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

Revision 963 - (view) (download)

1 :	edgomez	851	/**************************************************************************
2 :	Isibaar	3	*
3 :	edgomez	851	* XVID MPEG-4 VIDEO CODEC
4 :			* motion estimation
5 :	chl	259	*
6 :	edgomez	851	* This program is an implementation of a part of one or more MPEG-4
7 :			* Video tools as specified in ISO/IEC 14496-2 standard. Those intending
8 :			* to use this software module in hardware or software products are
9 :			* advised that its use may infringe existing patents or copyrights, and
10 :			* any such use would be at such party's own risk. The original
11 :			* developer of this software module and his/her company, and subsequent
12 :			* editors and their companies, will have no liability for use of this
13 :			* software or modifications or derivatives thereof.
14 :	chl	259	*
15 :	edgomez	851	* This program is free software; you can redistribute it and/or modify
16 :			* it under the terms of the GNU General Public License as published by
17 :			* the Free Software Foundation; either version 2 of the License, or
18 :			* (at your option) any later version.
19 :	chl	259	*
20 :	edgomez	851	* This program is distributed in the hope that it will be useful,
21 :			* but WITHOUT ANY WARRANTY; without even the implied warranty of
22 :			* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
23 :			* GNU General Public License for more details.
24 :	chl	259	*
25 :	edgomez	851	* You should have received a copy of the GNU General Public License
26 :			* along with this program; if not, write to the Free Software
27 :			* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
28 :	chl	259	*
29 :			*************************************************************************/
30 :
31 :	Isibaar	3	#include <assert.h>
32 :			#include <stdio.h>
33 :	chl	96	#include <stdlib.h>
34 :	edgomez	851	#include <string.h> // memcpy
35 :			#include <math.h> // lrint
36 :	Isibaar	3
37 :			#include "../encoder.h"
38 :			#include "../utils/mbfunctions.h"
39 :			#include "../prediction/mbprediction.h"
40 :			#include "../global.h"
41 :			#include "../utils/timer.h"
42 :	edgomez	851	#include "../image/interpolate8x8.h"
43 :			#include "motion_est.h"
44 :	suxen_drol	118	#include "motion.h"
45 :	Isibaar	3	#include "sad.h"
46 :	edgomez	851	#include "../utils/emms.h"
47 :			#include "../dct/fdct.h"
48 :	Isibaar	3
49 :	edgomez	872	/*****************************************************************************
50 :			* Modified rounding tables -- declared in motion.h
51 :			* Original tables see ISO spec tables 7-6 -> 7-9
52 :			****************************************************************************/
53 :
54 :			const uint32_t roundtab[16] =
55 :			{0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2 };
56 :
57 :			/* K = 4 */
58 :			const uint32_t roundtab_76[16] =
59 :			{ 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1 };
60 :
61 :			/* K = 2 */
62 :			const uint32_t roundtab_78[8] =
63 :			{ 0, 0, 1, 1, 0, 0, 0, 1 };
64 :
65 :			/* K = 1 */
66 :			const uint32_t roundtab_79[4] =
67 :			{ 0, 1, 0, 0 };
68 :
69 :	edgomez	851	#define INITIAL_SKIP_THRESH (10)
70 :			#define FINAL_SKIP_THRESH (50)
71 :			#define MAX_SAD00_FOR_SKIP (20)
72 :			#define MAX_CHROMA_SAD_FOR_SKIP (22)
73 :	Isibaar	3
74 :	edgomez	851	#define CHECK_CANDIDATE(X,Y,D) { \
75 :			CheckCandidate((X),(Y), (D), &iDirection, data ); }
76 :	Isibaar	3
77 :	edgomez	872	/*****************************************************************************
78 :			* Code
79 :			****************************************************************************/
80 :
81 :	edgomez	851	static __inline uint32_t
82 :			d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv)
83 :			{
84 :			int xb, yb;
85 :			x = qpel ? x<<1 : x;
86 :			y = qpel ? y<<1 : y;
87 :			if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); }
88 :	chl	141
89 :	edgomez	851	x -= pred.x;
90 :			y -= pred.y;
91 :	chl	141
92 :	edgomez	851	if (x) {
93 :			x = ABS(x);
94 :			x += (1 << (iFcode - 1)) - 1;
95 :			x >>= (iFcode - 1);
96 :			if (x > 32) x = 32;
97 :			xb = mvtab[x] + iFcode;
98 :			} else xb = 1;
99 :	chl	141
100 :	edgomez	851	if (y) {
101 :			y = ABS(y);
102 :			y += (1 << (iFcode - 1)) - 1;
103 :			y >>= (iFcode - 1);
104 :			if (y > 32) y = 32;
105 :			yb = mvtab[y] + iFcode;
106 :			} else yb = 1;
107 :			return xb + yb;
108 :			}
109 :	chl	141
110 :	edgomez	851	static int32_t ChromaSAD2(int fx, int fy, int bx, int by, const SearchData * const data)
111 :			{
112 :			int sad;
113 :			const uint32_t stride = data->iEdgedWidth/2;
114 :			uint8_t * f_refu = data->RefQ,
115 :			* f_refv = data->RefQ + 8,
116 :			* b_refu = data->RefQ + 16,
117 :			* b_refv = data->RefQ + 24;
118 :	Isibaar	3
119 :	edgomez	851	switch (((fx & 1) << 1) \| (fy & 1)) {
120 :			case 0:
121 :			fx = fx / 2; fy = fy / 2;
122 :			f_refu = (uint8_t)data->RefCU + fy stride + fx, stride;
123 :			f_refv = (uint8_t)data->RefCV + fy stride + fx, stride;
124 :			break;
125 :			case 1:
126 :			fx = fx / 2; fy = (fy - 1) / 2;
127 :			interpolate8x8_halfpel_v(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding);
128 :			interpolate8x8_halfpel_v(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding);
129 :			break;
130 :			case 2:
131 :			fx = (fx - 1) / 2; fy = fy / 2;
132 :			interpolate8x8_halfpel_h(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding);
133 :			interpolate8x8_halfpel_h(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding);
134 :			break;
135 :			default:
136 :			fx = (fx - 1) / 2; fy = (fy - 1) / 2;
137 :			interpolate8x8_halfpel_hv(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding);
138 :			interpolate8x8_halfpel_hv(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding);
139 :			break;
140 :			}
141 :	Isibaar	3
142 :	edgomez	851	switch (((bx & 1) << 1) \| (by & 1)) {
143 :			case 0:
144 :			bx = bx / 2; by = by / 2;
145 :			b_refu = (uint8_t)data->b_RefCU + by stride + bx, stride;
146 :			b_refv = (uint8_t)data->b_RefCV + by stride + bx, stride;
147 :			break;
148 :			case 1:
149 :			bx = bx / 2; by = (by - 1) / 2;
150 :			interpolate8x8_halfpel_v(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding);
151 :			interpolate8x8_halfpel_v(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding);
152 :			break;
153 :			case 2:
154 :			bx = (bx - 1) / 2; by = by / 2;
155 :			interpolate8x8_halfpel_h(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding);
156 :			interpolate8x8_halfpel_h(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding);
157 :			break;
158 :			default:
159 :			bx = (bx - 1) / 2; by = (by - 1) / 2;
160 :			interpolate8x8_halfpel_hv(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding);
161 :			interpolate8x8_halfpel_hv(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding);
162 :			break;
163 :			}
164 :	Isibaar	3
165 :	edgomez	851	sad = sad8bi(data->CurU, b_refu, f_refu, stride);
166 :			sad += sad8bi(data->CurV, b_refv, f_refv, stride);
167 :	Isibaar	3
168 :	edgomez	851	return sad;
169 :			}
170 :	Isibaar	3
171 :
172 :	edgomez	851	static int32_t
173 :			ChromaSAD(int dx, int dy, const SearchData * const data)
174 :			{
175 :			int sad;
176 :			const uint32_t stride = data->iEdgedWidth/2;
177 :	Isibaar	3
178 :	edgomez	851	if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently
179 :			data->temp[5] = dx; data->temp[6] = dy; // backup
180 :	Isibaar	3
181 :	edgomez	851	switch (((dx & 1) << 1) \| (dy & 1)) {
182 :			case 0:
183 :			dx = dx / 2; dy = dy / 2;
184 :			sad = sad8(data->CurU, data->RefCU + dy * stride + dx, stride);
185 :			sad += sad8(data->CurV, data->RefCV + dy * stride + dx, stride);
186 :			break;
187 :			case 1:
188 :			dx = dx / 2; dy = (dy - 1) / 2;
189 :			sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + (dy+1) * stride + dx, stride);
190 :			sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + (dy+1) * stride + dx, stride);
191 :			break;
192 :			case 2:
193 :			dx = (dx - 1) / 2; dy = dy / 2;
194 :			sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + dy * stride + dx+1, stride);
195 :			sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + dy * stride + dx+1, stride);
196 :			break;
197 :			default:
198 :			dx = (dx - 1) / 2; dy = (dy - 1) / 2;
199 :			interpolate8x8_halfpel_hv(data->RefQ, data->RefCU + dy * stride + dx, stride, data->rounding);
200 :			sad = sad8(data->CurU, data->RefQ, stride);
201 :
202 :			interpolate8x8_halfpel_hv(data->RefQ, data->RefCV + dy * stride + dx, stride, data->rounding);
203 :			sad += sad8(data->CurV, data->RefQ, stride);
204 :			break;
205 :			}
206 :			data->temp[7] = sad; //backup, part 2
207 :			return sad;
208 :	Isibaar	3	}
209 :
210 :	edgomez	851	static __inline const uint8_t *
211 :			GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data)
212 :	Isibaar	3	{
213 :	edgomez	851	// dir : 0 = forward, 1 = backward
214 :			switch ( (dir << 2) \| ((x&1)<<1) \| (y&1) ) {
215 :			case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth);
216 :			case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth);
217 :			case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth);
218 :			case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth);
219 :			case 4 : return data->bRef + x/2 + (y/2)*(data->iEdgedWidth);
220 :			case 5 : return data->bRefV + x/2 + ((y-1)/2)*(data->iEdgedWidth);
221 :			case 6 : return data->bRefH + (x-1)/2 + (y/2)*(data->iEdgedWidth);
222 :			default : return data->bRefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth);
223 :			}
224 :	Isibaar	3	}
225 :
226 :	edgomez	851	// this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate
227 :			static __inline const uint8_t *
228 :			GetReference(const int x, const int y, const SearchData * const data)
229 :	Isibaar	3	{
230 :	edgomez	851	switch ( ((x&1)<<1) \| (y&1) ) {
231 :			case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth);
232 :			case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth);
233 :			case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth);
234 :			default : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); //case 2
235 :			}
236 :	Isibaar	3	}
237 :	edgomez	851
238 :			static uint8_t *
239 :			Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data)
240 :	Isibaar	3	{
241 :	edgomez	851	// create or find a qpel-precision reference picture; return pointer to it
242 :			uint8_t * Reference = data->RefQ + 16*dir;
243 :			const uint32_t iEdgedWidth = data->iEdgedWidth;
244 :			const uint32_t rounding = data->rounding;
245 :			const int halfpel_x = x/2;
246 :			const int halfpel_y = y/2;
247 :			const uint8_t ref1, ref2, ref3, ref4;
248 :	suxen_drol	136
249 :	edgomez	851	ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
250 :			ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
251 :			switch( ((x&1)<<1) + (y&1) ) {
252 :	edgomez	963	case 3: // x and y in qpel resolution - the "corners" (top left/right and
253 :			// bottom left/right) during qpel refinement
254 :			ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
255 :			ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
256 :			ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
257 :			ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
258 :			ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
259 :			ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
260 :			interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
261 :	edgomez	851	break;
262 :	chl	184
263 :	edgomez	851	case 1: // x halfpel, y qpel - top or bottom during qpel refinement
264 :			ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
265 :			ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
266 :			interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
267 :			break;
268 :	suxen_drol	136
269 :	edgomez	851	case 2: // x qpel, y halfpel - left or right during qpel refinement
270 :			ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
271 :			ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
272 :			interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
273 :			break;
274 :	edgomez	195
275 :	edgomez	963	default: // pure halfpel position
276 :			return (uint8_t *) ref1;
277 :
278 :	edgomez	851	}
279 :			return Reference;
280 :			}
281 :	Isibaar	3
282 :	edgomez	851	static uint8_t *
283 :			Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data)
284 :			{
285 :			// create or find a qpel-precision reference picture; return pointer to it
286 :			uint8_t * Reference = data->RefQ + 16*dir;
287 :			const uint32_t iEdgedWidth = data->iEdgedWidth;
288 :			const uint32_t rounding = data->rounding;
289 :			const int halfpel_x = x/2;
290 :			const int halfpel_y = y/2;
291 :			const uint8_t ref1, ref2, ref3, ref4;
292 :	chl	326
293 :	edgomez	851	ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
294 :			switch( ((x&1)<<1) + (y&1) ) {
295 :			case 3: // x and y in qpel resolution - the "corners" (top left/right and
296 :			// bottom left/right) during qpel refinement
297 :			ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
298 :			ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
299 :			ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
300 :			interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
301 :			interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding);
302 :			interpolate8x8_avg4(Reference+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8iEdgedWidth, ref3+8iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding);
303 :			interpolate8x8_avg4(Reference+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8iEdgedWidth+8, ref3+8iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding);
304 :			break;
305 :	chl	184
306 :	edgomez	851	case 1: // x halfpel, y qpel - top or bottom during qpel refinement
307 :			ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
308 :			interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
309 :			interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
310 :			interpolate8x8_avg2(Reference+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
311 :			interpolate8x8_avg2(Reference+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
312 :			break;
313 :	chl	184
314 :	edgomez	851	case 2: // x qpel, y halfpel - left or right during qpel refinement
315 :			ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
316 :			interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
317 :			interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
318 :			interpolate8x8_avg2(Reference+8iEdgedWidth, ref1+8iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
319 :			interpolate8x8_avg2(Reference+8iEdgedWidth+8, ref1+8iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
320 :			break;
321 :	edgomez	195
322 :	edgomez	963	default: // pure halfpel position
323 :	edgomez	851	return (uint8_t *) ref1;
324 :			}
325 :			return Reference;
326 :			}
327 :	edgomez	195
328 :	edgomez	851	/* CHECK_CANDIATE FUNCTIONS START */
329 :	edgomez	195
330 :	edgomez	851	static void
331 :			CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
332 :			{
333 :			int xc, yc;
334 :			const uint8_t * Reference;
335 :			VECTOR * current;
336 :			int32_t sad; uint32_t t;
337 :	edgomez	195
338 :	edgomez	851	if ( (x > data->max_dx) \|\| (x < data->min_dx)
339 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
340 :	edgomez	195
341 :	edgomez	851	if (!data->qpel_precision) {
342 :			Reference = GetReference(x, y, data);
343 :			current = data->currentMV;
344 :			xc = x; yc = y;
345 :			} else { // x and y are in 1/4 precision
346 :			Reference = Interpolate16x16qpel(x, y, 0, data);
347 :			xc = x/2; yc = y/2; //for chroma sad
348 :			current = data->currentQMV;
349 :			}
350 :	edgomez	195
351 :	edgomez	851	sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
352 :			t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
353 :	edgomez	195
354 :	edgomez	851	sad += (data->lambda16 * t * sad)>>10;
355 :			data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
356 :	chl	175
357 :	edgomez	851	if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
358 :	edgomez	963	(yc >> 1) + roundtab_79[yc & 0x3], data);
359 :	edgomez	195
360 :	edgomez	851	if (sad < data->iMinSAD[0]) {
361 :			data->iMinSAD[0] = sad;
362 :			current[0].x = x; current[0].y = y;
363 :			*dir = Direction;
364 :			}
365 :	edgomez	195
366 :	edgomez	851	if (data->temp[1] < data->iMinSAD[1]) {
367 :			data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y = y; }
368 :			if (data->temp[2] < data->iMinSAD[2]) {
369 :			data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; }
370 :			if (data->temp[3] < data->iMinSAD[3]) {
371 :			data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; }
372 :			if (data->temp[4] < data->iMinSAD[4]) {
373 :			data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; }
374 :
375 :	Isibaar	3	}
376 :
377 :	edgomez	851	static void
378 :			CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
379 :			{
