1 |
/************************************************************************** |
/***************************************************************************** |
2 |
* |
* |
3 |
* XVID MPEG-4 VIDEO CODEC |
* XVID MPEG-4 VIDEO CODEC |
4 |
* motion estimation |
* - Motion Estimation related code - |
5 |
* |
* |
6 |
* This program is an implementation of a part of one or more MPEG-4 |
* Copyright(C) 2002 Christoph Lampert <gruel@web.de> |
7 |
* Video tools as specified in ISO/IEC 14496-2 standard. Those intending |
* 2002 Michael Militzer <michael@xvid.org> |
8 |
* to use this software module in hardware or software products are |
* 2002-2003 Radoslaw Czyz <xvid@syskin.cjb.net> |
|
* advised that its use may infringe existing patents or copyrights, and |
|
|
* any such use would be at such party's own risk. The original |
|
|
* developer of this software module and his/her company, and subsequent |
|
|
* editors and their companies, will have no liability for use of this |
|
|
* software or modifications or derivatives thereof. |
|
9 |
* |
* |
10 |
* This program is free software; you can redistribute it and/or modify |
* This program is free software; you can redistribute it and/or modify |
11 |
* it under the terms of the GNU General Public License as published by |
* it under the terms of the GNU General Public License as published by |
19 |
* |
* |
20 |
* You should have received a copy of the GNU General Public License |
* You should have received a copy of the GNU General Public License |
21 |
* along with this program; if not, write to the Free Software |
* along with this program; if not, write to the Free Software |
22 |
* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
23 |
|
* |
24 |
|
* $Id$ |
25 |
* |
* |
26 |
*************************************************************************/ |
****************************************************************************/ |
27 |
|
|
28 |
#include <assert.h> |
#include <assert.h> |
29 |
#include <stdio.h> |
#include <stdio.h> |
30 |
#include <stdlib.h> |
#include <stdlib.h> |
31 |
#include <string.h> // memcpy |
#include <string.h> /* memcpy */ |
32 |
#include <math.h> // lrint |
#include <math.h> /* lrint */ |
33 |
|
|
34 |
#include "../encoder.h" |
#include "../encoder.h" |
35 |
#include "../utils/mbfunctions.h" |
#include "../utils/mbfunctions.h" |
40 |
#include "motion_est.h" |
#include "motion_est.h" |
41 |
#include "motion.h" |
#include "motion.h" |
42 |
#include "sad.h" |
#include "sad.h" |
43 |
|
#include "gmc.h" |
44 |
#include "../utils/emms.h" |
#include "../utils/emms.h" |
45 |
#include "../dct/fdct.h" |
#include "../dct/fdct.h" |
46 |
|
|
70 |
#define MAX_CHROMA_SAD_FOR_SKIP (22) |
#define MAX_CHROMA_SAD_FOR_SKIP (22) |
71 |
|
|
72 |
#define CHECK_CANDIDATE(X,Y,D) { \ |
#define CHECK_CANDIDATE(X,Y,D) { \ |
73 |
CheckCandidate((X),(Y), (D), &iDirection, data ); } |
CheckCandidate((X),(Y), data, (D) ); } |
74 |
|
|
75 |
|
|
76 |
/***************************************************************************** |
/***************************************************************************** |
77 |
* Code |
* Code |
80 |
static __inline uint32_t |
static __inline uint32_t |
81 |
d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv) |
d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv) |
82 |
{ |
{ |
83 |
int xb, yb; |
int bits; |
84 |
x = qpel ? x<<1 : x; |
const int q = (1 << (iFcode - 1)) - 1; |
85 |
y = qpel ? y<<1 : y; |
|
86 |
|
x <<= qpel; |
87 |
|
y <<= qpel; |
88 |
if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); } |
if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); } |
89 |
|
|
90 |
x -= pred.x; |
x -= pred.x; |
91 |
y -= pred.y; |
bits = (x != 0 ? iFcode:0); |
92 |
|
x = abs(x); |
93 |
if (x) { |
x += q; |
|
x = ABS(x); |
|
|
x += (1 << (iFcode - 1)) - 1; |
|
94 |
x >>= (iFcode - 1); |
x >>= (iFcode - 1); |
95 |
if (x > 32) x = 32; |
bits += mvtab[x]; |
96 |
xb = mvtab[x] + iFcode; |
|
97 |
} else xb = 1; |
y -= pred.y; |
98 |
|
bits += (y != 0 ? iFcode:0); |
99 |
if (y) { |
y = abs(y); |
100 |
y = ABS(y); |
y += q; |
|
y += (1 << (iFcode - 1)) - 1; |
|
101 |
y >>= (iFcode - 1); |
y >>= (iFcode - 1); |
102 |
if (y > 32) y = 32; |
bits += mvtab[y]; |
103 |
yb = mvtab[y] + iFcode; |
|
104 |
} else yb = 1; |
return bits; |
|
return xb + yb; |
|
105 |
} |
} |
106 |
|
|
107 |
static int32_t ChromaSAD2(int fx, int fy, int bx, int by, const SearchData * const data) |
static int32_t ChromaSAD2(const int fx, const int fy, const int bx, const int by, |
108 |
|
const SearchData * const data) |
109 |
{ |
{ |
110 |
int sad; |
int sad; |
111 |
const uint32_t stride = data->iEdgedWidth/2; |
const uint32_t stride = data->iEdgedWidth/2; |
112 |
uint8_t * f_refu = data->RefQ, |
uint8_t *f_refu, *f_refv, *b_refu, *b_refv; |
|
* f_refv = data->RefQ + 8, |
|
|
* b_refu = data->RefQ + 16, |
|
|
* b_refv = data->RefQ + 24; |
|
113 |
|
|
114 |
switch (((fx & 1) << 1) | (fy & 1)) { |
const INTERPOLATE8X8_PTR interpolate8x8_halfpel[] = { |
115 |
case 0: |
NULL, |
116 |
fx = fx / 2; fy = fy / 2; |
interpolate8x8_halfpel_v, |
117 |
f_refu = (uint8_t*)data->RefCU + fy * stride + fx, stride; |
interpolate8x8_halfpel_h, |
118 |
f_refv = (uint8_t*)data->RefCV + fy * stride + fx, stride; |
interpolate8x8_halfpel_hv |
119 |
break; |
}; |
120 |
case 1: |
|
121 |
fx = fx / 2; fy = (fy - 1) / 2; |
int offset = (fx>>1) + (fy>>1)*stride; |
122 |
interpolate8x8_halfpel_v(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding); |
int filter = ((fx & 1) << 1) | (fy & 1); |
123 |
interpolate8x8_halfpel_v(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding); |
|
124 |
break; |
if (filter != 0) { |
125 |
case 2: |
f_refu = data->RefQ; |
126 |
fx = (fx - 1) / 2; fy = fy / 2; |
f_refv = data->RefQ + 8; |
127 |
interpolate8x8_halfpel_h(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding); |
interpolate8x8_halfpel[filter](f_refu, data->RefP[4] + offset, stride, data->rounding); |
128 |
interpolate8x8_halfpel_h(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding); |
interpolate8x8_halfpel[filter](f_refv, data->RefP[5] + offset, stride, data->rounding); |
129 |
break; |
} else { |
130 |
default: |
f_refu = (uint8_t*)data->RefP[4] + offset; |
131 |
fx = (fx - 1) / 2; fy = (fy - 1) / 2; |
f_refv = (uint8_t*)data->RefP[5] + offset; |
|
interpolate8x8_halfpel_hv(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding); |
|
|
interpolate8x8_halfpel_hv(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding); |
|
|
break; |
|
132 |
} |
} |
133 |
|
|
134 |
switch (((bx & 1) << 1) | (by & 1)) { |
offset = (bx>>1) + (by>>1)*stride; |
135 |
case 0: |
filter = ((bx & 1) << 1) | (by & 1); |
136 |
bx = bx / 2; by = by / 2; |
|
137 |
b_refu = (uint8_t*)data->b_RefCU + by * stride + bx, stride; |
if (filter != 0) { |
138 |
b_refv = (uint8_t*)data->b_RefCV + by * stride + bx, stride; |
b_refu = data->RefQ + 16; |
139 |
break; |
b_refv = data->RefQ + 24; |
140 |
case 1: |
interpolate8x8_halfpel[filter](b_refu, data->b_RefP[4] + offset, stride, data->rounding); |
141 |
bx = bx / 2; by = (by - 1) / 2; |
interpolate8x8_halfpel[filter](b_refv, data->b_RefP[5] + offset, stride, data->rounding); |
142 |
interpolate8x8_halfpel_v(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding); |
} else { |
143 |
interpolate8x8_halfpel_v(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding); |
b_refu = (uint8_t*)data->b_RefP[4] + offset; |
144 |
break; |
b_refv = (uint8_t*)data->b_RefP[5] + offset; |
|
case 2: |
|
|
bx = (bx - 1) / 2; by = by / 2; |
|
|
interpolate8x8_halfpel_h(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding); |
|
|
interpolate8x8_halfpel_h(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding); |
|
|
break; |
|
|
default: |
|
|
bx = (bx - 1) / 2; by = (by - 1) / 2; |
|
|
interpolate8x8_halfpel_hv(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding); |
|
|
interpolate8x8_halfpel_hv(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding); |
|
|
break; |
|
145 |
} |
} |
146 |
|
|
147 |
sad = sad8bi(data->CurU, b_refu, f_refu, stride); |
sad = sad8bi(data->CurU, b_refu, f_refu, stride); |
150 |
return sad; |
return sad; |
151 |
} |
} |
152 |
|
|
|
|
|
153 |
static int32_t |
static int32_t |
154 |
ChromaSAD(int dx, int dy, const SearchData * const data) |
ChromaSAD(const int dx, const int dy, const SearchData * const data) |
155 |
{ |
{ |
156 |
int sad; |
int sad; |
157 |
const uint32_t stride = data->iEdgedWidth/2; |
const uint32_t stride = data->iEdgedWidth/2; |
158 |
|
int offset = (dx>>1) + (dy>>1)*stride; |
159 |
|
int next = 1; |
160 |
|
|
161 |
if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently |
if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; /* it has been checked recently */ |
162 |
data->temp[5] = dx; data->temp[6] = dy; // backup |
data->temp[5] = dx; data->temp[6] = dy; /* backup */ |
163 |
|
|
164 |
switch (((dx & 1) << 1) | (dy & 1)) { |
switch (((dx & 1) << 1) | (dy & 1)) { |
165 |
case 0: |
case 0: |
166 |
dx = dx / 2; dy = dy / 2; |
sad = sad8(data->CurU, data->RefP[4] + offset, stride); |
167 |
sad = sad8(data->CurU, data->RefCU + dy * stride + dx, stride); |
sad += sad8(data->CurV, data->RefP[5] + offset, stride); |
|
sad += sad8(data->CurV, data->RefCV + dy * stride + dx, stride); |
|
168 |
break; |
break; |
169 |
case 1: |
case 1: |
170 |
dx = dx / 2; dy = (dy - 1) / 2; |
next = stride; |
|
sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + (dy+1) * stride + dx, stride); |
|
|
sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + (dy+1) * stride + dx, stride); |
|
|
break; |
|
171 |
case 2: |
case 2: |
172 |
dx = (dx - 1) / 2; dy = dy / 2; |
sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + next, stride); |
173 |
sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + dy * stride + dx+1, stride); |
sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + next, stride); |
|
sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + dy * stride + dx+1, stride); |
|
174 |
break; |
break; |
175 |
default: |
default: |
176 |
dx = (dx - 1) / 2; dy = (dy - 1) / 2; |
interpolate8x8_halfpel_hv(data->RefQ, data->RefP[4] + offset, stride, data->rounding); |
|
interpolate8x8_halfpel_hv(data->RefQ, data->RefCU + dy * stride + dx, stride, data->rounding); |
|
177 |
sad = sad8(data->CurU, data->RefQ, stride); |
sad = sad8(data->CurU, data->RefQ, stride); |
178 |
|
|
179 |
interpolate8x8_halfpel_hv(data->RefQ, data->RefCV + dy * stride + dx, stride, data->rounding); |
interpolate8x8_halfpel_hv(data->RefQ, data->RefP[5] + offset, stride, data->rounding); |
180 |
sad += sad8(data->CurV, data->RefQ, stride); |
sad += sad8(data->CurV, data->RefQ, stride); |
181 |
break; |
break; |
182 |
} |
} |
183 |
data->temp[7] = sad; //backup, part 2 |
data->temp[7] = sad; /* backup, part 2 */ |
184 |
return sad; |
return sad; |
185 |
} |
} |
186 |
|
|
187 |
static __inline const uint8_t * |
static __inline const uint8_t * |
188 |
GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data) |
GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data) |
189 |
{ |
{ |
190 |
// dir : 0 = forward, 1 = backward |
/* dir : 0 = forward, 1 = backward */ |
191 |
switch ( (dir << 2) | ((x&1)<<1) | (y&1) ) { |
const uint8_t *const *const direction = ( dir == 0 ? data->RefP : data->b_RefP ); |
192 |
case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth); |
const int picture = ((x&1)<<1) | (y&1); |
193 |
case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); |
const int offset = (x>>1) + (y>>1)*data->iEdgedWidth; |
194 |
case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); |
return direction[picture] + offset; |
|
case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); |
|
|
case 4 : return data->bRef + x/2 + (y/2)*(data->iEdgedWidth); |
|
|
case 5 : return data->bRefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); |
|
|
case 6 : return data->bRefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); |
|
|
default : return data->bRefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); |
|
|
} |
|
195 |
} |
} |
196 |
|
|
197 |
// this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate |
/* this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate */ |
198 |
static __inline const uint8_t * |
static __inline const uint8_t * |
199 |
GetReference(const int x, const int y, const SearchData * const data) |
GetReference(const int x, const int y, const SearchData * const data) |
200 |
{ |
{ |
201 |
switch ( ((x&1)<<1) | (y&1) ) { |
const int picture = ((x&1)<<1) | (y&1); |
202 |
case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth); |
const int offset = (x>>1) + (y>>1)*data->iEdgedWidth; |
203 |
case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); |
return data->RefP[picture] + offset; |
|
case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); |
|
|
default : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); //case 2 |
|
|
} |
|
204 |
} |
} |
205 |
|
|
206 |
static uint8_t * |
static uint8_t * |
207 |
Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data) |
Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data) |
208 |
{ |
{ |
209 |
// create or find a qpel-precision reference picture; return pointer to it |
/* create or find a qpel-precision reference picture; return pointer to it */ |
210 |
uint8_t * Reference = data->RefQ + 16*dir; |
uint8_t * Reference = data->RefQ + 16*dir; |
211 |
const uint32_t iEdgedWidth = data->iEdgedWidth; |
const uint32_t iEdgedWidth = data->iEdgedWidth; |
212 |
const uint32_t rounding = data->rounding; |
const uint32_t rounding = data->rounding; |
217 |
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
218 |
ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
219 |
switch( ((x&1)<<1) + (y&1) ) { |
switch( ((x&1)<<1) + (y&1) ) { |
220 |
case 0: // pure halfpel position |
case 3: /* x and y in qpel resolution - the "corners" (top left/right and */ |
221 |
return (uint8_t *) ref1; |
/* bottom left/right) during qpel refinement */ |
222 |
|
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
223 |
|
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
224 |
|
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
225 |
|
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
226 |
|
ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
227 |
|
ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
228 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
229 |
break; |
break; |
230 |
|
|
231 |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */ |
232 |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
233 |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
234 |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
