const uint32_t iEdgedWidth = data->iEdgedWidth;
213 :			const uint32_t rounding = data->rounding;
214 :			const int halfpel_x = x/2;
215 :			const int halfpel_y = y/2;
216 :			const uint8_t ref1, ref2, ref3, ref4;
217 :	suxen_drol	136
218 :	edgomez	851	ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
219 :			ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
220 :			switch( ((x&1)<<1) + (y&1) ) {
221 :	edgomez	1053	case 3: /* x and y in qpel resolution - the "corners" (top left/right and */
222 :			/* bottom left/right) during qpel refinement */
223 :	edgomez	963	ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
224 :			ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
225 :			ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
226 :			ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
227 :			ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
228 :			ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
229 :			interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
230 :	edgomez	851	break;
231 :	chl	184
232 :	edgomez	1053	case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */
233 :	edgomez	851	ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
234 :			ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
235 :			interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
236 :			break;
237 :	suxen_drol	136
238 :	edgomez	1053	case 2: /* x qpel, y halfpel - left or right during qpel refinement */
239 :	edgomez	851	ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
240 :			ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
241 :			interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
242 :			break;
243 :	edgomez	195
244 :	edgomez	1053	default: /* pure halfpel position */
245 :	edgomez	963	return (uint8_t *) ref1;
246 :	edgomez	978
247 :	edgomez	851	}
248 :			return Reference;
249 :			}
250 :	Isibaar	3
251 :	edgomez	851	static uint8_t *
252 :			Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data)
253 :			{
254 :	edgomez	1053	/* create or find a qpel-precision reference picture; return pointer to it */
255 :	edgomez	851	uint8_t * Reference = data->RefQ + 16*dir;
256 :			const uint32_t iEdgedWidth = data->iEdgedWidth;
257 :			const uint32_t rounding = data->rounding;
258 :			const int halfpel_x = x/2;
259 :			const int halfpel_y = y/2;
260 :			const uint8_t ref1, ref2, ref3, ref4;
261 :	chl	326
262 :	edgomez	851	ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
263 :			switch( ((x&1)<<1) + (y&1) ) {
264 :	edgomez	1053	case 3:
265 :			/*
266 :			* x and y in qpel resolution - the "corners" (top left/right and
267 :			* bottom left/right) during qpel refinement
268 :			*/
269 :	edgomez	851	ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
270 :			ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
271 :			ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
272 :			interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
273 :			interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding);
274 :			interpolate8x8_avg4(Reference+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8iEdgedWidth, ref3+8iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding);
275 :			interpolate8x8_avg4(Reference+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8iEdgedWidth+8, ref3+8iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding);
276 :			break;
277 :	chl	184
278 :	edgomez	1053	case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */
279 :	edgomez	851	ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
280 :			interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
281 :			interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
282 :			interpolate8x8_avg2(Reference+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
283 :			interpolate8x8_avg2(Reference+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
284 :			break;
285 :	chl	184
286 :	edgomez	1053	case 2: /* x qpel, y halfpel - left or right during qpel refinement */
287 :	edgomez	851	ref2 = GetReferenceB(x - halfpel_x, 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 :	edgomez	195
294 :	syskin	1129
295 :	edgomez	1053	default: /* pure halfpel position */
296 :	edgomez	851	return (uint8_t *) ref1;
297 :			}
298 :			return Reference;
299 :			}
300 :	edgomez	195
301 :	edgomez	851	/* CHECK_CANDIATE FUNCTIONS START */
302 :	edgomez	195
303 :	edgomez	851	static void
304 :			CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
305 :			{
306 :			int xc, yc;
307 :			const uint8_t * Reference;
308 :			VECTOR * current;
309 :			int32_t sad; uint32_t t;
310 :	edgomez	195
311 :	edgomez	851	if ( (x > data->max_dx) \|\| (x < data->min_dx)
312 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
313 :	edgomez	195
314 :	edgomez	851	if (!data->qpel_precision) {
315 :			Reference = GetReference(x, y, data);
316 :			current = data->currentMV;
317 :			xc = x; yc = y;
318 :	edgomez	1053	} else { /* x and y are in 1/4 precision */
319 :	edgomez	851	Reference = Interpolate16x16qpel(x, y, 0, data);
320 :	edgomez	1053	xc = x/2; yc = y/2; /* for chroma sad */
321 :	edgomez	851	current = data->currentQMV;
322 :			}
323 :	edgomez	195
324 :	edgomez	851	sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
325 :			t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
326 :	edgomez	195
327 :	edgomez	851	sad += (data->lambda16 * t * sad)>>10;
328 :			data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
329 :	chl	175
330 :	syskin	1084	if (data->chroma && sad < data->iMinSAD[0])
331 :			sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
332 :			(yc >> 1) + roundtab_79[yc & 0x3], data);
333 :	edgomez	195
334 :	edgomez	851	if (sad < data->iMinSAD[0]) {
335 :			data->iMinSAD[0] = sad;
336 :			current[0].x = x; current[0].y = y;
337 :			*dir = Direction;
338 :			}
339 :	edgomez	195
340 :	edgomez	851	if (data->temp[1] < data->iMinSAD[1]) {
341 :			data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y = y; }
342 :			if (data->temp[2] < data->iMinSAD[2]) {
343 :			data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; }
344 :			if (data->temp[3] < data->iMinSAD[3]) {
345 :			data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; }
346 :			if (data->temp[4] < data->iMinSAD[4]) {
347 :			data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; }
348 :	Isibaar	3	}
349 :
350 :	edgomez	851	static void
351 :			CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
352 :			{
353 :			int32_t sad; uint32_t t;
354 :			const uint8_t * Reference;
355 :	edgomez	974	VECTOR * current;
356 :	chl	326
357 :	edgomez	851	if ( (x > data->max_dx) \|\| (x < data->min_dx)
358 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
359 :	Isibaar	3
360 :	edgomez	974	if (!data->qpel_precision) {
361 :			Reference = GetReference(x, y, data);
362 :			current = data->currentMV;
363 :	edgomez	1053	} else { /* x and y are in 1/4 precision */
364 :	edgomez	974	Reference = Interpolate8x8qpel(x, y, 0, 0, data);
365 :			current = data->currentQMV;
366 :			}
367 :	Isibaar	3
368 :	edgomez	851	sad = sad8(data->Cur, Reference, data->iEdgedWidth);
369 :			t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
370 :	Isibaar	3
371 :	edgomez	851	sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10;
372 :	Isibaar	3
373 :	edgomez	851	if (sad < *(data->iMinSAD)) {
374 :			*(data->iMinSAD) = sad;
375 :	edgomez	974	current->x = x; current->y = y;
376 :	edgomez	851	*dir = Direction;
377 :			}
378 :	Isibaar	3	}
379 :
380 :	edgomez	851	static void
381 :			CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
382 :			{
383 :			uint32_t t;
384 :			const uint8_t * Reference;
385 :
386 :	edgomez	1053	if ( (!(x&1) && x !=0) \|\| (!(y&1) && y !=0) \|\| /* non-zero even value */
387 :	edgomez	851	(x > data->max_dx) \|\| (x < data->min_dx)
388 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
389 :
390 :			Reference = GetReference(x, y, data);
391 :			t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1);
392 :
393 :			data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
394 :
395 :			data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10;
396 :			data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
397 :
398 :			if (data->temp[0] < data->iMinSAD[0]) {
399 :			data->iMinSAD[0] = data->temp[0];
400 :			data->currentMV[0].x = x; data->currentMV[0].y = y;
401 :			*dir = Direction; }
402 :
403 :			if (data->temp[1] < data->iMinSAD[1]) {
404 :			data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
405 :			if (data->temp[2] < data->iMinSAD[2]) {
406 :			data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
407 :			if (data->temp[3] < data->iMinSAD[3]) {
408 :			data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
409 :			if (data->temp[4] < data->iMinSAD[4]) {
410 :			data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
411 :	Isibaar	3	}
412 :	edgomez	195
413 :	edgomez	851	static void
414 :			CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
415 :			{
416 :			int32_t sad, xc, yc;
417 :			const uint8_t * Reference;
418 :			uint32_t t;
419 :			VECTOR * current;
420 :
421 :	edgomez	963	if ( (x > data->max_dx) \|\| ( x < data->min_dx)
422 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
423 :	edgomez	851
424 :	edgomez	1053	if (data->rrv && (!(x&1) && x !=0) \| (!(y&1) && y !=0) ) return; /* non-zero even value */
425 :	edgomez	851
426 :	edgomez	1053	if (data->qpel_precision) { /* x and y are in 1/4 precision */
427 :	edgomez	851	Reference = Interpolate16x16qpel(x, y, 0, data);
428 :			current = data->currentQMV;
429 :			xc = x/2; yc = y/2;
430 :			} else {
431 :			Reference = GetReference(x, y, data);
432 :			current = data->currentMV;
433 :			xc = x; yc = y;
434 :			}
435 :			t = d_mv_bits(x, y, data->predMV, data->iFcode,
436 :			data->qpel^data->qpel_precision, data->rrv);
437 :
438 :			sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096);
439 :			sad += (data->lambda16 * t * sad)>>10;
440 :
441 :	syskin	1084	if (data->chroma && sad < *data->iMinSAD)
442 :			sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
443 :			(yc >> 1) + roundtab_79[yc & 0x3], data);
444 :	edgomez	851
445 :			if (sad < *(data->iMinSAD)) {
446 :			*(data->iMinSAD) = sad;
447 :			current->x = x; current->y = y;
448 :			*dir = Direction;
449 :			}
450 :	chl	96	}
451 :	Isibaar	3
452 :	edgomez	851	static void
453 :	chl	1077	CheckCandidate16I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
454 :			{
455 :			int sad;
456 :			// int xc, yc;
457 :			const uint8_t * Reference;
458 :			// VECTOR * current;
459 :
460 :			if ( (x > data->max_dx) \|\| ( x < data->min_dx)
461 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
462 :
463 :			Reference = GetReference(x, y, data);
464 :			// xc = x; yc = y;
465 :
466 :			sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096);
467 :			// sad += d_mv_bits(x, y, data->predMV, data->iFcode, 0, 0);
468 :
469 :			/* if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
470 :			(yc >> 1) + roundtab_79[yc & 0x3], data);
471 :			*/
472 :
473 :			if (sad < data->iMinSAD[0]) {
474 :			data->iMinSAD[0] = sad;
475 :			data->currentMV[0].x = x; data->currentMV[0].y = y;
476 :			*dir = Direction;
477 :			}
478 :			}
479 :
480 :			static void
481 :	edgomez	851	CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
482 :			{
483 :	edgomez	1053	/* maximum speed - for P/B/I decision */
484 :	edgomez	851	int32_t sad;
485 :
486 :			if ( (x > data->max_dx) \|\| (x < data->min_dx)
487 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
488 :
489 :	syskin	1084	sad = sad32v_c(data->Cur, data->RefP[0] + (x>>1) + (y>>1)*((int)data->iEdgedWidth),
490 :	edgomez	987	data->iEdgedWidth, data->temp+1);
491 :	edgomez	851
492 :			if (sad < *(data->iMinSAD)) {
493 :			*(data->iMinSAD) = sad;
494 :			data->currentMV[0].x = x; data->currentMV[0].y = y;
495 :			*dir = Direction;
496 :			}
497 :			if (data->temp[1] < data->iMinSAD[1]) {
498 :			data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
499 :			if (data->temp[2] < data->iMinSAD[2]) {
500 :			data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
501 :			if (data->temp[3] < data->iMinSAD[3]) {
502 :			data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
503 :			if (data->temp[4] < data->iMinSAD[4]) {
504 :			data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
505 :
506 :	Isibaar	3	}
507 :
508 :	edgomez	851	static void
509 :			CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data)
510 :			{
511 :			int32_t sad, xb, yb, xcf, ycf, xcb, ycb;
512 :			uint32_t t;
513 :			const uint8_t ReferenceF, ReferenceB;
514 :			VECTOR *current;
515 :
516 :	edgomez	953	if ((xf > data->max_dx) \|\| (xf < data->min_dx) \|\|
517 :			(yf > data->max_dy) \|\| (yf < data->min_dy))
518 :			return;
519 :	edgomez	851
520 :			if (!data->qpel_precision) {
521 :			ReferenceF = GetReference(xf, yf, data);
522 :			xb = data->currentMV[1].x; yb = data->currentMV[1].y;
523 :			ReferenceB = GetReferenceB(xb, yb, 1, data);
524 :			current = data->currentMV;
525 :			xcf = xf; ycf = yf;
526 :			xcb = xb; ycb = yb;
527 :			} else {
528 :			ReferenceF = Interpolate16x16qpel(xf, yf, 0, data);
529 :			xb = data->currentQMV[1].x; yb = data->currentQMV[1].y;
530 :			current = data->currentQMV;
531 :			ReferenceB = Interpolate16x16qpel(xb, yb, 1, data);
532 :			xcf = xf/2; ycf = yf/2;
533 :			xcb = xb/2; ycb = yb/2;
534 :			}
535 :
536 :			t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0)
537 :			+ d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0);
538 :
539 :			sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
540 :			sad += (data->lambda16 * t * sad)>>10;
541 :
542 :	syskin	1084	if (data->chroma && sad < *data->iMinSAD)
543 :			sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3],
544 :			(ycf >> 1) + roundtab_79[ycf & 0x3],
545 :			(xcb >> 1) + roundtab_79[xcb & 0x3],
546 :			(ycb >> 1) + roundtab_79[ycb & 0x3], data);
547 :	edgomez	851
548 :			if (sad < *(data->iMinSAD)) {
549 :			*(data->iMinSAD) = sad;
550 :			current->x = xf; current->y = yf;
551 :			*dir = Direction;
552 :			}
553 :	Isibaar	3	}
554 :
555 :	edgomez	851	static void
556 :			CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
557 :			{
558 :			int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0;
559 :			uint32_t k;
560 :			const uint8_t *ReferenceF;
561 :			const uint8_t *ReferenceB;
562 :			VECTOR mvs, b_mvs;
563 :
564 :	edgomez	953	if (( x > 31) \|\| ( x < -32) \|\| ( y > 31) \|\| (y < -32)) return;
565 :	edgomez	851
566 :			for (k = 0; k < 4; k++) {
567 :			mvs.x = data->directmvF[k].x + x;
568 :			b_mvs.x = ((x == 0) ?
569 :			data->directmvB[k].x
570 :			: mvs.x - data->referencemv[k].x);
571 :
572 :			mvs.y = data->directmvF[k].y + y;
573 :			b_mvs.y = ((y == 0) ?
574 :			data->directmvB[k].y
575 :			: mvs.y - data->referencemv[k].y);
576 :
577 :	edgomez	953	if ((mvs.x > data->max_dx) \|\| (mvs.x < data->min_dx) \|\|
578 :			(mvs.y > data->max_dy) \|\| (mvs.y < data->min_dy) \|\|
579 :			(b_mvs.x > data->max_dx) \|\| (b_mvs.x < data->min_dx) \|\|
580 :			(b_mvs.y > data->max_dy) \|\| (b_mvs.y < data->min_dy) )
581 :			return;
582 :	edgomez	851
583 :			if (data->qpel) {
584 :			xcf += mvs.x/2; ycf += mvs.y/2;
585 :			xcb += b_mvs.x/2; ycb += b_mvs.y/2;
586 :			} else {
587 :			xcf += mvs.x; ycf += mvs.y;
588 :			xcb += b_mvs.x; ycb += b_mvs.y;
589 :	edgomez	1053	mvs.x = 2; mvs.y = 2; /* we move to qpel precision anyway */
590 :	edgomez	851	b_mvs.x = 2; b_mvs.y = 2;
591 :			}
592 :
593 :			ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data);
594 :			ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data);
595 :
596 :			sad += sad8bi(data->Cur + 8(k&1) + 8(k>>1)*(data->iEdgedWidth),
597 :			ReferenceF, ReferenceB, data->iEdgedWidth);
598 :			if (sad > *(data->iMinSAD)) return;
599 :			}
600 :
601 :			sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
602 :
603 :	syskin	1084	if (data->chroma && sad < *data->iMinSAD)
604 :			sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
605 :			(ycf >> 3) + roundtab_76[ycf & 0xf],
606 :			(xcb >> 3) + roundtab_76[xcb & 0xf],
607 :			(ycb >> 3) + roundtab_76[ycb & 0xf], data);
608 :	edgomez	851
609 :			if (sad < *(data->iMinSAD)) {
610 :			*(data->iMinSAD) = sad;
611 :			data->currentMV->x = x; data->currentMV->y = y;
612 :			*dir = Direction;
613 :			}
614 :	Isibaar	3	}
615 :
616 :	edgomez	851	static void
617 :			CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
618 :	Isibaar	3	{
619 :	edgomez	851	int32_t sad, xcf, ycf, xcb, ycb;
620 :			const uint8_t *ReferenceF;
621 :			const uint8_t *ReferenceB;
622 :			VECTOR mvs, b_mvs;
623 :	Isibaar	3
624 :	edgomez	963	if (( x > 31) \|\| ( x < -32) \|\| ( y > 31) \|\| (y < -32)) return;
625 :	Isibaar	3
626 :	edgomez	851	mvs.x = data->directmvF[0].x + x;
627 :			b_mvs.x = ((x == 0) ?
628 :			data->directmvB[0].x
629 :			: mvs.x - data->referencemv[0].x);
630 :	Isibaar	3
631 :	edgomez	851	mvs.y = data->directmvF[0].y + y;
632 :			b_mvs.y = ((y == 0) ?