380 :			int32_t sad; uint32_t t;
381 :			const uint8_t * Reference;
382 :	chl	326
383 :	edgomez	851	if ( (x > data->max_dx) \|\| (x < data->min_dx)
384 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
385 :	Isibaar	3
386 :	edgomez	851	if (!data->qpel_precision) Reference = GetReference(x, y, data);
387 :			else Reference = Interpolate8x8qpel(x, y, 0, 0, data);
388 :	Isibaar	3
389 :	edgomez	851	sad = sad8(data->Cur, Reference, data->iEdgedWidth);
390 :			t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
391 :	Isibaar	3
392 :	edgomez	851	sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10;
393 :	Isibaar	3
394 :	edgomez	851	if (sad < *(data->iMinSAD)) {
395 :			*(data->iMinSAD) = sad;
396 :			data->currentMV->x = x; data->currentMV->y = y;
397 :			*dir = Direction;
398 :			}
399 :	Isibaar	3	}
400 :
401 :
402 :	edgomez	851	static void
403 :			CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
404 :			{
405 :			uint32_t t;
406 :			const uint8_t * Reference;
407 :
408 :	edgomez	963	if ( (!(x&1) && x !=0) \|\| (!(y&1) && y !=0) \|\| //non-zero even value
409 :	edgomez	851	(x > data->max_dx) \|\| (x < data->min_dx)
410 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
411 :
412 :			Reference = GetReference(x, y, data);
413 :			t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1);
414 :
415 :			data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
416 :
417 :			data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10;
418 :			data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
419 :
420 :			if (data->temp[0] < data->iMinSAD[0]) {
421 :			data->iMinSAD[0] = data->temp[0];
422 :			data->currentMV[0].x = x; data->currentMV[0].y = y;
423 :			*dir = Direction; }
424 :
425 :			if (data->temp[1] < data->iMinSAD[1]) {
426 :			data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
427 :			if (data->temp[2] < data->iMinSAD[2]) {
428 :			data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
429 :			if (data->temp[3] < data->iMinSAD[3]) {
430 :			data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
431 :			if (data->temp[4] < data->iMinSAD[4]) {
432 :			data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
433 :	Isibaar	3	}
434 :	edgomez	195
435 :	edgomez	851	static void
436 :			CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
437 :			{
438 :			int32_t sad, xc, yc;
439 :			const uint8_t * Reference;
440 :			uint32_t t;
441 :			VECTOR * current;
442 :
443 :	edgomez	963	if ( (x > data->max_dx) \|\| ( x < data->min_dx)
444 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
445 :	edgomez	851
446 :			if (data->rrv && (!(x&1) && x !=0) \| (!(y&1) && y !=0) ) return; //non-zero even value
447 :
448 :			if (data->qpel_precision) { // x and y are in 1/4 precision
449 :			Reference = Interpolate16x16qpel(x, y, 0, data);
450 :			current = data->currentQMV;
451 :			xc = x/2; yc = y/2;
452 :			} else {
453 :			Reference = GetReference(x, y, data);
454 :			current = data->currentMV;
455 :			xc = x; yc = y;
456 :			}
457 :			t = d_mv_bits(x, y, data->predMV, data->iFcode,
458 :			data->qpel^data->qpel_precision, data->rrv);
459 :
460 :			sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096);
461 :			sad += (data->lambda16 * t * sad)>>10;
462 :
463 :			if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
464 :			(yc >> 1) + roundtab_79[yc & 0x3], data);
465 :
466 :			if (sad < *(data->iMinSAD)) {
467 :			*(data->iMinSAD) = sad;
468 :			current->x = x; current->y = y;
469 :			*dir = Direction;
470 :			}
471 :	chl	96	}
472 :	Isibaar	3
473 :	edgomez	851	static void
474 :			CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
475 :			{
476 :			// maximum speed - for P/B/I decision
477 :			int32_t sad;
478 :
479 :			if ( (x > data->max_dx) \|\| (x < data->min_dx)
480 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
481 :
482 :			sad = sad32v_c(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth),
483 :			data->iEdgedWidth, data->temp+1);
484 :
485 :			if (sad < *(data->iMinSAD)) {
486 :			*(data->iMinSAD) = sad;
487 :			data->currentMV[0].x = x; data->currentMV[0].y = y;
488 :			*dir = Direction;
489 :			}
490 :			if (data->temp[1] < data->iMinSAD[1]) {
491 :			data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
492 :			if (data->temp[2] < data->iMinSAD[2]) {
493 :			data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
494 :			if (data->temp[3] < data->iMinSAD[3]) {
495 :			data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
496 :			if (data->temp[4] < data->iMinSAD[4]) {
497 :			data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
498 :
499 :	Isibaar	3	}
500 :
501 :	edgomez	851	static void
502 :			CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data)
503 :			{
504 :			int32_t sad, xb, yb, xcf, ycf, xcb, ycb;
505 :			uint32_t t;
506 :			const uint8_t ReferenceF, ReferenceB;
507 :			VECTOR *current;
508 :
509 :	edgomez	963	if ((xf > data->max_dx) \|\| (xf < data->min_dx) \|\|
510 :			(yf > data->max_dy) \|\| (yf < data->min_dy))
511 :			return;
512 :	edgomez	851
513 :			if (!data->qpel_precision) {
514 :			ReferenceF = GetReference(xf, yf, data);
515 :			xb = data->currentMV[1].x; yb = data->currentMV[1].y;
516 :			ReferenceB = GetReferenceB(xb, yb, 1, data);
517 :			current = data->currentMV;
518 :			xcf = xf; ycf = yf;
519 :			xcb = xb; ycb = yb;
520 :			} else {
521 :			ReferenceF = Interpolate16x16qpel(xf, yf, 0, data);
522 :			xb = data->currentQMV[1].x; yb = data->currentQMV[1].y;
523 :			current = data->currentQMV;
524 :			ReferenceB = Interpolate16x16qpel(xb, yb, 1, data);
525 :			xcf = xf/2; ycf = yf/2;
526 :			xcb = xb/2; ycb = yb/2;
527 :			}
528 :
529 :			t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0)
530 :			+ d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0);
531 :
532 :			sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
533 :			sad += (data->lambda16 * t * sad)>>10;
534 :
535 :			if (data->chroma) sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3],
536 :			(ycf >> 1) + roundtab_79[ycf & 0x3],
537 :			(xcb >> 1) + roundtab_79[xcb & 0x3],
538 :			(ycb >> 1) + roundtab_79[ycb & 0x3], data);
539 :
540 :			if (sad < *(data->iMinSAD)) {
541 :			*(data->iMinSAD) = sad;
542 :			current->x = xf; current->y = yf;
543 :			*dir = Direction;
544 :			}
545 :	Isibaar	3	}
546 :
547 :	edgomez	851	static void
548 :			CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
549 :			{
550 :			int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0;
551 :			uint32_t k;
552 :			const uint8_t *ReferenceF;
553 :			const uint8_t *ReferenceB;
554 :			VECTOR mvs, b_mvs;
555 :
556 :	syskin	904	if (( x > 31) \|\| ( x < -32) \|\| ( y > 31) \|\| (y < -32)) return;
557 :	edgomez	851
558 :			for (k = 0; k < 4; k++) {
559 :			mvs.x = data->directmvF[k].x + x;
560 :			b_mvs.x = ((x == 0) ?
561 :			data->directmvB[k].x
562 :			: mvs.x - data->referencemv[k].x);
563 :
564 :			mvs.y = data->directmvF[k].y + y;
565 :			b_mvs.y = ((y == 0) ?
566 :			data->directmvB[k].y
567 :			: mvs.y - data->referencemv[k].y);
568 :
569 :	edgomez	963	if ((mvs.x > data->max_dx) \|\| (mvs.x < data->min_dx) \|\|
570 :			(mvs.y > data->max_dy) \|\| (mvs.y < data->min_dy) \|\|
571 :			(b_mvs.x > data->max_dx) \|\| (b_mvs.x < data->min_dx) \|\|
572 :			(b_mvs.y > data->max_dy) \|\| (b_mvs.y < data->min_dy) )
573 :			return;
574 :	edgomez	851
575 :			if (data->qpel) {
576 :			xcf += mvs.x/2; ycf += mvs.y/2;
577 :			xcb += b_mvs.x/2; ycb += b_mvs.y/2;
578 :			} else {
579 :			xcf += mvs.x; ycf += mvs.y;
580 :			xcb += b_mvs.x; ycb += b_mvs.y;
581 :			mvs.x = 2; mvs.y = 2; //we move to qpel precision anyway
582 :			b_mvs.x = 2; b_mvs.y = 2;
583 :			}
584 :
585 :			ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data);
586 :			ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data);
587 :
588 :			sad += sad8bi(data->Cur + 8(k&1) + 8(k>>1)*(data->iEdgedWidth),
589 :			ReferenceF, ReferenceB, data->iEdgedWidth);
590 :			if (sad > *(data->iMinSAD)) return;
591 :			}
592 :
593 :			sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
594 :
595 :			if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
596 :			(ycf >> 3) + roundtab_76[ycf & 0xf],
597 :			(xcb >> 3) + roundtab_76[xcb & 0xf],
598 :			(ycb >> 3) + roundtab_76[ycb & 0xf], data);
599 :
600 :			if (sad < *(data->iMinSAD)) {
601 :			*(data->iMinSAD) = sad;
602 :			data->currentMV->x = x; data->currentMV->y = y;
603 :			*dir = Direction;
604 :			}
605 :	Isibaar	3	}
606 :
607 :	edgomez	851	static void
608 :			CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
609 :	Isibaar	3	{
610 :	edgomez	851	int32_t sad, xcf, ycf, xcb, ycb;
611 :			const uint8_t *ReferenceF;
612 :			const uint8_t *ReferenceB;
613 :			VECTOR mvs, b_mvs;
614 :	Isibaar	3
615 :	edgomez	963	if (( x > 31) \|\| ( x < -32) \|\| ( y > 31) \|\| (y < -32)) return;
616 :	Isibaar	3
617 :	edgomez	851	mvs.x = data->directmvF[0].x + x;
618 :			b_mvs.x = ((x == 0) ?
619 :			data->directmvB[0].x
620 :			: mvs.x - data->referencemv[0].x);
621 :	Isibaar	3
622 :	edgomez	851	mvs.y = data->directmvF[0].y + y;
623 :			b_mvs.y = ((y == 0) ?
624 :			data->directmvB[0].y
625 :			: mvs.y - data->referencemv[0].y);
626 :	Isibaar	3
627 :	edgomez	963	if ( (mvs.x > data->max_dx) \|\| (mvs.x < data->min_dx)
628 :			\|\| (mvs.y > data->max_dy) \|\| (mvs.y < data->min_dy)
629 :			\|\| (b_mvs.x > data->max_dx) \|\| (b_mvs.x < data->min_dx)
630 :			\|\| (b_mvs.y > data->max_dy) \|\| (b_mvs.y < data->min_dy) ) return;
631 :	Isibaar	3
632 :	edgomez	851	if (data->qpel) {
633 :			xcf = 4(mvs.x/2); ycf = 4(mvs.y/2);
634 :			xcb = 4(b_mvs.x/2); ycb = 4(b_mvs.y/2);
635 :			ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data);
636 :			ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data);
637 :			} else {
638 :			xcf = 4mvs.x; ycf = 4mvs.y;
639 :			xcb = 4b_mvs.x; ycb = 4b_mvs.y;
640 :			ReferenceF = GetReference(mvs.x, mvs.y, data);
641 :			ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data);
642 :			}
643 :
644 :			sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
645 :			sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
646 :
647 :			if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
648 :			(ycf >> 3) + roundtab_76[ycf & 0xf],
649 :			(xcb >> 3) + roundtab_76[xcb & 0xf],
650 :			(ycb >> 3) + roundtab_76[ycb & 0xf], data);
651 :
652 :			if (sad < *(data->iMinSAD)) {
653 :			*(data->iMinSAD) = sad;
654 :			data->currentMV->x = x; data->currentMV->y = y;
655 :			*dir = Direction;
656 :			}
657 :	Isibaar	3	}
658 :
659 :	edgomez	851
660 :			static void
661 :			CheckCandidateBits16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
662 :	Isibaar	3	{
663 :
664 :	edgomez	961	int16_t in = data->dctSpace, coeff = data->dctSpace + 64;
665 :	edgomez	851	int32_t bits = 0, sum;
666 :			VECTOR * current;
667 :			const uint8_t * ptr;
668 :			int i, cbp = 0, t, xc, yc;
669 :	edgomez	195
670 :	edgomez	851	if ( (x > data->max_dx) \|\| (x < data->min_dx)
671 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
672 :	edgomez	195
673 :	edgomez	851	if (!data->qpel_precision) {
674 :			ptr = GetReference(x, y, data);
675 :			current = data->currentMV;
676 :			xc = x; yc = y;
677 :			} else { // x and y are in 1/4 precision
678 :			ptr = Interpolate16x16qpel(x, y, 0, data);
679 :			current = data->currentQMV;
680 :			xc = x/2; yc = y/2;
681 :			}
682 :	Isibaar	3
683 :	edgomez	851	for(i = 0; i < 4; i++) {
684 :			int s = 8((i&1) + (i>>1)data->iEdgedWidth);
685 :			transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth);
686 :			fdct(in);
687 :			if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
688 :			else sum = quant4_inter(coeff, in, data->lambda16);
689 :			if (sum > 0) {
690 :			cbp \|= 1 << (5 - i);
691 :			bits += data->temp[i] = CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
692 :			} else data->temp[i] = 0;
693 :			}
694 :	Isibaar	3
695 :	edgomez	851	bits += t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
696 :	edgomez	195
697 :	edgomez	851	if (bits < data->iMinSAD[0]) { // there is still a chance, adding chroma
698 :			xc = (xc >> 1) + roundtab_79[xc & 0x3];
699 :			yc = (yc >> 1) + roundtab_79[yc & 0x3];
700 :
701 :			//chroma U
702 :			ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefCU, 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
703 :			transfer_8to16subro(in, ptr, data->CurU, data->iEdgedWidth/2);
704 :			fdct(in);
705 :			if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
706 :			else sum = quant4_inter(coeff, in, data->lambda16);
707 :			if (sum > 0) {
708 :			cbp \|= 1 << (5 - 4);
709 :			bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
710 :	chl	344	}
711 :	edgomez	851
712 :			if (bits < data->iMinSAD[0]) {
713 :			//chroma V
714 :			ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefCV, 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
715 :			transfer_8to16subro(in, ptr, data->CurV, data->iEdgedWidth/2);
716 :			fdct(in);
717 :			if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
718 :			else sum = quant4_inter(coeff, in, data->lambda16);
719 :			if (sum > 0) {
720 :			cbp \|= 1 << (5 - 5);
721 :			bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
722 :			}
723 :			}
724 :	edgomez	78	}
725 :	edgomez	851
726 :	chl	897	bits += xvid_cbpy_tab[15-(cbp>>2)].len;
727 :	edgomez	851	bits += mcbpc_inter_tab[(MODE_INTER & 7) \| ((cbp & 3) << 3)].len;
728 :
729 :			if (bits < data->iMinSAD[0]) {
730 :			data->iMinSAD[0] = bits;
731 :			current[0].x = x; current[0].y = y;
732 :			*dir = Direction;
733 :			}
734 :
735 :			if (data->temp[0] + t < data->iMinSAD[1]) {
736 :			data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; }
737 :			if (data->temp[1] < data->iMinSAD[2]) {
738 :			data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; }
739 :			if (data->temp[2] < data->iMinSAD[3]) {
740 :			data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; }
741 :			if (data->temp[3] < data->iMinSAD[4]) {
742 :			data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; }
743 :
744 :	Isibaar	3	}
745 :	edgomez	851	static void
746 :			CheckCandidateBits8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
747 :	chl	96	{
748 :
749 :	edgomez	961	int16_t in = data->dctSpace, coeff = data->dctSpace + 64;
750 :	edgomez	851	int32_t sum, bits;
751 :			VECTOR * current;
752 :			const uint8_t * ptr;
753 :			int cbp;
754 :	edgomez	195
755 :	edgomez	851	if ( (x > data->max_dx) \|\| (x < data->min_dx)
756 :			\|\| (y > data->max_dy) \|\| (y < data->min_dy) ) return;
757 :	edgomez	195
758 :	edgomez	851	if (!data->qpel_precision) {
759 :			ptr = GetReference(x, y, data);
760 :			current = data->currentMV;
761 :			} else { // x and y are in 1/4 precision
762 :			ptr = Interpolate8x8qpel(x, y, 0, 0, data);
763 :			current = data->currentQMV;
764 :			}
765 :	chl	96
766 :	edgomez	851	transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth);
767 :			fdct(in);
768 :			if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
769 :			else sum = quant4_inter(coeff, in, data->lambda16);
770 :			if (sum > 0) {
771 :			bits = CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
772 :			cbp = 1;
773 :			} else cbp = bits = 0;
774 :	chl	96
775 :	edgomez	851	bits += sum = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
776 :	chl	96
777 :	edgomez	851	if (bits < data->iMinSAD[0]) {
778 :			data->temp[0] = cbp;
779 :			data->iMinSAD[0] = bits;
780 :			current[0].x = x; current[0].y = y;
781 :			*dir = Direction;
782 :			}
783 :			}
784 :	chl	96
785 :	edgomez	851	/* CHECK_CANDIATE FUNCTIONS END */
786 :	edgomez	195
787 :	edgomez	851	/* MAINSEARCH FUNCTIONS START */
788 :	chl	96
789 :	edgomez	851	static void
790 :			AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection)
791 :			{
792 :	edgomez	195
793 :	edgomez	851	/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
794 :	edgomez	195
795 :	edgomez	851	int iDirection;
796 :	edgomez	195
797 :	edgomez	851	for(;;) { //forever
798 :			iDirection = 0;
799 :			if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
800 :			if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
801 :			if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
802 :			if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
803 :	edgomez	195
804 :	edgomez	851	/* now we're doing diagonal checks near our candidate */
805 :	edgomez	195
806 :	edgomez	851	if (iDirection) { //if anything found
807 :			bDirection = iDirection;
808 :			iDirection = 0;
809 :			x = data->currentMV->x; y = data->currentMV->y;
810 :			if (bDirection & 3) { //our candidate is left or right
811 :			CHECK_CANDIDATE(x, y + iDiamondSize, 8);
812 :			CHECK_CANDIDATE(x, y - iDiamondSize, 4);
813 :			} else { // what remains here is up or down
814 :			CHECK_CANDIDATE(x + iDiamondSize, y, 2);
815 :			CHECK_CANDIDATE(x - iDiamondSize, y, 1);
816 :			}
817 :	edgomez	195
818 :	edgomez	851	if (iDirection) {
819 :			bDirection += iDirection;
820 :			x = data->currentMV->x; y = data->currentMV->y;
821 :			}
822 :			} else { //about to quit, eh? not so fast....