235 |
break; |
break; |
236 |
|
|
237 |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
case 2: /* x qpel, y halfpel - left or right during qpel refinement */ |
238 |
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
239 |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
240 |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
241 |
break; |
break; |
242 |
|
|
243 |
default: // x and y in qpel resolution - the "corners" (top left/right and |
default: /* pure halfpel position */ |
244 |
// bottom left/right) during qpel refinement |
return (uint8_t *) ref1; |
245 |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
|
|
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
|
|
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
|
|
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
|
|
ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
|
|
ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
|
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
|
|
break; |
|
246 |
} |
} |
247 |
return Reference; |
return Reference; |
248 |
} |
} |
250 |
static uint8_t * |
static uint8_t * |
251 |
Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data) |
Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data) |
252 |
{ |
{ |
253 |
// create or find a qpel-precision reference picture; return pointer to it |
/* create or find a qpel-precision reference picture; return pointer to it */ |
254 |
uint8_t * Reference = data->RefQ + 16*dir; |
uint8_t * Reference = data->RefQ + 16*dir; |
255 |
const uint32_t iEdgedWidth = data->iEdgedWidth; |
const uint32_t iEdgedWidth = data->iEdgedWidth; |
256 |
const uint32_t rounding = data->rounding; |
const uint32_t rounding = data->rounding; |
260 |
|
|
261 |
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
262 |
switch( ((x&1)<<1) + (y&1) ) { |
switch( ((x&1)<<1) + (y&1) ) { |
263 |
case 3: // x and y in qpel resolution - the "corners" (top left/right and |
case 3: |
264 |
// bottom left/right) during qpel refinement |
/* |
265 |
|
* x and y in qpel resolution - the "corners" (top left/right and |
266 |
|
* bottom left/right) during qpel refinement |
267 |
|
*/ |
268 |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
269 |
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
270 |
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
274 |
interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); |
interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); |
275 |
break; |
break; |
276 |
|
|
277 |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */ |
278 |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
279 |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
280 |
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
282 |
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
283 |
break; |
break; |
284 |
|
|
285 |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
case 2: /* x qpel, y halfpel - left or right during qpel refinement */ |
286 |
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
287 |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
288 |
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
290 |
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
291 |
break; |
break; |
292 |
|
|
293 |
case 0: // pure halfpel position |
|
294 |
|
default: /* pure halfpel position */ |
295 |
return (uint8_t *) ref1; |
return (uint8_t *) ref1; |
296 |
} |
} |
297 |
return Reference; |
return Reference; |
300 |
/* CHECK_CANDIATE FUNCTIONS START */ |
/* CHECK_CANDIATE FUNCTIONS START */ |
301 |
|
|
302 |
static void |
static void |
303 |
CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidate16(const int x, const int y, const SearchData * const data, const int Direction) |
304 |
{ |
{ |
305 |
int xc, yc; |
int xc, yc; |
306 |
const uint8_t * Reference; |
const uint8_t * Reference; |
314 |
Reference = GetReference(x, y, data); |
Reference = GetReference(x, y, data); |
315 |
current = data->currentMV; |
current = data->currentMV; |
316 |
xc = x; yc = y; |
xc = x; yc = y; |
317 |
} else { // x and y are in 1/4 precision |
} else { /* x and y are in 1/4 precision */ |
318 |
Reference = Interpolate16x16qpel(x, y, 0, data); |
Reference = Interpolate16x16qpel(x, y, 0, data); |
319 |
xc = x/2; yc = y/2; //for chroma sad |
xc = x/2; yc = y/2; /* for chroma sad */ |
320 |
current = data->currentQMV; |
current = data->currentQMV; |
321 |
} |
} |
322 |
|
|
323 |
sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp); |
324 |
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
325 |
|
|
326 |
sad += (data->lambda16 * t * sad)>>10; |
sad += (data->lambda16 * t * sad)>>10; |
327 |
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; |
data->temp[0] += (data->lambda8 * t * (data->temp[0] + NEIGH_8X8_BIAS))>>10; |
328 |
|
|
329 |
if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
if (data->chroma && sad < data->iMinSAD[0]) |
330 |
|
sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
331 |
(yc >> 1) + roundtab_79[yc & 0x3], data); |
(yc >> 1) + roundtab_79[yc & 0x3], data); |
332 |
|
|
333 |
if (sad < data->iMinSAD[0]) { |
if (sad < data->iMinSAD[0]) { |
334 |
data->iMinSAD[0] = sad; |
data->iMinSAD[0] = sad; |
335 |
current[0].x = x; current[0].y = y; |
current[0].x = x; current[0].y = y; |
336 |
*dir = Direction; |
*data->dir = Direction; |
337 |
} |
} |
338 |
|
|
339 |
if (data->temp[1] < data->iMinSAD[1]) { |
if (data->temp[0] < data->iMinSAD[1]) { |
340 |
data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y = y; } |
data->iMinSAD[1] = data->temp[0]; current[1].x = x; current[1].y = y; } |
341 |
if (data->temp[2] < data->iMinSAD[2]) { |
if (data->temp[1] < data->iMinSAD[2]) { |
342 |
data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; } |
data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; } |
343 |
if (data->temp[3] < data->iMinSAD[3]) { |
if (data->temp[2] < data->iMinSAD[3]) { |
344 |
data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; } |
data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; } |
345 |
if (data->temp[4] < data->iMinSAD[4]) { |
if (data->temp[3] < data->iMinSAD[4]) { |
346 |
data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; } |
data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; } |
|
|
|
347 |
} |
} |
348 |
|
|
349 |
static void |
static void |
350 |
CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidate8(const int x, const int y, const SearchData * const data, const int Direction) |
351 |
{ |
{ |
352 |
int32_t sad; uint32_t t; |
int32_t sad; uint32_t t; |
353 |
const uint8_t * Reference; |
const uint8_t * Reference; |
354 |
|
VECTOR * current; |
355 |
|
|
356 |
if ( (x > data->max_dx) || (x < data->min_dx) |
if ( (x > data->max_dx) || (x < data->min_dx) |
357 |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
358 |
|
|
359 |
if (!data->qpel_precision) Reference = GetReference(x, y, data); |
if (!data->qpel_precision) { |
360 |
else Reference = Interpolate8x8qpel(x, y, 0, 0, data); |
Reference = GetReference(x, y, data); |
361 |
|
current = data->currentMV; |
362 |
|
} else { /* x and y are in 1/4 precision */ |
363 |
|
Reference = Interpolate8x8qpel(x, y, 0, 0, data); |
364 |
|
current = data->currentQMV; |
365 |
|
} |
366 |
|
|
367 |
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
368 |
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
371 |
|
|
372 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
373 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
374 |
data->currentMV->x = x; data->currentMV->y = y; |
current->x = x; current->y = y; |
375 |
*dir = Direction; |
*data->dir = Direction; |
376 |
} |
} |
377 |
} |
} |
378 |
|
|
|
|
|
379 |
static void |
static void |
380 |
CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidate32(const int x, const int y, const SearchData * const data, const int Direction) |
381 |
{ |
{ |
382 |
uint32_t t; |
uint32_t t; |
383 |
const uint8_t * Reference; |
const uint8_t * Reference; |
384 |
|
int sad; |
385 |
|
|
386 |
if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero integer value |
if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || /* non-zero even value */ |
387 |
(x > data->max_dx) || (x < data->min_dx) |
(x > data->max_dx) || (x < data->min_dx) |
388 |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
389 |
|
|
390 |
Reference = GetReference(x, y, data); |
Reference = GetReference(x, y, data); |
391 |
t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1); |
t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1); |
392 |
|
|
393 |
data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
sad = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp); |
394 |
|
|
395 |
data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10; |
sad += (data->lambda16 * t * sad) >> 10; |
396 |
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; |
data->temp[0] += (data->lambda8 * t * (data->temp[0] + NEIGH_8X8_BIAS))>>10; |
397 |
|
|
398 |
if (data->temp[0] < data->iMinSAD[0]) { |
if (sad < data->iMinSAD[0]) { |
399 |
data->iMinSAD[0] = data->temp[0]; |
data->iMinSAD[0] = sad; |
400 |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
401 |
*dir = Direction; } |
*data->dir = Direction; |
402 |
|
} |
403 |
|
|
404 |
if (data->temp[1] < data->iMinSAD[1]) { |
if (data->temp[0] < data->iMinSAD[1]) { |
405 |
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
data->iMinSAD[1] = data->temp[0]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
406 |
if (data->temp[2] < data->iMinSAD[2]) { |
if (data->temp[1] < data->iMinSAD[2]) { |
407 |
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
data->iMinSAD[2] = data->temp[1]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
408 |
if (data->temp[3] < data->iMinSAD[3]) { |
if (data->temp[2] < data->iMinSAD[3]) { |
409 |
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
data->iMinSAD[3] = data->temp[2]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
410 |
if (data->temp[4] < data->iMinSAD[4]) { |
if (data->temp[3] < data->iMinSAD[4]) { |
411 |
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
data->iMinSAD[4] = data->temp[3]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
412 |
} |
} |
413 |
|
|
414 |
static void |
static void |
415 |
CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidate16no4v(const int x, const int y, const SearchData * const data, const int Direction) |
416 |
{ |
{ |
417 |
int32_t sad, xc, yc; |
int32_t sad, xc, yc; |
418 |
const uint8_t * Reference; |
const uint8_t * Reference; |
419 |
uint32_t t; |
uint32_t t; |
420 |
VECTOR * current; |
VECTOR * current; |
421 |
|
|
422 |
if ( (x > data->max_dx) | ( x < data->min_dx) |
if ( (x > data->max_dx) || ( x < data->min_dx) |
423 |
| (y > data->max_dy) | (y < data->min_dy) ) return; |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
424 |
|
|
425 |
if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; //non-zero even value |
if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; /* non-zero even value */ |
426 |
|
|
427 |
if (data->qpel_precision) { // x and y are in 1/4 precision |
if (data->qpel_precision) { /* x and y are in 1/4 precision */ |
428 |
Reference = Interpolate16x16qpel(x, y, 0, data); |
Reference = Interpolate16x16qpel(x, y, 0, data); |
429 |
current = data->currentQMV; |
current = data->currentQMV; |
430 |
xc = x/2; yc = y/2; |
xc = x/2; yc = y/2; |
439 |
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
440 |
sad += (data->lambda16 * t * sad)>>10; |
sad += (data->lambda16 * t * sad)>>10; |
441 |
|
|
442 |
if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
if (data->chroma && sad < *data->iMinSAD) |
443 |
|
sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
444 |
(yc >> 1) + roundtab_79[yc & 0x3], data); |
(yc >> 1) + roundtab_79[yc & 0x3], data); |
445 |
|
|
446 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
447 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
448 |
current->x = x; current->y = y; |
current->x = x; current->y = y; |
449 |
*dir = Direction; |
*data->dir = Direction; |
450 |
|
} |
451 |
|
} |
452 |
|
|
453 |
|
static void |
454 |
|
CheckCandidate16I(const int x, const int y, const SearchData * const data, const int Direction) |
455 |
|
{ |
456 |
|
int sad; |
457 |
|
// int xc, yc; |
458 |
|
const uint8_t * Reference; |
459 |
|
// VECTOR * current; |
460 |
|
|
461 |
|
if ( (x > data->max_dx) || ( x < data->min_dx) |
462 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
463 |
|
|
464 |
|
Reference = GetReference(x, y, data); |
465 |
|
// xc = x; yc = y; |
466 |
|
|
467 |
|
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
468 |
|
// sad += d_mv_bits(x, y, data->predMV, data->iFcode, 0, 0); |
469 |
|
|
470 |
|
/* if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
471 |
|
(yc >> 1) + roundtab_79[yc & 0x3], data); |
472 |
|
*/ |
473 |
|
|
474 |
|
if (sad < data->iMinSAD[0]) { |
475 |
|
data->iMinSAD[0] = sad; |
476 |
|
data->currentMV[0].x = x; data->currentMV[0].y = y; |
477 |
|
*data->dir = Direction; |
478 |
} |
} |
479 |
} |
} |
480 |
|
|
481 |
static void |
static void |
482 |
CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidate32I(const int x, const int y, const SearchData * const data, const int Direction) |
483 |
{ |
{ |
484 |
// maximum speed - for P/B/I decision |
/* maximum speed - for P/B/I decision */ |
485 |
int32_t sad; |
int32_t sad; |
486 |
|
|
487 |
if ( (x > data->max_dx) || (x < data->min_dx) |
if ( (x > data->max_dx) || (x < data->min_dx) |
488 |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
489 |
|
|
490 |
sad = sad32v_c(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), |
sad = sad32v_c(data->Cur, data->RefP[0] + x + y*((int)data->iEdgedWidth), |
491 |
data->iEdgedWidth, data->temp+1); |
data->iEdgedWidth, data->temp); |
492 |
|
|
493 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
494 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
495 |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
496 |
*dir = Direction; |
*data->dir = Direction; |
497 |
} |
} |
498 |
if (data->temp[1] < data->iMinSAD[1]) { |
if (data->temp[0] < data->iMinSAD[1]) { |
499 |
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
data->iMinSAD[1] = data->temp[0]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
500 |
if (data->temp[2] < data->iMinSAD[2]) { |
if (data->temp[1] < data->iMinSAD[2]) { |
501 |
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
data->iMinSAD[2] = data->temp[1]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
502 |
if (data->temp[3] < data->iMinSAD[3]) { |
if (data->temp[2] < data->iMinSAD[3]) { |
503 |
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