633 :			data->directmvB[0].y
634 :			: mvs.y - data->referencemv[0].y);
635 :	Isibaar	3
636 :	edgomez	963	if ( (mvs.x > data->max_dx) \|\| (mvs.x < data->min_dx)
637 :			\|\| (mvs.y > data->max_dy) \|\| (mvs.y < data->min_dy)
638 :			\|\| (b_mvs.x > data->max_dx) \|\| (b_mvs.x < data->min_dx)
639 :			\|\| (b_mvs.y > data->max_dy) \|\| (b_mvs.y < data->min_dy) ) return;
640 :	Isibaar	3
641 :	edgomez	851	if (data->qpel) {
642 :			xcf = 4(mvs.x/2); ycf = 4(mvs.y/2);
643 :			xcb = 4(b_mvs.x/2); ycb = 4(b_mvs.y/2);
644 :			ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data);
645 :			ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data);
646 :			} else {
647 :			xcf = 4mvs.x; ycf = 4mvs.y;
648 :			xcb = 4b_mvs.x; ycb = 4b_mvs.y;
649 :			ReferenceF = GetReference(mvs.x, mvs.y, data);
650 :			ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data);
651 :			}
652 :
653 :			sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
654 :			sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
655 :
656 :	syskin	1084	if (data->chroma && sad < *data->iMinSAD)
657 :			sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
658 :			(ycf >> 3) + roundtab_76[ycf & 0xf],
659 :			(xcb >> 3) + roundtab_76[xcb & 0xf],
660 :			(ycb >> 3) + roundtab_76[ycb & 0xf], data);
661 :	edgomez	851
662 :			if (sad < *(data->iMinSAD)) {
663 :			*(data->iMinSAD) = sad;
664 :			data->currentMV->x = x; data->currentMV->y = y;
665 :			*dir = Direction;
666 :			}
667 :	Isibaar	3	}
668 :
669 :	edgomez	851
670 :			static void
671 :	syskin	1108	CheckCandidateRD16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
672 :	Isibaar	3	{
673 :
674 :	edgomez	962	int16_t in = data->dctSpace, coeff = data->dctSpace + 64;
675 :	syskin	1108	int32_t rd = 0;
676 :	edgomez	851	VECTOR * current;
677 :			const uint8_t * ptr;
678 :			int i, cbp = 0, t, xc, yc;
679 :	edgomez	195
680 :	edgomez	851	if ( (x > data->max_dx) \|\| (x < data->min_dx)
681 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
682 :	edgomez	195
683 :	edgomez	851	if (!data->qpel_precision) {
684 :			ptr = GetReference(x, y, data);
685 :			current = data->currentMV;
686 :			xc = x; yc = y;
687 :	edgomez	1053	} else { /* x and y are in 1/4 precision */
688 :	edgomez	851	ptr = Interpolate16x16qpel(x, y, 0, data);
689 :			current = data->currentQMV;
690 :			xc = x/2; yc = y/2;
691 :			}
692 :	Isibaar	3
693 :	edgomez	851	for(i = 0; i < 4; i++) {
694 :			int s = 8((i&1) + (i>>1)data->iEdgedWidth);
695 :			transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth);
696 :	syskin	1108	rd += data->temp[i] = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, i);
697 :	edgomez	851	}
698 :	Isibaar	3
699 :	syskin	1108	rd += t = BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
700 :	edgomez	195
701 :	syskin	1071	if (data->temp[0] + t < data->iMinSAD[1]) {
702 :	chl	1115	data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; data->cbp[1] = (data->cbp[1]&~32) \| (cbp&32); }
703 :	syskin	1071	if (data->temp[1] < data->iMinSAD[2]) {
704 :	chl	1115	data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; data->cbp[1] = (data->cbp[1]&~16) \| (cbp&16); }
705 :	syskin	1071	if (data->temp[2] < data->iMinSAD[3]) {
706 :	chl	1115	data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; data->cbp[1] = (data->cbp[1]&~8) \| (cbp&8); }
707 :	syskin	1071	if (data->temp[3] < data->iMinSAD[4]) {
708 :	chl	1115	data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; data->cbp[1] = (data->cbp[1]&~4) \| (cbp&4); }
709 :	syskin	1071
710 :	syskin	1108	rd += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len;
711 :	edgomez	851
712 :	syskin	1108	if (rd >= data->iMinSAD[0]) return;
713 :	edgomez	851
714 :	edgomez	1053	/* chroma */
715 :	edgomez	1022	xc = (xc >> 1) + roundtab_79[xc & 0x3];
716 :			yc = (yc >> 1) + roundtab_79[yc & 0x3];
717 :	edgomez	851
718 :	edgomez	1053	/* chroma U */
719 :	syskin	1084	ptr = interpolate8x8_switch2(data->RefQ, data->RefP[4], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
720 :			transfer_8to16subro(in, data->CurU, ptr, data->iEdgedWidth/2);
721 :	syskin	1108	rd += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 4);
722 :			if (rd >= data->iMinSAD[0]) return;
723 :	edgomez	1022
724 :	edgomez	1053	/* chroma V */
725 :	syskin	1084	ptr = interpolate8x8_switch2(data->RefQ, data->RefP[5], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
726 :			transfer_8to16subro(in, data->CurV, ptr, data->iEdgedWidth/2);
727 :	syskin	1108	rd += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5);
728 :	edgomez	1022
729 :	syskin	1108	rd += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) \| ((cbp & 3) << 3)].len;
730 :	edgomez	851
731 :	syskin	1108	if (rd < data->iMinSAD[0]) {
732 :			data->iMinSAD[0] = rd;
733 :	edgomez	851	current[0].x = x; current[0].y = y;
734 :			*dir = Direction;
735 :	syskin	1084	*data->cbp = cbp;
736 :	edgomez	851	}
737 :	syskin	1071	}
738 :	edgomez	851
739 :			static void
740 :	syskin	1108	CheckCandidateRD8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
741 :	chl	96	{
742 :
743 :	edgomez	962	int16_t in = data->dctSpace, coeff = data->dctSpace + 64;
744 :	syskin	1108	int32_t rd;
745 :	edgomez	851	VECTOR * current;
746 :			const uint8_t * ptr;
747 :	edgomez	1022	int cbp = 0;
748 :	edgomez	195
749 :	edgomez	851	if ( (x > data->max_dx) \|\| (x < data->min_dx)
750 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
751 :	edgomez	195
752 :	edgomez	851	if (!data->qpel_precision) {
753 :			ptr = GetReference(x, y, data);
754 :			current = data->currentMV;
755 :	edgomez	1053	} else { /* x and y are in 1/4 precision */
756 :	edgomez	851	ptr = Interpolate8x8qpel(x, y, 0, 0, data);
757 :			current = data->currentQMV;
758 :			}
759 :	chl	96
760 :	edgomez	851	transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth);
761 :	syskin	1108	rd = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5);
762 :			rd += BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
763 :	chl	96
764 :	syskin	1108	if (rd < data->iMinSAD[0]) {
765 :	syskin	1084	*data->cbp = cbp;
766 :	syskin	1108	data->iMinSAD[0] = rd;
767 :	edgomez	851	current[0].x = x; current[0].y = y;
768 :			*dir = Direction;
769 :			}
770 :			}
771 :	chl	96
772 :	edgomez	851	/* CHECK_CANDIATE FUNCTIONS END */
773 :	edgomez	195
774 :	edgomez	851	/* MAINSEARCH FUNCTIONS START */
775 :	chl	96
776 :	edgomez	851	static void
777 :			AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection)
778 :			{
779 :	edgomez	195
780 :	edgomez	851	/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
781 :	edgomez	195
782 :	edgomez	851	int iDirection;
783 :	edgomez	195
784 :	edgomez	1053	for(;;) { /* forever */
785 :	edgomez	851	iDirection = 0;
786 :			if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
787 :			if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
788 :			if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
789 :			if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
790 :	edgomez	195
791 :	edgomez	851	/* now we're doing diagonal checks near our candidate */
792 :	edgomez	195
793 :	edgomez	1053	if (iDirection) { /* if anything found */
794 :	edgomez	851	bDirection = iDirection;
795 :			iDirection = 0;
796 :			x = data->currentMV->x; y = data->currentMV->y;
797 :	edgomez	1053	if (bDirection & 3) { /* our candidate is left or right */
798 :	edgomez	851	CHECK_CANDIDATE(x, y + iDiamondSize, 8);
799 :			CHECK_CANDIDATE(x, y - iDiamondSize, 4);
800 :	edgomez	1053	} else { /* what remains here is up or down */
801 :	edgomez	851	CHECK_CANDIDATE(x + iDiamondSize, y, 2);
802 :			CHECK_CANDIDATE(x - iDiamondSize, y, 1);
803 :			}
804 :	edgomez	195
805 :	edgomez	851	if (iDirection) {
806 :			bDirection += iDirection;
807 :			x = data->currentMV->x; y = data->currentMV->y;
808 :			}
809 :	edgomez	1053	} else { /* about to quit, eh? not so fast.... */
810 :	edgomez	851	switch (bDirection) {
811 :			case 2:
812 :			CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
813 :			CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
814 :	edgomez	195	break;
815 :	edgomez	851	case 1:
816 :			CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
817 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
818 :	edgomez	195	break;
819 :	edgomez	851	case 2 + 4:
820 :			CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
821 :			CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
822 :			CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
823 :			break;
824 :			case 4:
825 :			CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
826 :			CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
827 :			break;
828 :	edgomez	195	case 8:
829 :	edgomez	851	CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
830 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
831 :	edgomez	195	break;
832 :	edgomez	851	case 1 + 4:
833 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
834 :			CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
835 :			CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
836 :	edgomez	195	break;
837 :	edgomez	851	case 2 + 8:
838 :			CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
839 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
840 :			CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
841 :			break;
842 :			case 1 + 8:
843 :			CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
844 :			CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
845 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
846 :			break;
847 :	edgomez	1053	default: /* 1+2+4+8 == we didn't find anything at all */
848 :	edgomez	851	CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
849 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
850 :			CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
851 :			CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
852 :			break;
853 :	chl	96	}
854 :	edgomez	1053	if (!iDirection) break; /* ok, the end. really */
855 :	edgomez	851	bDirection = iDirection;
856 :			x = data->currentMV->x; y = data->currentMV->y;
857 :	chl	345	}
858 :	edgomez	195	}
859 :	chl	96	}
860 :
861 :	edgomez	851	static void
862 :			SquareSearch(int x, int y, const SearchData * const data, int bDirection)
863 :	chl	96	{
864 :	edgomez	851	int iDirection;
865 :	edgomez	195
866 :	edgomez	851	do {
867 :			iDirection = 0;
868 :			if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64);
869 :			if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128);
870 :			if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32);
871 :			if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128);
872 :			if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64);
873 :			if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128);
874 :			if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128);
875 :			if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128);
876 :	chl	96
877 :	edgomez	851	bDirection = iDirection;
878 :			x = data->currentMV->x; y = data->currentMV->y;
879 :			} while (iDirection);
880 :	chl	96	}
881 :
882 :	edgomez	851	static void
883 :			DiamondSearch(int x, int y, const SearchData * const data, int bDirection)
884 :	chl	181	{
885 :
886 :			/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
887 :
888 :	edgomez	851	int iDirection;
889 :	edgomez	195
890 :	edgomez	851	do {
891 :			iDirection = 0;
892 :			if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
893 :			if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
894 :			if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
895 :			if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
896 :
897 :			/* now we're doing diagonal checks near our candidate */
898 :
899 :	edgomez	1053	if (iDirection) { /* checking if anything found */
900 :	edgomez	851	bDirection = iDirection;
901 :	chl	181	iDirection = 0;
902 :	edgomez	851	x = data->currentMV->x; y = data->currentMV->y;
903 :	edgomez	1053	if (bDirection & 3) { /* our candidate is left or right */
904 :	edgomez	851	CHECK_CANDIDATE(x, y + iDiamondSize, 8);
905 :			CHECK_CANDIDATE(x, y - iDiamondSize, 4);
906 :	edgomez	1053	} else { /* what remains here is up or down */
907 :	edgomez	851	CHECK_CANDIDATE(x + iDiamondSize, y, 2);
908 :			CHECK_CANDIDATE(x - iDiamondSize, y, 1);
909 :			}
910 :			bDirection += iDirection;
911 :			x = data->currentMV->x; y = data->currentMV->y;
912 :			}
913 :			}
914 :			while (iDirection);
915 :			}
916 :	chl	181
917 :	edgomez	851	/* MAINSEARCH FUNCTIONS END */
918 :	chl	181
919 :	edgomez	851	static void
920 :			SubpelRefine(const SearchData * const data)
921 :			{
922 :			/* Do a half-pel or q-pel refinement */
923 :			const VECTOR centerMV = data->qpel_precision ? data->currentQMV : data->currentMV;
924 :	edgomez	1053	int iDirection; /* only needed because macro expects it */
925 :	chl	181
926 :	edgomez	851	CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0);
927 :			CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0);
928 :			CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0);
929 :			CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0);
930 :			CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0);
931 :			CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0);
932 :			CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0);
933 :			CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0);
934 :			}
935 :	chl	181
936 :	edgomez	851	static __inline int
937 :			SkipDecisionP(const IMAGE * current, const IMAGE * reference,
938 :			const int x, const int y,
939 :			const uint32_t stride, const uint32_t iQuant, int rrv)
940 :	chl	181
941 :	edgomez	851	{
942 :	edgomez	963	int offset = (x + ystride)8;
943 :	edgomez	851	if(!rrv) {
944 :	edgomez	963	uint32_t sadC = sad8(current->u + offset,
945 :			reference->u + offset, stride);
946 :	edgomez	851	if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
947 :	edgomez	963	sadC += sad8(current->v + offset,
948 :			reference->v + offset, stride);
949 :	edgomez	851	if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
950 :			return 1;
951 :	chl	181
952 :	edgomez	851	} else {
953 :	edgomez	963	uint32_t sadC = sad16(current->u + 2*offset,
954 :			reference->u + 2offset, stride, 2564096);
955 :	edgomez	851	if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
956 :	edgomez	963	sadC += sad16(current->v + 2*offset,
957 :			reference->v + 2offset, stride, 2564096);
958 :	edgomez	851	if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
959 :			return 1;
960 :	chl	181	}
961 :	edgomez	851	}
962 :	edgomez	504
963 :	edgomez	851	static __inline void
964 :	Isibaar	1095	ZeroMacroblockP(MACROBLOCK *pMB, const int32_t sad)
965 :			{
966 :			pMB->mode = MODE_INTER;
967 :			pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV;
968 :			pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV;
969 :			pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
970 :			}
971 :
972 :			static __inline void
973 :	edgomez	1022	ModeDecision(SearchData * const Data,
974 :			MACROBLOCK * const pMB,
975 :			const MACROBLOCK * const pMBs,
976 :			const int x, const int y,
977 :			const MBParam * const pParam,
978 :			const uint32_t MotionFlags,
979 :			const uint32_t VopFlags,
980 :			const uint32_t VolFlags,
981 :			const IMAGE * const pCurrent,
982 :	syskin	1084	const IMAGE * const pRef,
983 :			const IMAGE * const vGMC,
984 :			const int coding_type)
985 :	edgomez	1022	{
986 :			int mode = MODE_INTER;
987 :	syskin	1084	int mcsel = 0;
988 :	edgomez	1022	int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0);
989 :			const uint32_t iQuant = pMB->quant;
990 :
991 :	syskin	1084	const int skip_possible = (coding_type == P_VOP) && (pMB->dquant == 0);
992 :	edgomez	1022
993 :	syskin	1084	pMB->mcsel = 0;
994 :
995 :	edgomez	1107	if (!(VopFlags & XVID_VOP_MODEDECISION_RD)) { /* normal, fast, SAD-based mode decision */
996 :	edgomez	1022	int sad;
997 :			int InterBias = MV16_INTER_BIAS;
998 :			if (inter4v == 0 \|\| Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] +
999 :			Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) {
1000 :			mode = MODE_INTER;
1001 :			sad = Data->iMinSAD[0];
1002 :			} else {
1003 :			mode = MODE_INTER4V;
1004 :			sad = Data->iMinSAD[1] + Data->iMinSAD[2] +
1005 :			Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant;
1006 :			Data->iMinSAD[0] = sad;
1007 :			}
1008 :
1009 :			/* final skip decision, a.k.a. "the vector you found, really that good?" */
1010 :			if (skip_possible && (pMB->sad16 < (int)iQuant * MAX_SAD00_FOR_SKIP))
1011 :			if ( (100*sad)/(pMB->sad16+1) > FINAL_SKIP_THRESH)
1012 :			if (Data->chroma \|\| SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant, Data->rrv)) {
1013 :			mode = MODE_NOT_CODED;
1014 :			sad = 0;
1015 :			}