823 :			switch (bDirection) {
824 :			case 2:
825 :			CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
826 :			CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
827 :	edgomez	195	break;
828 :	edgomez	851	case 1:
829 :			CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
830 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
831 :	edgomez	195	break;
832 :	edgomez	851	case 2 + 4:
833 :			CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
834 :			CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
835 :			CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
836 :			break;
837 :			case 4:
838 :			CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
839 :			CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
840 :			break;
841 :	edgomez	195	case 8:
842 :	edgomez	851	CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
843 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
844 :	edgomez	195	break;
845 :	edgomez	851	case 1 + 4:
846 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
847 :			CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
848 :			CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
849 :	edgomez	195	break;
850 :	edgomez	851	case 2 + 8:
851 :			CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
852 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
853 :			CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
854 :			break;
855 :			case 1 + 8:
856 :			CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
857 :			CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
858 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
859 :			break;
860 :			default: //1+2+4+8 == we didn't find anything at all
861 :			CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
862 :			CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
863 :			CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
864 :			CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
865 :			break;
866 :	chl	96	}
867 :	edgomez	851	if (!iDirection) break; //ok, the end. really
868 :			bDirection = iDirection;
869 :			x = data->currentMV->x; y = data->currentMV->y;
870 :	chl	345	}
871 :	edgomez	195	}
872 :	chl	96	}
873 :
874 :	edgomez	851	static void
875 :			SquareSearch(int x, int y, const SearchData * const data, int bDirection)
876 :	chl	96	{
877 :	edgomez	851	int iDirection;
878 :	edgomez	195
879 :	edgomez	851	do {
880 :			iDirection = 0;
881 :			if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64);
882 :			if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128);
883 :			if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32);
884 :			if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128);
885 :			if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64);
886 :			if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128);
887 :			if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128);
888 :			if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128);
889 :	chl	96
890 :	edgomez	851	bDirection = iDirection;
891 :			x = data->currentMV->x; y = data->currentMV->y;
892 :			} while (iDirection);
893 :	chl	96	}
894 :
895 :	edgomez	851	static void
896 :			DiamondSearch(int x, int y, const SearchData * const data, int bDirection)
897 :	chl	181	{
898 :
899 :			/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
900 :
901 :	edgomez	851	int iDirection;
902 :	edgomez	195
903 :	edgomez	851	do {
904 :			iDirection = 0;
905 :			if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
906 :			if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
907 :			if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
908 :			if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
909 :
910 :			/* now we're doing diagonal checks near our candidate */
911 :
912 :			if (iDirection) { //checking if anything found
913 :			bDirection = iDirection;
914 :	chl	181	iDirection = 0;
915 :	edgomez	851	x = data->currentMV->x; y = data->currentMV->y;
916 :			if (bDirection & 3) { //our candidate is left or right
917 :			CHECK_CANDIDATE(x, y + iDiamondSize, 8);
918 :			CHECK_CANDIDATE(x, y - iDiamondSize, 4);
919 :			} else { // what remains here is up or down
920 :			CHECK_CANDIDATE(x + iDiamondSize, y, 2);
921 :			CHECK_CANDIDATE(x - iDiamondSize, y, 1);
922 :			}
923 :			bDirection += iDirection;
924 :			x = data->currentMV->x; y = data->currentMV->y;
925 :			}
926 :			}
927 :			while (iDirection);
928 :			}
929 :	chl	181
930 :	edgomez	851	/* MAINSEARCH FUNCTIONS END */
931 :	chl	181
932 :	edgomez	851	static void
933 :			SubpelRefine(const SearchData * const data)
934 :			{
935 :			/* Do a half-pel or q-pel refinement */
936 :			const VECTOR centerMV = data->qpel_precision ? data->currentQMV : data->currentMV;
937 :			int iDirection; //only needed because macro expects it
938 :	chl	181
939 :	edgomez	851	CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0);
940 :			CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0);
941 :			CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0);
942 :			CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0);
943 :			CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0);
944 :			CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0);
945 :			CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0);
946 :			CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0);
947 :			}
948 :	chl	181
949 :	edgomez	851	static __inline int
950 :			SkipDecisionP(const IMAGE * current, const IMAGE * reference,
951 :			const int x, const int y,
952 :			const uint32_t stride, const uint32_t iQuant, int rrv)
953 :	chl	181
954 :	edgomez	851	{
955 :	edgomez	963	int offset = (x + ystride)8;
956 :	edgomez	851	if(!rrv) {
957 :	edgomez	963	uint32_t sadC = sad8(current->u + offset,
958 :			reference->u + offset, stride);
959 :	edgomez	851	if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
960 :	edgomez	963	sadC += sad8(current->v + offset,
961 :			reference->v + offset, stride);
962 :	edgomez	851	if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
963 :			return 1;
964 :	chl	181
965 :	edgomez	851	} else {
966 :	edgomez	963	uint32_t sadC = sad16(current->u + 2*offset,
967 :			reference->u + 2offset, stride, 2564096);
968 :	edgomez	851	if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
969 :	edgomez	963	sadC += sad16(current->v + 2*offset,
970 :			reference->v + 2offset, stride, 2564096);
971 :	edgomez	851	if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
972 :			return 1;
973 :	chl	181	}
974 :	edgomez	851	}
975 :	edgomez	504
976 :	edgomez	851	static __inline void
977 :			SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad)
978 :			{
979 :			pMB->mode = MODE_NOT_CODED;
980 :			pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV;
981 :			pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV;
982 :			pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
983 :	chl	181	}
984 :
985 :	edgomez	851	bool
986 :			MotionEstimation(MBParam * const pParam,
987 :			FRAMEINFO * const current,
988 :			FRAMEINFO * const reference,
989 :			const IMAGE * const pRefH,
990 :			const IMAGE * const pRefV,
991 :			const IMAGE * const pRefHV,
992 :			const uint32_t iLimit)
993 :	chl	181	{
994 :	edgomez	851	MACROBLOCK *const pMBs = current->mbs;
995 :			const IMAGE *const pCurrent = &current->image;
996 :			const IMAGE *const pRef = &reference->image;
997 :	chl	181
998 :	edgomez	851	uint32_t mb_width = pParam->mb_width;
999 :			uint32_t mb_height = pParam->mb_height;
1000 :			const uint32_t iEdgedWidth = pParam->edged_width;
1001 :			const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->global_flags);
1002 :	chl	181
1003 :	edgomez	851	uint32_t x, y;
1004 :			uint32_t iIntra = 0;
1005 :			int32_t quant = current->quant, sad00;
1006 :	syskin	904	int skip_thresh = INITIAL_SKIP_THRESH *
1007 :			(current->global_flags & XVID_REDUCED ? 4:1) *
1008 :			(current->global_flags & XVID_MODEDECISION_BITS ? 2:1);
1009 :	chl	181
1010 :	edgomez	851	// some pre-initialized thingies for SearchP
1011 :			int32_t temp[8];
1012 :			VECTOR currentMV[5];
1013 :			VECTOR currentQMV[5];
1014 :			int32_t iMinSAD[5];
1015 :	edgomez	961	DECLARE_ALIGNED_MATRIX(dct_space, 2, 64, int16_t, CACHE_LINE);
1016 :	edgomez	851	SearchData Data;
1017 :			memset(&Data, 0, sizeof(SearchData));
1018 :			Data.iEdgedWidth = iEdgedWidth;
1019 :			Data.currentMV = currentMV;
1020 :			Data.currentQMV = currentQMV;
1021 :			Data.iMinSAD = iMinSAD;
1022 :			Data.temp = temp;
1023 :			Data.iFcode = current->fcode;
1024 :			Data.rounding = pParam->m_rounding_type;
1025 :			Data.qpel = pParam->m_quarterpel;
1026 :			Data.chroma = MotionFlags & PMV_CHROMA16;
1027 :			Data.rrv = current->global_flags & XVID_REDUCED;
1028 :	edgomez	961	Data.dctSpace = dct_space;
1029 :	edgomez	195
1030 :	edgomez	851	if ((current->global_flags & XVID_REDUCED)) {
1031 :			mb_width = (pParam->width + 31) / 32;
1032 :			mb_height = (pParam->height + 31) / 32;
1033 :			Data.qpel = 0;
1034 :			}
1035 :	chl	181
1036 :	edgomez	851	Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose)
1037 :			if (sadInit) (*sadInit) ();
1038 :	chl	181
1039 :	edgomez	851	for (y = 0; y < mb_height; y++) {
1040 :			for (x = 0; x < mb_width; x++) {
1041 :			MACROBLOCK pMB = &pMBs[x + y pParam->mb_width];
1042 :	chl	181
1043 :	edgomez	851	if (!Data.rrv) pMB->sad16 =
1044 :			sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16,
1045 :			pRef->y + (x + y * iEdgedWidth) * 16,
1046 :			pParam->edged_width, pMB->sad8 );
1047 :	chl	181
1048 :	edgomez	851	else pMB->sad16 =
1049 :			sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32,
1050 :			pRef->y + (x + y * iEdgedWidth) * 32,
1051 :			pParam->edged_width, pMB->sad8 );
1052 :	chl	181
1053 :	edgomez	851	if (Data.chroma) {
1054 :			Data.temp[7] = sad8(pCurrent->u + x8 + y(iEdgedWidth/2)*8,
1055 :			pRef->u + x8 + y(iEdgedWidth/2)*8, iEdgedWidth/2)
1056 :			+ sad8(pCurrent->v + (x + y(iEdgedWidth/2))8,
1057 :			pRef->v + (x + y(iEdgedWidth/2))8, iEdgedWidth/2);
1058 :			pMB->sad16 += Data.temp[7];
1059 :			}
1060 :
1061 :			sad00 = pMB->sad16;
1062 :
1063 :			if (!(current->global_flags & XVID_LUMIMASKING)) {
1064 :			pMB->dquant = NO_CHANGE;
1065 :			} else {
1066 :			if (pMB->dquant != NO_CHANGE) {
1067 :			quant += DQtab[pMB->dquant];
1068 :			if (quant > 31) quant = 31;
1069 :			else if (quant < 1) quant = 1;
1070 :	chl	181	}
1071 :	edgomez	851	}
1072 :			pMB->quant = current->quant;
1073 :	chl	181
1074 :	edgomez	851	//initial skip decision
1075 :			/* no early skip for GMC (global vector = skip vector is unknown!) */
1076 :			if (!(current->global_flags & XVID_GMC)) { /* no fast SKIP for S(GMC)-VOPs */
1077 :	syskin	904	if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * skip_thresh)
1078 :	edgomez	851	if (Data.chroma \|\| SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) {
1079 :			SkipMacroblockP(pMB, sad00);
1080 :			continue;
1081 :			}
1082 :			}
1083 :
1084 :			SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
1085 :			y, MotionFlags, current->global_flags, pMB->quant,
1086 :			&Data, pParam, pMBs, reference->mbs,
1087 :			current->global_flags & XVID_INTER4V, pMB);
1088 :
1089 :			/* final skip decision, a.k.a. "the vector you found, really that good?" */
1090 :	syskin	904	if (!(current->global_flags & XVID_GMC \|\| current->global_flags & XVID_MODEDECISION_BITS)) {
1091 :	edgomez	851	if ( pMB->dquant == NO_CHANGE && sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) {
1092 :	syskin	904	if ( (100pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH (Data.rrv ? 4:1) )
1093 :			if (Data.chroma \|\| SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv))
1094 :			SkipMacroblockP(pMB, sad00);
1095 :	chl	181	}
1096 :			}
1097 :	edgomez	851	if (pMB->mode == MODE_INTRA)
1098 :			if (++iIntra > iLimit) return 1;
1099 :	chl	181	}
1100 :			}
1101 :	edgomez	851
1102 :			if (current->global_flags & XVID_GMC ) /* GMC only for S(GMC)-VOPs */
1103 :			{
1104 :			current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV);
1105 :			}
1106 :			return 0;
1107 :	chl	181	}
1108 :
1109 :
1110 :	edgomez	851	static __inline int
1111 :			make_mask(const VECTOR * const pmv, const int i)
1112 :	chl	96	{
1113 :	edgomez	851	int mask = 255, j;
1114 :			for (j = 0; j < i; j++) {
1115 :			if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already
1116 :			if (pmv[i].x == pmv[j].x) {
1117 :			if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4;
1118 :			else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8;
1119 :			} else
1120 :			if (pmv[i].y == pmv[j].y) {
1121 :			if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1;
1122 :			else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2;
1123 :			}
1124 :			}
1125 :			return mask;
1126 :			}
1127 :	edgomez	195
1128 :	edgomez	851	static __inline void
1129 :			PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount,
1130 :			int iHcount, const MACROBLOCK * const prevMB, int rrv)
1131 :			{
1132 :	chl	96
1133 :	edgomez	851	//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself
1134 :			if (rrv) { iWcount /= 2; iHcount /= 2; }
1135 :	edgomez	195
1136 :	edgomez	851	if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour
1137 :			pmv[5].x = EVEN(pmv[3].x);
1138 :			pmv[5].y = EVEN(pmv[3].y);
1139 :			} else pmv[5].x = pmv[5].y = 0;
1140 :	chl	96
1141 :	edgomez	851	if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour
1142 :			else pmv[3].x = pmv[3].y = 0;
1143 :	chl	96
1144 :	edgomez	851	if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour
1145 :			else pmv[4].x = pmv[4].y = 0;
1146 :	Isibaar	3
1147 :	edgomez	851	// [1] median prediction
1148 :			pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y);
1149 :	edgomez	195
1150 :	edgomez	851	pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask
1151 :	edgomez	195
1152 :	edgomez	851	pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame
1153 :			pmv[2].y = EVEN(prevMB->mvs[0].y);
1154 :
1155 :			if ((x < iWcount-1) && (y < iHcount-1)) {
1156 :			pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame
1157 :			pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y);
1158 :			} else pmv[6].x = pmv[6].y = 0;
1159 :
1160 :			if (rrv) {
1161 :			int i;
1162 :			for (i = 0; i < 7; i++) {