data->iMinSAD[3] = data->temp[2]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
504 |
if (data->temp[4] < data->iMinSAD[4]) { |
if (data->temp[3] < data->iMinSAD[4]) { |
505 |
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
data->iMinSAD[4] = data->temp[3]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
506 |
|
|
507 |
} |
} |
508 |
|
|
509 |
static void |
static void |
510 |
CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidateInt(const int xf, const int yf, const SearchData * const data, const int Direction) |
511 |
{ |
{ |
512 |
int32_t sad, xb, yb, xcf, ycf, xcb, ycb; |
int32_t sad, xb, yb, xcf, ycf, xcb, ycb; |
513 |
uint32_t t; |
uint32_t t; |
514 |
const uint8_t *ReferenceF, *ReferenceB; |
const uint8_t *ReferenceF, *ReferenceB; |
515 |
VECTOR *current; |
VECTOR *current; |
516 |
|
|
517 |
if ( (xf > data->max_dx) | (xf < data->min_dx) |
if ((xf > data->max_dx) || (xf < data->min_dx) || |
518 |
| (yf > data->max_dy) | (yf < data->min_dy) ) return; |
(yf > data->max_dy) || (yf < data->min_dy)) |
519 |
|
return; |
520 |
|
|
521 |
if (!data->qpel_precision) { |
if (!data->qpel_precision) { |
522 |
ReferenceF = GetReference(xf, yf, data); |
ReferenceF = GetReference(xf, yf, data); |
540 |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
541 |
sad += (data->lambda16 * t * sad)>>10; |
sad += (data->lambda16 * t * sad)>>10; |
542 |
|
|
543 |
if (data->chroma) sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3], |
if (data->chroma && sad < *data->iMinSAD) |
544 |
|
sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3], |
545 |
(ycf >> 1) + roundtab_79[ycf & 0x3], |
(ycf >> 1) + roundtab_79[ycf & 0x3], |
546 |
(xcb >> 1) + roundtab_79[xcb & 0x3], |
(xcb >> 1) + roundtab_79[xcb & 0x3], |
547 |
(ycb >> 1) + roundtab_79[ycb & 0x3], data); |
(ycb >> 1) + roundtab_79[ycb & 0x3], data); |
549 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
550 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
551 |
current->x = xf; current->y = yf; |
current->x = xf; current->y = yf; |
552 |
*dir = Direction; |
*data->dir = Direction; |
553 |
} |
} |
554 |
} |
} |
555 |
|
|
556 |
static void |
static void |
557 |
CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidateDirect(const int x, const int y, const SearchData * const data, const int Direction) |
558 |
{ |
{ |
559 |
int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0; |
int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0; |
560 |
uint32_t k; |
uint32_t k; |
562 |
const uint8_t *ReferenceB; |
const uint8_t *ReferenceB; |
563 |
VECTOR mvs, b_mvs; |
VECTOR mvs, b_mvs; |
564 |
|
|
565 |
if (( x > 31) | ( x < -32) | ( y > 31) | (y < -32)) return; |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
566 |
|
|
567 |
for (k = 0; k < 4; k++) { |
for (k = 0; k < 4; k++) { |
568 |
mvs.x = data->directmvF[k].x + x; |
mvs.x = data->directmvF[k].x + x; |
575 |
data->directmvB[k].y |
data->directmvB[k].y |
576 |
: mvs.y - data->referencemv[k].y); |
: mvs.y - data->referencemv[k].y); |
577 |
|
|
578 |
if ( (mvs.x > data->max_dx) | (mvs.x < data->min_dx) |
if ((mvs.x > data->max_dx) || (mvs.x < data->min_dx) || |
579 |
| (mvs.y > data->max_dy) | (mvs.y < data->min_dy) |
(mvs.y > data->max_dy) || (mvs.y < data->min_dy) || |
580 |
| (b_mvs.x > data->max_dx) | (b_mvs.x < data->min_dx) |
(b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx) || |
581 |
| (b_mvs.y > data->max_dy) | (b_mvs.y < data->min_dy) ) return; |
(b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) ) |
582 |
|
return; |
583 |
|
|
584 |
if (data->qpel) { |
if (data->qpel) { |
585 |
xcf += mvs.x/2; ycf += mvs.y/2; |
xcf += mvs.x/2; ycf += mvs.y/2; |
587 |
} else { |
} else { |
588 |
xcf += mvs.x; ycf += mvs.y; |
xcf += mvs.x; ycf += mvs.y; |
589 |
xcb += b_mvs.x; ycb += b_mvs.y; |
xcb += b_mvs.x; ycb += b_mvs.y; |
590 |
mvs.x *= 2; mvs.y *= 2; //we move to qpel precision anyway |
mvs.x *= 2; mvs.y *= 2; /* we move to qpel precision anyway */ |
591 |
b_mvs.x *= 2; b_mvs.y *= 2; |
b_mvs.x *= 2; b_mvs.y *= 2; |
592 |
} |
} |
593 |
|
|
601 |
|
|
602 |
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
603 |
|
|
604 |
if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
if (data->chroma && sad < *data->iMinSAD) |
605 |
|
sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
606 |
(ycf >> 3) + roundtab_76[ycf & 0xf], |
(ycf >> 3) + roundtab_76[ycf & 0xf], |
607 |
(xcb >> 3) + roundtab_76[xcb & 0xf], |
(xcb >> 3) + roundtab_76[xcb & 0xf], |
608 |
(ycb >> 3) + roundtab_76[ycb & 0xf], data); |
(ycb >> 3) + roundtab_76[ycb & 0xf], data); |
610 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
611 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
612 |
data->currentMV->x = x; data->currentMV->y = y; |
data->currentMV->x = x; data->currentMV->y = y; |
613 |
*dir = Direction; |
*data->dir = Direction; |
614 |
} |
} |
615 |
} |
} |
616 |
|
|
617 |
static void |
static void |
618 |
CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidateDirectno4v(const int x, const int y, const SearchData * const data, const int Direction) |
619 |
{ |
{ |
620 |
int32_t sad, xcf, ycf, xcb, ycb; |
int32_t sad, xcf, ycf, xcb, ycb; |
621 |
const uint8_t *ReferenceF; |
const uint8_t *ReferenceF; |
622 |
const uint8_t *ReferenceB; |
const uint8_t *ReferenceB; |
623 |
VECTOR mvs, b_mvs; |
VECTOR mvs, b_mvs; |
624 |
|
|
625 |
if (( x > 31) | ( x < -32) | ( y > 31) | (y < -32)) return; |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
626 |
|
|
627 |
mvs.x = data->directmvF[0].x + x; |
mvs.x = data->directmvF[0].x + x; |
628 |
b_mvs.x = ((x == 0) ? |
b_mvs.x = ((x == 0) ? |
634 |
data->directmvB[0].y |
data->directmvB[0].y |
635 |
: mvs.y - data->referencemv[0].y); |
: mvs.y - data->referencemv[0].y); |
636 |
|
|
637 |
if ( (mvs.x > data->max_dx) | (mvs.x < data->min_dx) |
if ( (mvs.x > data->max_dx) || (mvs.x < data->min_dx) |
638 |
| (mvs.y > data->max_dy) | (mvs.y < data->min_dy) |
|| (mvs.y > data->max_dy) || (mvs.y < data->min_dy) |
639 |
| (b_mvs.x > data->max_dx) | (b_mvs.x < data->min_dx) |
|| (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx) |
640 |
| (b_mvs.y > data->max_dy) | (b_mvs.y < data->min_dy) ) return; |
|| (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) ) return; |
641 |
|
|
642 |
if (data->qpel) { |
if (data->qpel) { |
643 |
xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2); |
xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2); |
654 |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
655 |
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
656 |
|
|
657 |
if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
if (data->chroma && sad < *data->iMinSAD) |
658 |
|
sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
659 |
(ycf >> 3) + roundtab_76[ycf & 0xf], |
(ycf >> 3) + roundtab_76[ycf & 0xf], |
660 |
(xcb >> 3) + roundtab_76[xcb & 0xf], |
(xcb >> 3) + roundtab_76[xcb & 0xf], |
661 |
(ycb >> 3) + roundtab_76[ycb & 0xf], data); |
(ycb >> 3) + roundtab_76[ycb & 0xf], data); |
663 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
664 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
665 |
data->currentMV->x = x; data->currentMV->y = y; |
data->currentMV->x = x; data->currentMV->y = y; |
666 |
*dir = Direction; |
*data->dir = Direction; |
667 |
} |
} |
668 |
} |
} |
669 |
|
|
670 |
|
|
671 |
static void |
static void |
672 |
CheckCandidateBits16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidateRD16(const int x, const int y, const SearchData * const data, const int Direction) |
673 |
{ |
{ |
674 |
|
|
675 |
static int16_t in[64], coeff[64]; |
int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; |
676 |
int32_t bits = 0, sum; |
int32_t rd = 0; |
677 |
VECTOR * current; |
VECTOR * current; |
678 |
const uint8_t * ptr; |
const uint8_t * ptr; |
679 |
int i, cbp = 0, t, xc, yc; |
int i, cbp = 0, t, xc, yc; |
685 |
ptr = GetReference(x, y, data); |
ptr = GetReference(x, y, data); |
686 |
current = data->currentMV; |
current = data->currentMV; |
687 |
xc = x; yc = y; |
xc = x; yc = y; |
688 |
} else { // x and y are in 1/4 precision |
} else { /* x and y are in 1/4 precision */ |
689 |
ptr = Interpolate16x16qpel(x, y, 0, data); |
ptr = Interpolate16x16qpel(x, y, 0, data); |
690 |
current = data->currentQMV; |
current = data->currentQMV; |
691 |
xc = x/2; yc = y/2; |
xc = x/2; yc = y/2; |
694 |
for(i = 0; i < 4; i++) { |
for(i = 0; i < 4; i++) { |
695 |
int s = 8*((i&1) + (i>>1)*data->iEdgedWidth); |
int s = 8*((i&1) + (i>>1)*data->iEdgedWidth); |
696 |
transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth); |
transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth); |
697 |
fdct(in); |
rd += data->temp[i] = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, i); |
|
if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); |
|
|
else sum = quant4_inter(coeff, in, data->lambda16); |
|
|
if (sum > 0) { |
|
|
cbp |= 1 << (5 - i); |
|
|
bits += data->temp[i] = CodeCoeffInter_CalcBits(coeff, scan_tables[0]); |
|
|
} else data->temp[i] = 0; |
|
698 |
} |
} |
699 |
|
|
700 |
bits += t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
rd += t = BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
701 |
|
|
702 |
if (bits < data->iMinSAD[0]) { // there is still a chance, adding chroma |
if (data->temp[0] + t < data->iMinSAD[1]) { |
703 |
|
data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; data->cbp[1] = (data->cbp[1]&~32) | (cbp&32); } |
704 |
|
if (data->temp[1] < data->iMinSAD[2]) { |
705 |
|
data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; data->cbp[1] = (data->cbp[1]&~16) | (cbp&16); } |
706 |
|
if (data->temp[2] < data->iMinSAD[3]) { |
707 |
|
data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; data->cbp[1] = (data->cbp[1]&~8) | (cbp&8); } |
708 |
|
if (data->temp[3] < data->iMinSAD[4]) { |
709 |
|
data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; data->cbp[1] = (data->cbp[1]&~4) | (cbp&4); } |
710 |
|
|
711 |
|
rd += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; |
712 |
|
|
713 |
|
if (rd >= data->iMinSAD[0]) return; |
714 |
|
|
715 |
|
/* chroma */ |
716 |
xc = (xc >> 1) + roundtab_79[xc & 0x3]; |
xc = (xc >> 1) + roundtab_79[xc & 0x3]; |
717 |
yc = (yc >> 1) + roundtab_79[yc & 0x3]; |
yc = (yc >> 1) + roundtab_79[yc & 0x3]; |
718 |
|
|
719 |
//chroma U |
/* chroma U */ |
720 |
ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefCU, 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); |
ptr = interpolate8x8_switch2(data->RefQ, data->RefP[4], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); |
721 |
transfer_8to16subro(in, ptr, data->CurU, data->iEdgedWidth/2); |
transfer_8to16subro(in, data->CurU, ptr, data->iEdgedWidth/2); |
722 |
fdct(in); |
rd += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 4); |
723 |
if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); |
if (rd >= data->iMinSAD[0]) return; |
724 |
else sum = quant4_inter(coeff, in, data->lambda16); |
|
725 |
if (sum > 0) { |
/* chroma V */ |
726 |
cbp |= 1 << (5 - 4); |
ptr = interpolate8x8_switch2(data->RefQ, data->RefP[5], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); |
727 |
bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); |
transfer_8to16subro(in, data->CurV, ptr, data->iEdgedWidth/2); |
728 |
} |
rd += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5); |
|
|
|
|
if (bits < data->iMinSAD[0]) { |
|
|
//chroma V |
|
|
ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefCV, 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); |
|
|
transfer_8to16subro(in, ptr, data->CurV, data->iEdgedWidth/2); |
|
|
fdct(in); |
|
|
if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); |
|
|
else sum = quant4_inter(coeff, in, data->lambda16); |
|
|
if (sum > 0) { |
|
|
cbp |= 1 << (5 - 5); |
|
|
bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); |
|
|
} |
|
|
} |
|
|
} |
|
729 |
|
|
730 |
bits += cbpy_tab[15-(cbp>>2)].len; |
rd += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; |
|
bits += mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; |
|
731 |
|
|
732 |
if (bits < data->iMinSAD[0]) { |
if (rd < data->iMinSAD[0]) { |
733 |
data->iMinSAD[0] = bits; |
data->iMinSAD[0] = rd; |
734 |
current[0].x = x; current[0].y = y; |
current[0].x = x; current[0].y = y; |
735 |
*dir = Direction; |
*data->dir = Direction; |
736 |
|
*data->cbp = cbp; |
737 |
} |
} |
|
|
|
|
if (data->temp[0] + t < data->iMinSAD[1]) { |
|
|
data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; } |
|
|
if (data->temp[1] < data->iMinSAD[2]) { |
|
|
data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; } |
|
|
if (data->temp[2] < data->iMinSAD[3]) { |
|
|
data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; } |
|
|
if (data->temp[3] < data->iMinSAD[4]) { |
|
|
data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; } |
|
|
|
|
738 |
} |
} |
739 |
|
|
740 |
static void |
static void |
741 |
CheckCandidateBits8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidateRD8(const int x, const int y, const SearchData * const data, const int Direction) |
742 |
{ |
{ |
743 |
|
|
744 |
static int16_t in[64], coeff[64]; |
int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; |
745 |
int32_t sum, bits; |
int32_t rd; |
746 |
VECTOR * current; |
VECTOR * current; |
747 |
const uint8_t * ptr; |
const uint8_t * ptr; |
748 |
int cbp; |
int cbp = 0; |
749 |
|
|
750 |
if ( (x > data->max_dx) || (x < data->min_dx) |
if ( (x > data->max_dx) || (x < data->min_dx) |
751 |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
753 |
if (!data->qpel_precision) { |
if (!data->qpel_precision) { |
754 |
ptr = GetReference(x, y, data); |
ptr = GetReference(x, y, data); |
755 |
current = data->currentMV; |
current = data->currentMV; |
756 |
} else { // x and y are in 1/4 precision |
} else { /* x and y are in 1/4 precision */ |
757 |
ptr = Interpolate8x8qpel(x, y, 0, 0, data); |
ptr = Interpolate8x8qpel(x, y, 0, 0, data); |
758 |
current = data->currentQMV; |
current = data->currentQMV; |
759 |
} |
} |
760 |
|
|
761 |
transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth); |
transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth); |
762 |
fdct(in); |
rd = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5); |
763 |
if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); |
rd += BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
764 |
else sum = quant4_inter(coeff, in, data->lambda16); |
|
765 |
if (sum > 0) { |
if (rd < data->iMinSAD[0]) { |
766 |
bits = CodeCoeffInter_CalcBits(coeff, scan_tables[0]); |
*data->cbp = cbp; |
767 |
cbp = 1; |
data->iMinSAD[0] = rd; |
|
} else cbp = bits = 0; |
|
|
|
|
|
bits += sum = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