1016 :
1017 :	syskin	1084	/* mcsel */
1018 :			if (coding_type == S_VOP) {
1019 :
1020 :			int32_t iSAD = sad16(Data->Cur,
1021 :			vGMC->y + 16yData->iEdgedWidth + 16*x, Data->iEdgedWidth, 65536);
1022 :
1023 :			if (Data->chroma) {
1024 :			iSAD += sad8(Data->CurU, vGMC->u + 8y(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2);
1025 :			iSAD += sad8(Data->CurV, vGMC->v + 8y(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2);
1026 :			}
1027 :
1028 :			if (iSAD <= sad) { /* mode decision GMC */
1029 :			mode = MODE_INTER;
1030 :			mcsel = 1;
1031 :			sad = iSAD;
1032 :			}
1033 :
1034 :			}
1035 :
1036 :	edgomez	1022	/* intra decision */
1037 :
1038 :	edgomez	1053	if (iQuant > 8) InterBias += 100 * (iQuant - 8); /* to make high quants work */
1039 :	edgomez	1022	if (y != 0)
1040 :			if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80;
1041 :			if (x != 0)
1042 :			if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80;
1043 :
1044 :	syskin	1084	if (Data->chroma) InterBias += 50; /* dev8(chroma) ??? <-- yes, we need dev8 (no big difference though) */
1045 :	edgomez	1022	if (Data->rrv) InterBias *= 4;
1046 :
1047 :	syskin	1084	if (InterBias < sad) {
1048 :	edgomez	1022	int32_t deviation;
1049 :	syskin	1084	if (!Data->rrv)
1050 :			deviation = dev16(Data->Cur, Data->iEdgedWidth);
1051 :			else
1052 :			deviation = dev16(Data->Cur, Data->iEdgedWidth) + /* dev32() */
1053 :			dev16(Data->Cur+16, Data->iEdgedWidth) +
1054 :			dev16(Data->Cur + 16*Data->iEdgedWidth, Data->iEdgedWidth) +
1055 :			dev16(Data->Cur+16+16*Data->iEdgedWidth, Data->iEdgedWidth);
1056 :	edgomez	1022
1057 :			if (deviation < (sad - InterBias)) mode = MODE_INTRA;
1058 :			}
1059 :
1060 :	syskin	1084	pMB->cbp = 63;
1061 :			pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
1062 :
1063 :	syskin	1108	} else { /* Rate-Distortion */
1064 :	edgomez	1022
1065 :	syskin	1108	int min_rd, intra_rd, i, cbp, c[2] = {0, 0};
1066 :	edgomez	1022	VECTOR backup[5], *v;
1067 :			Data->iQuant = iQuant;
1068 :	syskin	1084	Data->cbp = c;
1069 :	edgomez	1022
1070 :			v = Data->qpel ? Data->currentQMV : Data->currentMV;
1071 :			for (i = 0; i < 5; i++) {
1072 :			Data->iMinSAD[i] = 256*4096;
1073 :			backup[i] = v[i];
1074 :			}
1075 :
1076 :	syskin	1108	min_rd = findRDinter(Data, pMBs, x, y, pParam, MotionFlags);
1077 :	syskin	1084	cbp = *Data->cbp;
1078 :
1079 :			if (coding_type == S_VOP) {
1080 :	syskin	1108	int gmc_rd;
1081 :			Data->iMinSAD = min_rd += BITS_MULT1; /* mcsel */
1082 :			gmc_rd = findRDgmc(Data, vGMC, x, y);
1083 :			if (gmc_rd < min_rd) {
1084 :	syskin	1084	mcsel = 1;
1085 :	syskin	1108	*Data->iMinSAD = min_rd = gmc_rd;
1086 :	syskin	1084	mode = MODE_INTER;
1087 :			cbp = *Data->cbp;
1088 :	edgomez	1022	}
1089 :	syskin	1084	}
1090 :	edgomez	1022
1091 :	syskin	1084	if (inter4v) {
1092 :	syskin	1108	int v4_rd;
1093 :			v4_rd = findRDinter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup);
1094 :			if (v4_rd < min_rd) {
1095 :			Data->iMinSAD[0] = min_rd = v4_rd;
1096 :	syskin	1084	mode = MODE_INTER4V;
1097 :			cbp = *Data->cbp;
1098 :			}
1099 :			}
1100 :	edgomez	1022
1101 :	syskin	1108	intra_rd = findRDintra(Data);
1102 :			if (intra_rd < min_rd) {
1103 :			*Data->iMinSAD = min_rd = intra_rd;
1104 :	syskin	1084	mode = MODE_INTRA;
1105 :	edgomez	1022	}
1106 :	syskin	1084
1107 :			pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = 0;
1108 :			pMB->cbp = cbp;
1109 :	edgomez	1022	}
1110 :
1111 :			if (Data->rrv) {
1112 :			Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x);
1113 :			Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y);
1114 :			}
1115 :
1116 :	syskin	1084	if (mode == MODE_INTER && mcsel == 0) {
1117 :	edgomez	1022	pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
1118 :
1119 :			if(Data->qpel) {
1120 :			pMB->qmvs[0] = pMB->qmvs[1]
1121 :			= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
1122 :			pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x;
1123 :			pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y;
1124 :			} else {
1125 :			pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
1126 :			pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
1127 :			}
1128 :
1129 :	syskin	1084	} else if (mode == MODE_INTER ) { // but mcsel == 1
1130 :
1131 :			pMB->mcsel = 1;
1132 :			if (Data->qpel) {
1133 :			pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = pMB->amv;
1134 :			pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = pMB->amv.x/2;
1135 :			pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = pMB->amv.y/2;
1136 :			} else
1137 :			pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->amv;
1138 :
1139 :			} else
1140 :			if (mode == MODE_INTER4V) ; /* anything here? */
1141 :			else /* INTRA, NOT_CODED */
1142 :	syskin	1108	ZeroMacroblockP(pMB, 0);
1143 :	edgomez	1022
1144 :			pMB->mode = mode;
1145 :			}
1146 :
1147 :	edgomez	851	bool
1148 :			MotionEstimation(MBParam * const pParam,
1149 :			FRAMEINFO * const current,
1150 :			FRAMEINFO * const reference,
1151 :			const IMAGE * const pRefH,
1152 :			const IMAGE * const pRefV,
1153 :			const IMAGE * const pRefHV,
1154 :	chl	1077	const IMAGE * const pGMC,
1155 :	edgomez	851	const uint32_t iLimit)
1156 :	chl	181	{
1157 :	edgomez	851	MACROBLOCK *const pMBs = current->mbs;
1158 :			const IMAGE *const pCurrent = &current->image;
1159 :			const IMAGE *const pRef = &reference->image;
1160 :	chl	181
1161 :	edgomez	851	uint32_t mb_width = pParam->mb_width;
1162 :			uint32_t mb_height = pParam->mb_height;
1163 :			const uint32_t iEdgedWidth = pParam->edged_width;
1164 :	suxen_drol	890	const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->vop_flags, current->vol_flags);
1165 :	chl	181
1166 :	edgomez	851	uint32_t x, y;
1167 :			uint32_t iIntra = 0;
1168 :	syskin	1108	int32_t sad00;
1169 :	chl	1077	int skip_thresh = INITIAL_SKIP_THRESH * \
1170 :	edgomez	953	(current->vop_flags & XVID_VOP_REDUCED ? 4:1) * \
1171 :	edgomez	1107	(current->vop_flags & XVID_VOP_MODEDECISION_RD ? 2:1);
1172 :	chl	181
1173 :	edgomez	1053	/* some pre-initialized thingies for SearchP */
1174 :	edgomez	851	int32_t temp[8];
1175 :			VECTOR currentMV[5];
1176 :			VECTOR currentQMV[5];
1177 :			int32_t iMinSAD[5];
1178 :	syskin	1075	DECLARE_ALIGNED_MATRIX(dct_space, 3, 64, int16_t, CACHE_LINE);
1179 :	edgomez	851	SearchData Data;
1180 :			memset(&Data, 0, sizeof(SearchData));
1181 :			Data.iEdgedWidth = iEdgedWidth;
1182 :			Data.currentMV = currentMV;
1183 :			Data.currentQMV = currentQMV;
1184 :			Data.iMinSAD = iMinSAD;
1185 :			Data.temp = temp;
1186 :			Data.iFcode = current->fcode;
1187 :			Data.rounding = pParam->m_rounding_type;
1188 :	edgomez	963	Data.qpel = (current->vol_flags & XVID_VOL_QUARTERPEL ? 1:0);
1189 :	edgomez	1107	Data.chroma = MotionFlags & XVID_ME_CHROMA_PVOP;
1190 :	chl	1077	Data.rrv = (current->vop_flags & XVID_VOP_REDUCED) ? 1:0;
1191 :	edgomez	962	Data.dctSpace = dct_space;
1192 :	edgomez	1022	Data.quant_type = !(pParam->vol_flags & XVID_VOL_MPEGQUANT);
1193 :	edgomez	195
1194 :	edgomez	949	if ((current->vop_flags & XVID_VOP_REDUCED)) {
1195 :	edgomez	851	mb_width = (pParam->width + 31) / 32;
1196 :			mb_height = (pParam->height + 31) / 32;
1197 :			Data.qpel = 0;
1198 :			}
1199 :	chl	181
1200 :	edgomez	1053	Data.RefQ = pRefV->u; /* a good place, also used in MC (for similar purpose) */
1201 :	edgomez	851	if (sadInit) (*sadInit) ();
1202 :	chl	181
1203 :	edgomez	851	for (y = 0; y < mb_height; y++) {
1204 :			for (x = 0; x < mb_width; x++) {
1205 :			MACROBLOCK pMB = &pMBs[x + y pParam->mb_width];
1206 :	chl	181
1207 :	edgomez	851	if (!Data.rrv) pMB->sad16 =
1208 :			sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16,
1209 :			pRef->y + (x + y * iEdgedWidth) * 16,
1210 :			pParam->edged_width, pMB->sad8 );
1211 :	chl	181
1212 :	edgomez	851	else pMB->sad16 =
1213 :			sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32,
1214 :			pRef->y + (x + y * iEdgedWidth) * 32,
1215 :			pParam->edged_width, pMB->sad8 );
1216 :	chl	181
1217 :	edgomez	851	if (Data.chroma) {
1218 :			Data.temp[7] = sad8(pCurrent->u + x8 + y(iEdgedWidth/2)*8,
1219 :			pRef->u + x8 + y(iEdgedWidth/2)*8, iEdgedWidth/2)
1220 :			+ sad8(pCurrent->v + (x + y(iEdgedWidth/2))8,
1221 :			pRef->v + (x + y(iEdgedWidth/2))8, iEdgedWidth/2);
1222 :			pMB->sad16 += Data.temp[7];
1223 :			}
1224 :
1225 :			sad00 = pMB->sad16;
1226 :
1227 :	edgomez	1053	/* initial skip decision */
1228 :			/* no early skip for GMC (global vector = skip vector is unknown!) */
1229 :	syskin	1084	if (current->coding_type != S_VOP) { /* no fast SKIP for S(GMC)-VOPs */
1230 :	edgomez	953	if (pMB->dquant == 0 && sad00 < pMB->quant * skip_thresh)
1231 :	edgomez	851	if (Data.chroma \|\| SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) {
1232 :	syskin	1108	ZeroMacroblockP(pMB, sad00);
1233 :			pMB->mode = MODE_NOT_CODED;
1234 :	edgomez	851	continue;
1235 :			}
1236 :			}
1237 :
1238 :	Isibaar	1095	if ((current->vop_flags & XVID_VOP_CARTOON) &&
1239 :			(sad00 < pMB->quant * 4 * skip_thresh)) { /* favorize (0,0) vector for cartoons */
1240 :			ZeroMacroblockP(pMB, sad00);
1241 :			continue;
1242 :			}
1243 :
1244 :	edgomez	851	SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
1245 :	edgomez	1022	y, MotionFlags, current->vop_flags, current->vol_flags,
1246 :			&Data, pParam, pMBs, reference->mbs, pMB);
1247 :	edgomez	851
1248 :	edgomez	1022	ModeDecision(&Data, pMB, pMBs, x, y, pParam,
1249 :			MotionFlags, current->vop_flags, current->vol_flags,
1250 :	syskin	1084	pCurrent, pRef, pGMC, current->coding_type);
1251 :	edgomez	1022
1252 :	edgomez	851	if (pMB->mode == MODE_INTRA)
1253 :			if (++iIntra > iLimit) return 1;
1254 :	chl	181	}
1255 :			}
1256 :	edgomez	851
1257 :			return 0;
1258 :	chl	181	}
1259 :
1260 :
1261 :	edgomez	851	static __inline int
1262 :			make_mask(const VECTOR * const pmv, const int i)
1263 :	chl	96	{
1264 :	edgomez	851	int mask = 255, j;
1265 :			for (j = 0; j < i; j++) {
1266 :	edgomez	1053	if (MVequal(pmv[i], pmv[j])) return 0; /* same vector has been checked already */
1267 :	edgomez	851	if (pmv[i].x == pmv[j].x) {
1268 :			if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4;
1269 :			else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8;
1270 :			} else
1271 :			if (pmv[i].y == pmv[j].y) {
1272 :			if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1;
1273 :			else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2;
1274 :			}
1275 :			}
1276 :			return mask;
1277 :			}
1278 :	edgomez	195
1279 :	edgomez	851	static __inline void
1280 :			PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount,
1281 :			int iHcount, const MACROBLOCK * const prevMB, int rrv)
1282 :			{
1283 :	edgomez	1053	/* this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself */
1284 :	edgomez	851	if (rrv) { iWcount /= 2; iHcount /= 2; }
1285 :	edgomez	195
1286 :	edgomez	1053	if ( (y != 0) && (x < (iWcount-1)) ) { /* [5] top-right neighbour */
1287 :	edgomez	851	pmv[5].x = EVEN(pmv[3].x);
1288 :			pmv[5].y = EVEN(pmv[3].y);
1289 :			} else pmv[5].x = pmv[5].y = 0;
1290 :	chl	96
1291 :	edgomez	1053	if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }/* pmv[3] is left neighbour */
1292 :	edgomez	851	else pmv[3].x = pmv[3].y = 0;
1293 :	chl	96
1294 :	edgomez	1053	if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }/* [4] top neighbour */
1295 :	edgomez	851	else pmv[4].x = pmv[4].y = 0;
1296 :	Isibaar	3
1297 :	edgomez	1053	/* [1] median prediction */
1298 :	edgomez	851	pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y);
1299 :	edgomez	195
1300 :	edgomez	1053	pmv[0].x = pmv[0].y = 0; /* [0] is zero; not used in the loop (checked before) but needed here for make_mask */
1301 :	edgomez	195
1302 :	edgomez	1053	pmv[2].x = EVEN(prevMB->mvs[0].x); /* [2] is last frame */
1303 :	edgomez	851	pmv[2].y = EVEN(prevMB->mvs[0].y);
1304 :
1305 :			if ((x < iWcount-1) && (y < iHcount-1)) {
1306 :	edgomez	1053	pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); /* [6] right-down neighbour in last frame */
1307 :	edgomez	851	pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y);
1308 :			} else pmv[6].x = pmv[6].y = 0;
1309 :
1310 :			if (rrv) {
1311 :			int i;
1312 :			for (i = 0; i < 7; i++) {
1313 :			pmv[i].x = RRV_MV_SCALEUP(pmv[i].x);
1314 :			pmv[i].y = RRV_MV_SCALEUP(pmv[i].y);
1315 :			}
1316 :			}
1317 :	Isibaar	3	}
1318 :
1319 :	edgomez	851	static void
1320 :			SearchP(const IMAGE * const pRef,
1321 :			const uint8_t * const pRefH,
1322 :			const uint8_t * const pRefV,
1323 :			const uint8_t * const pRefHV,
1324 :			const IMAGE * const pCur,
1325 :			const int x,
1326 :			const int y,
1327 :			const uint32_t MotionFlags,
1328 :	suxen_drol	890	const uint32_t VopFlags,
1329 :	edgomez	1022	const uint32_t VolFlags,
1330 :	edgomez	851	SearchData * const Data,
1331 :			const MBParam * const pParam,
1332 :			const MACROBLOCK * const pMBs,
1333 :			const MACROBLOCK * const prevMBs,
1334 :			MACROBLOCK * const pMB)
1335 :			{
1336 :	Isibaar	3
1337 :	edgomez	851	int i, iDirection = 255, mask, threshA;
1338 :			VECTOR pmv[7];
1339 :	edgomez	1022	int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0);
1340 :	Isibaar	3
1341 :	syskin	1129	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
1342 :			pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, Data->rrv);
1343 :	edgomez	195
1344 :	edgomez	851	get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp);
1345 :	Isibaar	3
1346 :	edgomez	1053	Data->temp[5] = Data->temp[6] = 0; /* chroma-sad cache */
1347 :	edgomez	851	i = Data->rrv ? 2 : 1;
1348 :			Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i;
1349 :			Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1350 :			Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1351 :	chl	326
1352 :	edgomez	978	Data->RefP[0] = pRef->y + (x + Data->iEdgedWidthy) 16*i;
1353 :			Data->RefP[2] = pRefH + (x + Data->iEdgedWidthy) 16*i;
1354 :			Data->RefP[1] = pRefV + (x + Data->iEdgedWidthy) 16*i;
1355 :			Data->RefP[3] = pRefHV + (x + Data->iEdgedWidthy) 16*i;
1356 :			Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1357 :			Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1358 :	edgomez	195
1359 :	edgomez	1022	Data->lambda16 = lambda_vec16[pMB->quant];
1360 :			Data->lambda8 = lambda_vec8[pMB->quant];
1361 :	edgomez	851	Data->qpel_precision = 0;
1362 :	edgomez	195
1363 :	edgomez	963	memset(Data->currentMV, 0, 5*sizeof(VECTOR));
1364 :	chl	169
1365 :	edgomez	851	if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0);
1366 :			else Data->predMV = pmv[0];
1367 :	Isibaar	3
1368 :	edgomez	851	i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0);
1369 :			Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10);
1370 :			Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10);
1371 :			Data->iMinSAD[2] = pMB->sad8[1];
1372 :			Data->iMinSAD[3] = pMB->sad8[2];
1373 :			Data->iMinSAD[4] = pMB->sad8[3];
1374 :	chl	169
1375 :	edgomez	1107	if ((!(VopFlags & XVID_VOP_MODEDECISION_RD)) && (x \| y)) {
1376 :	edgomez	1053	threshA = Data->temp[0]; /* that's where we keep this SAD atm */
1377 :	edgomez	851	if (threshA < 512) threshA = 512;
1378 :			else if (threshA > 1024) threshA = 1024;
1379 :			} else
1380 :			threshA = 512;
1381 :	chl	169
1382 :	edgomez	851	PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height,
1383 :			prevMBs + x + y * pParam->mb_width, Data->rrv);
1384 :	chl	169
1385 :	edgomez	851	if (!Data->rrv) {
1386 :			if (inter4v \| Data->chroma) CheckCandidate = CheckCandidate16;
1387 :	edgomez	1053	else CheckCandidate = CheckCandidate16no4v; /* for extra speed */
1388 :	edgomez	851	} else CheckCandidate = CheckCandidate32;
1389 :	chl	169
1390 :	edgomez	851	/* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/
1391 :	chl	169
1392 :	edgomez	851	for (i = 1; i < 7; i++) {
1393 :			if (!(mask = make_mask(pmv, i)) ) continue;
1394 :			CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1395 :			if (Data->iMinSAD[0] <= threshA) break;
1396 :			}
1397 :	chl	169
1398 :	edgomez	851	if ((Data->iMinSAD[0] <= threshA) \|\|
1399 :			(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) &&
1400 :	edgomez	1022	(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16)))
1401 :			inter4v = 0;
1402 :	edgomez	851	else {
1403 :	Isibaar	3
1404 :	edgomez	851	MainSearchFunc * MainSearchPtr;
1405 :	edgomez	949	if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch;
1406 :			else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1407 :	edgomez	851	else MainSearchPtr = DiamondSearch;
1408 :	Isibaar	3
1409 :	edgomez	851	MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1410 :	Isibaar	3
1411 :	edgomez	851	/* extended search, diamond starting in 0,0 and in prediction.