1163 :			pmv[i].x = RRV_MV_SCALEUP(pmv[i].x);
1164 :			pmv[i].y = RRV_MV_SCALEUP(pmv[i].y);
1165 :			}
1166 :			}
1167 :	Isibaar	3	}
1168 :
1169 :	edgomez	851	static int
1170 :			ModeDecision(const uint32_t iQuant, SearchData * const Data,
1171 :			int inter4v,
1172 :			MACROBLOCK * const pMB,
1173 :			const MACROBLOCK * const pMBs,
1174 :			const int x, const int y,
1175 :			const MBParam * const pParam,
1176 :			const uint32_t MotionFlags,
1177 :			const uint32_t GlobalFlags)
1178 :			{
1179 :	Isibaar	3
1180 :	edgomez	851	int mode = MODE_INTER;
1181 :	chl	96
1182 :	edgomez	851	if (!(GlobalFlags & XVID_MODEDECISION_BITS)) { //normal, fast, SAD-based mode decision
1183 :			int sad;
1184 :			int InterBias = MV16_INTER_BIAS;
1185 :			if (inter4v == 0 \|\| Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] +
1186 :			Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) {
1187 :	edgomez	963	mode = MODE_INTER;
1188 :			sad = Data->iMinSAD[0];
1189 :	edgomez	851	} else {
1190 :			mode = MODE_INTER4V;
1191 :			sad = Data->iMinSAD[1] + Data->iMinSAD[2] +
1192 :			Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant;
1193 :			Data->iMinSAD[0] = sad;
1194 :			}
1195 :	chl	326
1196 :	edgomez	851	/* intra decision */
1197 :	Isibaar	3
1198 :	edgomez	851	if (iQuant > 8) InterBias += 100 * (iQuant - 8); // to make high quants work
1199 :			if (y != 0)
1200 :			if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80;
1201 :			if (x != 0)
1202 :			if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80;
1203 :	Isibaar	3
1204 :	edgomez	851	if (Data->chroma) InterBias += 50; // to compensate bigger SAD
1205 :			if (Data->rrv) InterBias *= 4;
1206 :	edgomez	195
1207 :	edgomez	851	if (InterBias < pMB->sad16) {
1208 :			int32_t deviation;
1209 :			if (!Data->rrv) deviation = dev16(Data->Cur, Data->iEdgedWidth);
1210 :			else deviation = dev16(Data->Cur, Data->iEdgedWidth) +
1211 :			dev16(Data->Cur+8, Data->iEdgedWidth) +
1212 :			dev16(Data->Cur + 8*Data->iEdgedWidth, Data->iEdgedWidth) +
1213 :			dev16(Data->Cur+8+8*Data->iEdgedWidth, Data->iEdgedWidth);
1214 :	edgomez	195
1215 :	edgomez	963	if (deviation < (sad - InterBias)) return MODE_INTRA;
1216 :	edgomez	851	}
1217 :			return mode;
1218 :	Isibaar	3
1219 :	edgomez	851	} else {
1220 :	edgomez	195
1221 :	edgomez	851	int bits, intra, i;
1222 :			VECTOR backup[5], *v;
1223 :			Data->lambda16 = iQuant;
1224 :			Data->lambda8 = pParam->m_quant_type;
1225 :	chl	181
1226 :	edgomez	851	v = Data->qpel ? Data->currentQMV : Data->currentMV;
1227 :			for (i = 0; i < 5; i++) {
1228 :			Data->iMinSAD[i] = 256*4096;
1229 :			backup[i] = v[i];
1230 :			}
1231 :	Isibaar	3
1232 :	edgomez	851	bits = CountMBBitsInter(Data, pMBs, x, y, pParam, MotionFlags);
1233 :			if (bits == 0) return MODE_INTER; // quick stop
1234 :	Isibaar	3
1235 :	edgomez	851	if (inter4v) {
1236 :	edgomez	963	int bits_inter4v = CountMBBitsInter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup);
1237 :			if (bits_inter4v < bits) { Data->iMinSAD[0] = bits = bits_inter4v; mode = MODE_INTER4V; }
1238 :	edgomez	851	}
1239 :	edgomez	195
1240 :	Isibaar	3
1241 :	edgomez	851	intra = CountMBBitsIntra(Data);
1242 :	Isibaar	3
1243 :	edgomez	851	if (intra < bits) { *Data->iMinSAD = bits = intra; return MODE_INTRA; }
1244 :
1245 :			return mode;
1246 :	edgomez	195	}
1247 :	edgomez	851	}
1248 :	Isibaar	3
1249 :	edgomez	851	static void
1250 :			SearchP(const IMAGE * const pRef,
1251 :			const uint8_t * const pRefH,
1252 :			const uint8_t * const pRefV,
1253 :			const uint8_t * const pRefHV,
1254 :			const IMAGE * const pCur,
1255 :			const int x,
1256 :			const int y,
1257 :			const uint32_t MotionFlags,
1258 :			const uint32_t GlobalFlags,
1259 :			const uint32_t iQuant,
1260 :			SearchData * const Data,
1261 :			const MBParam * const pParam,
1262 :			const MACROBLOCK * const pMBs,
1263 :			const MACROBLOCK * const prevMBs,
1264 :			int inter4v,
1265 :			MACROBLOCK * const pMB)
1266 :			{
1267 :	Isibaar	3
1268 :	edgomez	851	int i, iDirection = 255, mask, threshA;
1269 :			VECTOR pmv[7];
1270 :	Isibaar	3
1271 :	edgomez	851	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1272 :			pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1273 :	edgomez	195
1274 :	edgomez	851	get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp);
1275 :	Isibaar	3
1276 :	edgomez	851	Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache
1277 :			i = Data->rrv ? 2 : 1;
1278 :			Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i;
1279 :			Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1280 :			Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1281 :	chl	326
1282 :	edgomez	851	Data->Ref = pRef->y + (x + Data->iEdgedWidthy) 16*i;
1283 :			Data->RefH = pRefH + (x + Data->iEdgedWidthy) 16*i;
1284 :			Data->RefV = pRefV + (x + Data->iEdgedWidthy) 16*i;
1285 :			Data->RefHV = pRefHV + (x + Data->iEdgedWidthy) 16*i;
1286 :			Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1287 :			Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1288 :	edgomez	195
1289 :	edgomez	851	Data->lambda16 = lambda_vec16[iQuant];
1290 :			Data->lambda8 = lambda_vec8[iQuant];
1291 :			Data->qpel_precision = 0;
1292 :	edgomez	195
1293 :	edgomez	851	if (pMB->dquant != NO_CHANGE) inter4v = 0;
1294 :	edgomez	195
1295 :	edgomez	963	memset(Data->currentMV, 0, 5*sizeof(VECTOR));
1296 :	chl	169
1297 :	edgomez	851	if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0);
1298 :			else Data->predMV = pmv[0];
1299 :	Isibaar	3
1300 :	edgomez	851	i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0);
1301 :			Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10);
1302 :			Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10);
1303 :			Data->iMinSAD[2] = pMB->sad8[1];
1304 :			Data->iMinSAD[3] = pMB->sad8[2];
1305 :			Data->iMinSAD[4] = pMB->sad8[3];
1306 :	chl	169
1307 :	edgomez	851	if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) \|\| (x \| y)) {
1308 :			threshA = Data->temp[0]; // that's where we keep this SAD atm
1309 :			if (threshA < 512) threshA = 512;
1310 :			else if (threshA > 1024) threshA = 1024;
1311 :			} else
1312 :			threshA = 512;
1313 :	chl	169
1314 :	edgomez	851	PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height,
1315 :			prevMBs + x + y * pParam->mb_width, Data->rrv);
1316 :	chl	169
1317 :	edgomez	851	if (!Data->rrv) {
1318 :			if (inter4v \| Data->chroma) CheckCandidate = CheckCandidate16;
1319 :			else CheckCandidate = CheckCandidate16no4v; //for extra speed
1320 :			} else CheckCandidate = CheckCandidate32;
1321 :	chl	169
1322 :	edgomez	851	/* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/
1323 :	chl	169
1324 :	edgomez	851	for (i = 1; i < 7; i++) {
1325 :			if (!(mask = make_mask(pmv, i)) ) continue;
1326 :			CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1327 :			if (Data->iMinSAD[0] <= threshA) break;
1328 :			}
1329 :	chl	169
1330 :	edgomez	851	if ((Data->iMinSAD[0] <= threshA) \|\|
1331 :			(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) &&
1332 :			(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) {
1333 :			if (!(GlobalFlags & XVID_MODEDECISION_BITS)) inter4v = 0; }
1334 :			else {
1335 :	Isibaar	3
1336 :	edgomez	851	MainSearchFunc * MainSearchPtr;
1337 :			if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1338 :			else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1339 :			else MainSearchPtr = DiamondSearch;
1340 :	Isibaar	3
1341 :	edgomez	851	MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1342 :	Isibaar	3
1343 :	edgomez	851	/* extended search, diamond starting in 0,0 and in prediction.
1344 :			note that this search is/might be done in halfpel positions,
1345 :			which makes it more different than the diamond above */
1346 :	chl	169
1347 :	edgomez	851	if (MotionFlags & PMV_EXTSEARCH16) {
1348 :			int32_t bSAD;
1349 :			VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0];
1350 :			if (Data->rrv) {
1351 :			startMV.x = RRV_MV_SCALEUP(startMV.x);
1352 :			startMV.y = RRV_MV_SCALEUP(startMV.y);
1353 :			}
1354 :			if (!(MVequal(startMV, backupMV))) {
1355 :			bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1356 :	edgomez	195
1357 :	edgomez	851	CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1358 :			MainSearchPtr(startMV.x, startMV.y, Data, 255);
1359 :			if (bSAD < Data->iMinSAD[0]) {
1360 :			Data->currentMV[0] = backupMV;
1361 :			Data->iMinSAD[0] = bSAD; }
1362 :			}
1363 :	chl	169
1364 :	edgomez	851	backupMV = Data->currentMV[0];
1365 :			startMV.x = startMV.y = 1;
1366 :			if (!(MVequal(startMV, backupMV))) {
1367 :			bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1368 :	chl	169
1369 :	edgomez	851	CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1370 :			MainSearchPtr(startMV.x, startMV.y, Data, 255);
1371 :			if (bSAD < Data->iMinSAD[0]) {
1372 :			Data->currentMV[0] = backupMV;
1373 :			Data->iMinSAD[0] = bSAD; }
1374 :	edgomez	195	}
1375 :	edgomez	851	}
1376 :			}
1377 :	edgomez	195
1378 :	edgomez	851	if (MotionFlags & PMV_HALFPELREFINE16)
1379 :			if ((!(MotionFlags & HALFPELREFINE16_BITS)) \|\| Data->iMinSAD[0] < 200*(int)iQuant)
1380 :			SubpelRefine(Data);
1381 :	edgomez	195
1382 :	edgomez	851	for(i = 0; i < 5; i++) {
1383 :			Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors
1384 :			Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
1385 :			}
1386 :	Isibaar	3
1387 :	syskin	904	if (MotionFlags & PMV_QUARTERPELREFINE16) {
1388 :
1389 :			get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1390 :			pParam->width, pParam->height, Data->iFcode, 1, 0);
1391 :
1392 :	edgomez	851	if ((!(MotionFlags & QUARTERPELREFINE16_BITS)) \|\| (Data->iMinSAD[0] < 200*(int)iQuant)) {
1393 :			Data->qpel_precision = 1;
1394 :			SubpelRefine(Data);
1395 :			}
1396 :	syskin	904	}
1397 :	Isibaar	3
1398 :	edgomez	851	if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) && (Data->iMinSAD[0] < (int32_t)iQuant * 30)) inter4v = 0;
1399 :	Isibaar	3
1400 :	edgomez	851	if (inter4v && (!(GlobalFlags & XVID_MODEDECISION_BITS) \|\|
1401 :			(!(MotionFlags & QUARTERPELREFINE8_BITS)) \|\| (!(MotionFlags & HALFPELREFINE8_BITS)) \|\|
1402 :			((!(MotionFlags & EXTSEARCH_BITS)) && (!(MotionFlags&PMV_EXTSEARCH8)) ))) {
1403 :			// if decision is BITS-based and all refinement steps will be done in BITS domain, there is no reason to call this loop
1404 :	Isibaar	3
1405 :	edgomez	851	SearchData Data8;
1406 :			memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data
1407 :	Isibaar	3
1408 :	edgomez	851	Search8(Data, 2x, 2y, MotionFlags, pParam, pMB, pMBs, 0, &Data8);
1409 :			Search8(Data, 2x + 1, 2y, MotionFlags, pParam, pMB, pMBs, 1, &Data8);
1410 :			Search8(Data, 2x, 2y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8);
1411 :			Search8(Data, 2x + 1, 2y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8);
1412 :	chl	181
1413 :	edgomez	851	if ((Data->chroma) && (!(GlobalFlags & XVID_MODEDECISION_BITS))) {
1414 :			// chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, there is no reason to compute it
1415 :			int sumx = 0, sumy = 0;
1416 :			const int div = 1 + Data->qpel;
1417 :			const VECTOR * const mv = Data->qpel ? pMB->qmvs : pMB->mvs;
1418 :	Isibaar	3
1419 :	edgomez	851	for (i = 0; i < 4; i++) {
1420 :			sumx += mv[i].x / div;
1421 :			sumy += mv[i].y / div;
1422 :			}
1423 :	chl	259
1424 :	edgomez	851	Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf],
1425 :			(sumy >> 3) + roundtab_76[sumy & 0xf], Data);
1426 :			}
1427 :			}
1428 :	edgomez	195
1429 :	edgomez	851	inter4v = ModeDecision(iQuant, Data, inter4v, pMB, pMBs, x, y, pParam, MotionFlags, GlobalFlags);
1430 :
1431 :			if (Data->rrv) {
1432 :			Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x);
1433 :			Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y);
1434 :	Isibaar	3	}
1435 :
1436 :	edgomez	851	if (inter4v == MODE_INTER) {
1437 :			pMB->mode = MODE_INTER;
1438 :			pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
1439 :			pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0];
1440 :	Isibaar	3
1441 :	edgomez	851	if(Data->qpel) {
1442 :			pMB->qmvs[0] = pMB->qmvs[1]
1443 :			= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
1444 :			pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x;
1445 :			pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y;
1446 :			} else {
1447 :			pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
1448 :			pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
1449 :	Isibaar	3	}
1450 :
1451 :	edgomez	851	} else if (inter4v == MODE_INTER4V) {
1452 :			pMB->mode = MODE_INTER4V;
1453 :			pMB->sad16 = Data->iMinSAD[0];
1454 :			} else { // INTRA mode
1455 :			SkipMacroblockP(pMB, 0); // not skip, but similar enough
1456 :			pMB->mode = MODE_INTRA;
1457 :			}
1458 :	edgomez	195
1459 :	edgomez	851	}
1460 :
1461 :			static void
1462 :			Search8(const SearchData * const OldData,
1463 :			const int x, const int y,
1464 :			const uint32_t MotionFlags,
1465 :			const MBParam * const pParam,
1466 :			MACROBLOCK * const pMB,
1467 :			const MACROBLOCK * const pMBs,
1468 :			const int block,
1469 :			SearchData * const Data)
1470 :			{
1471 :			int i = 0;
1472 :			Data->iMinSAD = OldData->iMinSAD + 1 + block;
1473 :			Data->currentMV = OldData->currentMV + 1 + block;
1474 :			Data->currentQMV = OldData->currentQMV + 1 + block;
1475 :
1476 :			if(Data->qpel) {
1477 :			Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1478 :			if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y,
1479 :			Data->predMV, Data->iFcode, 0, 0);
1480 :			} else {
1481 :			Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1482 :			if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y,
1483 :			Data->predMV, Data->iFcode, 0, Data->rrv);
1484 :	Isibaar	3	}
1485 :
1486 :	edgomez	851	(Data->iMinSAD) += (Data->lambda8 i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10;
1487 :	Isibaar	3
1488 :	edgomez	851	if (MotionFlags & (PMV_EXTSEARCH8\|PMV_HALFPELREFINE8\|PMV_QUARTERPELREFINE8)) {
1489 :			if (Data->rrv) i = 2; else i = 1;
1490 :	Isibaar	3