|
|
|
|
|
if (bits < data->iMinSAD[0]) { |
|
|
data->temp[0] = cbp; |
|
|
data->iMinSAD[0] = bits; |
|
768 |
current[0].x = x; current[0].y = y; |
current[0].x = x; current[0].y = y; |
769 |
*dir = Direction; |
*data->dir = Direction; |
770 |
} |
} |
771 |
} |
} |
772 |
|
|
775 |
/* MAINSEARCH FUNCTIONS START */ |
/* MAINSEARCH FUNCTIONS START */ |
776 |
|
|
777 |
static void |
static void |
778 |
AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection) |
AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection, CheckFunc * const CheckCandidate) |
779 |
{ |
{ |
780 |
|
|
781 |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
782 |
|
|
783 |
int iDirection; |
unsigned int * const iDirection = data->dir; |
784 |
|
|
785 |
for(;;) { //forever |
for(;;) { /* forever */ |
786 |
iDirection = 0; |
*iDirection = 0; |
787 |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
788 |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
789 |
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
791 |
|
|
792 |
/* now we're doing diagonal checks near our candidate */ |
/* now we're doing diagonal checks near our candidate */ |
793 |
|
|
794 |
if (iDirection) { //if anything found |
if (*iDirection) { /* if anything found */ |
795 |
bDirection = iDirection; |
bDirection = *iDirection; |
796 |
iDirection = 0; |
*iDirection = 0; |
797 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
798 |
if (bDirection & 3) { //our candidate is left or right |
if (bDirection & 3) { /* our candidate is left or right */ |
799 |
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
800 |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
801 |
} else { // what remains here is up or down |
} else { /* what remains here is up or down */ |
802 |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
803 |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
804 |
} |
} |
805 |
|
|
806 |
if (iDirection) { |
if (*iDirection) { |
807 |
bDirection += iDirection; |
bDirection += *iDirection; |
808 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
809 |
} |
} |
810 |
} else { //about to quit, eh? not so fast.... |
} else { /* about to quit, eh? not so fast.... */ |
811 |
switch (bDirection) { |
switch (bDirection) { |
812 |
case 2: |
case 2: |
813 |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
845 |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
846 |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
847 |
break; |
break; |
848 |
default: //1+2+4+8 == we didn't find anything at all |
default: /* 1+2+4+8 == we didn't find anything at all */ |
849 |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
850 |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
851 |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
852 |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
853 |
break; |
break; |
854 |
} |
} |
855 |
if (!iDirection) break; //ok, the end. really |
if (!*iDirection) break; /* ok, the end. really */ |
856 |
bDirection = iDirection; |
bDirection = *iDirection; |
857 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
858 |
} |
} |
859 |
} |
} |
860 |
} |
} |
861 |
|
|
862 |
static void |
static void |
863 |
SquareSearch(int x, int y, const SearchData * const data, int bDirection) |
SquareSearch(int x, int y, const SearchData * const data, int bDirection, CheckFunc * const CheckCandidate) |
864 |
{ |
{ |
865 |
int iDirection; |
unsigned int * const iDirection = data->dir; |
866 |
|
|
867 |
do { |
do { |
868 |
iDirection = 0; |
*iDirection = 0; |
869 |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64); |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64); |
870 |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128); |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128); |
871 |
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32); |
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32); |
875 |
if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128); |
if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128); |
876 |
if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128); |
if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128); |
877 |
|
|
878 |
bDirection = iDirection; |
bDirection = *iDirection; |
879 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
880 |
} while (iDirection); |
} while (*iDirection); |
881 |
} |
} |
882 |
|
|
883 |
static void |
static void |
884 |
DiamondSearch(int x, int y, const SearchData * const data, int bDirection) |
DiamondSearch(int x, int y, const SearchData * const data, int bDirection, CheckFunc * const CheckCandidate) |
885 |
{ |
{ |
886 |
|
|
887 |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
888 |
|
|
889 |
int iDirection; |
unsigned int * const iDirection = data->dir; |
890 |
|
|
891 |
do { |
do { |
892 |
iDirection = 0; |
*iDirection = 0; |
893 |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
894 |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
895 |
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
897 |
|
|
898 |
/* now we're doing diagonal checks near our candidate */ |
/* now we're doing diagonal checks near our candidate */ |
899 |
|
|
900 |
if (iDirection) { //checking if anything found |
if (*iDirection) { /* checking if anything found */ |
901 |
bDirection = iDirection; |
bDirection = *iDirection; |
902 |
iDirection = 0; |
*iDirection = 0; |
903 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
904 |
if (bDirection & 3) { //our candidate is left or right |
if (bDirection & 3) { /* our candidate is left or right */ |
905 |
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
906 |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
907 |
} else { // what remains here is up or down |
} else { /* what remains here is up or down */ |
908 |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
909 |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
910 |
} |
} |
911 |
bDirection += iDirection; |
bDirection += *iDirection; |
912 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
913 |
} |
} |
914 |
} |
} |
915 |
while (iDirection); |
while (*iDirection); |
916 |
} |
} |
917 |
|
|
918 |
/* MAINSEARCH FUNCTIONS END */ |
/* MAINSEARCH FUNCTIONS END */ |
919 |
|
|
920 |
static void |
static void |
921 |
SubpelRefine(const SearchData * const data) |
SubpelRefine(const SearchData * const data, CheckFunc * const CheckCandidate) |
922 |
{ |
{ |
923 |
/* Do a half-pel or q-pel refinement */ |
/* Do a half-pel or q-pel refinement */ |
924 |
const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV; |
const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV; |
|
int iDirection; //only needed because macro expects it |
|
925 |
|
|
926 |
CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0); |
CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0); |
927 |
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0); |
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0); |
939 |
const uint32_t stride, const uint32_t iQuant, int rrv) |
const uint32_t stride, const uint32_t iQuant, int rrv) |
940 |
|
|
941 |
{ |
{ |
942 |
|
int offset = (x + y*stride)*8; |
943 |
if(!rrv) { |
if(!rrv) { |
944 |
uint32_t sadC = sad8(current->u + x*8 + y*stride*8, |
uint32_t sadC = sad8(current->u + offset, |
945 |
reference->u + x*8 + y*stride*8, stride); |
reference->u + offset, stride); |
946 |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
947 |
sadC += sad8(current->v + (x + y*stride)*8, |
sadC += sad8(current->v + offset, |
948 |
reference->v + (x + y*stride)*8, stride); |
reference->v + offset, stride); |
949 |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
950 |
return 1; |
return 1; |
951 |
|
|
952 |
} else { |
} else { |
953 |
uint32_t sadC = sad16(current->u + x*16 + y*stride*16, |
uint32_t sadC = sad16(current->u + 2*offset, |
954 |
reference->u + x*16 + y*stride*16, stride, 256*4096); |
reference->u + 2*offset, stride, 256*4096); |
955 |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
956 |
sadC += sad16(current->v + (x + y*stride)*16, |
sadC += sad16(current->v + 2*offset, |
957 |
reference->v + (x + y*stride)*16, stride, 256*4096); |
reference->v + 2*offset, stride, 256*4096); |
958 |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
959 |
return 1; |
return 1; |
960 |
} |
} |
961 |
} |
} |
962 |
|
|
963 |
static __inline void |
static __inline void |
964 |
SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
ZeroMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
965 |
{ |
{ |
966 |
pMB->mode = MODE_NOT_CODED; |
pMB->mode = MODE_INTER; |
967 |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV; |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV; |
968 |
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV; |
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV; |
969 |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
970 |
} |
} |
971 |
|
|
972 |
|
static __inline void |
973 |
|
ModeDecision(SearchData * const Data, |
974 |
|
MACROBLOCK * const pMB, |
975 |
|
const MACROBLOCK * const pMBs, |
976 |
|
const int x, const int y, |
977 |
|
const MBParam * const pParam, |
978 |
|
const uint32_t MotionFlags, |
979 |
|
const uint32_t VopFlags, |
980 |
|
const uint32_t VolFlags, |
981 |
|
const IMAGE * const pCurrent, |
982 |
|
const IMAGE * const pRef, |
983 |
|
const IMAGE * const vGMC, |
984 |
|
const int coding_type) |
985 |
|
{ |
986 |
|
int mode = MODE_INTER; |
987 |
|
int mcsel = 0; |
988 |
|
int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0); |
989 |
|
const uint32_t iQuant = pMB->quant; |
990 |
|
|
991 |
|
const int skip_possible = (coding_type == P_VOP) && (pMB->dquant == 0); |
992 |
|
|
993 |
|
pMB->mcsel = 0; |
994 |
|
|
995 |
|
if (!(VopFlags & XVID_VOP_MODEDECISION_RD)) { /* normal, fast, SAD-based mode decision */ |
996 |
|
int sad; |
997 |
|
int InterBias = MV16_INTER_BIAS; |
998 |
|
if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
999 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) { |
1000 |
|
mode = MODE_INTER; |
1001 |
|
sad = Data->iMinSAD[0]; |
1002 |
|
} else { |
1003 |
|
mode = MODE_INTER4V; |
1004 |
|
sad = Data->iMinSAD[1] + Data->iMinSAD[2] + |
1005 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant; |
1006 |
|
Data->iMinSAD[0] = sad; |
1007 |
|
} |
1008 |
|
|
1009 |
|
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
1010 |
|
if (skip_possible && (pMB->sad16 < (int)iQuant * MAX_SAD00_FOR_SKIP)) |
1011 |
|
if ( (100*sad)/(pMB->sad16+1) > FINAL_SKIP_THRESH) |
1012 |
|
if (Data->chroma || SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant, Data->rrv)) { |
1013 |
|
mode = MODE_NOT_CODED; |
1014 |
|
sad = 0; |
1015 |
|
} |
1016 |
|
|
1017 |
|
/* mcsel */ |
1018 |
|
if (coding_type == S_VOP) { |
1019 |
|
|
1020 |
|
int32_t iSAD = sad16(Data->Cur, |
1021 |
|
vGMC->y + 16*y*Data->iEdgedWidth + 16*x, Data->iEdgedWidth, 65536); |
1022 |
|
|
1023 |
|
if (Data->chroma) { |
1024 |
|
iSAD += sad8(Data->CurU, vGMC->u + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); |
1025 |
|
iSAD += sad8(Data->CurV, vGMC->v + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); |
1026 |
|
} |
1027 |
|
|
1028 |
|
if (iSAD <= sad) { /* mode decision GMC */ |
1029 |
|
mode = MODE_INTER; |
1030 |
|
mcsel = 1; |
1031 |
|
sad = iSAD; |
1032 |
|
} |
1033 |
|
|
1034 |
|
} |
1035 |
|
|
1036 |
|
/* intra decision */ |
1037 |
|
|
1038 |
|
if (iQuant > 8) InterBias += 100 * (iQuant - 8); /* to make high quants work */ |
1039 |
|
if (y != 0) |
1040 |
|
if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; |
1041 |
|
if (x != 0) |
1042 |
|
if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; |
1043 |
|
|
1044 |
|
if (Data->chroma) InterBias += 50; /* dev8(chroma) ??? <-- yes, we need dev8 (no big difference though) */ |
1045 |
|
if (Data->rrv) InterBias *= 4; |
1046 |
|
|
1047 |
|
if (InterBias < sad) { |
1048 |
|
int32_t deviation; |
1049 |
|
if (!Data->rrv) |
1050 |
|
deviation = dev16(Data->Cur, Data->iEdgedWidth); |
1051 |
|
else |
1052 |
|
deviation = dev16(Data->Cur, Data->iEdgedWidth) + /* dev32() */ |
1053 |
|
dev16(Data->Cur+16, Data->iEdgedWidth) + |
1054 |
|
dev16(Data->Cur + 16*Data->iEdgedWidth, Data->iEdgedWidth) + |
1055 |
|
dev16(Data->Cur+16+16*Data->iEdgedWidth, Data->iEdgedWidth); |
1056 |
|
|
1057 |
|
if (deviation < (sad - InterBias)) mode = MODE_INTRA; |
1058 |
|
} |
1059 |
|
|
1060 |
|
pMB->cbp = 63; |
1061 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
1062 |
|
|
1063 |
|
} else { /* Rate-Distortion */ |
1064 |
|
|
1065 |
|
int min_rd, intra_rd, i, cbp, c[2] = {0, 0}; |
1066 |
|
VECTOR backup[5], *v; |
1067 |
|
Data->iQuant = iQuant; |
1068 |
|
Data->cbp = c; |
1069 |
|
|
1070 |
|
v = Data->qpel ? Data->currentQMV : Data->currentMV; |
1071 |
|
for (i = 0; i < 5; i++) { |
1072 |
|
Data->iMinSAD[i] = 256*4096; |
1073 |
|
backup[i] = v[i]; |
1074 |
|
} |
1075 |
|
|
1076 |
|
min_rd = findRDinter(Data, pMBs, x, y, pParam, MotionFlags); |
1077 |
|
cbp = *Data->cbp; |
1078 |
|
|
1079 |
|
if (coding_type == S_VOP) { |
1080 |
|
int gmc_rd; |
1081 |
|
*Data->iMinSAD = min_rd += BITS_MULT*1; /* mcsel */ |
1082 |
|
gmc_rd = findRDgmc(Data, vGMC, x, y); |
1083 |
|
if (gmc_rd < min_rd) { |
1084 |
|
mcsel = 1; |
1085 |
|
*Data->iMinSAD = min_rd = gmc_rd; |
1086 |
|
mode = MODE_INTER; |
1087 |
|
cbp = *Data->cbp; |
1088 |
|
} |
1089 |
|
} |
1090 |
|
|
1091 |
|
if (inter4v) { |
1092 |
|
int v4_rd; |
1093 |
|
v4_rd = findRDinter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup); |
1094 |
|
if (v4_rd < min_rd) { |
1095 |
|
Data->iMinSAD[0] = min_rd = v4_rd; |
1096 |
|
mode = MODE_INTER4V; |
1097 |
|
cbp = *Data->cbp; |
1098 |
|
} |
1099 |
|
} |
1100 |
|
|
1101 |
|
intra_rd = findRDintra(Data); |
1102 |
|
if (intra_rd < min_rd) { |
1103 |
|
*Data->iMinSAD = min_rd = intra_rd; |
1104 |
|
mode = MODE_INTRA; |
1105 |
|
} |
1106 |
|
|
1107 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = 0; |
1108 |
|
pMB->cbp = cbp; |
1109 |
|
} |
1110 |
|
|
1111 |
|
if (Data->rrv) { |
1112 |
|
Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); |
1113 |
|
Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); |
1114 |
|
} |
1115 |
|
|
1116 |
|
if (mode == MODE_INTER && mcsel == 0) { |
1117 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
1118 |
|
|
1119 |
|
if(Data->qpel) { |
1120 |
|
pMB->qmvs[0] = pMB->qmvs[1] |
1121 |
|
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
1122 |
|
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; |
1123 |
|
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; |
1124 |
|
} else { |
1125 |
|
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
1126 |
|
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
1127 |
|
} |
1128 |
|
|
1129 |
|