1412 :			note that this search is/might be done in halfpel positions,
1413 :			which makes it more different than the diamond above */
1414 :	chl	169
1415 :	edgomez	949	if (MotionFlags & XVID_ME_EXTSEARCH16) {
1416 :	edgomez	851	int32_t bSAD;
1417 :			VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0];
1418 :			if (Data->rrv) {
1419 :			startMV.x = RRV_MV_SCALEUP(startMV.x);
1420 :			startMV.y = RRV_MV_SCALEUP(startMV.y);
1421 :			}
1422 :			if (!(MVequal(startMV, backupMV))) {
1423 :			bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1424 :	edgomez	195
1425 :	edgomez	851	CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1426 :			MainSearchPtr(startMV.x, startMV.y, Data, 255);
1427 :			if (bSAD < Data->iMinSAD[0]) {
1428 :			Data->currentMV[0] = backupMV;
1429 :			Data->iMinSAD[0] = bSAD; }
1430 :			}
1431 :	chl	169
1432 :	edgomez	851	backupMV = Data->currentMV[0];
1433 :			startMV.x = startMV.y = 1;
1434 :			if (!(MVequal(startMV, backupMV))) {
1435 :			bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1436 :	chl	169
1437 :	edgomez	851	CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1438 :			MainSearchPtr(startMV.x, startMV.y, Data, 255);
1439 :			if (bSAD < Data->iMinSAD[0]) {
1440 :			Data->currentMV[0] = backupMV;
1441 :			Data->iMinSAD[0] = bSAD; }
1442 :	edgomez	195	}
1443 :	edgomez	851	}
1444 :			}
1445 :	edgomez	195
1446 :	edgomez	949	if (MotionFlags & XVID_ME_HALFPELREFINE16)
1447 :	edgomez	851	SubpelRefine(Data);
1448 :	edgomez	195
1449 :	edgomez	851	for(i = 0; i < 5; i++) {
1450 :	edgomez	1053	Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* initialize qpel vectors */
1451 :	edgomez	851	Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
1452 :			}
1453 :	Isibaar	3
1454 :	edgomez	1023	if (Data->qpel) {
1455 :	syskin	1129	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
1456 :			pParam->width, pParam->height, Data->iFcode, 2, 0);
1457 :	edgomez	987	Data->qpel_precision = 1;
1458 :	edgomez	1023	if (MotionFlags & XVID_ME_QUARTERPELREFINE16)
1459 :			SubpelRefine(Data);
1460 :	edgomez	953	}
1461 :	Isibaar	3
1462 :	edgomez	1023	if (Data->iMinSAD[0] < (int32_t)pMB->quant * 30)
1463 :	edgomez	1022	inter4v = 0;
1464 :	Isibaar	3
1465 :	edgomez	987	if (inter4v) {
1466 :	edgomez	851	SearchData Data8;
1467 :	edgomez	1053	memcpy(&Data8, Data, sizeof(SearchData)); /* quick copy of common data */
1468 :	Isibaar	3
1469 :	edgomez	851	Search8(Data, 2x, 2y, MotionFlags, pParam, pMB, pMBs, 0, &Data8);
1470 :			Search8(Data, 2x + 1, 2y, MotionFlags, pParam, pMB, pMBs, 1, &Data8);
1471 :			Search8(Data, 2x, 2y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8);
1472 :			Search8(Data, 2x + 1, 2y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8);
1473 :	chl	181
1474 :	edgomez	1107	if ((Data->chroma) && (!(VopFlags & XVID_VOP_MODEDECISION_RD))) {
1475 :	edgomez	1053	/* chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, it will not be used */
1476 :	edgomez	851	int sumx = 0, sumy = 0;
1477 :	Isibaar	3
1478 :	edgomez	1022	if (Data->qpel)
1479 :			for (i = 1; i < 5; i++) {
1480 :			sumx += Data->currentQMV[i].x/2;
1481 :			sumy += Data->currentQMV[i].y/2;
1482 :			}
1483 :			else
1484 :			for (i = 1; i < 5; i++) {
1485 :			sumx += Data->currentMV[i].x;
1486 :			sumy += Data->currentMV[i].y;
1487 :			}
1488 :	chl	259
1489 :	edgomez	851	Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf],
1490 :			(sumy >> 3) + roundtab_76[sumy & 0xf], Data);
1491 :			}
1492 :	edgomez	1022	} else Data->iMinSAD[1] = 4096*256;
1493 :	edgomez	851	}
1494 :
1495 :			static void
1496 :			Search8(const SearchData * const OldData,
1497 :			const int x, const int y,
1498 :			const uint32_t MotionFlags,
1499 :			const MBParam * const pParam,
1500 :			MACROBLOCK * const pMB,
1501 :			const MACROBLOCK * const pMBs,
1502 :			const int block,
1503 :			SearchData * const Data)
1504 :			{
1505 :			int i = 0;
1506 :			Data->iMinSAD = OldData->iMinSAD + 1 + block;
1507 :			Data->currentMV = OldData->currentMV + 1 + block;
1508 :			Data->currentQMV = OldData->currentQMV + 1 + block;
1509 :
1510 :			if(Data->qpel) {
1511 :			Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1512 :			if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y,
1513 :			Data->predMV, Data->iFcode, 0, 0);
1514 :			} else {
1515 :			Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1516 :			if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y,
1517 :			Data->predMV, Data->iFcode, 0, Data->rrv);
1518 :	Isibaar	3	}
1519 :
1520 :	edgomez	851	(Data->iMinSAD) += (Data->lambda8 i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10;
1521 :	Isibaar	3
1522 :	edgomez	949	if (MotionFlags & (XVID_ME_EXTSEARCH8\|XVID_ME_HALFPELREFINE8\|XVID_ME_QUARTERPELREFINE8)) {
1523 :	Isibaar	3
1524 :	edgomez	978	if (Data->rrv) i = 16; else i = 8;
1525 :	Isibaar	3
1526 :	edgomez	978	Data->RefP[0] = OldData->RefP[0] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1527 :			Data->RefP[1] = OldData->RefP[1] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1528 :			Data->RefP[2] = OldData->RefP[2] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1529 :			Data->RefP[3] = OldData->RefP[3] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1530 :
1531 :			Data->Cur = OldData->Cur + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1532 :	edgomez	851	Data->qpel_precision = 0;
1533 :	Isibaar	3
1534 :	syskin	1129	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 3,
1535 :			pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, Data->rrv);
1536 :	Isibaar	3
1537 :	edgomez	851	if (!Data->rrv) CheckCandidate = CheckCandidate8;
1538 :			else CheckCandidate = CheckCandidate16no4v;
1539 :	Isibaar	3
1540 :	edgomez	1107	if (MotionFlags & XVID_ME_EXTSEARCH8 && (!(MotionFlags & XVID_ME_EXTSEARCH_RD))) {
1541 :	edgomez	1053	int32_t temp_sad = (Data->iMinSAD); / store current MinSAD */
1542 :	Isibaar	3
1543 :	edgomez	851	MainSearchFunc *MainSearchPtr;
1544 :	edgomez	949	if (MotionFlags & XVID_ME_USESQUARES8) MainSearchPtr = SquareSearch;
1545 :			else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch;
1546 :	edgomez	851	else MainSearchPtr = DiamondSearch;
1547 :	Isibaar	3
1548 :	edgomez	851	MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255);
1549 :	Isibaar	3
1550 :	edgomez	851	if(*(Data->iMinSAD) < temp_sad) {
1551 :	edgomez	1053	Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */
1552 :	edgomez	851	Data->currentQMV->y = 2 * Data->currentMV->y;
1553 :			}
1554 :			}
1555 :	edgomez	195
1556 :	edgomez	949	if (MotionFlags & XVID_ME_HALFPELREFINE8) {
1557 :	edgomez	1053	int32_t temp_sad = (Data->iMinSAD); / store current MinSAD */
1558 :	edgomez	195
1559 :	edgomez	1053	SubpelRefine(Data); /* perform halfpel refine of current best vector */
1560 :	Isibaar	3
1561 :	edgomez	1053	if((Data->iMinSAD) < temp_sad) { / we have found a better match */
1562 :			Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */
1563 :	edgomez	851	Data->currentQMV->y = 2 * Data->currentMV->y;
1564 :			}
1565 :			}
1566 :	Isibaar	3
1567 :	edgomez	949	if (Data->qpel && MotionFlags & XVID_ME_QUARTERPELREFINE8) {
1568 :	edgomez	851	Data->qpel_precision = 1;
1569 :	syskin	1129	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 3,
1570 :			pParam->width, pParam->height, Data->iFcode, 2, 0);
1571 :	edgomez	851	SubpelRefine(Data);
1572 :			}
1573 :			}
1574 :	edgomez	195
1575 :	edgomez	851	if (Data->rrv) {
1576 :			Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x);
1577 :			Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y);
1578 :			}
1579 :
1580 :			if(Data->qpel) {
1581 :			pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x;
1582 :			pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y;
1583 :			pMB->qmvs[block] = *Data->currentQMV;
1584 :	edgomez	195	} else {
1585 :	edgomez	851	pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x;
1586 :			pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y;
1587 :	edgomez	78	}
1588 :	edgomez	851
1589 :			pMB->mvs[block] = *Data->currentMV;
1590 :			pMB->sad8[block] = 4 * *Data->iMinSAD;
1591 :	Isibaar	3	}
1592 :
1593 :	edgomez	851	/* motion estimation for B-frames */
1594 :	chl	345
1595 :	edgomez	851	static __inline VECTOR
1596 :			ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode)
1597 :			{
1598 :			/* the stupidiest function ever */
1599 :			return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]);
1600 :			}
1601 :	chl	345
1602 :	edgomez	851	static void __inline
1603 :			PreparePredictionsBF(VECTOR * const pmv, const int x, const int y,
1604 :			const uint32_t iWcount,
1605 :			const MACROBLOCK * const pMB,
1606 :			const uint32_t mode_curr)
1607 :	chl	345	{
1608 :
1609 :	edgomez	1053	/* [0] is prediction */
1610 :	edgomez	851	pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y);
1611 :	chl	345
1612 :	edgomez	1053	pmv[1].x = pmv[1].y = 0; /* [1] is zero */
1613 :	chl	345
1614 :	edgomez	851	pmv[2] = ChoosePred(pMB, mode_curr);
1615 :			pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y);
1616 :	chl	345
1617 :	edgomez	1053	if ((y != 0)&&(x != (int)(iWcount+1))) { /* [3] top-right neighbour */
1618 :	edgomez	851	pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr);
1619 :			pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y);
1620 :			} else pmv[3].x = pmv[3].y = 0;
1621 :	chl	345
1622 :	edgomez	851	if (y != 0) {
1623 :			pmv[4] = ChoosePred(pMB-iWcount, mode_curr);
1624 :			pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y);
1625 :			} else pmv[4].x = pmv[4].y = 0;
1626 :	chl	345
1627 :	edgomez	851	if (x != 0) {
1628 :			pmv[5] = ChoosePred(pMB-1, mode_curr);
1629 :			pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y);
1630 :			} else pmv[5].x = pmv[5].y = 0;
1631 :	chl	345
1632 :	edgomez	851	if (x != 0 && y != 0) {
1633 :			pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr);
1634 :			pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y);
1635 :			} else pmv[6].x = pmv[6].y = 0;
1636 :			}
1637 :	chl	345
1638 :
1639 :	edgomez	851	/* search backward or forward */
1640 :			static void
1641 :			SearchBF( const IMAGE * const pRef,
1642 :			const uint8_t * const pRefH,
1643 :			const uint8_t * const pRefV,
1644 :			const uint8_t * const pRefHV,
1645 :			const IMAGE * const pCur,
1646 :			const int x, const int y,
1647 :			const uint32_t MotionFlags,
1648 :			const uint32_t iFcode,
1649 :			const MBParam * const pParam,
1650 :			MACROBLOCK * const pMB,
1651 :			const VECTOR * const predMV,
1652 :			int32_t * const best_sad,
1653 :			const int32_t mode_current,
1654 :			SearchData * const Data)
1655 :			{
1656 :	chl	345
1657 :	edgomez	851	int i, iDirection = 255, mask;
1658 :			VECTOR pmv[7];
1659 :			MainSearchFunc *MainSearchPtr;
1660 :			*Data->iMinSAD = MV_MAX_ERROR;
1661 :			Data->iFcode = iFcode;
1662 :			Data->qpel_precision = 0;
1663 :	edgomez	1053	Data->temp[5] = Data->temp[6] = Data->temp[7] = 2564096; / reset chroma-sad cache */
1664 :	chl	345
1665 :	edgomez	978	Data->RefP[0] = pRef->y + (x + Data->iEdgedWidthy) 16;
1666 :			Data->RefP[2] = pRefH + (x + Data->iEdgedWidthy) 16;
1667 :			Data->RefP[1] = pRefV + (x + Data->iEdgedWidthy) 16;
1668 :			Data->RefP[3] = pRefHV + (x + Data->iEdgedWidthy) 16;
1669 :			Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8;
1670 :			Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8;
1671 :	chl	345
1672 :	edgomez	851	Data->predMV = *predMV;
1673 :	chl	345
1674 :	syskin	1129	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
1675 :			pParam->width, pParam->height, iFcode - Data->qpel, 1, 0);
1676 :	chl	345
1677 :	edgomez	851	pmv[0] = Data->predMV;
1678 :			if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; }
1679 :	chl	345
1680 :	edgomez	851	PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current);
1681 :	chl	345
1682 :	edgomez	851	Data->currentMV->x = Data->currentMV->y = 0;
1683 :			CheckCandidate = CheckCandidate16no4v;
1684 :	chl	345
1685 :	edgomez	1053	/* main loop. checking all predictions */
1686 :	edgomez	851	for (i = 0; i < 7; i++) {
1687 :			if (!(mask = make_mask(pmv, i)) ) continue;
1688 :			CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1689 :	chl	345	}
1690 :
1691 :	edgomez	949	if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch;
1692 :			else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1693 :	edgomez	851	else MainSearchPtr = DiamondSearch;
1694 :	chl	345
1695 :	edgomez	851	MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1696 :	chl	345
1697 :	edgomez	851	SubpelRefine(Data);
1698 :	chl	345
1699 :	edgomez	851	if (Data->qpel && Data->iMinSAD < best_sad + 300) {
1700 :			Data->currentQMV->x = 2*Data->currentMV->x;
1701 :			Data->currentQMV->y = 2*Data->currentMV->y;
1702 :			Data->qpel_precision = 1;
1703 :	syskin	1129	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
1704 :			pParam->width, pParam->height, iFcode, 2, 0);
1705 :	edgomez	851	SubpelRefine(Data);
1706 :			}
1707 :	chl	345
1708 :	edgomez	1053	/* three bits are needed to code backward mode. four for forward */
1709 :	Isibaar	3
1710 :	edgomez	851	if (mode_current == MODE_FORWARD) Data->iMinSAD += 4 Data->lambda16;
1711 :			else Data->iMinSAD += 3 Data->lambda16;
1712 :	edgomez	195
1713 :	edgomez	851	if (Data->iMinSAD < best_sad) {
1714 :			best_sad = Data->iMinSAD;
1715 :			pMB->mode = mode_current;
1716 :			if (Data->qpel) {
1717 :			pMB->pmvs[0].x = Data->currentQMV->x - predMV->x;
1718 :			pMB->pmvs[0].y = Data->currentQMV->y - predMV->y;
1719 :			if (mode_current == MODE_FORWARD)
1720 :			pMB->qmvs[0] = *Data->currentQMV;
1721 :			else
1722 :			pMB->b_qmvs[0] = *Data->currentQMV;
1723 :			} else {
1724 :			pMB->pmvs[0].x = Data->currentMV->x - predMV->x;
1725 :			pMB->pmvs[0].y = Data->currentMV->y - predMV->y;
1726 :			}
1727 :			if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV;
1728 :			else pMB->b_mvs[0] = *Data->currentMV;
1729 :			}
1730 :	edgomez	195
1731 :	edgomez	851	if (mode_current == MODE_FORWARD) (Data->currentMV+2) = Data->currentMV;
1732 :	edgomez	1053	else (Data->currentMV+1) = Data->currentMV; /* we store currmv for interpolate search */
1733 :	Isibaar	3	}
1734 :
1735 :	edgomez	851	static void
1736 :			SkipDecisionB(const IMAGE * const pCur,
1737 :			const IMAGE * const f_Ref,
1738 :			const IMAGE * const b_Ref,
1739 :			MACROBLOCK * const pMB,
1740 :			const uint32_t x, const uint32_t y,
1741 :			const SearchData * const Data)
1742 :			{
1743 :			int dx = 0, dy = 0, b_dx = 0, b_dy = 0;
1744 :			int32_t sum;
1745 :			const int div = 1 + Data->qpel;
1746 :			int k;
1747 :			const uint32_t stride = Data->iEdgedWidth/2;
1748 :	edgomez	1053	/* this is not full chroma compensation, only it's fullpel approximation. should work though */