1491 :	edgomez	851	Data->Ref = OldData->Ref + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1492 :			Data->RefH = OldData->RefH + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1493 :			Data->RefV = OldData->RefV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1494 :			Data->RefHV = OldData->RefHV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1495 :	Isibaar	3
1496 :	edgomez	851	Data->Cur = OldData->Cur + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1497 :			Data->qpel_precision = 0;
1498 :	Isibaar	3
1499 :	edgomez	851	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1500 :			pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1501 :	Isibaar	3
1502 :	edgomez	851	if (!Data->rrv) CheckCandidate = CheckCandidate8;
1503 :			else CheckCandidate = CheckCandidate16no4v;
1504 :	Isibaar	3
1505 :	edgomez	851	if (MotionFlags & PMV_EXTSEARCH8 && (!(MotionFlags & EXTSEARCH_BITS))) {
1506 :			int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1507 :	Isibaar	3
1508 :	edgomez	851	MainSearchFunc *MainSearchPtr;
1509 :			if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch;
1510 :			else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch;
1511 :			else MainSearchPtr = DiamondSearch;
1512 :	Isibaar	3
1513 :	edgomez	851	MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255);
1514 :	Isibaar	3
1515 :	edgomez	851	if(*(Data->iMinSAD) < temp_sad) {
1516 :			Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1517 :			Data->currentQMV->y = 2 * Data->currentMV->y;
1518 :			}
1519 :			}
1520 :	edgomez	195
1521 :	edgomez	851	if (MotionFlags & PMV_HALFPELREFINE8) {
1522 :			int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1523 :	edgomez	195
1524 :	edgomez	851	SubpelRefine(Data); // perform halfpel refine of current best vector
1525 :	Isibaar	3
1526 :	edgomez	851	if(*(Data->iMinSAD) < temp_sad) { // we have found a better match
1527 :			Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1528 :			Data->currentQMV->y = 2 * Data->currentMV->y;
1529 :			}
1530 :			}
1531 :	Isibaar	3
1532 :	edgomez	851	if (Data->qpel && MotionFlags & PMV_QUARTERPELREFINE8) {
1533 :			Data->qpel_precision = 1;
1534 :			get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1535 :			pParam->width, pParam->height, Data->iFcode, 1, 0);
1536 :			SubpelRefine(Data);
1537 :			}
1538 :			}
1539 :	edgomez	195
1540 :	edgomez	851	if (Data->rrv) {
1541 :			Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x);
1542 :			Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y);
1543 :			}
1544 :
1545 :			if(Data->qpel) {
1546 :			pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x;
1547 :			pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y;
1548 :			pMB->qmvs[block] = *Data->currentQMV;
1549 :	edgomez	195	} else {
1550 :	edgomez	851	pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x;
1551 :			pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y;
1552 :	edgomez	78	}
1553 :	edgomez	851
1554 :			pMB->mvs[block] = *Data->currentMV;
1555 :			pMB->sad8[block] = 4 * *Data->iMinSAD;
1556 :	Isibaar	3	}
1557 :
1558 :	edgomez	851	/* motion estimation for B-frames */
1559 :	chl	345
1560 :	edgomez	851	static __inline VECTOR
1561 :			ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode)
1562 :			{
1563 :			/* the stupidiest function ever */
1564 :			return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]);
1565 :			}
1566 :	chl	345
1567 :	edgomez	851	static void __inline
1568 :			PreparePredictionsBF(VECTOR * const pmv, const int x, const int y,
1569 :			const uint32_t iWcount,
1570 :			const MACROBLOCK * const pMB,
1571 :			const uint32_t mode_curr)
1572 :	chl	345	{
1573 :
1574 :	edgomez	851	// [0] is prediction
1575 :			pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y);
1576 :	chl	345
1577 :	edgomez	851	pmv[1].x = pmv[1].y = 0; // [1] is zero
1578 :	chl	345
1579 :	edgomez	851	pmv[2] = ChoosePred(pMB, mode_curr);
1580 :			pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y);
1581 :	chl	345
1582 :	edgomez	851	if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour
1583 :			pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr);
1584 :			pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y);
1585 :			} else pmv[3].x = pmv[3].y = 0;
1586 :	chl	345
1587 :	edgomez	851	if (y != 0) {
1588 :			pmv[4] = ChoosePred(pMB-iWcount, mode_curr);
1589 :			pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y);
1590 :			} else pmv[4].x = pmv[4].y = 0;
1591 :	chl	345
1592 :	edgomez	851	if (x != 0) {
1593 :			pmv[5] = ChoosePred(pMB-1, mode_curr);
1594 :			pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y);
1595 :			} else pmv[5].x = pmv[5].y = 0;
1596 :	chl	345
1597 :	edgomez	851	if (x != 0 && y != 0) {
1598 :			pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr);
1599 :			pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y);
1600 :			} else pmv[6].x = pmv[6].y = 0;
1601 :			}
1602 :	chl	345
1603 :
1604 :	edgomez	851	/* search backward or forward */
1605 :			static void
1606 :			SearchBF( const IMAGE * const pRef,
1607 :			const uint8_t * const pRefH,
1608 :			const uint8_t * const pRefV,
1609 :			const uint8_t * const pRefHV,
1610 :			const IMAGE * const pCur,
1611 :			const int x, const int y,
1612 :			const uint32_t MotionFlags,
1613 :			const uint32_t iFcode,
1614 :			const MBParam * const pParam,
1615 :			MACROBLOCK * const pMB,
1616 :			const VECTOR * const predMV,
1617 :			int32_t * const best_sad,
1618 :			const int32_t mode_current,
1619 :			SearchData * const Data)
1620 :			{
1621 :	chl	345
1622 :	edgomez	851	int i, iDirection = 255, mask;
1623 :			VECTOR pmv[7];
1624 :			MainSearchFunc *MainSearchPtr;
1625 :			*Data->iMinSAD = MV_MAX_ERROR;
1626 :			Data->iFcode = iFcode;
1627 :			Data->qpel_precision = 0;
1628 :			Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; // reset chroma-sad cache
1629 :	chl	345
1630 :	edgomez	851	Data->Ref = pRef->y + (x + y * Data->iEdgedWidth) * 16;
1631 :			Data->RefH = pRefH + (x + y * Data->iEdgedWidth) * 16;
1632 :			Data->RefV = pRefV + (x + y * Data->iEdgedWidth) * 16;
1633 :			Data->RefHV = pRefHV + (x + y * Data->iEdgedWidth) * 16;
1634 :			Data->RefCU = pRef->u + (x + y * Data->iEdgedWidth/2) * 8;
1635 :			Data->RefCV = pRef->v + (x + y * Data->iEdgedWidth/2) * 8;
1636 :	chl	345
1637 :	edgomez	851	Data->predMV = *predMV;
1638 :	chl	345
1639 :	edgomez	851	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1640 :			pParam->width, pParam->height, iFcode - Data->qpel, 0, 0);
1641 :	chl	345
1642 :	edgomez	851	pmv[0] = Data->predMV;
1643 :			if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; }
1644 :	chl	345
1645 :	edgomez	851	PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current);
1646 :	chl	345
1647 :	edgomez	851	Data->currentMV->x = Data->currentMV->y = 0;
1648 :			CheckCandidate = CheckCandidate16no4v;
1649 :	chl	345
1650 :	edgomez	851	// main loop. checking all predictions
1651 :			for (i = 0; i < 7; i++) {
1652 :			if (!(mask = make_mask(pmv, i)) ) continue;
1653 :			CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1654 :	chl	345	}
1655 :
1656 :	edgomez	851	if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1657 :			else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1658 :			else MainSearchPtr = DiamondSearch;
1659 :	chl	345
1660 :	edgomez	851	MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1661 :	chl	345
1662 :	edgomez	851	SubpelRefine(Data);
1663 :	chl	345
1664 :	edgomez	851	if (Data->qpel && Data->iMinSAD < best_sad + 300) {
1665 :			Data->currentQMV->x = 2*Data->currentMV->x;
1666 :			Data->currentQMV->y = 2*Data->currentMV->y;
1667 :			Data->qpel_precision = 1;
1668 :			get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1669 :			pParam->width, pParam->height, iFcode, 1, 0);
1670 :			SubpelRefine(Data);
1671 :			}
1672 :	chl	345
1673 :	edgomez	851	// three bits are needed to code backward mode. four for forward
1674 :	Isibaar	3
1675 :	edgomez	851	if (mode_current == MODE_FORWARD) Data->iMinSAD += 4 Data->lambda16;
1676 :			else Data->iMinSAD += 3 Data->lambda16;
1677 :	edgomez	195
1678 :	edgomez	851	if (Data->iMinSAD < best_sad) {
1679 :			best_sad = Data->iMinSAD;
1680 :			pMB->mode = mode_current;
1681 :			if (Data->qpel) {
1682 :			pMB->pmvs[0].x = Data->currentQMV->x - predMV->x;
1683 :			pMB->pmvs[0].y = Data->currentQMV->y - predMV->y;
1684 :			if (mode_current == MODE_FORWARD)
1685 :			pMB->qmvs[0] = *Data->currentQMV;
1686 :			else
1687 :			pMB->b_qmvs[0] = *Data->currentQMV;
1688 :			} else {
1689 :			pMB->pmvs[0].x = Data->currentMV->x - predMV->x;
1690 :			pMB->pmvs[0].y = Data->currentMV->y - predMV->y;
1691 :			}
1692 :			if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV;
1693 :			else pMB->b_mvs[0] = *Data->currentMV;
1694 :			}
1695 :	edgomez	195
1696 :	edgomez	851	if (mode_current == MODE_FORWARD) (Data->currentMV+2) = Data->currentMV;
1697 :			else (Data->currentMV+1) = Data->currentMV; //we store currmv for interpolate search
1698 :	Isibaar	3	}
1699 :
1700 :	edgomez	851	static void
1701 :			SkipDecisionB(const IMAGE * const pCur,
1702 :			const IMAGE * const f_Ref,
1703 :			const IMAGE * const b_Ref,
1704 :			MACROBLOCK * const pMB,
1705 :			const uint32_t x, const uint32_t y,
1706 :			const SearchData * const Data)
1707 :			{
1708 :			int dx = 0, dy = 0, b_dx = 0, b_dy = 0;
1709 :			int32_t sum;
1710 :			const int div = 1 + Data->qpel;
1711 :			int k;
1712 :			const uint32_t stride = Data->iEdgedWidth/2;
1713 :			//this is not full chroma compensation, only it's fullpel approximation. should work though
1714 :	Isibaar	3
1715 :	edgomez	851	for (k = 0; k < 4; k++) {
1716 :			dy += Data->directmvF[k].y / div;
1717 :	edgomez	963	dx += Data->directmvF[k].x / div;
1718 :			b_dy += Data->directmvB[k].y / div;
1719 :			b_dx += Data->directmvB[k].x / div;
1720 :	edgomez	851	}
1721 :	Isibaar	3
1722 :	edgomez	851	dy = (dy >> 3) + roundtab_76[dy & 0xf];
1723 :			dx = (dx >> 3) + roundtab_76[dx & 0xf];
1724 :			b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf];
1725 :			b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf];
1726 :	Isibaar	3
1727 :	edgomez	851	sum = sad8bi(pCur->u + 8 * x + 8 * y * stride,
1728 :			f_Ref->u + (y8 + dy/2) stride + x*8 + dx/2,
1729 :			b_Ref->u + (y8 + b_dy/2) stride + x*8 + b_dx/2,
1730 :			stride);
1731 :	Isibaar	3
1732 :	edgomez	851	if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; //no skip
1733 :	Isibaar	3
1734 :	edgomez	851	sum += sad8bi(pCur->v + 8x + 8 y * stride,
1735 :			f_Ref->v + (y8 + dy/2) stride + x*8 + dx/2,
1736 :			b_Ref->v + (y8 + b_dy/2) stride + x*8 + b_dx/2,
1737 :			stride);
1738 :	edgomez	195
1739 :	edgomez	963	if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) {
1740 :			pMB->mode = MODE_DIRECT_NONE_MV; //skipped
1741 :			for (k = 0; k < 4; k++) {
1742 :			pMB->qmvs[k] = pMB->mvs[k];
1743 :			pMB->b_qmvs[k] = pMB->b_mvs[k];
1744 :			}
1745 :			}
1746 :	edgomez	851	}
1747 :	Isibaar	3
1748 :	edgomez	851	static __inline uint32_t
1749 :			SearchDirect(const IMAGE * const f_Ref,
1750 :			const uint8_t * const f_RefH,
1751 :			const uint8_t * const f_RefV,
1752 :			const uint8_t * const f_RefHV,
1753 :			const IMAGE * const b_Ref,
1754 :			const uint8_t * const b_RefH,
1755 :			const uint8_t * const b_RefV,
1756 :			const uint8_t * const b_RefHV,
1757 :			const IMAGE * const pCur,
1758 :			const int x, const int y,
1759 :			const uint32_t MotionFlags,
1760 :			const int32_t TRB, const int32_t TRD,
1761 :			const MBParam * const pParam,
1762 :			MACROBLOCK * const pMB,
1763 :			const MACROBLOCK * const b_mb,
1764 :			int32_t * const best_sad,
1765 :			SearchData * const Data)
1766 :	Isibaar	3
1767 :	edgomez	851	{
1768 :			int32_t skip_sad;
1769 :			int k = (x + Data->iEdgedWidthy) 16;
1770 :			MainSearchFunc *MainSearchPtr;
1771 :	Isibaar	3
1772 :	edgomez	851	Data->iMinSAD = 2564096;
1773 :			Data->Ref = f_Ref->y + k;
1774 :			Data->RefH = f_RefH + k;
1775 :			Data->RefV = f_RefV + k;
1776 :			Data->RefHV = f_RefHV + k;
1777 :			Data->bRef = b_Ref->y + k;
1778 :			Data->bRefH = b_RefH + k;
1779 :			Data->bRefV = b_RefV + k;
1780 :			Data->bRefHV = b_RefHV + k;
1781 :			Data->RefCU = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1782 :			Data->RefCV = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1783 :			Data->b_RefCU = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1784 :			Data->b_RefCV = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1785 :	edgomez	195
1786 :	edgomez	851	k = Data->qpel ? 4 : 2;
1787 :			Data->max_dx = k * (pParam->width - x * 16);
1788 :			Data->max_dy = k * (pParam->height - y * 16);
1789 :			Data->min_dx = -k * (16 + x * 16);
1790 :			Data->min_dy = -k * (16 + y * 16);
1791 :	chl	181
1792 :	edgomez	851	Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs;
1793 :			Data->qpel_precision = 0;
1794 :	edgomez	170
1795 :	edgomez	851	for (k = 0; k < 4; k++) {
1796 :			pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD);
1797 :			pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD;
1798 :			pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD);
1799 :			pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD;
1800 :	Isibaar	3
1801 :	edgomez	851	if ( (pMB->b_mvs[k].x > Data->max_dx) \| (pMB->b_mvs[k].x < Data->min_dx)
1802 :			\| (pMB->b_mvs[k].y > Data->max_dy) \| (pMB->b_mvs[k].y < Data->min_dy) ) {
1803 :
1804 :			best_sad = 2564096; // in that case, we won't use direct mode
1805 :			pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV"
1806 :			pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0;
1807 :			return 256*4096;
1808 :			}
1809 :			if (b_mb->mode != MODE_INTER4V) {
1810 :			pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0];
1811 :			pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0];
1812 :			Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0];
1813 :			Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0];
1814 :			break;
1815 :			}
1816 :	edgomez	195	}
1817 :	Isibaar	3
1818 :	edgomez	851	CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v;
1819 :	Isibaar	3
1820 :	edgomez	851	CheckCandidate(0, 0, 255, &k, Data);
1821 :	edgomez	195
1822 :	edgomez	851	// initial (fast) skip decision
1823 :			if (Data->iMinSAD < pMB->quant INITIAL_SKIP_THRESH * (2 + Data->chroma?1:0)) {
1824 :			//possible skip
1825 :			if (Data->chroma) {
1826 :			pMB->mode = MODE_DIRECT_NONE_MV;
1827 :			return *Data->iMinSAD; // skip.