} else if (mode == MODE_INTER ) { // but mcsel == 1 |
1130 |
|
|
1131 |
|
pMB->mcsel = 1; |
1132 |
|
if (Data->qpel) { |
1133 |
|
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = pMB->amv; |
1134 |
|
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = pMB->amv.x/2; |
1135 |
|
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = pMB->amv.y/2; |
1136 |
|
} else |
1137 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->amv; |
1138 |
|
|
1139 |
|
} else |
1140 |
|
if (mode == MODE_INTER4V) ; /* anything here? */ |
1141 |
|
else /* INTRA, NOT_CODED */ |
1142 |
|
ZeroMacroblockP(pMB, 0); |
1143 |
|
|
1144 |
|
pMB->mode = mode; |
1145 |
|
} |
1146 |
|
|
1147 |
bool |
bool |
1148 |
MotionEstimation(MBParam * const pParam, |
MotionEstimation(MBParam * const pParam, |
1149 |
FRAMEINFO * const current, |
FRAMEINFO * const current, |
1151 |
const IMAGE * const pRefH, |
const IMAGE * const pRefH, |
1152 |
const IMAGE * const pRefV, |
const IMAGE * const pRefV, |
1153 |
const IMAGE * const pRefHV, |
const IMAGE * const pRefHV, |
1154 |
|
const IMAGE * const pGMC, |
1155 |
const uint32_t iLimit) |
const uint32_t iLimit) |
1156 |
{ |
{ |
1157 |
MACROBLOCK *const pMBs = current->mbs; |
MACROBLOCK *const pMBs = current->mbs; |
1161 |
uint32_t mb_width = pParam->mb_width; |
uint32_t mb_width = pParam->mb_width; |
1162 |
uint32_t mb_height = pParam->mb_height; |
uint32_t mb_height = pParam->mb_height; |
1163 |
const uint32_t iEdgedWidth = pParam->edged_width; |
const uint32_t iEdgedWidth = pParam->edged_width; |
1164 |
const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->global_flags); |
const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->vop_flags, current->vol_flags); |
1165 |
|
|
1166 |
uint32_t x, y; |
uint32_t x, y; |
1167 |
uint32_t iIntra = 0; |
uint32_t iIntra = 0; |
1168 |
int32_t quant = current->quant, sad00; |
int32_t sad00; |
1169 |
|
int skip_thresh = INITIAL_SKIP_THRESH * \ |
1170 |
|
(current->vop_flags & XVID_VOP_REDUCED ? 4:1) * \ |
1171 |
|
(current->vop_flags & XVID_VOP_MODEDECISION_RD ? 2:1); |
1172 |
|
|
1173 |
// some pre-initialized thingies for SearchP |
/* some pre-initialized thingies for SearchP */ |
1174 |
int32_t temp[8]; |
int32_t temp[8]; uint32_t dir; |
1175 |
VECTOR currentMV[5]; |
VECTOR currentMV[5]; |
1176 |
VECTOR currentQMV[5]; |
VECTOR currentQMV[5]; |
1177 |
int32_t iMinSAD[5]; |
int32_t iMinSAD[5]; |
1178 |
|
DECLARE_ALIGNED_MATRIX(dct_space, 3, 64, int16_t, CACHE_LINE); |
1179 |
SearchData Data; |
SearchData Data; |
1180 |
memset(&Data, 0, sizeof(SearchData)); |
memset(&Data, 0, sizeof(SearchData)); |
1181 |
Data.iEdgedWidth = iEdgedWidth; |
Data.iEdgedWidth = iEdgedWidth; |
1183 |
Data.currentQMV = currentQMV; |
Data.currentQMV = currentQMV; |
1184 |
Data.iMinSAD = iMinSAD; |
Data.iMinSAD = iMinSAD; |
1185 |
Data.temp = temp; |
Data.temp = temp; |
1186 |
|
Data.dir = &dir; |
1187 |
Data.iFcode = current->fcode; |
Data.iFcode = current->fcode; |
1188 |
Data.rounding = pParam->m_rounding_type; |
Data.rounding = pParam->m_rounding_type; |
1189 |
Data.qpel = pParam->m_quarterpel; |
Data.qpel = (current->vol_flags & XVID_VOL_QUARTERPEL ? 1:0); |
1190 |
Data.chroma = MotionFlags & PMV_CHROMA16; |
Data.chroma = MotionFlags & XVID_ME_CHROMA_PVOP; |
1191 |
Data.rrv = current->global_flags & XVID_REDUCED; |
Data.rrv = (current->vop_flags & XVID_VOP_REDUCED) ? 1:0; |
1192 |
|
Data.dctSpace = dct_space; |
1193 |
|
Data.quant_type = !(pParam->vol_flags & XVID_VOL_MPEGQUANT); |
1194 |
|
|
1195 |
if ((current->global_flags & XVID_REDUCED)) { |
if ((current->vop_flags & XVID_VOP_REDUCED)) { |
1196 |
mb_width = (pParam->width + 31) / 32; |
mb_width = (pParam->width + 31) / 32; |
1197 |
mb_height = (pParam->height + 31) / 32; |
mb_height = (pParam->height + 31) / 32; |
1198 |
Data.qpel = 0; |
Data.qpel = 0; |
1199 |
} |
} |
1200 |
|
|
1201 |
Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose) |
Data.RefQ = pRefV->u; /* a good place, also used in MC (for similar purpose) */ |
1202 |
if (sadInit) (*sadInit) (); |
if (sadInit) (*sadInit) (); |
1203 |
|
|
1204 |
for (y = 0; y < mb_height; y++) { |
for (y = 0; y < mb_height; y++) { |
1225 |
|
|
1226 |
sad00 = pMB->sad16; |
sad00 = pMB->sad16; |
1227 |
|
|
1228 |
if (!(current->global_flags & XVID_LUMIMASKING)) { |
/* initial skip decision */ |
|
pMB->dquant = NO_CHANGE; |
|
|
} else { |
|
|
if (pMB->dquant != NO_CHANGE) { |
|
|
quant += DQtab[pMB->dquant]; |
|
|
if (quant > 31) quant = 31; |
|
|
else if (quant < 1) quant = 1; |
|
|
} |
|
|
} |
|
|
pMB->quant = current->quant; |
|
|
|
|
|
//initial skip decision |
|
1229 |
/* no early skip for GMC (global vector = skip vector is unknown!) */ |
/* no early skip for GMC (global vector = skip vector is unknown!) */ |
1230 |
if (!(current->global_flags & XVID_GMC)) { /* no fast SKIP for S(GMC)-VOPs */ |
if (current->coding_type != S_VOP) { /* no fast SKIP for S(GMC)-VOPs */ |
1231 |
if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH * (Data.rrv ? 4:1) ) |
if (pMB->dquant == 0 && sad00 < pMB->quant * skip_thresh) |
1232 |
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) { |
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) { |
1233 |
SkipMacroblockP(pMB, sad00); |
ZeroMacroblockP(pMB, sad00); |
1234 |
|
pMB->mode = MODE_NOT_CODED; |
1235 |
continue; |
continue; |
1236 |
} |
} |
1237 |
} |
} |
1238 |
|
|
1239 |
|
if ((current->vop_flags & XVID_VOP_CARTOON) && |
1240 |
|
(sad00 < pMB->quant * 4 * skip_thresh)) { /* favorize (0,0) vector for cartoons */ |
1241 |
|
ZeroMacroblockP(pMB, sad00); |
1242 |
|
continue; |
1243 |
|
} |
1244 |
|
|
1245 |
SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
1246 |
y, MotionFlags, current->global_flags, pMB->quant, |
y, MotionFlags, current->vop_flags, current->vol_flags, |
1247 |
&Data, pParam, pMBs, reference->mbs, |
&Data, pParam, pMBs, reference->mbs, pMB); |
|
current->global_flags & XVID_INTER4V, pMB); |
|
1248 |
|
|
1249 |
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
ModeDecision(&Data, pMB, pMBs, x, y, pParam, |
1250 |
if (!(current->global_flags & XVID_GMC)) { |
MotionFlags, current->vop_flags, current->vol_flags, |
1251 |
if ( pMB->dquant == NO_CHANGE && sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) { |
pCurrent, pRef, pGMC, current->coding_type); |
|
if (!(current->global_flags & XVID_MODEDECISION_BITS)) { |
|
|
if ( (100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1) ) |
|
|
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) |
|
|
SkipMacroblockP(pMB, sad00); |
|
|
} else { // BITS mode decision |
|
|
if (pMB->sad16 > 10) |
|
|
SkipMacroblockP(pMB, sad00); // more than 10 bits would be used for this MB - skip |
|
1252 |
|
|
|
} |
|
|
} |
|
|
} |
|
1253 |
if (pMB->mode == MODE_INTRA) |
if (pMB->mode == MODE_INTRA) |
1254 |
if (++iIntra > iLimit) return 1; |
if (++iIntra > iLimit) return 1; |
1255 |
} |
} |
1256 |
} |
} |
1257 |
|
|
|
if (current->global_flags & XVID_GMC ) /* GMC only for S(GMC)-VOPs */ |
|
|
{ |
|
|
current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV); |
|
|
} |
|
1258 |
return 0; |
return 0; |
1259 |
} |
} |
1260 |
|
|
1261 |
|
/* check if given vector is equal to any vector checked before */ |
1262 |
|
static __inline int |
1263 |
|
vector_repeats(const VECTOR * const pmv, const int i) |
1264 |
|
{ |
1265 |
|
unsigned int j; |
1266 |
|
for (j = 0; j < i; j++) |
1267 |
|
if (MVequal(pmv[i], pmv[j])) return 1; /* same vector has been checked already */ |
1268 |
|
return 0; |
1269 |
|
} |
1270 |
|
|
1271 |
|
/* make a binary mask that prevents diamonds/squares |
1272 |
|
from checking a vector which has been checked as a prediction */ |
1273 |
static __inline int |
static __inline int |
1274 |
make_mask(const VECTOR * const pmv, const int i) |
make_mask(const VECTOR * const pmv, const int i, const int current) |
1275 |
{ |
{ |
1276 |
int mask = 255, j; |
unsigned int mask = 255, j; |
1277 |
for (j = 0; j < i; j++) { |
for (j = 0; j < i; j++) { |
1278 |
if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already |
if (pmv[current].x == pmv[j].x) { |
1279 |
if (pmv[i].x == pmv[j].x) { |
if (pmv[current].y == pmv[j].y + iDiamondSize) mask &= ~4; |
1280 |
if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4; |
else if (pmv[current].y == pmv[j].y - iDiamondSize) mask &= ~8; |
|
else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8; |
|
1281 |
} else |
} else |
1282 |
if (pmv[i].y == pmv[j].y) { |
if (pmv[current].y == pmv[j].y) { |
1283 |
if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1; |
if (pmv[current].x == pmv[j].x + iDiamondSize) mask &= ~1; |
1284 |
else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2; |
else if (pmv[current].x == pmv[j].x - iDiamondSize) mask &= ~2; |
1285 |
} |
} |
1286 |
} |
} |
1287 |
return mask; |
return mask; |
1291 |
PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, |
PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, |
1292 |
int iHcount, const MACROBLOCK * const prevMB, int rrv) |
int iHcount, const MACROBLOCK * const prevMB, int rrv) |
1293 |
{ |
{ |
1294 |
|
/* this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself */ |
|
//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself |
|
1295 |
if (rrv) { iWcount /= 2; iHcount /= 2; } |
if (rrv) { iWcount /= 2; iHcount /= 2; } |
1296 |
|
|
1297 |
if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour |
if ( (y != 0) && (x < (iWcount-1)) ) { /* [5] top-right neighbour */ |
1298 |
pmv[5].x = EVEN(pmv[3].x); |
pmv[5].x = EVEN(pmv[3].x); |
1299 |
pmv[5].y = EVEN(pmv[3].y); |
pmv[5].y = EVEN(pmv[3].y); |
1300 |
} else pmv[5].x = pmv[5].y = 0; |
} else pmv[5].x = pmv[5].y = 0; |
1301 |
|
|
1302 |
if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour |
if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }/* pmv[3] is left neighbour */ |
1303 |
else pmv[3].x = pmv[3].y = 0; |
else pmv[3].x = pmv[3].y = 0; |
1304 |
|
|
1305 |
if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour |
if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }/* [4] top neighbour */ |
1306 |
else pmv[4].x = pmv[4].y = 0; |
else pmv[4].x = pmv[4].y = 0; |
1307 |
|
|
1308 |
// [1] median prediction |
/* [1] median prediction */ |
1309 |
pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); |
pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); |
1310 |
|
|
1311 |
pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask |
pmv[0].x = pmv[0].y = 0; /* [0] is zero; not used in the loop (checked before) but needed here for make_mask */ |
1312 |
|
|
1313 |
pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
pmv[2].x = EVEN(prevMB->mvs[0].x); /* [2] is last frame */ |
1314 |
pmv[2].y = EVEN(prevMB->mvs[0].y); |
pmv[2].y = EVEN(prevMB->mvs[0].y); |
1315 |
|
|
1316 |
if ((x < iWcount-1) && (y < iHcount-1)) { |
if ((x < iWcount-1) && (y < iHcount-1)) { |
1317 |
pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame |
pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); /* [6] right-down neighbour in last frame */ |
1318 |
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
1319 |
} else pmv[6].x = pmv[6].y = 0; |
} else pmv[6].x = pmv[6].y = 0; |
1320 |
|
|
1327 |
} |
} |
1328 |
} |
} |
1329 |
|
|
|
static int |
|
|
ModeDecision(const uint32_t iQuant, SearchData * const Data, |
|
|
int inter4v, |
|
|
MACROBLOCK * const pMB, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const int x, const int y, |
|
|
const MBParam * const pParam, |
|
|
const uint32_t MotionFlags, |
|
|
const uint32_t GlobalFlags) |
|
|
{ |
|
|
|
|
|
int mode = MODE_INTER; |
|
|
|
|
|
if (!(GlobalFlags & XVID_MODEDECISION_BITS)) { //normal, fast, SAD-based mode decision |
|
|
// int intra = 0; |
|
|
int sad; |
|
|
int InterBias = MV16_INTER_BIAS; |
|
|
if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
|
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) { |
|
|
mode = 0; //inter |
|
|
sad = Data->iMinSAD[0]; |
|
|
} else { |
|
|
mode = MODE_INTER4V; |
|
|
sad = Data->iMinSAD[1] + Data->iMinSAD[2] + |
|
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant; |
|
|
Data->iMinSAD[0] = sad; |
|
|
} |
|
|
|
|
|
/* intra decision */ |
|
|
|
|
|
if (iQuant > 8) InterBias += 100 * (iQuant - 8); // to make high quants work |
|
|
if (y != 0) |
|
|
if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; |
|
|
if (x != 0) |
|
|
if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; |
|
|
|
|
|
if (Data->chroma) InterBias += 50; // to compensate bigger SAD |
|
|
if (Data->rrv) InterBias *= 4; |
|
|
|
|
|
if (InterBias < pMB->sad16) { |
|
|
int32_t deviation; |
|
|
if (!Data->rrv) deviation = dev16(Data->Cur, Data->iEdgedWidth); |
|
|
else deviation = dev16(Data->Cur, Data->iEdgedWidth) + |
|
|
dev16(Data->Cur+8, Data->iEdgedWidth) + |
|
|
dev16(Data->Cur + 8*Data->iEdgedWidth, Data->iEdgedWidth) + |
|
|
dev16(Data->Cur+8+8*Data->iEdgedWidth, Data->iEdgedWidth); |
|
|
|
|
|
if (deviation < (sad - InterBias)) return MODE_INTRA;// intra |
|
|
} |
|
|
return mode; |
|
|
|
|
|
} else { |
|
|
|
|
|
int bits, intra, i; |
|
|
VECTOR backup[5], *v; |
|
|
Data->lambda16 = iQuant; |
|
|
Data->lambda8 = pParam->m_quant_type; |
|
|
|
|
|
v = Data->qpel ? Data->currentQMV : Data->currentMV; |
|
|
for (i = 0; i < 5; i++) { |
|
|
Data->iMinSAD[i] = 256*4096; |
|
|
backup[i] = v[i]; |
|
|
} |
|
|
|
|
|
bits = CountMBBitsInter(Data, pMBs, x, y, pParam, MotionFlags); |
|
|
if (bits == 0) return MODE_INTER; // quick stop |
|
|
|
|
|
if (inter4v) { |
|
|
int inter4v = CountMBBitsInter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup); |
|
|
if (inter4v < bits) { Data->iMinSAD[0] = bits = inter4v; mode = MODE_INTER4V; } |
|
|
} |
|
|
|
|
|
|
|
|
intra = CountMBBitsIntra(Data); |
|
|
|
|
|
if (intra < bits) { *Data->iMinSAD = bits = intra; return MODE_INTRA; } |
|
|
|
|
|
return mode; |
|
|
} |
|
|
} |
|
|
|
|
1330 |
static void |
static void |
1331 |
SearchP(const IMAGE * const pRef, |
SearchP(const IMAGE * const pRef, |
1332 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1336 |
const int x, |
const int x, |
1337 |
const int y, |
const int y, |
1338 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
1339 |
const uint32_t GlobalFlags, |
const uint32_t VopFlags, |
1340 |
const uint32_t iQuant, |
const uint32_t VolFlags, |
1341 |
SearchData * const Data, |
SearchData * const Data, |
1342 |
const MBParam * const pParam, |
const MBParam * const pParam, |
1343 |
const MACROBLOCK * const pMBs, |
const MACROBLOCK * const pMBs, |
1344 |
const MACROBLOCK * const prevMBs, |
const MACROBLOCK * const prevMBs, |
|
int inter4v, |
|
1345 |
MACROBLOCK * const pMB) |
MACROBLOCK * const pMB) |
1346 |
{ |
{ |
1347 |
|
|
1348 |
int i, iDirection = 255, mask, threshA; |