1749 :	Isibaar	3
1750 :	edgomez	851	for (k = 0; k < 4; k++) {
1751 :			dy += Data->directmvF[k].y / div;
1752 :	edgomez	963	dx += Data->directmvF[k].x / div;
1753 :			b_dy += Data->directmvB[k].y / div;
1754 :			b_dx += Data->directmvB[k].x / div;
1755 :	edgomez	851	}
1756 :	Isibaar	3
1757 :	edgomez	851	dy = (dy >> 3) + roundtab_76[dy & 0xf];
1758 :			dx = (dx >> 3) + roundtab_76[dx & 0xf];
1759 :			b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf];
1760 :			b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf];
1761 :	Isibaar	3
1762 :	edgomez	851	sum = sad8bi(pCur->u + 8 * x + 8 * y * stride,
1763 :			f_Ref->u + (y8 + dy/2) stride + x*8 + dx/2,
1764 :			b_Ref->u + (y8 + b_dy/2) stride + x*8 + b_dx/2,
1765 :			stride);
1766 :	Isibaar	3
1767 :	syskin	1112	if (sum >= MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; /* no skip */
1768 :	Isibaar	3
1769 :	edgomez	851	sum += sad8bi(pCur->v + 8x + 8 y * stride,
1770 :			f_Ref->v + (y8 + dy/2) stride + x*8 + dx/2,
1771 :			b_Ref->v + (y8 + b_dy/2) stride + x*8 + b_dx/2,
1772 :			stride);
1773 :	edgomez	195
1774 :	syskin	1112	if (sum < MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) {
1775 :	edgomez	1053	pMB->mode = MODE_DIRECT_NONE_MV; /* skipped */
1776 :	edgomez	978	for (k = 0; k < 4; k++) {
1777 :	edgomez	963	pMB->qmvs[k] = pMB->mvs[k];
1778 :			pMB->b_qmvs[k] = pMB->b_mvs[k];
1779 :			}
1780 :			}
1781 :	edgomez	851	}
1782 :	Isibaar	3
1783 :	edgomez	851	static __inline uint32_t
1784 :			SearchDirect(const IMAGE * const f_Ref,
1785 :			const uint8_t * const f_RefH,
1786 :			const uint8_t * const f_RefV,
1787 :			const uint8_t * const f_RefHV,
1788 :			const IMAGE * const b_Ref,
1789 :			const uint8_t * const b_RefH,
1790 :			const uint8_t * const b_RefV,
1791 :			const uint8_t * const b_RefHV,
1792 :			const IMAGE * const pCur,
1793 :			const int x, const int y,
1794 :			const uint32_t MotionFlags,
1795 :			const int32_t TRB, const int32_t TRD,
1796 :			const MBParam * const pParam,
1797 :			MACROBLOCK * const pMB,
1798 :			const MACROBLOCK * const b_mb,
1799 :			int32_t * const best_sad,
1800 :			SearchData * const Data)
1801 :	Isibaar	3
1802 :	edgomez	851	{
1803 :			int32_t skip_sad;
1804 :			int k = (x + Data->iEdgedWidthy) 16;
1805 :			MainSearchFunc *MainSearchPtr;
1806 :	Isibaar	3
1807 :	edgomez	851	Data->iMinSAD = 2564096;
1808 :	edgomez	978	Data->RefP[0] = f_Ref->y + k;
1809 :			Data->RefP[2] = f_RefH + k;
1810 :			Data->RefP[1] = f_RefV + k;
1811 :			Data->RefP[3] = f_RefHV + k;
1812 :			Data->b_RefP[0] = b_Ref->y + k;
1813 :			Data->b_RefP[2] = b_RefH + k;
1814 :			Data->b_RefP[1] = b_RefV + k;
1815 :			Data->b_RefP[3] = b_RefHV + k;
1816 :			Data->RefP[4] = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1817 :			Data->RefP[5] = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1818 :			Data->b_RefP[4] = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1819 :			Data->b_RefP[5] = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1820 :	edgomez	195
1821 :	edgomez	851	k = Data->qpel ? 4 : 2;
1822 :			Data->max_dx = k * (pParam->width - x * 16);
1823 :			Data->max_dy = k * (pParam->height - y * 16);
1824 :			Data->min_dx = -k * (16 + x * 16);
1825 :			Data->min_dy = -k * (16 + y * 16);
1826 :	chl	181
1827 :	edgomez	851	Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs;
1828 :			Data->qpel_precision = 0;
1829 :	edgomez	170
1830 :	edgomez	851	for (k = 0; k < 4; k++) {
1831 :			pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD);
1832 :			pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD;
1833 :			pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD);
1834 :			pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD;
1835 :	Isibaar	3
1836 :	edgomez	851	if ( (pMB->b_mvs[k].x > Data->max_dx) \| (pMB->b_mvs[k].x < Data->min_dx)
1837 :			\| (pMB->b_mvs[k].y > Data->max_dy) \| (pMB->b_mvs[k].y < Data->min_dy) ) {
1838 :
1839 :	edgomez	1053	best_sad = 2564096; /* in that case, we won't use direct mode */
1840 :			pMB->mode = MODE_DIRECT; /* just to make sure it doesn't say "MODE_DIRECT_NONE_MV" */
1841 :	edgomez	851	pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0;
1842 :			return 256*4096;
1843 :			}
1844 :			if (b_mb->mode != MODE_INTER4V) {
1845 :			pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0];
1846 :			pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0];
1847 :			Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0];
1848 :			Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0];
1849 :			break;
1850 :			}
1851 :	edgomez	195	}
1852 :	Isibaar	3
1853 :	edgomez	851	CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v;
1854 :	Isibaar	3
1855 :	edgomez	851	CheckCandidate(0, 0, 255, &k, Data);
1856 :	edgomez	195
1857 :	edgomez	1053	/* initial (fast) skip decision */
1858 :	edgomez	1023	if (Data->iMinSAD < pMB->quant INITIAL_SKIP_THRESH * (Data->chroma?3:2)) {
1859 :	edgomez	1053	/* possible skip */
1860 :	edgomez	851	if (Data->chroma) {
1861 :			pMB->mode = MODE_DIRECT_NONE_MV;
1862 :	edgomez	1053	return Data->iMinSAD; / skip. */
1863 :	edgomez	851	} else {
1864 :			SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data);
1865 :	edgomez	1053	if (pMB->mode == MODE_DIRECT_NONE_MV) return Data->iMinSAD; / skip. */
1866 :	edgomez	851	}
1867 :	Isibaar	3	}
1868 :
1869 :	edgomez	953	*Data->iMinSAD += Data->lambda16;
1870 :	edgomez	851	skip_sad = *Data->iMinSAD;
1871 :	syskin	1084
1872 :	edgomez	1053	/*
1873 :			* DIRECT MODE DELTA VECTOR SEARCH.
1874 :			* This has to be made more effective, but at the moment I'm happy it's running at all
1875 :			*/
1876 :	edgomez	195
1877 :	edgomez	949	if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch;
1878 :			else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1879 :	edgomez	851	else MainSearchPtr = DiamondSearch;
1880 :	Isibaar	3
1881 :	edgomez	851	MainSearchPtr(0, 0, Data, 255);
1882 :	Isibaar	3
1883 :	edgomez	851	SubpelRefine(Data);
1884 :	chl	181
1885 :	edgomez	851	best_sad = Data->iMinSAD;
1886 :	chl	181
1887 :	edgomez	851	if (Data->qpel \|\| b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT;
1888 :	edgomez	1053	else pMB->mode = MODE_DIRECT_NO4V; /* for faster compensation */
1889 :	chl	181
1890 :	edgomez	851	pMB->pmvs[3] = *Data->currentMV;
1891 :	edgomez	195
1892 :	edgomez	851	for (k = 0; k < 4; k++) {
1893 :			pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x;
1894 :			pMB->b_mvs[k].x = ( (Data->currentMV->x == 0)
1895 :			? Data->directmvB[k].x
1896 :			:pMB->mvs[k].x - Data->referencemv[k].x);
1897 :			pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y);
1898 :			pMB->b_mvs[k].y = ((Data->currentMV->y == 0)
1899 :			? Data->directmvB[k].y
1900 :			: pMB->mvs[k].y - Data->referencemv[k].y);
1901 :			if (Data->qpel) {
1902 :			pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2;
1903 :			pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2;
1904 :			pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2;
1905 :			pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2;
1906 :			}
1907 :	edgomez	195
1908 :	edgomez	851	if (b_mb->mode != MODE_INTER4V) {
1909 :			pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0];
1910 :			pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0];
1911 :			pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0];
1912 :			pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0];
1913 :			break;
1914 :			}
1915 :	edgomez	78	}
1916 :	edgomez	851	return skip_sad;
1917 :			}
1918 :	Isibaar	3
1919 :	edgomez	851	static void
1920 :			SearchInterpolate(const IMAGE * const f_Ref,
1921 :			const uint8_t * const f_RefH,
1922 :			const uint8_t * const f_RefV,
1923 :			const uint8_t * const f_RefHV,
1924 :			const IMAGE * const b_Ref,
1925 :			const uint8_t * const b_RefH,
1926 :			const uint8_t * const b_RefV,
1927 :			const uint8_t * const b_RefHV,
1928 :			const IMAGE * const pCur,
1929 :			const int x, const int y,
1930 :			const uint32_t fcode,
1931 :			const uint32_t bcode,
1932 :			const uint32_t MotionFlags,
1933 :			const MBParam * const pParam,
1934 :			const VECTOR * const f_predMV,
1935 :			const VECTOR * const b_predMV,
1936 :			MACROBLOCK * const pMB,
1937 :			int32_t * const best_sad,
1938 :			SearchData * const fData)
1939 :	chl	96
1940 :	edgomez	851	{
1941 :	chl	169
1942 :	edgomez	851	int iDirection, i, j;
1943 :			SearchData bData;
1944 :	chl	169
1945 :	edgomez	851	fData->qpel_precision = 0;
1946 :	edgomez	1053	memcpy(&bData, fData, sizeof(SearchData)); /* quick copy of common data */
1947 :	edgomez	851	fData->iMinSAD = 4096256;
1948 :			bData.currentMV++; bData.currentQMV++;
1949 :			fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode;
1950 :	chl	169
1951 :	edgomez	851	i = (x + y * fData->iEdgedWidth) * 16;
1952 :	chl	169
1953 :	edgomez	978	bData.b_RefP[0] = fData->RefP[0] = f_Ref->y + i;
1954 :			bData.b_RefP[2] = fData->RefP[2] = f_RefH + i;
1955 :			bData.b_RefP[1] = fData->RefP[1] = f_RefV + i;
1956 :			bData.b_RefP[3] = fData->RefP[3] = f_RefHV + i;
1957 :			bData.RefP[0] = fData->b_RefP[0] = b_Ref->y + i;
1958 :			bData.RefP[2] = fData->b_RefP[2] = b_RefH + i;
1959 :			bData.RefP[1] = fData->b_RefP[1] = b_RefV + i;
1960 :			bData.RefP[3] = fData->b_RefP[3] = b_RefHV + i;
1961 :			bData.b_RefP[4] = fData->RefP[4] = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1962 :			bData.b_RefP[5] = fData->RefP[5] = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1963 :			bData.RefP[4] = fData->b_RefP[4] = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1964 :			bData.RefP[5] = fData->b_RefP[5] = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1965 :	chl	169
1966 :	edgomez	851	bData.bpredMV = fData->predMV = *f_predMV;
1967 :			fData->bpredMV = bData.predMV = *b_predMV;
1968 :			fData->currentMV[0] = fData->currentMV[2];
1969 :	Isibaar	3
1970 :	syskin	1129	get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 4, pParam->width, pParam->height, fcode - fData->qpel, 1, 0);
1971 :			get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 4, pParam->width, pParam->height, bcode - fData->qpel, 1, 0);
1972 :	Isibaar	3
1973 :	edgomez	851	if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx;
1974 :			if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx;
1975 :			if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy;
1976 :			if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy;
1977 :	chl	169
1978 :	edgomez	851	if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx;
1979 :			if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx;
1980 :			if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy;
1981 :			if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy;
1982 :	Isibaar	3
1983 :	edgomez	851	CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData);
1984 :	chl	169
1985 :	edgomez	1053	/* diamond */
1986 :	edgomez	851	do {
1987 :			iDirection = 255;
1988 :	edgomez	1053	/* forward MV moves */
1989 :	edgomez	851	i = fData->currentMV[0].x; j = fData->currentMV[0].y;
1990 :	chl	169
1991 :	edgomez	851	CheckCandidateInt(i + 1, j, 0, &iDirection, fData);
1992 :			CheckCandidateInt(i, j + 1, 0, &iDirection, fData);
1993 :			CheckCandidateInt(i - 1, j, 0, &iDirection, fData);
1994 :			CheckCandidateInt(i, j - 1, 0, &iDirection, fData);
1995 :	edgomez	195
1996 :	edgomez	1053	/* backward MV moves */
1997 :	edgomez	851	i = fData->currentMV[1].x; j = fData->currentMV[1].y;
1998 :			fData->currentMV[2] = fData->currentMV[0];
1999 :			CheckCandidateInt(i + 1, j, 0, &iDirection, &bData);
2000 :			CheckCandidateInt(i, j + 1, 0, &iDirection, &bData);
2001 :			CheckCandidateInt(i - 1, j, 0, &iDirection, &bData);
2002 :			CheckCandidateInt(i, j - 1, 0, &iDirection, &bData);
2003 :	Isibaar	3
2004 :	edgomez	851	} while (!(iDirection));
2005 :	chl	140
2006 :	edgomez	1053	/* qpel refinement */
2007 :	edgomez	851	if (fData->qpel) {
2008 :			if (fData->iMinSAD > best_sad + 500) return;
2009 :			CheckCandidate = CheckCandidateInt;
2010 :			fData->qpel_precision = bData.qpel_precision = 1;
2011 :	syskin	1129	get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 4, pParam->width, pParam->height, fcode, 2, 0);
2012 :			get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 4, pParam->width, pParam->height, bcode, 2, 0);
2013 :	edgomez	851	fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x;
2014 :			fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y;
2015 :			fData->currentQMV[1].x = 2 * fData->currentMV[1].x;
2016 :			fData->currentQMV[1].y = 2 * fData->currentMV[1].y;
2017 :			SubpelRefine(fData);
2018 :			if (fData->iMinSAD > best_sad + 300) return;
2019 :			fData->currentQMV[2] = fData->currentQMV[0];
2020 :			SubpelRefine(&bData);
2021 :			}
2022 :	chl	140
2023 :	edgomez	1053	fData->iMinSAD += (2+3) fData->lambda16; /* two bits are needed to code interpolate mode. */
2024 :	Isibaar	3
2025 :	edgomez	851	if (fData->iMinSAD < best_sad) {
2026 :			best_sad = fData->iMinSAD;
2027 :			pMB->mvs[0] = fData->currentMV[0];
2028 :			pMB->b_mvs[0] = fData->currentMV[1];
2029 :			pMB->mode = MODE_INTERPOLATE;
2030 :			if (fData->qpel) {
2031 :			pMB->qmvs[0] = fData->currentQMV[0];
2032 :			pMB->b_qmvs[0] = fData->currentQMV[1];
2033 :			pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x;
2034 :			pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y;
2035 :			pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x;
2036 :			pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y;
2037 :			} else {
2038 :			pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x;
2039 :			pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y;
2040 :			pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x;
2041 :			pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y;
2042 :			}
2043 :	edgomez	78	}
2044 :	edgomez	851	}
2045 :	Isibaar	3
2046 :	edgomez	851	void
2047 :			MotionEstimationBVOP(MBParam * const pParam,
2048 :			FRAMEINFO * const frame,
2049 :			const int32_t time_bp,
2050 :			const int32_t time_pp,
2051 :	edgomez	1053	/* forward (past) reference */
2052 :	edgomez	851	const MACROBLOCK * const f_mbs,
2053 :			const IMAGE * const f_ref,
2054 :			const IMAGE * const f_refH,
2055 :			const IMAGE * const f_refV,
2056 :			const IMAGE * const f_refHV,
2057 :	edgomez	1053	/* backward (future) reference */
2058 :	edgomez	851	const FRAMEINFO * const b_reference,
2059 :			const IMAGE * const b_ref,
2060 :			const IMAGE * const b_refH,
2061 :			const IMAGE * const b_refV,
2062 :			const IMAGE * const b_refHV)
2063 :			{
2064 :			uint32_t i, j;
2065 :			int32_t best_sad;
2066 :			uint32_t skip_sad;
2067 :			int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0;
2068 :			const MACROBLOCK * const b_mbs = b_reference->mbs;