1828 :			} else {
1829 :			SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data);
1830 :			if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip.
1831 :			}
1832 :	Isibaar	3	}
1833 :
1834 :	syskin	904	*Data->iMinSAD += Data->lambda16;
1835 :	edgomez	851	skip_sad = *Data->iMinSAD;
1836 :	edgomez	195
1837 :	edgomez	851	// DIRECT MODE DELTA VECTOR SEARCH.
1838 :			// This has to be made more effective, but at the moment I'm happy it's running at all
1839 :	Isibaar	3
1840 :	edgomez	851	if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1841 :			else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1842 :			else MainSearchPtr = DiamondSearch;
1843 :	Isibaar	3
1844 :	edgomez	851	MainSearchPtr(0, 0, Data, 255);
1845 :	Isibaar	3
1846 :	edgomez	851	SubpelRefine(Data);
1847 :	chl	181
1848 :	edgomez	851	best_sad = Data->iMinSAD;
1849 :	chl	181
1850 :	edgomez	851	if (Data->qpel \|\| b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT;
1851 :			else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation
1852 :	chl	181
1853 :	edgomez	851	pMB->pmvs[3] = *Data->currentMV;
1854 :	edgomez	195
1855 :	edgomez	851	for (k = 0; k < 4; k++) {
1856 :			pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x;
1857 :			pMB->b_mvs[k].x = ( (Data->currentMV->x == 0)
1858 :			? Data->directmvB[k].x
1859 :			:pMB->mvs[k].x - Data->referencemv[k].x);
1860 :			pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y);
1861 :			pMB->b_mvs[k].y = ((Data->currentMV->y == 0)
1862 :			? Data->directmvB[k].y
1863 :			: pMB->mvs[k].y - Data->referencemv[k].y);
1864 :			if (Data->qpel) {
1865 :			pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2;
1866 :			pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2;
1867 :			pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2;
1868 :			pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2;
1869 :			}
1870 :	edgomez	195
1871 :	edgomez	851	if (b_mb->mode != MODE_INTER4V) {
1872 :			pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0];
1873 :			pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0];
1874 :			pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0];
1875 :			pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0];
1876 :			break;
1877 :			}
1878 :	edgomez	78	}
1879 :	edgomez	851	return skip_sad;
1880 :			}
1881 :	Isibaar	3
1882 :	edgomez	851	static void
1883 :			SearchInterpolate(const IMAGE * const f_Ref,
1884 :			const uint8_t * const f_RefH,
1885 :			const uint8_t * const f_RefV,
1886 :			const uint8_t * const f_RefHV,
1887 :			const IMAGE * const b_Ref,
1888 :			const uint8_t * const b_RefH,
1889 :			const uint8_t * const b_RefV,
1890 :			const uint8_t * const b_RefHV,
1891 :			const IMAGE * const pCur,
1892 :			const int x, const int y,
1893 :			const uint32_t fcode,
1894 :			const uint32_t bcode,
1895 :			const uint32_t MotionFlags,
1896 :			const MBParam * const pParam,
1897 :			const VECTOR * const f_predMV,
1898 :			const VECTOR * const b_predMV,
1899 :			MACROBLOCK * const pMB,
1900 :			int32_t * const best_sad,
1901 :			SearchData * const fData)
1902 :	chl	96
1903 :	edgomez	851	{
1904 :	chl	169
1905 :	edgomez	851	int iDirection, i, j;
1906 :			SearchData bData;
1907 :	chl	169
1908 :	edgomez	851	fData->qpel_precision = 0;
1909 :			memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data
1910 :			fData->iMinSAD = 4096256;
1911 :			bData.currentMV++; bData.currentQMV++;
1912 :			fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode;
1913 :	chl	169
1914 :	edgomez	851	i = (x + y * fData->iEdgedWidth) * 16;
1915 :			bData.bRef = fData->Ref = f_Ref->y + i;
1916 :			bData.bRefH = fData->RefH = f_RefH + i;
1917 :			bData.bRefV = fData->RefV = f_RefV + i;
1918 :			bData.bRefHV = fData->RefHV = f_RefHV + i;
1919 :			bData.Ref = fData->bRef = b_Ref->y + i;
1920 :			bData.RefH = fData->bRefH = b_RefH + i;
1921 :			bData.RefV = fData->bRefV = b_RefV + i;
1922 :			bData.RefHV = fData->bRefHV = b_RefHV + i;
1923 :			bData.b_RefCU = fData->RefCU = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1924 :			bData.b_RefCV = fData->RefCV = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1925 :			bData.RefCU = fData->b_RefCU = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1926 :			bData.RefCV = fData->b_RefCV = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1927 :	chl	169
1928 :
1929 :	edgomez	851	bData.bpredMV = fData->predMV = *f_predMV;
1930 :			fData->bpredMV = bData.predMV = *b_predMV;
1931 :			fData->currentMV[0] = fData->currentMV[2];
1932 :	Isibaar	3
1933 :	edgomez	851	get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode - fData->qpel, 0, 0);
1934 :			get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode - fData->qpel, 0, 0);
1935 :	Isibaar	3
1936 :	edgomez	851	if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx;
1937 :			if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx;
1938 :			if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy;
1939 :			if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy;
1940 :	chl	169
1941 :	edgomez	851	if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx;
1942 :			if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx;
1943 :			if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy;
1944 :			if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy;
1945 :	Isibaar	3
1946 :	edgomez	851	CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData);
1947 :	chl	169
1948 :	edgomez	851	//diamond
1949 :			do {
1950 :			iDirection = 255;
1951 :			// forward MV moves
1952 :			i = fData->currentMV[0].x; j = fData->currentMV[0].y;
1953 :	chl	169
1954 :	edgomez	851	CheckCandidateInt(i + 1, j, 0, &iDirection, fData);
1955 :			CheckCandidateInt(i, j + 1, 0, &iDirection, fData);
1956 :			CheckCandidateInt(i - 1, j, 0, &iDirection, fData);
1957 :			CheckCandidateInt(i, j - 1, 0, &iDirection, fData);
1958 :	edgomez	195
1959 :	edgomez	851	// backward MV moves
1960 :			i = fData->currentMV[1].x; j = fData->currentMV[1].y;
1961 :			fData->currentMV[2] = fData->currentMV[0];
1962 :			CheckCandidateInt(i + 1, j, 0, &iDirection, &bData);
1963 :			CheckCandidateInt(i, j + 1, 0, &iDirection, &bData);
1964 :			CheckCandidateInt(i - 1, j, 0, &iDirection, &bData);
1965 :			CheckCandidateInt(i, j - 1, 0, &iDirection, &bData);
1966 :	Isibaar	3
1967 :	edgomez	851	} while (!(iDirection));
1968 :	chl	140
1969 :	edgomez	851	//qpel refinement
1970 :			if (fData->qpel) {
1971 :			if (fData->iMinSAD > best_sad + 500) return;
1972 :			CheckCandidate = CheckCandidateInt;
1973 :			fData->qpel_precision = bData.qpel_precision = 1;
1974 :			get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0);
1975 :			get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0);
1976 :			fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x;
1977 :			fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y;
1978 :			fData->currentQMV[1].x = 2 * fData->currentMV[1].x;
1979 :			fData->currentQMV[1].y = 2 * fData->currentMV[1].y;
1980 :			SubpelRefine(fData);
1981 :			if (fData->iMinSAD > best_sad + 300) return;
1982 :			fData->currentQMV[2] = fData->currentQMV[0];
1983 :			SubpelRefine(&bData);
1984 :			}
1985 :	chl	140
1986 :	edgomez	851	fData->iMinSAD += (2+3) fData->lambda16; // two bits are needed to code interpolate mode.
1987 :	Isibaar	3
1988 :	edgomez	851	if (fData->iMinSAD < best_sad) {
1989 :			best_sad = fData->iMinSAD;
1990 :			pMB->mvs[0] = fData->currentMV[0];
1991 :			pMB->b_mvs[0] = fData->currentMV[1];
1992 :			pMB->mode = MODE_INTERPOLATE;
1993 :			if (fData->qpel) {
1994 :			pMB->qmvs[0] = fData->currentQMV[0];
1995 :			pMB->b_qmvs[0] = fData->currentQMV[1];
1996 :			pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x;
1997 :			pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y;
1998 :			pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x;
1999 :			pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y;
2000 :			} else {
2001 :			pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x;
2002 :			pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y;
2003 :			pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x;
2004 :			pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y;
2005 :			}
2006 :	edgomez	78	}
2007 :	edgomez	851	}
2008 :	Isibaar	3
2009 :	edgomez	851	void
2010 :			MotionEstimationBVOP(MBParam * const pParam,
2011 :			FRAMEINFO * const frame,
2012 :			const int32_t time_bp,
2013 :			const int32_t time_pp,
2014 :			// forward (past) reference
2015 :			const MACROBLOCK * const f_mbs,
2016 :			const IMAGE * const f_ref,
2017 :			const IMAGE * const f_refH,
2018 :			const IMAGE * const f_refV,
2019 :			const IMAGE * const f_refHV,
2020 :			// backward (future) reference
2021 :			const FRAMEINFO * const b_reference,
2022 :			const IMAGE * const b_ref,
2023 :			const IMAGE * const b_refH,
2024 :			const IMAGE * const b_refV,
2025 :			const IMAGE * const b_refHV)
2026 :			{
2027 :			uint32_t i, j;
2028 :			int32_t best_sad;
2029 :			uint32_t skip_sad;
2030 :			int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0;
2031 :			const MACROBLOCK * const b_mbs = b_reference->mbs;
2032 :	Isibaar	3
2033 :	edgomez	851	VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/
2034 :	Isibaar	3
2035 :	edgomez	851	const int32_t TRB = time_pp - time_bp;
2036 :			const int32_t TRD = time_pp;
2037 :	edgomez	195
2038 :	edgomez	851	// some pre-inintialized data for the rest of the search
2039 :	Isibaar	3
2040 :	edgomez	851	SearchData Data;
2041 :			int32_t iMinSAD;
2042 :			VECTOR currentMV[3];
2043 :			VECTOR currentQMV[3];
2044 :			int32_t temp[8];
2045 :			memset(&Data, 0, sizeof(SearchData));
2046 :			Data.iEdgedWidth = pParam->edged_width;
2047 :			Data.currentMV = currentMV; Data.currentQMV = currentQMV;
2048 :			Data.iMinSAD = &iMinSAD;
2049 :			Data.lambda16 = lambda_vec16[frame->quant];
2050 :			Data.qpel = pParam->m_quarterpel;
2051 :			Data.rounding = 0;
2052 :			Data.chroma = frame->motion_flags & PMV_CHROMA8;
2053 :			Data.temp = temp;
2054 :	Isibaar	3
2055 :	edgomez	851	Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose)
2056 :			// note: i==horizontal, j==vertical
2057 :			for (j = 0; j < pParam->mb_height; j++) {
2058 :	edgomez	195
2059 :	edgomez	851	f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */
2060 :	Isibaar	3
2061 :	edgomez	851	for (i = 0; i < pParam->mb_width; i++) {
2062 :			MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width;
2063 :			const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width;
2064 :	edgomez	195
2065 :	edgomez	851	/* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */
2066 :			if (b_reference->coding_type != S_VOP)
2067 :			if (b_mb->mode == MODE_NOT_CODED) {
2068 :			pMB->mode = MODE_NOT_CODED;
2069 :			continue;
2070 :			}
2071 :	Isibaar	3
2072 :	edgomez	851	Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16;
2073 :			Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8;
2074 :			Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8;
2075 :			pMB->quant = frame->quant;
2076 :	Isibaar	3
2077 :	edgomez	851	/* direct search comes first, because it (1) checks for SKIP-mode
2078 :			and (2) sets very good predictions for forward and backward search */
2079 :			skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2080 :			b_ref, b_refH->y, b_refV->y, b_refHV->y,
2081 :			&frame->image,
2082 :			i, j,
2083 :			frame->motion_flags,
2084 :			TRB, TRD,
2085 :			pParam,
2086 :			pMB, b_mb,
2087 :			&best_sad,
2088 :			&Data);
2089 :	Isibaar	3
2090 :	edgomez	851	if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; }
2091 :	edgomez	195
2092 :	edgomez	851	// forward search
2093 :			SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2094 :			&frame->image, i, j,
2095 :			frame->motion_flags,
2096 :			frame->fcode, pParam,
2097 :			pMB, &f_predMV, &best_sad,
2098 :			MODE_FORWARD, &Data);
2099 :	edgomez	195
2100 :	edgomez	851	// backward search
2101 :			SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y,
2102 :			&frame->image, i, j,
2103 :			frame->motion_flags,
2104 :			frame->bcode, pParam,
2105 :			pMB, &b_predMV, &best_sad,
2106 :			MODE_BACKWARD, &Data);
2107 :
2108 :			// interpolate search comes last, because it uses data from forward and backward as prediction
2109 :			SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2110 :			b_ref, b_refH->y, b_refV->y, b_refHV->y,
2111 :			&frame->image,
2112 :			i, j,
2113 :			frame->fcode, frame->bcode,
2114 :			frame->motion_flags,
2115 :			pParam,
2116 :			&f_predMV, &b_predMV,
2117 :			pMB, &best_sad,
2118 :			&Data);
2119 :
2120 :			// final skip decision
2121 :			if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2)
2122 :			&& ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) )
2123 :			SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data);
2124 :
2125 :			switch (pMB->mode) {
2126 :			case MODE_FORWARD:
2127 :			f_count++;
2128 :			f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2129 :			break;
2130 :			case MODE_BACKWARD:
2131 :			b_count++;
2132 :			b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2133 :			break;
2134 :			case MODE_INTERPOLATE:
2135 :			i_count++;
2136 :			f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2137 :			b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2138 :			break;
2139 :			case MODE_DIRECT:
2140 :			case MODE_DIRECT_NO4V:
2141 :			d_count++;
2142 :			default:
2143 :			break;
2144 :			}
2145 :			}
2146 :			}
2147 :	Isibaar	3	}
2148 :	chl	96
2149 :	edgomez	851	static __inline void
2150 :			MEanalyzeMB ( const uint8_t * const pRef,
2151 :			const uint8_t * const pCur,
2152 :			const int x,
2153 :			const int y,