int i, threshA; |
1349 |
VECTOR pmv[7]; |
VECTOR pmv[7]; |
1350 |
|
int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0); |
1351 |
|
CheckFunc * CheckCandidate; |
1352 |
|
|
1353 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
1354 |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, Data->rrv); |
1355 |
|
|
1356 |
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); |
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); |
1357 |
|
|
1358 |
Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache |
Data->temp[5] = Data->temp[6] = 0; /* chroma-sad cache */ |
1359 |
i = Data->rrv ? 2 : 1; |
i = Data->rrv ? 2 : 1; |
1360 |
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i; |
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i; |
1361 |
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1362 |
Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1363 |
|
|
1364 |
Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16*i; |
Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16*i; |
1365 |
Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16*i; |
Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16*i; |
1366 |
Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16*i; |
Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16*i; |
1367 |
Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16*i; |
Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16*i; |
1368 |
Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1369 |
Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1370 |
|
|
1371 |
Data->lambda16 = lambda_vec16[iQuant]; |
Data->lambda16 = lambda_vec16[pMB->quant]; |
1372 |
Data->lambda8 = lambda_vec8[iQuant]; |
Data->lambda8 = lambda_vec8[pMB->quant]; |
1373 |
Data->qpel_precision = 0; |
Data->qpel_precision = 0; |
1374 |
|
*Data->dir = 0; |
1375 |
|
|
1376 |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
memset(Data->currentMV, 0, 5*sizeof(VECTOR)); |
|
|
|
|
for(i = 0; i < 5; i++) |
|
|
Data->currentMV[i].x = Data->currentMV[i].y = 0; |
|
1377 |
|
|
1378 |
if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
1379 |
else Data->predMV = pmv[0]; |
else Data->predMV = pmv[0]; |
1385 |
Data->iMinSAD[3] = pMB->sad8[2]; |
Data->iMinSAD[3] = pMB->sad8[2]; |
1386 |
Data->iMinSAD[4] = pMB->sad8[3]; |
Data->iMinSAD[4] = pMB->sad8[3]; |
1387 |
|
|
1388 |
if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) || (x | y)) { |
if ((!(VopFlags & XVID_VOP_MODEDECISION_RD)) && (x | y)) { |
1389 |
threshA = Data->temp[0]; // that's where we keep this SAD atm |
threshA = Data->temp[0]; /* that's where we keep this SAD atm */ |
1390 |
if (threshA < 512) threshA = 512; |
if (threshA < 512) threshA = 512; |
1391 |
else if (threshA > 1024) threshA = 1024; |
else if (threshA > 1024) threshA = 1024; |
1392 |
} else |
} else |
1396 |
prevMBs + x + y * pParam->mb_width, Data->rrv); |
prevMBs + x + y * pParam->mb_width, Data->rrv); |
1397 |
|
|
1398 |
if (!Data->rrv) { |
if (!Data->rrv) { |
1399 |
if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16; |
if (inter4v) CheckCandidate = CheckCandidate16; |
1400 |
else CheckCandidate = CheckCandidate16no4v; //for extra speed |
else CheckCandidate = CheckCandidate16no4v; /* for extra speed */ |
1401 |
} else CheckCandidate = CheckCandidate32; |
} else CheckCandidate = CheckCandidate32; |
1402 |
|
|
1403 |
/* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/ |
/* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/ |
1404 |
|
|
1405 |
for (i = 1; i < 7; i++) { |
for (i = 1; i < 7; i++) |
1406 |
if (!(mask = make_mask(pmv, i)) ) continue; |
if (!vector_repeats(pmv, i)) { |
1407 |
CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
CheckCandidate(pmv[i].x, pmv[i].y, Data, i); |
1408 |
if (Data->iMinSAD[0] <= threshA) break; |
if (Data->iMinSAD[0] <= threshA) { i++; break; } |
1409 |
} |
} |
1410 |
|
|
1411 |
if ((Data->iMinSAD[0] <= threshA) || |
if ((Data->iMinSAD[0] <= threshA) || |
1412 |
(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
1413 |
(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { |
(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) |
1414 |
if (!(GlobalFlags & XVID_MODEDECISION_BITS)) inter4v = 0; } |
inter4v = 0; |
1415 |
else { |
else { |
1416 |
|
|
1417 |
MainSearchFunc * MainSearchPtr; |
MainSearchFunc * MainSearchPtr; |
1418 |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
int mask = make_mask(pmv, i, *Data->dir); // all vectors pmv[0..i-1] have been checked |
1419 |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
|
1420 |
|
if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch; |
1421 |
|
else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1422 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1423 |
|
|
1424 |
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate); |
1425 |
|
|
1426 |
/* extended search, diamond starting in 0,0 and in prediction. |
/* extended search, diamond starting in 0,0 and in prediction. |
1427 |
note that this search is/might be done in halfpel positions, |
note that this search is/might be done in halfpel positions, |
1428 |
which makes it more different than the diamond above */ |
which makes it more different than the diamond above */ |
1429 |
|
|
1430 |
if (MotionFlags & PMV_EXTSEARCH16) { |
if (MotionFlags & XVID_ME_EXTSEARCH16) { |
1431 |
int32_t bSAD; |
int32_t bSAD; |
1432 |
VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; |
VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; |
1433 |
if (Data->rrv) { |
if (Data->rrv) { |
1437 |
if (!(MVequal(startMV, backupMV))) { |
if (!(MVequal(startMV, backupMV))) { |
1438 |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1439 |
|
|
1440 |
CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data); |
CheckCandidate(startMV.x, startMV.y, Data, 255); |
1441 |
MainSearchPtr(startMV.x, startMV.y, Data, 255); |
MainSearchPtr(startMV.x, startMV.y, Data, 255, CheckCandidate); |
1442 |
if (bSAD < Data->iMinSAD[0]) { |
if (bSAD < Data->iMinSAD[0]) { |
1443 |
Data->currentMV[0] = backupMV; |
Data->currentMV[0] = backupMV; |
1444 |
Data->iMinSAD[0] = bSAD; } |
Data->iMinSAD[0] = bSAD; } |
1449 |
if (!(MVequal(startMV, backupMV))) { |
if (!(MVequal(startMV, backupMV))) { |
1450 |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1451 |
|
|
1452 |
CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data); |
CheckCandidate(startMV.x, startMV.y, Data, 255); |
1453 |
MainSearchPtr(startMV.x, startMV.y, Data, 255); |
MainSearchPtr(startMV.x, startMV.y, Data, 255, CheckCandidate); |
1454 |
if (bSAD < Data->iMinSAD[0]) { |
if (bSAD < Data->iMinSAD[0]) { |
1455 |
Data->currentMV[0] = backupMV; |
Data->currentMV[0] = backupMV; |
1456 |
Data->iMinSAD[0] = bSAD; } |
Data->iMinSAD[0] = bSAD; |
1457 |
|
} |
1458 |
} |
} |
1459 |
} |
} |
1460 |
} |
} |
1461 |
|
|
1462 |
if (MotionFlags & PMV_HALFPELREFINE16) |
if (MotionFlags & XVID_ME_HALFPELREFINE16) |
1463 |
if ((!(MotionFlags & HALFPELREFINE16_BITS)) || Data->iMinSAD[0] < 200*(int)iQuant) |
SubpelRefine(Data, CheckCandidate); |
|
SubpelRefine(Data); |
|
1464 |
|
|
1465 |
for(i = 0; i < 5; i++) { |
for(i = 0; i < 5; i++) { |
1466 |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* initialize qpel vectors */ |
1467 |
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
1468 |
} |
} |
1469 |
|
|
1470 |
if (MotionFlags & PMV_QUARTERPELREFINE16) |
if (Data->qpel) { |
1471 |
if ((!(MotionFlags & QUARTERPELREFINE16_BITS)) || (Data->iMinSAD[0] < 200*(int)iQuant)) { |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
1472 |
|
pParam->width, pParam->height, Data->iFcode, 2, 0); |
1473 |
Data->qpel_precision = 1; |
Data->qpel_precision = 1; |
1474 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
if (MotionFlags & XVID_ME_QUARTERPELREFINE16) |
1475 |
pParam->width, pParam->height, Data->iFcode, 1, 0); |
SubpelRefine(Data, CheckCandidate); |
|
|
|
|
SubpelRefine(Data); |
|
1476 |
} |
} |
1477 |
|
|
1478 |
if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) && (Data->iMinSAD[0] < (int32_t)iQuant * 30)) inter4v = 0; |
if (Data->iMinSAD[0] < (int32_t)pMB->quant * 30) |
1479 |
|
inter4v = 0; |
|
if (inter4v && (!(GlobalFlags & XVID_MODEDECISION_BITS) || |
|
|
(!(MotionFlags & QUARTERPELREFINE8_BITS)) || (!(MotionFlags & HALFPELREFINE8_BITS)) || |
|
|
((!(MotionFlags & EXTSEARCH_BITS)) && (!(MotionFlags&PMV_EXTSEARCH8)) ))) { |
|
|
// if decision is BITS-based and all refinement steps will be done in BITS domain, there is no reason to call this loop |
|
1480 |
|
|
1481 |
|
if (inter4v) { |
1482 |
SearchData Data8; |
SearchData Data8; |
1483 |
memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data |
memcpy(&Data8, Data, sizeof(SearchData)); /* quick copy of common data */ |
1484 |
|
|
1485 |
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
1486 |
Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); |
Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); |
1487 |
Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); |
Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); |
1488 |
Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); |
Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); |
1489 |
|
|
1490 |
if ((Data->chroma) && (!(GlobalFlags & XVID_MODEDECISION_BITS))) { |
if ((Data->chroma) && (!(VopFlags & XVID_VOP_MODEDECISION_RD))) { |
1491 |
// chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, there is no reason to compute it |
/* chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, it will not be used */ |
1492 |
int sumx = 0, sumy = 0; |
int sumx = 0, sumy = 0; |
|
const int div = 1 + Data->qpel; |
|
|
const VECTOR * const mv = Data->qpel ? pMB->qmvs : pMB->mvs; |
|
1493 |
|
|
1494 |
for (i = 0; i < 4; i++) { |
if (Data->qpel) |
1495 |
sumx += mv[i].x / div; |
for (i = 1; i < 5; i++) { |
1496 |
sumy += mv[i].y / div; |
sumx += Data->currentQMV[i].x/2; |
1497 |
|
sumy += Data->currentQMV[i].y/2; |
1498 |
|
} |
1499 |
|
else |
1500 |
|
for (i = 1; i < 5; i++) { |
1501 |
|
sumx += Data->currentMV[i].x; |
1502 |
|
sumy += Data->currentMV[i].y; |
1503 |
} |
} |
1504 |
|
|
1505 |
Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf], |
Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf], |
1506 |
(sumy >> 3) + roundtab_76[sumy & 0xf], Data); |
(sumy >> 3) + roundtab_76[sumy & 0xf], Data); |
1507 |
} |
} |
1508 |
} |
} else Data->iMinSAD[1] = 4096*256; |
|
|
|
|
inter4v = ModeDecision(iQuant, Data, inter4v, pMB, pMBs, x, y, pParam, MotionFlags, GlobalFlags); |
|
|
|
|
|
if (Data->rrv) { |
|
|
Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); |
|
|
Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); |
|
|
} |
|
|
|
|
|
if (inter4v == MODE_INTER) { |
|
|
pMB->mode = MODE_INTER; |
|
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
|
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
|
|
|
|
|
if(Data->qpel) { |
|
|
pMB->qmvs[0] = pMB->qmvs[1] |
|
|
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
|
|
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; |
|
|
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; |
|
|
} else { |
|
|
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
|
|
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
|
|
} |
|
|
|
|
|
} else if (inter4v == MODE_INTER4V) { |
|
|
pMB->mode = MODE_INTER4V; |
|
|
pMB->sad16 = Data->iMinSAD[0]; |
|
|
} else { // INTRA mode |
|
|
SkipMacroblockP(pMB, 0); // not skip, but similar enough |
|
|
pMB->mode = MODE_INTRA; |
|
|
} |
|
|
|
|
1509 |
} |
} |
1510 |
|
|
1511 |
static void |
static void |
1519 |
SearchData * const Data) |
SearchData * const Data) |
1520 |
{ |
{ |
1521 |
int i = 0; |
int i = 0; |
1522 |
|
CheckFunc * CheckCandidate; |
1523 |
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
1524 |
Data->currentMV = OldData->currentMV + 1 + block; |
Data->currentMV = OldData->currentMV + 1 + block; |
1525 |
Data->currentQMV = OldData->currentQMV + 1 + block; |
Data->currentQMV = OldData->currentQMV + 1 + block; |
1536 |
|
|
1537 |
*(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10; |
*(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10; |
1538 |
|
|
1539 |
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8|PMV_QUARTERPELREFINE8)) { |
if (MotionFlags & (XVID_ME_EXTSEARCH8|XVID_ME_HALFPELREFINE8|XVID_ME_QUARTERPELREFINE8)) { |
|
if (Data->rrv) i = 2; else i = 1; |
|
1540 |
|
|
1541 |
Data->Ref = OldData->Ref + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
if (Data->rrv) i = 16; else i = 8; |
|
Data->RefH = OldData->RefH + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
|
|
Data->RefV = OldData->RefV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
|
|
Data->RefHV = OldData->RefHV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
|
1542 |
|
|
1543 |
Data->Cur = OldData->Cur + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
Data->RefP[0] = OldData->RefP[0] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1544 |
|
Data->RefP[1] = OldData->RefP[1] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1545 |
|
Data->RefP[2] = OldData->RefP[2] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1546 |
|
Data->RefP[3] = OldData->RefP[3] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1547 |
|
|
1548 |
|
Data->Cur = OldData->Cur + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1549 |
Data->qpel_precision = 0; |
Data->qpel_precision = 0; |
1550 |
|
|
1551 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 3, |
1552 |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, Data->rrv); |
1553 |
|
|
1554 |
if (!Data->rrv) CheckCandidate = CheckCandidate8; |
if (!Data->rrv) CheckCandidate = CheckCandidate8; |
1555 |
else CheckCandidate = CheckCandidate16no4v; |
else CheckCandidate = CheckCandidate16no4v; |
1556 |
|
|
1557 |
if (MotionFlags & PMV_EXTSEARCH8 && (!(MotionFlags & EXTSEARCH_BITS))) { |
if (MotionFlags & XVID_ME_EXTSEARCH8 && (!(MotionFlags & XVID_ME_EXTSEARCH_RD))) { |
1558 |
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */ |
1559 |
|
|
1560 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1561 |
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
if (MotionFlags & XVID_ME_USESQUARES8) MainSearchPtr = SquareSearch; |
1562 |
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
1563 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1564 |
|
|
1565 |
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255); |
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255, CheckCandidate); |
1566 |
|
|
1567 |
if(*(Data->iMinSAD) < temp_sad) { |
if(*(Data->iMinSAD) < temp_sad) { |
1568 |
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */ |
1569 |
Data->currentQMV->y = 2 * Data->currentMV->y; |
Data->currentQMV->y = 2 * Data->currentMV->y; |