2069 :	Isibaar	3
2070 :	edgomez	851	VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/
2071 :	Isibaar	3
2072 :	edgomez	851	const int32_t TRB = time_pp - time_bp;
2073 :			const int32_t TRD = time_pp;
2074 :	edgomez	195
2075 :	edgomez	1053	/* some pre-inintialized data for the rest of the search */
2076 :	Isibaar	3
2077 :	edgomez	851	SearchData Data;
2078 :			int32_t iMinSAD;
2079 :			VECTOR currentMV[3];
2080 :			VECTOR currentQMV[3];
2081 :			int32_t temp[8];
2082 :			memset(&Data, 0, sizeof(SearchData));
2083 :			Data.iEdgedWidth = pParam->edged_width;
2084 :			Data.currentMV = currentMV; Data.currentQMV = currentQMV;
2085 :			Data.iMinSAD = &iMinSAD;
2086 :			Data.lambda16 = lambda_vec16[frame->quant];
2087 :	syskin	1112	Data.qpel = pParam->vol_flags & XVID_VOL_QUARTERPEL ? 1 : 0;
2088 :	edgomez	851	Data.rounding = 0;
2089 :	edgomez	1107	Data.chroma = frame->motion_flags & XVID_ME_CHROMA_BVOP;
2090 :	edgomez	851	Data.temp = temp;
2091 :	Isibaar	3
2092 :	edgomez	1053	Data.RefQ = f_refV->u; /* a good place, also used in MC (for similar purpose) */
2093 :
2094 :			/* note: i==horizontal, j==vertical */
2095 :	edgomez	851	for (j = 0; j < pParam->mb_height; j++) {
2096 :	edgomez	195
2097 :	edgomez	851	f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */
2098 :	Isibaar	3
2099 :	edgomez	851	for (i = 0; i < pParam->mb_width; i++) {
2100 :			MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width;
2101 :			const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width;
2102 :	edgomez	195
2103 :	edgomez	851	/* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */
2104 :			if (b_reference->coding_type != S_VOP)
2105 :			if (b_mb->mode == MODE_NOT_CODED) {
2106 :			pMB->mode = MODE_NOT_CODED;
2107 :			continue;
2108 :			}
2109 :	Isibaar	3
2110 :	edgomez	851	Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16;
2111 :			Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8;
2112 :			Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8;
2113 :	edgomez	953	pMB->quant = frame->quant;
2114 :	Isibaar	3
2115 :	edgomez	851	/* direct search comes first, because it (1) checks for SKIP-mode
2116 :			and (2) sets very good predictions for forward and backward search */
2117 :			skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2118 :			b_ref, b_refH->y, b_refV->y, b_refHV->y,
2119 :			&frame->image,
2120 :			i, j,
2121 :			frame->motion_flags,
2122 :			TRB, TRD,
2123 :			pParam,
2124 :			pMB, b_mb,
2125 :			&best_sad,
2126 :			&Data);
2127 :	Isibaar	3
2128 :	edgomez	851	if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; }
2129 :	edgomez	195
2130 :	edgomez	1053	/* forward search */
2131 :	edgomez	851	SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2132 :			&frame->image, i, j,
2133 :			frame->motion_flags,
2134 :			frame->fcode, pParam,
2135 :			pMB, &f_predMV, &best_sad,
2136 :			MODE_FORWARD, &Data);
2137 :	edgomez	195
2138 :	edgomez	1053	/* backward search */
2139 :	edgomez	851	SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y,
2140 :			&frame->image, i, j,
2141 :			frame->motion_flags,
2142 :			frame->bcode, pParam,
2143 :			pMB, &b_predMV, &best_sad,
2144 :			MODE_BACKWARD, &Data);
2145 :
2146 :	edgomez	1053	/* interpolate search comes last, because it uses data from forward and backward as prediction */
2147 :	edgomez	851	SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2148 :			b_ref, b_refH->y, b_refV->y, b_refHV->y,
2149 :			&frame->image,
2150 :			i, j,
2151 :			frame->fcode, frame->bcode,
2152 :			frame->motion_flags,
2153 :			pParam,
2154 :			&f_predMV, &b_predMV,
2155 :			pMB, &best_sad,
2156 :			&Data);
2157 :
2158 :	edgomez	1053	/* final skip decision */
2159 :	edgomez	851	if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2)
2160 :			&& ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) )
2161 :			SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data);
2162 :
2163 :			switch (pMB->mode) {
2164 :			case MODE_FORWARD:
2165 :			f_count++;
2166 :			f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2167 :			break;
2168 :			case MODE_BACKWARD:
2169 :			b_count++;
2170 :			b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2171 :			break;
2172 :			case MODE_INTERPOLATE:
2173 :			i_count++;
2174 :			f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2175 :			b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2176 :			break;
2177 :			case MODE_DIRECT:
2178 :			case MODE_DIRECT_NO4V:
2179 :			d_count++;
2180 :			default:
2181 :			break;
2182 :			}
2183 :			}
2184 :			}
2185 :	Isibaar	3	}
2186 :	chl	96
2187 :	edgomez	851	static __inline void
2188 :			MEanalyzeMB ( const uint8_t * const pRef,
2189 :			const uint8_t * const pCur,
2190 :			const int x,
2191 :			const int y,
2192 :			const MBParam * const pParam,
2193 :			MACROBLOCK * const pMBs,
2194 :			SearchData * const Data)
2195 :	chl	96	{
2196 :
2197 :	edgomez	851	int i, mask;
2198 :	edgomez	953	int quarterpel = (pParam->vol_flags & XVID_VOL_QUARTERPEL)? 1: 0;
2199 :	edgomez	851	VECTOR pmv[3];
2200 :	edgomez	953	MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width];
2201 :	chl	96
2202 :	edgomez	851	for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR;
2203 :	chl	96
2204 :	edgomez	1053	/* median is only used as prediction. it doesn't have to be real */
2205 :	edgomez	851	if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0;
2206 :			else
2207 :	edgomez	1053	if (x == 1) /* left macroblock does not have any vector now */
2208 :			Data->predMV = (pMB - pParam->mb_width)->mvs[0]; /* top instead of median */
2209 :			else if (y == 1) /* top macroblock doesn't have it's vector */
2210 :			Data->predMV = (pMB - 1)->mvs[0]; /* left instead of median */
2211 :			else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); /* else median */
2212 :	edgomez	195
2213 :	syskin	1129	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
2214 :			pParam->width, pParam->height, Data->iFcode - quarterpel, 1, 0);
2215 :	edgomez	195
2216 :	edgomez	851	Data->Cur = pCur + (x + y * pParam->edged_width) * 16;
2217 :	edgomez	978	Data->RefP[0] = pRef + (x + y * pParam->edged_width) * 16;
2218 :	chl	96
2219 :	edgomez	851	pmv[1].x = EVEN(pMB->mvs[0].x);
2220 :			pmv[1].y = EVEN(pMB->mvs[0].y);
2221 :			pmv[2].x = EVEN(Data->predMV.x);
2222 :			pmv[2].y = EVEN(Data->predMV.y);
2223 :			pmv[0].x = pmv[0].y = 0;
2224 :	edgomez	195
2225 :	edgomez	851	CheckCandidate32I(0, 0, 255, &i, Data);
2226 :	edgomez	195
2227 :	edgomez	953	if (Data->iMinSAD > 4 MAX_SAD00_FOR_SKIP) {
2228 :	chl	96
2229 :	edgomez	851	if (!(mask = make_mask(pmv, 1)))
2230 :			CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data);
2231 :			if (!(mask = make_mask(pmv, 2)))
2232 :			CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data);
2233 :	chl	96
2234 :	edgomez	1053	if (Data->iMinSAD > 4 MAX_SAD00_FOR_SKIP) /* diamond only if needed */
2235 :	edgomez	851	DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i);
2236 :	edgomez	953	}
2237 :	edgomez	851
2238 :	edgomez	953	for (i = 0; i < 4; i++) {
2239 :			MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width];
2240 :			MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i];
2241 :			MB->mode = MODE_INTER;
2242 :			MB->sad16 = Data->iMinSAD[i+1];
2243 :	chl	96	}
2244 :	edgomez	851	}
2245 :	chl	96
2246 :	syskin	1075	#define INTRA_THRESH 2200
2247 :			#define INTER_THRESH 50
2248 :			#define INTRA_THRESH2 95
2249 :	chl	96
2250 :	edgomez	851	int
2251 :			MEanalysis( const IMAGE * const pRef,
2252 :	edgomez	953	const FRAMEINFO * const Current,
2253 :			const MBParam * const pParam,
2254 :	syskin	1075	const int maxIntra, //maximum number if non-I frames
2255 :			const int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame
2256 :	syskin	1084	const int bCount, // number of B frames in a row
2257 :	edgomez	953	const int b_thresh)
2258 :	edgomez	851	{
2259 :			uint32_t x, y, intra = 0;
2260 :			int sSAD = 0;
2261 :			MACROBLOCK * const pMBs = Current->mbs;
2262 :			const IMAGE * const pCurrent = &Current->image;
2263 :	syskin	1075	int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + b_thresh;
2264 :	chl	1077	int blocks = 0;
2265 :	syskin	1075	int complexity = 0;
2266 :	chl	96
2267 :	edgomez	851	int32_t iMinSAD[5], temp[5];
2268 :			VECTOR currentMV[5];
2269 :			SearchData Data;
2270 :			Data.iEdgedWidth = pParam->edged_width;
2271 :			Data.currentMV = currentMV;
2272 :			Data.iMinSAD = iMinSAD;
2273 :			Data.iFcode = Current->fcode;
2274 :			Data.temp = temp;
2275 :			CheckCandidate = CheckCandidate32I;
2276 :	chl	96
2277 :	edgomez	1023	if (intraCount != 0) {
2278 :	syskin	1075	if (intraCount < 10) // we're right after an I frame
2279 :	edgomez	1023	IntraThresh += 15* (intraCount - 10) * (intraCount - 10);
2280 :			else
2281 :	syskin	1075	if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec
2282 :	edgomez	1023	IntraThresh -= (IntraThresh * (maxIntra - 8*(maxIntra - intraCount)))/maxIntra;
2283 :			}
2284 :	chl	96
2285 :	syskin	1075	InterThresh -= 12 * bCount;
2286 :			if (InterThresh < 15 + b_thresh) InterThresh = 15 + b_thresh;
2287 :	chl	326
2288 :	edgomez	851	if (sadInit) (*sadInit) ();
2289 :	chl	96
2290 :	edgomez	851	for (y = 1; y < pParam->mb_height-1; y += 2) {
2291 :			for (x = 1; x < pParam->mb_width-1; x += 2) {
2292 :			int i;
2293 :	syskin	1075	blocks += 10;
2294 :	edgomez	195
2295 :	edgomez	851	if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV;
2296 :	syskin	1075	else { //extrapolation of the vector found for last frame
2297 :	edgomez	978	pMBs[x + y * pParam->mb_width].mvs[0].x =
2298 :	edgomez	953	(pMBs[x + y * pParam->mb_width].mvs[0].x * (bCount+1) ) / bCount;
2299 :	edgomez	978	pMBs[x + y * pParam->mb_width].mvs[0].y =
2300 :	edgomez	953	(pMBs[x + y * pParam->mb_width].mvs[0].y * (bCount+1) ) / bCount;
2301 :			}
2302 :	edgomez	195
2303 :	edgomez	851	MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data);
2304 :	edgomez	195
2305 :	edgomez	851	for (i = 0; i < 4; i++) {
2306 :			int dev;
2307 :			MACROBLOCK pMB = &pMBs[x+(i&1) + (y+(i>>1)) pParam->mb_width];
2308 :	syskin	1075	dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16,
2309 :			pParam->edged_width);
2310 :	syskin	1084
2311 :	syskin	1108	complexity += MAX(dev, 300);
2312 :	syskin	1075	if (dev + IntraThresh < pMB->sad16) {
2313 :			pMB->mode = MODE_INTRA;
2314 :			if (++intra > ((pParam->mb_height-2)*(pParam->mb_width-2))/2) return I_VOP;
2315 :	edgomez	851	}
2316 :	edgomez	953
2317 :	syskin	1084	if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0)
2318 :	syskin	1075	if (dev > 500 && pMB->sad16 < 1000)
2319 :			sSAD += 1000;
2320 :
2321 :	syskin	1108	sSAD += (dev < 3000) ? pMB->sad16 : pMB->sad16/2; /* blocks with big contrast differences usually have large SAD - while they look very good in b-frames */
2322 :	edgomez	851	}
2323 :			}
2324 :	edgomez	195	}
2325 :	syskin	1075	complexity >>= 7;
2326 :	edgomez	953
2327 :	syskin	1075	sSAD /= complexity + 4*blocks;
2328 :	edgomez	953
2329 :	syskin	1108	if (intraCount > 80 && sSAD > INTRA_THRESH2 ) return I_VOP;
2330 :	edgomez	851	if (sSAD > InterThresh ) return P_VOP;
2331 :			emms();
2332 :			return B_VOP;
2333 :			}
2334 :	chl	96
2335 :
2336 :	edgomez	1053	/* functions which perform BITS-based search/bitcount */
2337 :	chl	96
2338 :	edgomez	851	static int
2339 :	syskin	1108	findRDinter(SearchData * const Data,
2340 :			const MACROBLOCK * const pMBs, const int x, const int y,
2341 :			const MBParam * const pParam,
2342 :			const uint32_t MotionFlags)
2343 :	chl	96	{
2344 :	edgomez	851	int i, iDirection;
2345 :			int32_t bsad[5];
2346 :	chl	141
2347 :	syskin	1108	CheckCandidate = CheckCandidateRD16;
2348 :	chl	96
2349 :	edgomez	851	if (Data->qpel) {
2350 :			for(i = 0; i < 5; i++) {
2351 :			Data->currentMV[i].x = Data->currentQMV[i].x/2;
2352 :			Data->currentMV[i].y = Data->currentQMV[i].y/2;
2353 :			}
2354 :			Data->qpel_precision = 1;
2355 :	syskin	1108	CheckCandidateRD16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data);
2356 :	chl	96
2357 :	edgomez	1107	if (MotionFlags & (XVID_ME_HALFPELREFINE16_RD \| XVID_ME_EXTSEARCH_RD)) { /* we have to prepare for halfpixel-precision search */
2358 :	edgomez	851	for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i];
2359 :	syskin	1129	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
2360 :			pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, Data->rrv);
2361 :	edgomez	851	Data->qpel_precision = 0;
2362 :			if (Data->currentQMV->x & 1 \|\| Data->currentQMV->y & 1)
2363 :	syskin	1108	CheckCandidateRD16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2364 :	edgomez	851	}
2365 :	edgomez	195
2366 :	edgomez	1053	} else { /* not qpel */
2367 :	edgomez	195
2368 :	syskin	1108	CheckCandidateRD16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2369 :	edgomez	851	}
2370 :	chl	96
2371 :	edgomez	1107	if (MotionFlags&XVID_ME_EXTSEARCH_RD) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
2372 :	chl	96
2373 :	edgomez	1107	if (MotionFlags&XVID_ME_HALFPELREFINE16_RD) SubpelRefine(Data);
2374 :	chl	96
2375 :	edgomez	851	if (Data->qpel) {
2376 :	edgomez	1107	if (MotionFlags&(XVID_ME_EXTSEARCH_RD \| XVID_ME_HALFPELREFINE16_RD)) { /* there was halfpel-precision search */
2377 :	edgomez	851	for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) {
2378 :	edgomez	1053	Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* we have found a better match */
2379 :	edgomez	851	Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
2380 :			}
2381 :	edgomez	195
2382 :	edgomez	1053	/* preparing for qpel-precision search */
2383 :	edgomez	851	Data->qpel_precision = 1;
2384 :	syskin	1129	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
2385 :			pParam->width, pParam->height, Data->iFcode, 2, 0);
2386 :	edgomez	851	}
2387 :	edgomez	1107	if (MotionFlags&XVID_ME_QUARTERPELREFINE16_RD) SubpelRefine(Data);
2388 :	edgomez	851	}
2389 :	chl	96
2390 :	edgomez	1107	if (MotionFlags&XVID_ME_CHECKPREDICTION_RD) { /* let's check vector equal to prediction */
2391 :	edgomez	851	VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV;
2392 :			if (!(Data->predMV.x == v->x && Data->predMV.y == v->y))
2393 :	syskin	1108	CheckCandidateRD16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data);
2394 :	edgomez	851	}
2395 :			return Data->iMinSAD[0];
2396 :			}
2397 :	chl	96
2398 :	edgomez	851	static int
2399 :	syskin	1108	findRDinter4v(const SearchData * const Data,
2400 :			MACROBLOCK * const pMB, const MACROBLOCK * const pMBs,
2401 :			const int x, const int y,
2402 :			const MBParam * const pParam, const uint32_t MotionFlags,
2403 :			const VECTOR * const backup)
2404 :	edgomez	851	{
2405 :	chl	96
2406 :	edgomez	851	int cbp = 0, bits = 0, t = 0, i, iDirection;