2154 :			const MBParam * const pParam,
2155 :			MACROBLOCK * const pMBs,
2156 :			SearchData * const Data)
2157 :	chl	96	{
2158 :
2159 :	edgomez	851	int i, mask;
2160 :			VECTOR pmv[3];
2161 :	syskin	904	MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width];
2162 :	chl	96
2163 :	edgomez	851	for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR;
2164 :	chl	96
2165 :	edgomez	851	//median is only used as prediction. it doesn't have to be real
2166 :			if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0;
2167 :			else
2168 :			if (x == 1) //left macroblock does not have any vector now
2169 :			Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median
2170 :			else if (y == 1) // top macroblock doesn't have it's vector
2171 :			Data->predMV = (pMB - 1)->mvs[0]; // left instead of median
2172 :			else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median
2173 :	edgomez	195
2174 :	edgomez	851	get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2175 :	syskin	904	pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, 0);
2176 :	edgomez	195
2177 :	edgomez	851	Data->Cur = pCur + (x + y * pParam->edged_width) * 16;
2178 :			Data->Ref = pRef + (x + y * pParam->edged_width) * 16;
2179 :	chl	96
2180 :	edgomez	851	pmv[1].x = EVEN(pMB->mvs[0].x);
2181 :			pmv[1].y = EVEN(pMB->mvs[0].y);
2182 :			pmv[2].x = EVEN(Data->predMV.x);
2183 :			pmv[2].y = EVEN(Data->predMV.y);
2184 :			pmv[0].x = pmv[0].y = 0;
2185 :	edgomez	195
2186 :	edgomez	851	CheckCandidate32I(0, 0, 255, &i, Data);
2187 :	syskin	904	Data->iMinSAD[1] -= 50;
2188 :			Data->iMinSAD[2] -= 50;
2189 :			Data->iMinSAD[3] -= 50;
2190 :			Data->iMinSAD[4] -= 50;
2191 :	edgomez	195
2192 :	syskin	904	if (Data->iMinSAD > 4 MAX_SAD00_FOR_SKIP) {
2193 :	chl	96
2194 :	edgomez	851	if (!(mask = make_mask(pmv, 1)))
2195 :			CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data);
2196 :			if (!(mask = make_mask(pmv, 2)))
2197 :			CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data);
2198 :	chl	96
2199 :	syskin	904	if (Data->iMinSAD > 4 MAX_SAD00_FOR_SKIP) // diamond only if needed
2200 :	edgomez	851	DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i);
2201 :	syskin	904	}
2202 :	edgomez	851
2203 :	syskin	904	for (i = 0; i < 4; i++) {
2204 :			MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width];
2205 :			MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i];
2206 :			MB->mode = MODE_INTER;
2207 :			MB->sad16 = Data->iMinSAD[i+1];
2208 :	chl	96	}
2209 :	edgomez	851	}
2210 :	chl	96
2211 :	syskin	904	#define INTRA_THRESH 2400
2212 :			#define INTER_THRESH 1100
2213 :	chl	96
2214 :	edgomez	851	int
2215 :			MEanalysis( const IMAGE * const pRef,
2216 :	syskin	904	const FRAMEINFO * const Current,
2217 :			const MBParam * const pParam,
2218 :			const int maxIntra, //maximum number if non-I frames
2219 :			const int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame
2220 :			const int bCount, // number of B frames in a row
2221 :			const int b_thresh)
2222 :	edgomez	851	{
2223 :			uint32_t x, y, intra = 0;
2224 :			int sSAD = 0;
2225 :			MACROBLOCK * const pMBs = Current->mbs;
2226 :			const IMAGE * const pCurrent = &Current->image;
2227 :	syskin	904	int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + 10*b_thresh;
2228 :			int s = 0, blocks = 0;
2229 :	chl	96
2230 :	edgomez	851	int32_t iMinSAD[5], temp[5];
2231 :			VECTOR currentMV[5];
2232 :			SearchData Data;
2233 :			Data.iEdgedWidth = pParam->edged_width;
2234 :			Data.currentMV = currentMV;
2235 :			Data.iMinSAD = iMinSAD;
2236 :			Data.iFcode = Current->fcode;
2237 :			Data.temp = temp;
2238 :			CheckCandidate = CheckCandidate32I;
2239 :	chl	96
2240 :	edgomez	851	if (intraCount != 0 && intraCount < 10) // we're right after an I frame
2241 :	syskin	904	IntraThresh += 8 * (intraCount - 10) * (intraCount - 10);
2242 :	edgomez	851	else
2243 :			if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec
2244 :			IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra;
2245 :	chl	96
2246 :	syskin	904	InterThresh -= (350 - 8b_thresh) bCount;
2247 :			if (InterThresh < 300 + 5b_thresh) InterThresh = 300 + 5b_thresh;
2248 :	chl	326
2249 :	edgomez	851	if (sadInit) (*sadInit) ();
2250 :	chl	96
2251 :	edgomez	851	for (y = 1; y < pParam->mb_height-1; y += 2) {
2252 :			for (x = 1; x < pParam->mb_width-1; x += 2) {
2253 :			int i;
2254 :	syskin	904	blocks += 4;
2255 :	edgomez	195
2256 :	edgomez	851	if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV;
2257 :	syskin	904	else { //extrapolation of the vector found for last frame
2258 :			pMBs[x + y * pParam->mb_width].mvs[0].x =
2259 :			(pMBs[x + y * pParam->mb_width].mvs[0].x * (bCount+1) ) / bCount;
2260 :			pMBs[x + y * pParam->mb_width].mvs[0].y =
2261 :			(pMBs[x + y * pParam->mb_width].mvs[0].y * (bCount+1) ) / bCount;
2262 :			}
2263 :	edgomez	195
2264 :	edgomez	851	MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data);
2265 :	edgomez	195
2266 :	edgomez	851	for (i = 0; i < 4; i++) {
2267 :			int dev;
2268 :			MACROBLOCK pMB = &pMBs[x+(i&1) + (y+(i>>1)) pParam->mb_width];
2269 :			if (pMB->sad16 > IntraThresh) {
2270 :			dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16,
2271 :			pParam->edged_width);
2272 :			if (dev + IntraThresh < pMB->sad16) {
2273 :			pMB->mode = MODE_INTRA;
2274 :	syskin	904	if (++intra > ((pParam->mb_height-2)*(pParam->mb_width-2))/2) return I_VOP;
2275 :	edgomez	851	}
2276 :			}
2277 :	syskin	904	if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0) s++;
2278 :
2279 :	edgomez	851	sSAD += pMB->sad16;
2280 :			}
2281 :			}
2282 :	edgomez	195	}
2283 :	syskin	904
2284 :			sSAD /= blocks;
2285 :			s = (10*s) / blocks;
2286 :
2287 :			if (s > 5) sSAD += (s - 4) * (180 - 2*b_thresh); //static block - looks bad when in bframe...
2288 :
2289 :	edgomez	851	if (sSAD > InterThresh ) return P_VOP;
2290 :			emms();
2291 :			return B_VOP;
2292 :			}
2293 :	chl	96
2294 :
2295 :	edgomez	851	static WARPPOINTS
2296 :			GlobalMotionEst(const MACROBLOCK * const pMBs,
2297 :			const MBParam * const pParam,
2298 :			const FRAMEINFO * const current,
2299 :			const FRAMEINFO * const reference,
2300 :			const IMAGE * const pRefH,
2301 :			const IMAGE * const pRefV,
2302 :			const IMAGE * const pRefHV )
2303 :			{
2304 :	chl	96
2305 :	edgomez	851	const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs
2306 :			const int deltay=8;
2307 :			const int grad=512; // lower bound for deviation in MB
2308 :	chl	96
2309 :	edgomez	851	WARPPOINTS gmc;
2310 :	chl	96
2311 :	edgomez	851	uint32_t mx, my;
2312 :	chl	96
2313 :	edgomez	851	int MBh = pParam->mb_height;
2314 :			int MBw = pParam->mb_width;
2315 :	chl	96
2316 :	edgomez	851	int MBmask= calloc(MBhMBw,sizeof(int));
2317 :			double DtimesF[4] = { 0.,0., 0., 0. };
2318 :			double sol[4] = { 0., 0., 0., 0. };
2319 :			double a,b,c,n,denom;
2320 :			double meanx,meany;
2321 :			int num,oldnum;
2322 :	edgomez	195
2323 :	chl	871	if (!MBmask) { fprintf(stderr,"Mem error\n");
2324 :			gmc.duv[0].x= gmc.duv[0].y =
2325 :			gmc.duv[1].x= gmc.duv[1].y =
2326 :			gmc.duv[2].x= gmc.duv[2].y = 0;
2327 :			return gmc; }
2328 :	edgomez	195
2329 :	edgomez	851	// filter mask of all blocks
2330 :
2331 :	edgomez	876	for (my = 1; my < (uint32_t)MBh-1; my++)
2332 :			for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2333 :	edgomez	851	{
2334 :			const int mbnum = mx + my * MBw;
2335 :			const MACROBLOCK *pMB = &pMBs[mbnum];
2336 :			const VECTOR mv = pMB->mvs[0];
2337 :
2338 :			if (pMB->mode == MODE_INTRA \|\| pMB->mode == MODE_NOT_CODED)
2339 :			continue;
2340 :
2341 :			if ( ( (ABS(mv.x - (pMB-1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-1)->mvs[0].y) < deltay) )
2342 :			&& ( (ABS(mv.x - (pMB+1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+1)->mvs[0].y) < deltay) )
2343 :			&& ( (ABS(mv.x - (pMB-MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-MBw)->mvs[0].y) < deltay) )
2344 :			&& ( (ABS(mv.x - (pMB+MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) )
2345 :			MBmask[mbnum]=1;
2346 :	chl	96	}
2347 :
2348 :	edgomez	876	for (my = 1; my < (uint32_t)MBh-1; my++)
2349 :			for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2350 :	edgomez	851	{
2351 :			const uint8_t const pCur = current->image.y + 16mypParam->edged_width + 16mx;
2352 :	chl	96
2353 :	edgomez	851	const int mbnum = mx + my * MBw;
2354 :			if (!MBmask[mbnum])
2355 :			continue;
2356 :
2357 :	edgomez	876	if (sad16 ( pCur, pCur+1 , pParam->edged_width, 65536) <= (uint32_t)grad )
2358 :	edgomez	851	MBmask[mbnum] = 0;
2359 :	edgomez	876	if (sad16 ( pCur, pCur+pParam->edged_width, pParam->edged_width, 65536) <= (uint32_t)grad )
2360 :	edgomez	851	MBmask[mbnum] = 0;
2361 :
2362 :	edgomez	195	}
2363 :	chl	96
2364 :	edgomez	851	emms();
2365 :	chl	96
2366 :	edgomez	851	do { /* until convergence */
2367 :	chl	96
2368 :	edgomez	851	a = b = c = n = 0;
2369 :			DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.;
2370 :	edgomez	876	for (my = 0; my < (uint32_t)MBh; my++)
2371 :			for (mx = 0; mx < (uint32_t)MBw; mx++)
2372 :	edgomez	851	{
2373 :			const int mbnum = mx + my * MBw;
2374 :			const MACROBLOCK *pMB = &pMBs[mbnum];
2375 :			const VECTOR mv = pMB->mvs[0];
2376 :	chl	96
2377 :	edgomez	851	if (!MBmask[mbnum])
2378 :			continue;
2379 :	chl	96
2380 :	edgomez	851	n++;
2381 :			a += 16*mx+8;
2382 :			b += 16*my+8;
2383 :			c += (16mx+8)(16mx+8)+(16my+8)(16my+8);
2384 :	chl	96
2385 :	edgomez	851	DtimesF[0] += (double)mv.x;
2386 :			DtimesF[1] += (double)mv.x(16mx+8) + (double)mv.y(16my+8);
2387 :			DtimesF[2] += (double)mv.x(16my+8) - (double)mv.y(16mx+8);
2388 :			DtimesF[3] += (double)mv.y;
2389 :			}
2390 :	chl	96
2391 :	edgomez	851	denom = aa+bb-c*n;
2392 :	chl	96
2393 :	edgomez	851	/* Solve the system: sol = (D'ED)^{-1} D'EF */
2394 :			/* D'EF has been calculated in the same loop as matrix */
2395 :	chl	96
2396 :	edgomez	851	sol[0] = -cDtimesF[0] + aDtimesF[1] + b*DtimesF[2];
2397 :			sol[1] = aDtimesF[0] - nDtimesF[1] + b*DtimesF[3];
2398 :			sol[2] = bDtimesF[0] - nDtimesF[2] - a*DtimesF[3];
2399 :			sol[3] = bDtimesF[1] - aDtimesF[2] - c*DtimesF[3];
2400 :	chl	96
2401 :	edgomez	851	sol[0] /= denom;
2402 :			sol[1] /= denom;
2403 :			sol[2] /= denom;
2404 :			sol[3] /= denom;
2405 :	chl	96
2406 :	edgomez	851	meanx = meany = 0.;
2407 :			oldnum = 0;
2408 :	edgomez	876	for (my = 0; my < (uint32_t)MBh; my++)
2409 :			for (mx = 0; mx < (uint32_t)MBw; mx++)
2410 :	edgomez	851	{
2411 :			const int mbnum = mx + my * MBw;
2412 :			const MACROBLOCK *pMB = &pMBs[mbnum];
2413 :			const VECTOR mv = pMB->mvs[0];
2414 :	chl	96
2415 :	edgomez	851	if (!MBmask[mbnum])
2416 :			continue;
2417 :	chl	184
2418 :	edgomez	851	oldnum++;
2419 :			meanx += ABS(( sol[0] + (16mx+8)sol[1] + (16my+8)sol[2] ) - mv.x );
2420 :			meany += ABS(( sol[3] - (16mx+8)sol[2] + (16my+8)sol[1] ) - mv.y );
2421 :			}
2422 :	chl	96
2423 :	edgomez	851	if (4meanx > oldnum) / better fit than 0.25 is useless */
2424 :			meanx /= oldnum;
2425 :			else
2426 :			meanx = 0.25;
2427 :	edgomez	195
2428 :	edgomez	851	if (4*meany > oldnum)
2429 :			meany /= oldnum;
2430 :			else
2431 :			meany = 0.25;
2432 :	chl	96
2433 :	edgomez	851	/* fprintf(stderr,"sol = (%8.5f, %8.5f, %8.5f, %8.5f)\n",sol[0],sol[1],sol[2],sol[3]);
2434 :			fprintf(stderr,"meanx = %8.5f meany = %8.5f %d\n",meanx,meany, oldnum);
2435 :			*/
2436 :			num = 0;
2437 :	edgomez	876	for (my = 0; my < (uint32_t)MBh; my++)
2438 :			for (mx = 0; mx < (uint32_t)MBw; mx++)
2439 :	edgomez	851	{
2440 :			const int mbnum = mx + my * MBw;
2441 :			const MACROBLOCK *pMB = &pMBs[mbnum];
2442 :			const VECTOR mv = pMB->mvs[0];
2443 :	chl	96
2444 :	edgomez	851	if (!MBmask[mbnum])
2445 :			continue;
2446 :	chl	96
2447 :	edgomez	851	if ( ( ABS(( sol[0] + (16mx+8)sol[1] + (16my+8)sol[2] ) - mv.x ) > meanx )
2448 :			\|\| ( ABS(( sol[3] - (16mx+8)sol[2] + (16my+8)sol[1] ) - mv.y ) > meany ) )
2449 :			MBmask[mbnum]=0;
2450 :			else
2451 :			num++;
2452 :	chl	96	}
2453 :	edgomez	195
2454 :	edgomez	851	} while ( (oldnum != num) && (num>=4) );
2455 :	edgomez	195
2456 :	edgomez	851	if (num < 4)
2457 :			{
2458 :			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;
2459 :			} else {
2460 :	edgomez	195
2461 :	edgomez	851	gmc.duv[0].x=(int)(sol[0]+0.5);
2462 :			gmc.duv[0].y=(int)(sol[3]+0.5);
2463 :	chl	96
2464 :	edgomez	851	gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5);
2465 :			gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5);
2466 :	chl	96
2467 :	edgomez	851	gmc.duv[2].x=0;
2468 :			gmc.duv[2].y=0;
2469 :			}
2470 :			// fprintf(stderr,"wp1 = ( %4d, %4d) wp2 = ( %4d, %4d) \n", gmc.duv[0].x, gmc.duv[0].y, gmc.duv[1].x, gmc.duv[1].y);
2471 :	chl	96
2472 :	edgomez	851	free(MBmask);
2473 :	chl	96
2474 :	edgomez	851	return gmc;
2475 :	chl	96	}
2476 :
2477 :	edgomez	851	// functions which perform BITS-based search/bitcount
2478 :	chl	96
2479 :	edgomez	851	static int
2480 :			CountMBBitsInter(SearchData * const Data,
2481 :			const MACROBLOCK * const pMBs, const int x, const int y,
2482 :			const MBParam * const pParam,
2483 :			const uint32_t MotionFlags)
2484 :	chl	96	{
2485 :	edgomez	851	int i, iDirection;
2486 :			int32_t bsad[5];
2487 :	chl	141
2488 :	edgomez	851	CheckCandidate = CheckCandidateBits16;
2489 :	chl	96
2490 :	edgomez	851	if (Data->qpel) {
2491 :			for(i = 0; i < 5; i++) {
2492 :			Data->currentMV[i].x = Data->currentQMV[i].x/2;
2493 :			Data->currentMV[i].y = Data->currentQMV[i].y/2;
2494 :			}
2495 :			Data->qpel_precision = 1;
2496 :			CheckCandidateBits16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data);
2497 :	chl	96
2498 :	edgomez	851	//checking if this vector is perfect. if it is, we stop.