1570 |
} |
} |
1571 |
} |
} |
1572 |
|
|
1573 |
if (MotionFlags & PMV_HALFPELREFINE8) { |
if (MotionFlags & XVID_ME_HALFPELREFINE8) { |
1574 |
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */ |
1575 |
|
|
1576 |
SubpelRefine(Data); // perform halfpel refine of current best vector |
SubpelRefine(Data, CheckCandidate); /* perform halfpel refine of current best vector */ |
1577 |
|
|
1578 |
if(*(Data->iMinSAD) < temp_sad) { // we have found a better match |
if(*(Data->iMinSAD) < temp_sad) { /* we have found a better match */ |
1579 |
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */ |
1580 |
Data->currentQMV->y = 2 * Data->currentMV->y; |
Data->currentQMV->y = 2 * Data->currentMV->y; |
1581 |
} |
} |
1582 |
} |
} |
1583 |
|
|
1584 |
if (Data->qpel && MotionFlags & PMV_QUARTERPELREFINE8) { |
if (Data->qpel && MotionFlags & XVID_ME_QUARTERPELREFINE8) { |
1585 |
Data->qpel_precision = 1; |
Data->qpel_precision = 1; |
1586 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 3, |
1587 |
pParam->width, pParam->height, Data->iFcode, 1, 0); |
pParam->width, pParam->height, Data->iFcode, 2, 0); |
1588 |
SubpelRefine(Data); |
SubpelRefine(Data, CheckCandidate); |
1589 |
} |
} |
1590 |
} |
} |
1591 |
|
|
1623 |
const uint32_t mode_curr) |
const uint32_t mode_curr) |
1624 |
{ |
{ |
1625 |
|
|
1626 |
// [0] is prediction |
/* [0] is prediction */ |
1627 |
pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); |
pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); |
1628 |
|
|
1629 |
pmv[1].x = pmv[1].y = 0; // [1] is zero |
pmv[1].x = pmv[1].y = 0; /* [1] is zero */ |
1630 |
|
|
1631 |
pmv[2] = ChoosePred(pMB, mode_curr); |
pmv[2] = ChoosePred(pMB, mode_curr); |
1632 |
pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); |
pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); |
1633 |
|
|
1634 |
if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour |
if ((y != 0)&&(x != (int)(iWcount+1))) { /* [3] top-right neighbour */ |
1635 |
pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); |
pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); |
1636 |
pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); |
pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); |
1637 |
} else pmv[3].x = pmv[3].y = 0; |
} else pmv[3].x = pmv[3].y = 0; |
1671 |
SearchData * const Data) |
SearchData * const Data) |
1672 |
{ |
{ |
1673 |
|
|
1674 |
int i, iDirection = 255, mask; |
int i; |
1675 |
VECTOR pmv[7]; |
VECTOR pmv[7]; |
1676 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1677 |
*Data->iMinSAD = MV_MAX_ERROR; |
*Data->iMinSAD = MV_MAX_ERROR; |
1678 |
Data->iFcode = iFcode; |
Data->iFcode = iFcode; |
1679 |
Data->qpel_precision = 0; |
Data->qpel_precision = 0; |
1680 |
Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; // reset chroma-sad cache |
Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; /* reset chroma-sad cache */ |
1681 |
|
|
1682 |
Data->Ref = pRef->y + (x + y * Data->iEdgedWidth) * 16; |
Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16; |
1683 |
Data->RefH = pRefH + (x + y * Data->iEdgedWidth) * 16; |
Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16; |
1684 |
Data->RefV = pRefV + (x + y * Data->iEdgedWidth) * 16; |
Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16; |
1685 |
Data->RefHV = pRefHV + (x + y * Data->iEdgedWidth) * 16; |
Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16; |
1686 |
Data->RefCU = pRef->u + (x + y * Data->iEdgedWidth/2) * 8; |
Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
1687 |
Data->RefCV = pRef->v + (x + y * Data->iEdgedWidth/2) * 8; |
Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
1688 |
|
|
1689 |
Data->predMV = *predMV; |
Data->predMV = *predMV; |
1690 |
|
|
1691 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
1692 |
pParam->width, pParam->height, iFcode - Data->qpel, 0, 0); |
pParam->width, pParam->height, iFcode - Data->qpel, 1, 0); |
1693 |
|
|
1694 |
pmv[0] = Data->predMV; |
pmv[0] = Data->predMV; |
1695 |
if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } |
if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } |
1697 |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
1698 |
|
|
1699 |
Data->currentMV->x = Data->currentMV->y = 0; |
Data->currentMV->x = Data->currentMV->y = 0; |
|
CheckCandidate = CheckCandidate16no4v; |
|
1700 |
|
|
1701 |
// main loop. checking all predictions |
/* main loop. checking all predictions */ |
1702 |
for (i = 0; i < 7; i++) { |
for (i = 0; i < 7; i++) |
1703 |
if (!(mask = make_mask(pmv, i)) ) continue; |
if (!vector_repeats(pmv, i) ) |
1704 |
CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
CheckCandidate16no4v(pmv[i].x, pmv[i].y, Data, i); |
|
} |
|
1705 |
|
|
1706 |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch; |
1707 |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1708 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1709 |
|
|
1710 |
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
if (*Data->iMinSAD > 512) { |
1711 |
|
unsigned int mask = make_mask(pmv, 7, *Data->dir); |
1712 |
|
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate16no4v); |
1713 |
|
} |
1714 |
|
|
1715 |
SubpelRefine(Data); |
SubpelRefine(Data, CheckCandidate16no4v); |
1716 |
|
|
1717 |
if (Data->qpel && *Data->iMinSAD < *best_sad + 300) { |
if (Data->qpel && *Data->iMinSAD < *best_sad + 300) { |
1718 |
Data->currentQMV->x = 2*Data->currentMV->x; |
Data->currentQMV->x = 2*Data->currentMV->x; |
1719 |
Data->currentQMV->y = 2*Data->currentMV->y; |
Data->currentQMV->y = 2*Data->currentMV->y; |
1720 |
Data->qpel_precision = 1; |
Data->qpel_precision = 1; |
1721 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
1722 |
pParam->width, pParam->height, iFcode, 1, 0); |
pParam->width, pParam->height, iFcode, 2, 0); |
1723 |
SubpelRefine(Data); |
SubpelRefine(Data, CheckCandidate16no4v); |
1724 |
} |
} |
1725 |
|
|
1726 |
// three bits are needed to code backward mode. four for forward |
/* three bits are needed to code backward mode. four for forward */ |
1727 |
|
|
1728 |
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
1729 |
else *Data->iMinSAD += 3 * Data->lambda16; |
else *Data->iMinSAD += 3 * Data->lambda16; |
1747 |
} |
} |
1748 |
|
|
1749 |
if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; |
if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; |
1750 |
else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search |
else *(Data->currentMV+1) = *Data->currentMV; /* we store currmv for interpolate search */ |
1751 |
} |
} |
1752 |
|
|
1753 |
static void |
static void |
1760 |
{ |
{ |
1761 |
int dx = 0, dy = 0, b_dx = 0, b_dy = 0; |
int dx = 0, dy = 0, b_dx = 0, b_dy = 0; |
1762 |
int32_t sum; |
int32_t sum; |
|
const int div = 1 + Data->qpel; |
|
1763 |
int k; |
int k; |
1764 |
const uint32_t stride = Data->iEdgedWidth/2; |
const uint32_t stride = Data->iEdgedWidth/2; |
1765 |
//this is not full chroma compensation, only it's fullpel approximation. should work though |
/* this is not full chroma compensation, only it's fullpel approximation. should work though */ |
1766 |
|
|
1767 |
for (k = 0; k < 4; k++) { |
for (k = 0; k < 4; k++) { |
1768 |
dy += Data->directmvF[k].y / div; |
dy += Data->directmvF[k].y >> Data->qpel; |
1769 |
dx += Data->directmvF[0].x / div; |
dx += Data->directmvF[k].x >> Data->qpel; |
1770 |
b_dy += Data->directmvB[0].y / div; |
b_dy += Data->directmvB[k].y >> Data->qpel; |
1771 |
b_dx += Data->directmvB[0].x / div; |
b_dx += Data->directmvB[k].x >> Data->qpel; |
1772 |
} |
} |
1773 |
|
|
1774 |
dy = (dy >> 3) + roundtab_76[dy & 0xf]; |
dy = (dy >> 3) + roundtab_76[dy & 0xf]; |
1781 |
b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, |
b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, |
1782 |
stride); |
stride); |
1783 |
|
|
1784 |
if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; //no skip |
if (sum >= MAX_CHROMA_SAD_FOR_SKIP * Data->iQuant) return; /* no skip */ |
1785 |
|
|
1786 |
sum += sad8bi(pCur->v + 8*x + 8 * y * stride, |
sum += sad8bi(pCur->v + 8*x + 8 * y * stride, |
1787 |
f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2, |
f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2, |
1788 |
b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, |
b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, |
1789 |
stride); |
stride); |
1790 |
|
|
1791 |
if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) pMB->mode = MODE_DIRECT_NONE_MV; //skipped |
if (sum < MAX_CHROMA_SAD_FOR_SKIP * Data->iQuant) { |
1792 |
|
pMB->mode = MODE_DIRECT_NONE_MV; /* skipped */ |
1793 |
|
for (k = 0; k < 4; k++) { |
1794 |
|
pMB->qmvs[k] = pMB->mvs[k]; |
1795 |
|
pMB->b_qmvs[k] = pMB->b_mvs[k]; |
1796 |
|
} |
1797 |
|
} |
1798 |
} |
} |
1799 |
|
|
1800 |
static __inline uint32_t |
static __inline uint32_t |
1820 |
int32_t skip_sad; |
int32_t skip_sad; |
1821 |
int k = (x + Data->iEdgedWidth*y) * 16; |
int k = (x + Data->iEdgedWidth*y) * 16; |
1822 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1823 |
|
CheckFunc * CheckCandidate; |
1824 |
|
|
1825 |
*Data->iMinSAD = 256*4096; |
*Data->iMinSAD = 256*4096; |
1826 |
Data->Ref = f_Ref->y + k; |
Data->RefP[0] = f_Ref->y + k; |
1827 |
Data->RefH = f_RefH + k; |
Data->RefP[2] = f_RefH + k; |
1828 |
Data->RefV = f_RefV + k; |
Data->RefP[1] = f_RefV + k; |
1829 |
Data->RefHV = f_RefHV + k; |
Data->RefP[3] = f_RefHV + k; |
1830 |
Data->bRef = b_Ref->y + k; |
Data->b_RefP[0] = b_Ref->y + k; |
1831 |
Data->bRefH = b_RefH + k; |
Data->b_RefP[2] = b_RefH + k; |
1832 |
Data->bRefV = b_RefV + k; |
Data->b_RefP[1] = b_RefV + k; |
1833 |
Data->bRefHV = b_RefHV + k; |
Data->b_RefP[3] = b_RefHV + k; |
1834 |
Data->RefCU = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; |
Data->RefP[4] = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; |
1835 |
Data->RefCV = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; |
Data->RefP[5] = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; |
1836 |
Data->b_RefCU = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; |
Data->b_RefP[4] = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; |
1837 |
Data->b_RefCV = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; |
Data->b_RefP[5] = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; |
1838 |
|
|
1839 |
k = Data->qpel ? 4 : 2; |
k = Data->qpel ? 4 : 2; |
1840 |
Data->max_dx = k * (pParam->width - x * 16); |
Data->max_dx = k * (pParam->width - x * 16); |
1854 |
if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx) |
if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx) |
1855 |
| (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) { |
| (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) { |
1856 |
|
|
1857 |
*best_sad = 256*4096; // in that case, we won't use direct mode |
*best_sad = 256*4096; /* in that case, we won't use direct mode */ |
1858 |
pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV" |
pMB->mode = MODE_DIRECT; /* just to make sure it doesn't say "MODE_DIRECT_NONE_MV" */ |
1859 |
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
1860 |
return 256*4096; |
return 256*4096; |
1861 |
} |
} |
1870 |
|
|
1871 |
CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v; |
CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v; |
1872 |
|
|
1873 |
CheckCandidate(0, 0, 255, &k, Data); |
CheckCandidate(0, 0, Data, 255); |
1874 |
|
|
1875 |
// initial (fast) skip decision |
/* initial (fast) skip decision */ |
1876 |
if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (2 + Data->chroma?1:0)) { |
if (*Data->iMinSAD < Data->iQuant * INITIAL_SKIP_THRESH * (Data->chroma?3:2)) { |
1877 |
//possible skip |
/* possible skip */ |
1878 |
if (Data->chroma) { |
if (Data->chroma) { |
1879 |
pMB->mode = MODE_DIRECT_NONE_MV; |
pMB->mode = MODE_DIRECT_NONE_MV; |
1880 |
return *Data->iMinSAD; // skip. |
return *Data->iMinSAD; /* skip. */ |
1881 |
} else { |
} else { |
1882 |
SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data); |
SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data); |
1883 |
if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip. |
if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; /* skip. */ |
1884 |
} |
} |
1885 |
} |
} |
1886 |
|
|
1887 |
|
*Data->iMinSAD += Data->lambda16; |
1888 |
skip_sad = *Data->iMinSAD; |
skip_sad = *Data->iMinSAD; |
1889 |
|
|
1890 |
// DIRECT MODE DELTA VECTOR SEARCH. |
/* |
1891 |
// This has to be made more effective, but at the moment I'm happy it's running at all |
* DIRECT MODE DELTA VECTOR SEARCH. |
1892 |
|
* This has to be made more effective, but at the moment I'm happy it's running at all |
1893 |
|
*/ |
1894 |
|
|
1895 |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch; |
1896 |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1897 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1898 |
|
|
1899 |
MainSearchPtr(0, 0, Data, 255); |
MainSearchPtr(0, 0, Data, 255, CheckCandidate); |
1900 |
|
|
1901 |
SubpelRefine(Data); |
SubpelRefine(Data, CheckCandidate); |
1902 |
|
|
1903 |
*best_sad = *Data->iMinSAD; |
*best_sad = *Data->iMinSAD; |
1904 |
|
|
1905 |
if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT; |
if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT; |
1906 |
else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
else pMB->mode = MODE_DIRECT_NO4V; /* for faster compensation */ |
1907 |
|
|
1908 |
pMB->pmvs[3] = *Data->currentMV; |
pMB->pmvs[3] = *Data->currentMV; |
1909 |
|
|
1957 |
|
|
1958 |
{ |
{ |
1959 |
|
|
1960 |
int iDirection, i, j; |
int i, j; |
1961 |
SearchData bData; |
SearchData bData; |
1962 |
|
|
1963 |
fData->qpel_precision = 0; |
fData->qpel_precision = 0; |
1964 |
memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data |
memcpy(&bData, fData, sizeof(SearchData)); /* quick copy of common data */ |
1965 |
*fData->iMinSAD = 4096*256; |
*fData->iMinSAD = 4096*256; |
1966 |
bData.currentMV++; bData.currentQMV++; |
bData.currentMV++; bData.currentQMV++; |
1967 |
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
1968 |
|
|
1969 |
i = (x + y * fData->iEdgedWidth) * 16; |
i = (x + y * fData->iEdgedWidth) * 16; |
|
bData.bRef = fData->Ref = f_Ref->y + i; |
|
|
bData.bRefH = fData->RefH = f_RefH + i; |
|
|
bData.bRefV = fData->RefV = f_RefV + i; |
|
|
bData.bRefHV = fData->RefHV = f_RefHV + i; |
|
|
bData.Ref = fData->bRef = b_Ref->y + i; |
|
|
bData.RefH = fData->bRefH = b_RefH + i; |
|
|
bData.RefV = fData->bRefV = b_RefV + i; |
|
|
bData.RefHV = fData->bRefHV = b_RefHV + i; |
|
|
bData.b_RefCU = fData->RefCU = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; |
|
|
bData.b_RefCV = fData->RefCV = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; |
|
|