2407 :			SearchData Data2, *Data8 = &Data2;
2408 :			int sumx = 0, sumy = 0;
2409 :	edgomez	962	int16_t in = Data->dctSpace, coeff = Data->dctSpace + 64;
2410 :	edgomez	1022	uint8_t * ptr;
2411 :	chl	96
2412 :	edgomez	851	memcpy(Data8, Data, sizeof(SearchData));
2413 :	syskin	1108	CheckCandidate = CheckCandidateRD8;
2414 :	chl	96
2415 :	edgomez	1053	for (i = 0; i < 4; i++) { /* for all luma blocks */
2416 :	edgomez	1022
2417 :	edgomez	851	Data8->iMinSAD = Data->iMinSAD + i + 1;
2418 :			Data8->currentMV = Data->currentMV + i + 1;
2419 :			Data8->currentQMV = Data->currentQMV + i + 1;
2420 :			Data8->Cur = Data->Cur + 8((i&1) + (i>>1)Data->iEdgedWidth);
2421 :	edgomez	978	Data8->RefP[0] = Data->RefP[0] + 8((i&1) + (i>>1)Data->iEdgedWidth);
2422 :			Data8->RefP[2] = Data->RefP[2] + 8((i&1) + (i>>1)Data->iEdgedWidth);
2423 :			Data8->RefP[1] = Data->RefP[1] + 8((i&1) + (i>>1)Data->iEdgedWidth);
2424 :			Data8->RefP[3] = Data->RefP[3] + 8((i&1) + (i>>1)Data->iEdgedWidth);
2425 :	syskin	1084	*Data8->cbp = (Data->cbp[1] & (1<<(5-i))) ? 1:0; // copy corresponding cbp bit
2426 :	edgomez	195
2427 :	edgomez	851	if(Data->qpel) {
2428 :			Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i);
2429 :			if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y,
2430 :			Data8->predMV, Data8->iFcode, 0, 0);
2431 :			} else {
2432 :			Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i);
2433 :			if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y,
2434 :			Data8->predMV, Data8->iFcode, 0, 0);
2435 :			}
2436 :	chl	96
2437 :	syskin	1129	get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2x + (i&1), 2y + (i>>1), 3,
2438 :			pParam->width, pParam->height, Data8->iFcode, Data8->qpel+1, 0);
2439 :	chl	96
2440 :	edgomez	1023	Data8->iMinSAD += BITS_MULTt;
2441 :	chl	96
2442 :	edgomez	851	Data8->qpel_precision = Data8->qpel;
2443 :	edgomez	1053	/* checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far) */
2444 :	edgomez	987	{
2445 :			VECTOR *v = Data8->qpel ? Data8->currentQMV : Data8->currentMV;
2446 :	edgomez	1022	if (!MVequal (*v, backup[i+1]) )
2447 :	syskin	1108	CheckCandidateRD8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8);
2448 :	edgomez	851	}
2449 :	edgomez	195
2450 :	edgomez	851	if (Data8->qpel) {
2451 :	edgomez	1107	if (MotionFlags&XVID_ME_HALFPELREFINE8_RD \|\| (MotionFlags&XVID_ME_EXTSEARCH8 && MotionFlags&XVID_ME_EXTSEARCH_RD)) { /* halfpixel motion search follows */
2452 :	edgomez	851	int32_t s = *Data8->iMinSAD;
2453 :			Data8->currentMV->x = Data8->currentQMV->x/2;
2454 :			Data8->currentMV->y = Data8->currentQMV->y/2;
2455 :			Data8->qpel_precision = 0;
2456 :	syskin	1129	get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2x + (i&1), 2y + (i>>1), 3,
2457 :			pParam->width, pParam->height, Data8->iFcode - 1, 1, 0);
2458 :	edgomez	195
2459 :	edgomez	851	if (Data8->currentQMV->x & 1 \|\| Data8->currentQMV->y & 1)
2460 :	syskin	1108	CheckCandidateRD8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8);
2461 :	edgomez	195
2462 :	edgomez	1107	if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_RD)
2463 :	edgomez	851	SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255);
2464 :	chl	96
2465 :	edgomez	1107	if (MotionFlags & XVID_ME_HALFPELREFINE8_RD)
2466 :	edgomez	1022	SubpelRefine(Data8);
2467 :	chl	96
2468 :	edgomez	1053	if(s > Data8->iMinSAD) { / we have found a better match */
2469 :	edgomez	851	Data8->currentQMV->x = 2*Data8->currentMV->x;
2470 :			Data8->currentQMV->y = 2*Data8->currentMV->y;
2471 :			}
2472 :	chl	96
2473 :	edgomez	851	Data8->qpel_precision = 1;
2474 :	syskin	1129	get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2x + (i&1), 2y + (i>>1), 3,
2475 :			pParam->width, pParam->height, Data8->iFcode, 2, 0);
2476 :	chl	96
2477 :	edgomez	851	}
2478 :	edgomez	1107	if (MotionFlags & XVID_ME_QUARTERPELREFINE8_RD) SubpelRefine(Data8);
2479 :	edgomez	195
2480 :	edgomez	1053	} else { /* not qpel */
2481 :	edgomez	851
2482 :	edgomez	1107	if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_RD) /* extsearch */
2483 :	edgomez	1022	SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255);
2484 :
2485 :	edgomez	1107	if (MotionFlags & XVID_ME_HALFPELREFINE8_RD)
2486 :	edgomez	1053	SubpelRefine(Data8); /* halfpel refinement */
2487 :	edgomez	1022	}
2488 :
2489 :	edgomez	1053	/* checking vector equal to predicion */
2490 :	edgomez	1107	if (i != 0 && MotionFlags & XVID_ME_CHECKPREDICTION_RD) {
2491 :	edgomez	851	const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV;
2492 :	edgomez	1022	if (!MVequal(*v, Data8->predMV))
2493 :	syskin	1108	CheckCandidateRD8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8);
2494 :	chl	141	}
2495 :	edgomez	195
2496 :	edgomez	851	bits += *Data8->iMinSAD;
2497 :	edgomez	1053	if (bits >= Data->iMinSAD[0]) return bits; /* no chances for INTER4V */
2498 :	edgomez	851
2499 :	edgomez	1053	/* MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else */
2500 :	edgomez	851	if(Data->qpel) {
2501 :			pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x;
2502 :			pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y;
2503 :			pMB->qmvs[i] = *Data8->currentQMV;
2504 :			sumx += Data8->currentQMV->x/2;
2505 :			sumy += Data8->currentQMV->y/2;
2506 :			} else {
2507 :			pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x;
2508 :			pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y;
2509 :			sumx += Data8->currentMV->x;
2510 :			sumy += Data8->currentMV->y;
2511 :	chl	141	}
2512 :	edgomez	851	pMB->mvs[i] = *Data8->currentMV;
2513 :			pMB->sad8[i] = 4 * *Data8->iMinSAD;
2514 :	syskin	1084	if (Data8->cbp[0]) cbp \|= 1 << (5 - i);
2515 :	chl	141
2516 :	syskin	1084	} /* end - for all luma blocks */
2517 :	chl	141
2518 :	edgomez	1023	bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len;
2519 :	chl	141
2520 :	edgomez	1053	/* let's check chroma */
2521 :	edgomez	1022	sumx = (sumx >> 3) + roundtab_76[sumx & 0xf];
2522 :			sumy = (sumy >> 3) + roundtab_76[sumy & 0xf];
2523 :	chl	96
2524 :	edgomez	1053	/* chroma U */
2525 :	edgomez	1022	ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[4], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2526 :			transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2);
2527 :	syskin	1075	bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4);
2528 :	edgomez	1022
2529 :			if (bits >= *Data->iMinSAD) return bits;
2530 :
2531 :	edgomez	1053	/* chroma V */
2532 :	edgomez	1022	ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[5], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2533 :			transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2);
2534 :	syskin	1075	bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5);
2535 :	edgomez	1022
2536 :	edgomez	1023	bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER4V & 7) \| ((cbp & 3) << 3)].len;
2537 :	edgomez	1022
2538 :	syskin	1084	*Data->cbp = cbp;
2539 :	edgomez	851	return bits;
2540 :			}
2541 :	chl	96
2542 :	edgomez	851	static int
2543 :	syskin	1108	findRDintra(const SearchData * const Data)
2544 :	edgomez	851	{
2545 :	edgomez	1053	int bits = BITS_MULT1; / this one is ac/dc prediction flag bit */
2546 :	edgomez	1023	int cbp = 0, i, dc = 0;
2547 :	edgomez	962	int16_t in = Data->dctSpace, coeff = Data->dctSpace + 64;
2548 :	edgomez	195
2549 :	edgomez	851	for(i = 0; i < 4; i++) {
2550 :			int s = 8((i&1) + (i>>1)Data->iEdgedWidth);
2551 :			transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth);
2552 :	syskin	1075	bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i, &dc);
2553 :	chl	141
2554 :	edgomez	1022	if (bits >= Data->iMinSAD[0]) return bits;
2555 :	edgomez	851	}
2556 :	chl	181
2557 :	edgomez	1023	bits += BITS_MULT*xvid_cbpy_tab[cbp>>2].len;
2558 :	chl	96
2559 :	edgomez	1053	/chroma U /
2560 :	edgomez	1022	transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2);
2561 :	syskin	1075	bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4, &dc);
2562 :	syskin	1084
2563 :	edgomez	1022	if (bits >= Data->iMinSAD[0]) return bits;
2564 :
2565 :	edgomez	1053	/* chroma V */
2566 :	edgomez	1022	transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2);
2567 :	syskin	1075	bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5, &dc);
2568 :	edgomez	1022
2569 :	edgomez	1023	bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTRA & 7) \| ((cbp & 3) << 3)].len;
2570 :	edgomez	1022
2571 :	edgomez	851	return bits;
2572 :	chl	96	}
2573 :	chl	1077
2574 :	syskin	1084	static int
2575 :	syskin	1108	findRDgmc(const SearchData * const Data, const IMAGE * const vGMC, const int x, const int y)
2576 :	syskin	1084	{
2577 :			int bits = BITS_MULT1; / this one is mcsel */
2578 :			int cbp = 0, i;
2579 :			int16_t in = Data->dctSpace, coeff = Data->dctSpace + 64;
2580 :	chl	1077
2581 :	syskin	1084	for(i = 0; i < 4; i++) {
2582 :			int s = 8((i&1) + (i>>1)Data->iEdgedWidth);
2583 :			transfer_8to16subro(in, Data->Cur + s, vGMC->y + s + 16(x+yData->iEdgedWidth), Data->iEdgedWidth);
2584 :			bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i);
2585 :			if (bits >= Data->iMinSAD[0]) return bits;
2586 :			}
2587 :	chl	1077
2588 :	syskin	1084	bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len;
2589 :	chl	1077
2590 :	syskin	1084	/chroma U /
2591 :			transfer_8to16subro(in, Data->CurU, vGMC->u + 8(x+y(Data->iEdgedWidth/2)), Data->iEdgedWidth/2);
2592 :			bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4);
2593 :	chl	1077
2594 :	syskin	1084	if (bits >= Data->iMinSAD[0]) return bits;
2595 :
2596 :			/* chroma V */
2597 :			transfer_8to16subro(in, Data->CurV , vGMC->v + 8(x+y(Data->iEdgedWidth/2)), Data->iEdgedWidth/2);
2598 :			bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5);
2599 :
2600 :			bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) \| ((cbp & 3) << 3)].len;
2601 :
2602 :			*Data->cbp = cbp;
2603 :
2604 :			return bits;
2605 :			}
2606 :
2607 :
2608 :
2609 :
2610 :	chl	1077	static __inline void
2611 :			GMEanalyzeMB ( const uint8_t * const pCur,
2612 :			const uint8_t * const pRef,
2613 :			const uint8_t * const pRefH,
2614 :			const uint8_t * const pRefV,
2615 :			const uint8_t * const pRefHV,
2616 :			const int x,
2617 :			const int y,
2618 :			const MBParam * const pParam,
2619 :			MACROBLOCK * const pMBs,
2620 :			SearchData * const Data)
2621 :			{
2622 :
2623 :			int i=0;
2624 :			MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width];
2625 :
2626 :			Data->iMinSAD[0] = MV_MAX_ERROR;
2627 :
2628 :	syskin	1084	Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0);
2629 :	chl	1077
2630 :	syskin	1129	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
2631 :			pParam->width, pParam->height, 16, 1, 0);
2632 :	chl	1077
2633 :			Data->Cur = pCur + 16(x + y pParam->edged_width);
2634 :			Data->RefP[0] = pRef + 16(x + y pParam->edged_width);
2635 :			Data->RefP[1] = pRefV + 16(x + y pParam->edged_width);
2636 :			Data->RefP[2] = pRefH + 16(x + y pParam->edged_width);
2637 :			Data->RefP[3] = pRefHV + 16(x + y pParam->edged_width);
2638 :
2639 :			Data->currentMV[0].x = Data->currentMV[0].y = 0;
2640 :			CheckCandidate16I(0, 0, 255, &i, Data);
2641 :
2642 :			if ( (Data->predMV.x !=0) \|\| (Data->predMV.y != 0) )
2643 :			CheckCandidate16I(Data->predMV.x, Data->predMV.y, 255, &i, Data);
2644 :
2645 :	syskin	1084	AdvDiamondSearch(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255);
2646 :	chl	1077
2647 :	syskin	1084	SubpelRefine(Data);
2648 :
2649 :
2650 :	chl	1077	/* for QPel halfpel positions are worse than in halfpel mode :( */
2651 :			/* if (Data->qpel) {
2652 :			Data->currentQMV->x = 2*Data->currentMV->x;
2653 :			Data->currentQMV->y = 2*Data->currentMV->y;
2654 :			Data->qpel_precision = 1;
2655 :	syskin	1129	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
2656 :			pParam->width, pParam->height, iFcode, 2, 0);
2657 :	chl	1077	SubpelRefine(Data);
2658 :			}
2659 :			*/
2660 :
2661 :			pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
2662 :			pMB->sad16 = Data->iMinSAD[0];
2663 :	syskin	1084	pMB->mode = MODE_INTER;
2664 :			pMB->sad16 += 10*d_mv_bits(pMB->mvs[0].x, pMB->mvs[0].y, Data->predMV, Data->iFcode, 0, 0);
2665 :	chl	1077	return;
2666 :			}
2667 :
2668 :			void
2669 :			GMEanalysis(const MBParam * const pParam,
2670 :			const FRAMEINFO * const current,
2671 :			const FRAMEINFO * const reference,
2672 :			const IMAGE * const pRefH,
2673 :			const IMAGE * const pRefV,
2674 :			const IMAGE * const pRefHV)
2675 :			{
2676 :			uint32_t x, y;
2677 :			MACROBLOCK * const pMBs = current->mbs;
2678 :			const IMAGE * const pCurrent = &current->image;
2679 :			const IMAGE * const pReference = &reference->image;
2680 :
2681 :			int32_t iMinSAD[5], temp[5];
2682 :			VECTOR currentMV[5];
2683 :			SearchData Data;
2684 :			memset(&Data, 0, sizeof(SearchData));
2685 :	syskin	1084
2686 :	chl	1077	Data.iEdgedWidth = pParam->edged_width;
2687 :			Data.rounding = pParam->m_rounding_type;
2688 :	syskin	1084
2689 :	chl	1077	Data.currentMV = &currentMV[0];
2690 :			Data.iMinSAD = &iMinSAD[0];
2691 :			Data.iFcode = current->fcode;
2692 :			Data.temp = temp;
2693 :
2694 :			CheckCandidate = CheckCandidate16I;
2695 :	syskin	1084
2696 :	chl	1077	if (sadInit) (*sadInit) ();
2697 :
2698 :			for (y = 0; y < pParam->mb_height; y ++) {
2699 :			for (x = 0; x < pParam->mb_width; x ++) {
2700 :
2701 :			GMEanalyzeMB(pCurrent->y, pReference->y, pRefH->y, pRefV->y, pRefHV->y, x, y, pParam, pMBs, &Data);
2702 :			}
2703 :			}
2704 :			return;
2705 :			}
2706 :
2707 :
2708 :			WARPPOINTS
2709 :			GlobalMotionEst(MACROBLOCK * const pMBs,
2710 :			const MBParam * const pParam,
2711 :			const FRAMEINFO * const current,
2712 :			const FRAMEINFO * const reference,
2713 :			const IMAGE * const pRefH,
2714 :			const IMAGE * const pRefV,
2715 :			const IMAGE * const pRefHV)
2716 :			{
2717 :
2718 :	syskin	1084	const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs
2719 :			const int deltay=8;
2720 :	chl	1077	const unsigned int gradx=512; // lower bound for gradient in MB (ignore "flat" blocks)
2721 :	syskin	1084	const unsigned int grady=512;
2722 :	chl	1077
2723 :			double sol[4] = { 0., 0., 0., 0. };
2724 :	syskin	1084
2725 :	chl	1077	WARPPOINTS gmc;
2726 :
2727 :			uint32_t mx, my;
2728 :
2729 :			int MBh = pParam->mb_height;
2730 :			int MBw = pParam->mb_width;
2731 :			const int minblocks = 9; //MBhMBw/32+3; / just some reasonable number 3% + 3 */
2732 :			const int maxblocks = MBhMBw/4; / just some reasonable number 3% + 3 */
2733 :
2734 :			int num=0;
2735 :			int oldnum;
2736 :
2737 :			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;
2738 :
2739 :			GMEanalysis(pParam,current, reference, pRefH, pRefV, pRefHV);
2740 :
2741 :			/* block based ME isn't done, yet, so do a quick presearch */