2499 :			if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0)
2500 :			return 0; //quick stop
2501 :	edgomez	195
2502 :	edgomez	851	if (MotionFlags & (HALFPELREFINE16_BITS \| EXTSEARCH_BITS)) { //we have to prepare for halfpixel-precision search
2503 :			for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i];
2504 :			get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2505 :			pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
2506 :			Data->qpel_precision = 0;
2507 :			if (Data->currentQMV->x & 1 \|\| Data->currentQMV->y & 1)
2508 :			CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2509 :			}
2510 :	edgomez	195
2511 :	edgomez	851	} else { // not qpel
2512 :	edgomez	195
2513 :	edgomez	851	CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2514 :			//checking if this vector is perfect. if it is, we stop.
2515 :			if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0) {
2516 :			return 0; //inter
2517 :			}
2518 :			}
2519 :	chl	96
2520 :	edgomez	851	if (MotionFlags&EXTSEARCH_BITS) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
2521 :	chl	96
2522 :	edgomez	851	if (MotionFlags&HALFPELREFINE16_BITS) SubpelRefine(Data);
2523 :	chl	96
2524 :	edgomez	851	if (Data->qpel) {
2525 :			if (MotionFlags&(EXTSEARCH_BITS \| HALFPELREFINE16_BITS)) { // there was halfpel-precision search
2526 :			for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) {
2527 :			Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // we have found a better match
2528 :			Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
2529 :			}
2530 :	edgomez	195
2531 :	edgomez	851	// preparing for qpel-precision search
2532 :			Data->qpel_precision = 1;
2533 :			get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2534 :			pParam->width, pParam->height, Data->iFcode, 1, 0);
2535 :			}
2536 :			if (MotionFlags&QUARTERPELREFINE16_BITS) SubpelRefine(Data);
2537 :			}
2538 :	chl	96
2539 :	edgomez	851	if (MotionFlags&CHECKPREDICTION_BITS) { //let's check vector equal to prediction
2540 :			VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV;
2541 :			if (!(Data->predMV.x == v->x && Data->predMV.y == v->y))
2542 :			CheckCandidateBits16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data);
2543 :			}
2544 :			return Data->iMinSAD[0];
2545 :			}
2546 :	chl	96
2547 :
2548 :	edgomez	851	static int
2549 :			CountMBBitsInter4v(const SearchData * const Data,
2550 :			MACROBLOCK * const pMB, const MACROBLOCK * const pMBs,
2551 :			const int x, const int y,
2552 :			const MBParam * const pParam, const uint32_t MotionFlags,
2553 :			const VECTOR * const backup)
2554 :			{
2555 :	chl	96
2556 :	edgomez	851	int cbp = 0, bits = 0, t = 0, i, iDirection;
2557 :			SearchData Data2, *Data8 = &Data2;
2558 :			int sumx = 0, sumy = 0;
2559 :	edgomez	961	int16_t in = Data->dctSpace, coeff = Data->dctSpace + 64;
2560 :	chl	96
2561 :	edgomez	851	memcpy(Data8, Data, sizeof(SearchData));
2562 :			CheckCandidate = CheckCandidateBits8;
2563 :	chl	96
2564 :	edgomez	851	for (i = 0; i < 4; i++) {
2565 :			Data8->iMinSAD = Data->iMinSAD + i + 1;
2566 :			Data8->currentMV = Data->currentMV + i + 1;
2567 :			Data8->currentQMV = Data->currentQMV + i + 1;
2568 :			Data8->Cur = Data->Cur + 8((i&1) + (i>>1)Data->iEdgedWidth);
2569 :			Data8->Ref = Data->Ref + 8((i&1) + (i>>1)Data->iEdgedWidth);
2570 :			Data8->RefH = Data->RefH + 8((i&1) + (i>>1)Data->iEdgedWidth);
2571 :			Data8->RefV = Data->RefV + 8((i&1) + (i>>1)Data->iEdgedWidth);
2572 :			Data8->RefHV = Data->RefHV + 8((i&1) + (i>>1)Data->iEdgedWidth);
2573 :	edgomez	195
2574 :	edgomez	851	if(Data->qpel) {
2575 :			Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i);
2576 :			if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y,
2577 :			Data8->predMV, Data8->iFcode, 0, 0);
2578 :			} else {
2579 :			Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i);
2580 :			if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y,
2581 :			Data8->predMV, Data8->iFcode, 0, 0);
2582 :			}
2583 :	chl	96
2584 :	edgomez	851	get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2x + (i&1), 2y + (i>>1), 8,
2585 :			pParam->width, pParam->height, Data8->iFcode, Data8->qpel, 0);
2586 :	chl	96
2587 :	edgomez	851	*Data8->iMinSAD += t;
2588 :	chl	96
2589 :	edgomez	851	Data8->qpel_precision = Data8->qpel;
2590 :			// checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far)
2591 :			if (Data8->qpel) {
2592 :			if (!(Data8->currentQMV->x == backup[i+1].x && Data8->currentQMV->y == backup[i+1].y))
2593 :			CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8);
2594 :			} else {
2595 :			if (!(Data8->currentMV->x == backup[i+1].x && Data8->currentMV->y == backup[i+1].y))
2596 :			CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8);
2597 :			}
2598 :	edgomez	195
2599 :	edgomez	851	if (Data8->qpel) {
2600 :			if (MotionFlags&HALFPELREFINE8_BITS \|\| (MotionFlags&PMV_EXTSEARCH8 && MotionFlags&EXTSEARCH_BITS)) { // halfpixel motion search follows
2601 :			int32_t s = *Data8->iMinSAD;
2602 :			Data8->currentMV->x = Data8->currentQMV->x/2;
2603 :			Data8->currentMV->y = Data8->currentQMV->y/2;
2604 :			Data8->qpel_precision = 0;
2605 :			get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2x + (i&1), 2y + (i>>1), 8,
2606 :			pParam->width, pParam->height, Data8->iFcode - 1, 0, 0);
2607 :	edgomez	195
2608 :	edgomez	851	if (Data8->currentQMV->x & 1 \|\| Data8->currentQMV->y & 1)
2609 :			CheckCandidateBits8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8);
2610 :	edgomez	195
2611 :	edgomez	851	if (MotionFlags & PMV_EXTSEARCH8 && MotionFlags & EXTSEARCH_BITS)
2612 :			SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255);
2613 :	chl	96
2614 :	edgomez	851	if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8);
2615 :	chl	96
2616 :	edgomez	851	if(s > *Data8->iMinSAD) { //we have found a better match
2617 :			Data8->currentQMV->x = 2*Data8->currentMV->x;
2618 :			Data8->currentQMV->y = 2*Data8->currentMV->y;
2619 :			}
2620 :	chl	96
2621 :	edgomez	851	Data8->qpel_precision = 1;
2622 :			get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2x + (i&1), 2y + (i>>1), 8,
2623 :			pParam->width, pParam->height, Data8->iFcode, 1, 0);
2624 :	chl	96
2625 :	edgomez	851	}
2626 :			if (MotionFlags & QUARTERPELREFINE8_BITS) SubpelRefine(Data8);
2627 :	edgomez	195
2628 :	edgomez	851	} else // not qpel
2629 :			if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8); //halfpel mode, halfpel refinement
2630 :
2631 :			//checking vector equal to predicion
2632 :			if (i != 0 && MotionFlags & CHECKPREDICTION_BITS) {
2633 :			const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV;
2634 :			if (!(Data8->predMV.x == v->x && Data8->predMV.y == v->y))
2635 :			CheckCandidateBits8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8);
2636 :	chl	141	}
2637 :	edgomez	195
2638 :	edgomez	851	bits += *Data8->iMinSAD;
2639 :			if (bits >= Data->iMinSAD[0]) break; // no chances for INTER4V
2640 :
2641 :			// MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else
2642 :			if(Data->qpel) {
2643 :			pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x;
2644 :			pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y;
2645 :			pMB->qmvs[i] = *Data8->currentQMV;
2646 :			sumx += Data8->currentQMV->x/2;
2647 :			sumy += Data8->currentQMV->y/2;
2648 :			} else {
2649 :			pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x;
2650 :			pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y;
2651 :			sumx += Data8->currentMV->x;
2652 :			sumy += Data8->currentMV->y;
2653 :	chl	141	}
2654 :	edgomez	851	pMB->mvs[i] = *Data8->currentMV;
2655 :			pMB->sad8[i] = 4 * *Data8->iMinSAD;
2656 :			if (Data8->temp[0]) cbp \|= 1 << (5 - i);
2657 :	chl	141	}
2658 :
2659 :	edgomez	851	if (bits < *Data->iMinSAD) { // there is still a chance for inter4v mode. let's check chroma
2660 :			const uint8_t * ptr;
2661 :			sumx = (sumx >> 3) + roundtab_76[sumx & 0xf];
2662 :			sumy = (sumy >> 3) + roundtab_76[sumy & 0xf];
2663 :	chl	141
2664 :	edgomez	851	//chroma U
2665 :			ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefCU, 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2666 :			transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2);
2667 :			fdct(in);
2668 :			if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16);
2669 :			else i = quant4_inter(coeff, in, Data->lambda16);
2670 :			if (i > 0) {
2671 :			bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
2672 :			cbp \|= 1 << (5 - 4);
2673 :			}
2674 :	chl	141
2675 :	edgomez	851	if (bits < *Data->iMinSAD) { // still possible
2676 :			//chroma V
2677 :			ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefCV, 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2678 :			transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2);
2679 :			fdct(in);
2680 :			if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16);
2681 :			else i = quant4_inter(coeff, in, Data->lambda16);
2682 :			if (i > 0) {
2683 :			bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
2684 :			cbp \|= 1 << (5 - 5);
2685 :			}
2686 :	chl	897	bits += xvid_cbpy_tab[15-(cbp>>2)].len;
2687 :	edgomez	851	bits += mcbpc_inter_tab[(MODE_INTER4V & 7) \| ((cbp & 3) << 3)].len;
2688 :			}
2689 :	edgomez	195	}
2690 :	chl	96
2691 :	edgomez	851	return bits;
2692 :			}
2693 :	chl	96
2694 :
2695 :	edgomez	851	static int
2696 :			CountMBBitsIntra(const SearchData * const Data)
2697 :			{
2698 :			int bits = 1; //this one is ac/dc prediction flag. always 1.
2699 :			int cbp = 0, i, t, dc = 0, b_dc = 1024;
2700 :			const uint32_t iQuant = Data->lambda16;
2701 :	edgomez	961	int16_t in = Data->dctSpace, coeff = Data->dctSpace + 64;
2702 :	edgomez	195
2703 :	edgomez	851	for(i = 0; i < 4; i++) {
2704 :			uint32_t iDcScaler = get_dc_scaler(iQuant, 1);
2705 :	chl	96
2706 :	edgomez	851	int s = 8((i&1) + (i>>1)Data->iEdgedWidth);
2707 :			transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth);
2708 :			fdct(in);
2709 :			b_dc = dc;
2710 :			dc = in[0];
2711 :			in[0] -= b_dc;
2712 :			if (Data->lambda8 == 0) quant_intra_c(coeff, in, iQuant, iDcScaler);
2713 :			else quant4_intra_c(coeff, in, iQuant, iDcScaler);
2714 :	chl	141
2715 :	edgomez	851	b_dc = dc;
2716 :			dc = coeff[0];
2717 :			if (i != 0) coeff[0] -= b_dc;
2718 :	chl	181
2719 :	edgomez	851	bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcy_tab[coeff[0] + 255].len;;
2720 :			Data->temp[i] = t;
2721 :			if (t != 0) cbp \|= 1 << (5 - i);
2722 :			if (bits >= Data->iMinSAD[0]) break;
2723 :			}
2724 :	chl	181
2725 :	edgomez	851	if (bits < Data->iMinSAD[0]) { // INTRA still looks good, let's add chroma
2726 :			uint32_t iDcScaler = get_dc_scaler(iQuant, 0);
2727 :			//chroma U
2728 :			transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2);
2729 :			fdct(in);
2730 :			in[0] -= 1024;
2731 :			if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler);
2732 :			else quant4_intra(coeff, in, iQuant, iDcScaler);
2733 :	chl	96
2734 :	edgomez	851	bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len;
2735 :			if (t != 0) cbp \|= 1 << (5 - 4);
2736 :	chl	141
2737 :	edgomez	851	if (bits < Data->iMinSAD[0]) {
2738 :	syskin	904	iDcScaler = get_dc_scaler(iQuant, 1);
2739 :	edgomez	851	//chroma V
2740 :			transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2);
2741 :			fdct(in);
2742 :			in[0] -= 1024;
2743 :			if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler);
2744 :			else quant4_intra(coeff, in, iQuant, iDcScaler);
2745 :	chl	96
2746 :	edgomez	851	bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len;
2747 :			if (t != 0) cbp \|= 1 << (5 - 5);
2748 :	chl	96
2749 :	syskin	904	bits += xvid_cbpy_tab[cbp>>2].len;
2750 :			bits += mcbpc_inter_tab[(MODE_INTRA & 7) \| ((cbp & 3) << 3)].len;
2751 :	chl	96	}
2752 :			}
2753 :	edgomez	851	return bits;
2754 :	chl	96	}

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