2742 :
2743 :			// filter mask of all blocks
2744 :
2745 :			for (my = 0; my < (uint32_t)MBh; my++)
2746 :			for (mx = 0; mx < (uint32_t)MBw; mx++)
2747 :			{
2748 :			const int mbnum = mx + my * MBw;
2749 :			pMBs[mbnum].mcsel = 0;
2750 :			}
2751 :	syskin	1084
2752 :
2753 :			for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */
2754 :	chl	1077	for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */
2755 :			{
2756 :			const int mbnum = mx + my * MBw;
2757 :			MACROBLOCK *const pMB = &pMBs[mbnum];
2758 :			const VECTOR mv = pMB->mvs[0];
2759 :
2760 :			/* don't use object boundaries */
2761 :	syskin	1084	if ( (abs(mv.x - (pMB-1)->mvs[0].x) < deltax)
2762 :	chl	1077	&& (abs(mv.y - (pMB-1)->mvs[0].y) < deltay)
2763 :	syskin	1084	&& (abs(mv.x - (pMB+1)->mvs[0].x) < deltax)
2764 :	chl	1077	&& (abs(mv.y - (pMB+1)->mvs[0].y) < deltay)
2765 :	syskin	1084	&& (abs(mv.x - (pMB-MBw)->mvs[0].x) < deltax)
2766 :	chl	1077	&& (abs(mv.y - (pMB-MBw)->mvs[0].y) < deltay)
2767 :	syskin	1084	&& (abs(mv.x - (pMB+MBw)->mvs[0].x) < deltax)
2768 :	chl	1077	&& (abs(mv.y - (pMB+MBw)->mvs[0].y) < deltay) )
2769 :			{ const int iEdgedWidth = pParam->edged_width;
2770 :			const uint8_t const pCur = current->image.y + 16(my*iEdgedWidth + mx);
2771 :			if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx )
2772 :			&& (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) )
2773 :	syskin	1084	{ pMB->mcsel = 1;
2774 :	chl	1077	num++;
2775 :			}
2776 :
2777 :			/* only use "structured" blocks */
2778 :			}
2779 :			}
2780 :			emms();
2781 :
2782 :	syskin	1084	/* further filtering would be possible, but during iteration, remaining
2783 :	chl	1077	outliers usually are removed, too */
2784 :
2785 :			if (num>= minblocks)
2786 :			do { /* until convergence */
2787 :			double DtimesF[4];
2788 :			double a,b,c,n,invdenom;
2789 :			double meanx,meany;
2790 :
2791 :			a = b = c = n = 0;
2792 :			DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.;
2793 :			for (my = 1; my < (uint32_t)MBh-1; my++)
2794 :			for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2795 :			{
2796 :			const int mbnum = mx + my * MBw;
2797 :			const VECTOR mv = pMBs[mbnum].mvs[0];
2798 :
2799 :			if (!pMBs[mbnum].mcsel)
2800 :			continue;
2801 :
2802 :			n++;
2803 :			a += 16*mx+8;
2804 :			b += 16*my+8;
2805 :			c += (16mx+8)(16mx+8)+(16my+8)(16my+8);
2806 :
2807 :			DtimesF[0] += (double)mv.x;
2808 :			DtimesF[1] += (double)mv.x(16mx+8) + (double)mv.y(16my+8);
2809 :			DtimesF[2] += (double)mv.x(16my+8) - (double)mv.y(16mx+8);
2810 :			DtimesF[3] += (double)mv.y;
2811 :			}
2812 :	syskin	1084
2813 :	chl	1077	invdenom = aa+bb-c*n;
2814 :
2815 :			/* Solve the system: sol = (D'ED)^{-1} D'EF */
2816 :			/* D'EF has been calculated in the same loop as matrix */
2817 :
2818 :			sol[0] = -cDtimesF[0] + aDtimesF[1] + b*DtimesF[2];
2819 :			sol[1] = aDtimesF[0] - nDtimesF[1] + b*DtimesF[3];
2820 :			sol[2] = bDtimesF[0] - nDtimesF[2] - a*DtimesF[3];
2821 :			sol[3] = bDtimesF[1] - aDtimesF[2] - c*DtimesF[3];
2822 :
2823 :			sol[0] /= invdenom;
2824 :			sol[1] /= invdenom;
2825 :			sol[2] /= invdenom;
2826 :			sol[3] /= invdenom;
2827 :
2828 :			meanx = meany = 0.;
2829 :			oldnum = 0;
2830 :			for (my = 1; my < (uint32_t)MBh-1; my++)
2831 :			for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2832 :			{
2833 :			const int mbnum = mx + my * MBw;
2834 :			const VECTOR mv = pMBs[mbnum].mvs[0];
2835 :
2836 :			if (!pMBs[mbnum].mcsel)
2837 :			continue;
2838 :
2839 :			oldnum++;
2840 :			meanx += fabs(( sol[0] + (16mx+8)sol[1] + (16my+8)sol[2] ) - (double)mv.x );
2841 :			meany += fabs(( sol[3] - (16mx+8)sol[2] + (16my+8)sol[1] ) - (double)mv.y );
2842 :			}
2843 :
2844 :			if (4meanx > oldnum) / better fit than 0.25 (=1/4pel) is useless */
2845 :			meanx /= oldnum;
2846 :			else
2847 :			meanx = 0.25;
2848 :
2849 :			if (4*meany > oldnum)
2850 :			meany /= oldnum;
2851 :			else
2852 :			meany = 0.25;
2853 :
2854 :			num = 0;
2855 :			for (my = 0; my < (uint32_t)MBh; my++)
2856 :			for (mx = 0; mx < (uint32_t)MBw; mx++)
2857 :			{
2858 :			const int mbnum = mx + my * MBw;
2859 :			const VECTOR mv = pMBs[mbnum].mvs[0];
2860 :
2861 :			if (!pMBs[mbnum].mcsel)
2862 :			continue;
2863 :
2864 :			if ( ( fabs(( sol[0] + (16mx+8)sol[1] + (16my+8)sol[2] ) - (double)mv.x ) > meanx )
2865 :			\|\| ( fabs(( sol[3] - (16mx+8)sol[2] + (16my+8)sol[1] ) - (double)mv.y ) > meany ) )
2866 :			pMBs[mbnum].mcsel=0;
2867 :			else
2868 :			num++;
2869 :			}
2870 :
2871 :			} while ( (oldnum != num) && (num>= minblocks) );
2872 :
2873 :			if (num < minblocks)
2874 :			{
2875 :			const int iEdgedWidth = pParam->edged_width;
2876 :			num = 0;
2877 :
2878 :			/* fprintf(stderr,"Warning! Unreliable GME (%d/%d blocks), falling back to translation.\n",num,MBh*MBw);
2879 :	syskin	1084	*/
2880 :	chl	1077	gmc.duv[0].x= gmc.duv[0].y= gmc.duv[1].x= gmc.duv[1].y= gmc.duv[2].x= gmc.duv[2].y=0;
2881 :
2882 :	edgomez	1107	if (!(current->motion_flags & XVID_ME_GME_REFINE))
2883 :	chl	1077	return gmc;
2884 :	syskin	1084
2885 :			for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */
2886 :	chl	1077	for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */
2887 :			{
2888 :			const int mbnum = mx + my * MBw;
2889 :			MACROBLOCK *const pMB = &pMBs[mbnum];
2890 :			const uint8_t const pCur = current->image.y + 16(my*iEdgedWidth + mx);
2891 :			if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx )
2892 :			&& (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) )
2893 :	syskin	1084	{ pMB->mcsel = 1;
2894 :	chl	1077	gmc.duv[0].x += pMB->mvs[0].x;
2895 :			gmc.duv[0].y += pMB->mvs[0].y;
2896 :			num++;
2897 :			}
2898 :			}
2899 :	syskin	1084
2900 :	chl	1077	if (gmc.duv[0].x)
2901 :			gmc.duv[0].x /= num;
2902 :	syskin	1084	if (gmc.duv[0].y)
2903 :	chl	1077	gmc.duv[0].y /= num;
2904 :			} else {
2905 :
2906 :			gmc.duv[0].x=(int)(sol[0]+0.5);
2907 :			gmc.duv[0].y=(int)(sol[3]+0.5);
2908 :
2909 :			gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5);
2910 :			gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5);
2911 :
2912 :			gmc.duv[2].x=-gmc.duv[1].y; /* two warp points only */
2913 :			gmc.duv[2].y=gmc.duv[1].x;
2914 :			}
2915 :			if (num>maxblocks)
2916 :			{ for (my = 1; my < (uint32_t)MBh-1; my++)
2917 :			for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2918 :			{
2919 :			const int mbnum = mx + my * MBw;
2920 :			if (pMBs[mbnum-1].mcsel)
2921 :			pMBs[mbnum].mcsel=0;
2922 :	syskin	1084	else
2923 :	chl	1077	if (pMBs[mbnum-MBw].mcsel)
2924 :			pMBs[mbnum].mcsel=0;
2925 :			}
2926 :			}
2927 :			return gmc;
2928 :			}
2929 :
2930 :			int
2931 :			GlobalMotionEstRefine(
2932 :			WARPPOINTS *const startwp,
2933 :			MACROBLOCK * const pMBs,
2934 :			const MBParam * const pParam,
2935 :			const FRAMEINFO * const current,
2936 :			const FRAMEINFO * const reference,
2937 :			const IMAGE * const pCurr,
2938 :			const IMAGE * const pRef,
2939 :			const IMAGE * const pRefH,
2940 :			const IMAGE * const pRefV,
2941 :			const IMAGE * const pRefHV)
2942 :			{
2943 :			uint8_t* GMCblock = (uint8_t)malloc(16pParam->edged_width);
2944 :			WARPPOINTS bestwp=*startwp;
2945 :			WARPPOINTS centerwp,currwp;
2946 :			int gmcminSAD=0;
2947 :			int gmcSAD=0;
2948 :			int direction;
2949 :			// int mx,my;
2950 :
2951 :			/* use many blocks... */
2952 :			/* for (my = 0; my < (uint32_t)pParam->mb_height; my++)
2953 :			for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++)
2954 :			{
2955 :			const int mbnum = mx + my * pParam->mb_width;
2956 :	syskin	1084	pMBs[mbnum].mcsel=1;
2957 :	chl	1077	}
2958 :			*/
2959 :
2960 :			/* or rather don't use too many blocks... */
2961 :			/*
2962 :			for (my = 1; my < (uint32_t)MBh-1; my++)
2963 :			for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2964 :			{
2965 :			const int mbnum = mx + my * MBw;
2966 :			if (MBmask[mbnum-1])
2967 :			MBmask[mbnum-1]=0;
2968 :	syskin	1084	else
2969 :	chl	1077	if (MBmask[mbnum-MBw])
2970 :			MBmask[mbnum-1]=0;
2971 :	syskin	1084
2972 :	chl	1077	}
2973 :			*/
2974 :			gmcminSAD = globalSAD(&bestwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
2975 :
2976 :	syskin	1084	if ( (reference->coding_type == S_VOP)
2977 :			&& ( (reference->warp.duv[1].x != bestwp.duv[1].x)
2978 :			\|\| (reference->warp.duv[1].y != bestwp.duv[1].y)
2979 :			\|\| (reference->warp.duv[0].x != bestwp.duv[0].x)
2980 :			\|\| (reference->warp.duv[0].y != bestwp.duv[0].y)
2981 :			\|\| (reference->warp.duv[2].x != bestwp.duv[2].x)
2982 :			\|\| (reference->warp.duv[2].y != bestwp.duv[2].y) ) )
2983 :	chl	1077	{
2984 :	syskin	1084	gmcSAD = globalSAD(&reference->warp, pParam, pMBs,
2985 :	chl	1077	current, pRef, pCurr, GMCblock);
2986 :	syskin	1084
2987 :	chl	1077	if (gmcSAD < gmcminSAD)
2988 :			{ bestwp = reference->warp;
2989 :			gmcminSAD = gmcSAD;
2990 :			}
2991 :			}
2992 :
2993 :			do {
2994 :			direction = 0;
2995 :			centerwp = bestwp;
2996 :
2997 :	syskin	1084	currwp = centerwp;
2998 :
2999 :	chl	1077	currwp.duv[0].x--;
3000 :			gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3001 :			if (gmcSAD < gmcminSAD)
3002 :			{ bestwp = currwp;
3003 :			gmcminSAD = gmcSAD;
3004 :	syskin	1084	direction = 1;
3005 :	chl	1077	}
3006 :	syskin	1084	else
3007 :			{
3008 :	chl	1077	currwp = centerwp; currwp.duv[0].x++;
3009 :			gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3010 :			if (gmcSAD < gmcminSAD)
3011 :			{ bestwp = currwp;
3012 :			gmcminSAD = gmcSAD;
3013 :	syskin	1084	direction = 2;
3014 :	chl	1077	}
3015 :			}
3016 :			if (direction) continue;
3017 :
3018 :			currwp = centerwp; currwp.duv[0].y--;
3019 :			gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3020 :			if (gmcSAD < gmcminSAD)
3021 :			{ bestwp = currwp;
3022 :			gmcminSAD = gmcSAD;
3023 :	syskin	1084	direction = 4;
3024 :			}
3025 :			else
3026 :	chl	1077	{
3027 :			currwp = centerwp; currwp.duv[0].y++;
3028 :			gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3029 :			if (gmcSAD < gmcminSAD)
3030 :			{ bestwp = currwp;
3031 :			gmcminSAD = gmcSAD;
3032 :	syskin	1084	direction = 8;
3033 :	chl	1077	}
3034 :	syskin	1084	}
3035 :	chl	1077	if (direction) continue;
3036 :	syskin	1084
3037 :	chl	1077	currwp = centerwp; currwp.duv[1].x++;
3038 :			gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3039 :			if (gmcSAD < gmcminSAD)
3040 :			{ bestwp = currwp;
3041 :			gmcminSAD = gmcSAD;
3042 :	syskin	1084	direction = 32;
3043 :			}
3044 :	chl	1077	currwp.duv[2].y++;
3045 :			gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3046 :			if (gmcSAD < gmcminSAD)
3047 :			{ bestwp = currwp;
3048 :			gmcminSAD = gmcSAD;
3049 :	syskin	1084	direction = 1024;
3050 :			}
3051 :	chl	1077
3052 :			currwp = centerwp; currwp.duv[1].x--;
3053 :			gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3054 :			if (gmcSAD < gmcminSAD)
3055 :			{ bestwp = currwp;
3056 :			gmcminSAD = gmcSAD;
3057 :	syskin	1084	direction = 16;
3058 :			}
3059 :			else
3060 :	chl	1077	{
3061 :			currwp = centerwp; currwp.duv[1].x++;
3062 :			gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3063 :			if (gmcSAD < gmcminSAD)
3064 :			{ bestwp = currwp;
3065 :			gmcminSAD = gmcSAD;
3066 :	syskin	1084	direction = 32;
3067 :	chl	1077	}
3068 :	syskin	1084	}
3069 :	chl	1077	if (direction) continue;
3070 :
3071 :
3072 :			currwp = centerwp; currwp.duv[1].y--;
3073 :			gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3074 :			if (gmcSAD < gmcminSAD)
3075 :			{ bestwp = currwp;
3076 :			gmcminSAD = gmcSAD;
3077 :	syskin	1084	direction = 64;
3078 :			}
3079 :			else
3080 :	chl	1077	{
3081 :			currwp = centerwp; currwp.duv[1].y++;
3082 :			gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3083 :			if (gmcSAD < gmcminSAD)
3084 :			{ bestwp = currwp;
3085 :			gmcminSAD = gmcSAD;
3086 :	syskin	1084	direction = 128;
3087 :	chl	1077	}
3088 :			}
3089 :			if (direction) continue;
3090 :	syskin	1084
3091 :	chl	1077	currwp = centerwp; currwp.duv[2].x--;
3092 :			gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3093 :			if (gmcSAD < gmcminSAD)
3094 :			{ bestwp = currwp;
3095 :			gmcminSAD = gmcSAD;
3096 :	syskin	1084	direction = 256;
3097 :			}
3098 :			else
3099 :	chl	1077	{
3100 :			currwp = centerwp; currwp.duv[2].x++;
3101 :			gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3102 :			if (gmcSAD < gmcminSAD)
3103 :			{ bestwp = currwp;
3104 :			gmcminSAD = gmcSAD;
3105 :	syskin	1084	direction = 512;
3106 :	chl	1077	}
3107 :			}
3108 :			if (direction) continue;
3109 :	syskin	1084
3110 :	chl	1077	currwp = centerwp; currwp.duv[2].y--;
3111 :			gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3112 :			if (gmcSAD < gmcminSAD)
3113 :			{ bestwp = currwp;
3114 :			gmcminSAD = gmcSAD;
3115 :	syskin	1084	direction = 1024;
3116 :			}
3117 :	chl	1077	else
3118 :	syskin	1084	{
3119 :	chl	1077	currwp = centerwp; currwp.duv[2].y++;
3120 :			gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3121 :			if (gmcSAD < gmcminSAD)
3122 :			{ bestwp = currwp;
3123 :			gmcminSAD = gmcSAD;
3124 :	syskin	1084	direction = 2048;
3125 :	chl	1077	}
3126 :			}
3127 :			} while (direction);
3128 :			free(GMCblock);
3129 :	syskin	1084
3130 :	chl	1077	*startwp = bestwp;
3131 :
3132 :			return gmcminSAD;
3133 :			}
3134 :
3135 :			int
3136 :	syskin	1084	globalSAD(const WARPPOINTS *const wp,
3137 :			const MBParam * const pParam,
3138 :	chl	1077	const MACROBLOCK * const pMBs,
3139 :			const FRAMEINFO * const current,
3140 :			const IMAGE * const pRef,
3141 :			const IMAGE * const pCurr,
3142 :			uint8_t *const GMCblock)
3143 :			{
3144 :			NEW_GMC_DATA gmc_data;
3145 :			int iSAD, gmcSAD=0;
3146 :			int num=0;
3147 :			unsigned int mx, my;
3148 :	syskin	1084
3149 :	chl	1077	generate_GMCparameters( 3, 3, wp, pParam->width, pParam->height, &gmc_data);
3150 :	syskin	1084
3151 :	chl	1077	for (my = 0; my < (uint32_t)pParam->mb_height; my++)
3152 :			for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++) {
3153 :
3154 :			const int mbnum = mx + my * pParam->mb_width;
3155 :			const int iEdgedWidth = pParam->edged_width;
3156 :	syskin	1084
3157 :	chl	1077	if (!pMBs[mbnum].mcsel)
3158 :			continue;
3159 :	syskin	1084
3160 :			gmc_data.predict_16x16(&gmc_data, GMCblock,
3161 :	chl	1077	pRef->y,
3162 :	syskin	1084	iEdgedWidth,
3163 :			iEdgedWidth,
3164 :			mx, my,
3165 :	chl	1077	pParam->m_rounding_type);
3166 :
3167 :			iSAD = sad16 ( pCurr->y + 16(myiEdgedWidth + mx),
3168 :	syskin	1084	GMCblock , iEdgedWidth, 65536);
3169 :	chl	1077	iSAD -= pMBs[mbnum].sad16;
3170 :	syskin	1084
3171 :			if (iSAD<0)
3172 :	chl	1077	gmcSAD += iSAD;
3173 :			num++;
3174 :			}
3175 :			return gmcSAD;
3176 :			}
3177 :

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