bData.RefCU = fData->b_RefCU = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; |
|
|
bData.RefCV = fData->b_RefCV = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; |
|
1970 |
|
|
1971 |
|
bData.b_RefP[0] = fData->RefP[0] = f_Ref->y + i; |
1972 |
|
bData.b_RefP[2] = fData->RefP[2] = f_RefH + i; |
1973 |
|
bData.b_RefP[1] = fData->RefP[1] = f_RefV + i; |
1974 |
|
bData.b_RefP[3] = fData->RefP[3] = f_RefHV + i; |
1975 |
|
bData.RefP[0] = fData->b_RefP[0] = b_Ref->y + i; |
1976 |
|
bData.RefP[2] = fData->b_RefP[2] = b_RefH + i; |
1977 |
|
bData.RefP[1] = fData->b_RefP[1] = b_RefV + i; |
1978 |
|
bData.RefP[3] = fData->b_RefP[3] = b_RefHV + i; |
1979 |
|
bData.b_RefP[4] = fData->RefP[4] = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; |
1980 |
|
bData.b_RefP[5] = fData->RefP[5] = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; |
1981 |
|
bData.RefP[4] = fData->b_RefP[4] = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; |
1982 |
|
bData.RefP[5] = fData->b_RefP[5] = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; |
1983 |
|
bData.dir = fData->dir; |
1984 |
|
|
1985 |
bData.bpredMV = fData->predMV = *f_predMV; |
bData.bpredMV = fData->predMV = *f_predMV; |
1986 |
fData->bpredMV = bData.predMV = *b_predMV; |
fData->bpredMV = bData.predMV = *b_predMV; |
1987 |
fData->currentMV[0] = fData->currentMV[2]; |
fData->currentMV[0] = fData->currentMV[2]; |
1988 |
|
|
1989 |
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); |
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 4, pParam->width, pParam->height, fcode - fData->qpel, 1, 0); |
1990 |
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode - fData->qpel, 0, 0); |
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 4, pParam->width, pParam->height, bcode - fData->qpel, 1, 0); |
1991 |
|
|
1992 |
if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; |
if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; |
1993 |
if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx; |
if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx; |
1999 |
if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy; |
if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy; |
2000 |
if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy; |
if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy; |
2001 |
|
|
2002 |
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, fData, 255); |
2003 |
|
|
2004 |
//diamond |
/* diamond */ |
2005 |
do { |
do { |
2006 |
iDirection = 255; |
*fData->dir = 255; |
2007 |
// forward MV moves |
/* forward MV moves */ |
2008 |
i = fData->currentMV[0].x; j = fData->currentMV[0].y; |
i = fData->currentMV[0].x; j = fData->currentMV[0].y; |
2009 |
|
|
2010 |
CheckCandidateInt(i + 1, j, 0, &iDirection, fData); |
CheckCandidateInt(i + 1, j, fData, 0); |
2011 |
CheckCandidateInt(i, j + 1, 0, &iDirection, fData); |
CheckCandidateInt(i, j + 1, fData, 0); |
2012 |
CheckCandidateInt(i - 1, j, 0, &iDirection, fData); |
CheckCandidateInt(i - 1, j, fData, 0); |
2013 |
CheckCandidateInt(i, j - 1, 0, &iDirection, fData); |
CheckCandidateInt(i, j - 1, fData, 0); |
2014 |
|
|
2015 |
// backward MV moves |
/* backward MV moves */ |
2016 |
i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
2017 |
fData->currentMV[2] = fData->currentMV[0]; |
fData->currentMV[2] = fData->currentMV[0]; |
2018 |
CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); |
CheckCandidateInt(i + 1, j, &bData, 0); |
2019 |
CheckCandidateInt(i, j + 1, 0, &iDirection, &bData); |
CheckCandidateInt(i, j + 1, &bData, 0); |
2020 |
CheckCandidateInt(i - 1, j, 0, &iDirection, &bData); |
CheckCandidateInt(i - 1, j, &bData, 0); |
2021 |
CheckCandidateInt(i, j - 1, 0, &iDirection, &bData); |
CheckCandidateInt(i, j - 1, &bData, 0); |
2022 |
|
|
2023 |
} while (!(iDirection)); |
} while (!(*fData->dir)); |
2024 |
|
|
2025 |
//qpel refinement |
/* qpel refinement */ |
2026 |
if (fData->qpel) { |
if (fData->qpel) { |
2027 |
if (*fData->iMinSAD > *best_sad + 500) return; |
if (*fData->iMinSAD > *best_sad + 500) return; |
|
CheckCandidate = CheckCandidateInt; |
|
2028 |
fData->qpel_precision = bData.qpel_precision = 1; |
fData->qpel_precision = bData.qpel_precision = 1; |
2029 |
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0); |
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 4, pParam->width, pParam->height, fcode, 2, 0); |
2030 |
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0); |
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 4, pParam->width, pParam->height, bcode, 2, 0); |
2031 |
fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x; |
fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x; |
2032 |
fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y; |
fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y; |
2033 |
fData->currentQMV[1].x = 2 * fData->currentMV[1].x; |
fData->currentQMV[1].x = 2 * fData->currentMV[1].x; |
2034 |
fData->currentQMV[1].y = 2 * fData->currentMV[1].y; |
fData->currentQMV[1].y = 2 * fData->currentMV[1].y; |
2035 |
SubpelRefine(fData); |
SubpelRefine(fData, CheckCandidateInt); |
2036 |
if (*fData->iMinSAD > *best_sad + 300) return; |
if (*fData->iMinSAD > *best_sad + 300) return; |
2037 |
fData->currentQMV[2] = fData->currentQMV[0]; |
fData->currentQMV[2] = fData->currentQMV[0]; |
2038 |
SubpelRefine(&bData); |
SubpelRefine(&bData, CheckCandidateInt); |
2039 |
} |
} |
2040 |
|
|
2041 |
*fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode. |
*fData->iMinSAD += (2+3) * fData->lambda16; /* two bits are needed to code interpolate mode. */ |
2042 |
|
|
2043 |
if (*fData->iMinSAD < *best_sad) { |
if (*fData->iMinSAD < *best_sad) { |
2044 |
*best_sad = *fData->iMinSAD; |
*best_sad = *fData->iMinSAD; |
2066 |
FRAMEINFO * const frame, |
FRAMEINFO * const frame, |
2067 |
const int32_t time_bp, |
const int32_t time_bp, |
2068 |
const int32_t time_pp, |
const int32_t time_pp, |
2069 |
// forward (past) reference |
/* forward (past) reference */ |
2070 |
const MACROBLOCK * const f_mbs, |
const MACROBLOCK * const f_mbs, |
2071 |
const IMAGE * const f_ref, |
const IMAGE * const f_ref, |
2072 |
const IMAGE * const f_refH, |
const IMAGE * const f_refH, |
2073 |
const IMAGE * const f_refV, |
const IMAGE * const f_refV, |
2074 |
const IMAGE * const f_refHV, |
const IMAGE * const f_refHV, |
2075 |
// backward (future) reference |
/* backward (future) reference */ |
2076 |
const FRAMEINFO * const b_reference, |
const FRAMEINFO * const b_reference, |
2077 |
const IMAGE * const b_ref, |
const IMAGE * const b_ref, |
2078 |
const IMAGE * const b_refH, |
const IMAGE * const b_refH, |
2090 |
const int32_t TRB = time_pp - time_bp; |
const int32_t TRB = time_pp - time_bp; |
2091 |
const int32_t TRD = time_pp; |
const int32_t TRD = time_pp; |
2092 |
|
|
2093 |
// some pre-inintialized data for the rest of the search |
/* some pre-inintialized data for the rest of the search */ |
2094 |
|
|
2095 |
SearchData Data; |
SearchData Data; |
2096 |
int32_t iMinSAD; |
int32_t iMinSAD; |
2097 |
|
uint32_t dir; |
2098 |
VECTOR currentMV[3]; |
VECTOR currentMV[3]; |
2099 |
VECTOR currentQMV[3]; |
VECTOR currentQMV[3]; |
2100 |
int32_t temp[8]; |
int32_t temp[8]; |
2102 |
Data.iEdgedWidth = pParam->edged_width; |
Data.iEdgedWidth = pParam->edged_width; |
2103 |
Data.currentMV = currentMV; Data.currentQMV = currentQMV; |
Data.currentMV = currentMV; Data.currentQMV = currentQMV; |
2104 |
Data.iMinSAD = &iMinSAD; |
Data.iMinSAD = &iMinSAD; |
2105 |
Data.lambda16 = lambda_vec16[frame->quant]; |
Data.lambda16 = lambda_vec16[MAX(frame->quant-2, 2)]; |
2106 |
Data.qpel = pParam->m_quarterpel; |
Data.qpel = pParam->vol_flags & XVID_VOL_QUARTERPEL ? 1 : 0; |
2107 |
Data.rounding = 0; |
Data.rounding = 0; |
2108 |
Data.chroma = frame->motion_flags & PMV_CHROMA8; |
Data.chroma = frame->motion_flags & XVID_ME_CHROMA_BVOP; |
2109 |
Data.temp = temp; |
Data.temp = temp; |
2110 |
|
Data.dir = &dir; |
2111 |
|
Data.iQuant = frame->quant; |
2112 |
|
|
2113 |
|
Data.RefQ = f_refV->u; /* a good place, also used in MC (for similar purpose) */ |
2114 |
|
|
2115 |
Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose) |
/* note: i==horizontal, j==vertical */ |
|
// note: i==horizontal, j==vertical |
|
2116 |
for (j = 0; j < pParam->mb_height; j++) { |
for (j = 0; j < pParam->mb_height; j++) { |
2117 |
|
|
2118 |
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
2131 |
Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
2132 |
Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8; |
Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8; |
2133 |
Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8; |
Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8; |
|
pMB->quant = frame->quant; |
|
2134 |
|
|
2135 |
/* direct search comes first, because it (1) checks for SKIP-mode |
/* direct search comes first, because it (1) checks for SKIP-mode |
2136 |
and (2) sets very good predictions for forward and backward search */ |
and (2) sets very good predictions for forward and backward search */ |
2147 |
|
|
2148 |
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
2149 |
|
|
2150 |
// forward search |
/* forward search */ |
2151 |
SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
2152 |
&frame->image, i, j, |
&frame->image, i, j, |
2153 |
frame->motion_flags, |
frame->motion_flags, |
2155 |
pMB, &f_predMV, &best_sad, |
pMB, &f_predMV, &best_sad, |
2156 |
MODE_FORWARD, &Data); |
MODE_FORWARD, &Data); |
2157 |
|
|
2158 |
// backward search |
/* backward search */ |
2159 |
SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y, |
SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y, |
2160 |
&frame->image, i, j, |
&frame->image, i, j, |
2161 |
frame->motion_flags, |
frame->motion_flags, |
2163 |
pMB, &b_predMV, &best_sad, |
pMB, &b_predMV, &best_sad, |
2164 |
MODE_BACKWARD, &Data); |
MODE_BACKWARD, &Data); |
2165 |
|
|
2166 |
// interpolate search comes last, because it uses data from forward and backward as prediction |
/* interpolate search comes last, because it uses data from forward and backward as prediction */ |
2167 |
SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
2168 |
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
2169 |
&frame->image, |
&frame->image, |
2175 |
pMB, &best_sad, |
pMB, &best_sad, |
2176 |
&Data); |
&Data); |
2177 |
|
|
2178 |
// final skip decision |
/* final skip decision */ |
2179 |
if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2) |
if ( (skip_sad < Data.iQuant * MAX_SAD00_FOR_SKIP * 2) |
2180 |
&& ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) |
&& ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) |
2181 |
SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data); |
SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data); |
2182 |
|
|
2214 |
SearchData * const Data) |
SearchData * const Data) |
2215 |
{ |
{ |
2216 |
|
|
2217 |
int i, mask; |
int i; |
2218 |
VECTOR pmv[3]; |
VECTOR pmv[3]; |
2219 |
MACROBLOCK * pMB = &pMBs[x + y * pParam->mb_width]; |
MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width]; |
2220 |
|
|
2221 |
for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
unsigned int simplicity = 0; |
2222 |
|
|
2223 |
//median is only used as prediction. it doesn't have to be real |
for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
|
if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; |
|
|
else |
|
|
if (x == 1) //left macroblock does not have any vector now |
|
|
Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median |
|
|
else if (y == 1) // top macroblock doesn't have it's vector |
|
|
Data->predMV = (pMB - 1)->mvs[0]; // left instead of median |
|
|
else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median |
|
2224 |
|
|
2225 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
2226 |
pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, Data->rrv); |
pParam->width, pParam->height, Data->iFcode - Data->qpel - 1, 0, 0); |
2227 |
|
|
2228 |
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
2229 |
Data->Ref = pRef + (x + y * pParam->edged_width) * 16; |
Data->RefP[0] = pRef + (x + y * pParam->edged_width) * 16; |
2230 |
|
|
2231 |
pmv[1].x = EVEN(pMB->mvs[0].x); |
pmv[0].x = pMB->mvs[0].x; |
2232 |
pmv[1].y = EVEN(pMB->mvs[0].y); |
pmv[0].y = pMB->mvs[0].y; |
|
pmv[2].x = EVEN(Data->predMV.x); |
|
|
pmv[2].y = EVEN(Data->predMV.y); |
|
|
pmv[0].x = pmv[0].y = 0; |
|
2233 |
|
|
2234 |
CheckCandidate32I(0, 0, 255, &i, Data); |
CheckCandidate32I(pmv[0].x, pmv[0].y, Data, 0); |
2235 |
|
|
2236 |
if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) { |
if (*Data->iMinSAD > 200) { |
2237 |
|
|
2238 |
if (!(mask = make_mask(pmv, 1))) |
pmv[1].x = pmv[1].y = 0; |
|
CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data); |
|
|
if (!(mask = make_mask(pmv, 2))) |
|
|
CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data); |
|
2239 |
|
|
2240 |
if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) // diamond only if needed |
/* median is only used as prediction. it doesn't have to be real */ |
2241 |
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; |
2242 |
|
else |
2243 |
|
if (x == 1) /* left macroblock does not have any vector now */ |
2244 |
|
Data->predMV = (pMB - pParam->mb_width)->mvs[0]; /* top instead of median */ |
2245 |
|
else if (y == 1) /* top macroblock doesn't have it's vector */ |
2246 |
|
Data->predMV = (pMB - 1)->mvs[0]; /* left instead of median */ |
2247 |
|
else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); /* else median */ |
2248 |
|
|
2249 |
|
pmv[2].x = Data->predMV.x; |
2250 |
|
pmv[2].y = Data->predMV.y; |
2251 |
|
|
2252 |
|
if (!vector_repeats(pmv, 1)) |
2253 |
|
CheckCandidate32I(pmv[1].x, pmv[1].y, Data, 1); |
2254 |
|
if (!vector_repeats(pmv, 2)) |
2255 |
|
CheckCandidate32I(pmv[2].x, pmv[2].y, Data, 2); |
2256 |
|
|
2257 |
|
if (*Data->iMinSAD > 500) { /* diamond only if needed */ |
2258 |
|
unsigned int mask = make_mask(pmv, 3, *Data->dir); |
2259 |
|
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate32I); |
2260 |
|
} else simplicity++; |
2261 |
|
|
2262 |
|
if (*Data->iMinSAD > 500) /* refinement from 2-pixel to 1-pixel */ |
2263 |
|
SubpelRefine(Data, CheckCandidate32I); |
2264 |
|
else simplicity++; |
2265 |
|
} else simplicity++; |
2266 |
|
|
2267 |
for (i = 0; i < 4; i++) { |
for (i = 0; i < 4; i++) { |
2268 |
MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width]; |
MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width]; |
2269 |
MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i]; |
MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i]; |
2270 |
MB->mode = MODE_INTER; |
MB->mode = MODE_INTER; |
2271 |
MB->sad16 = Data->iMinSAD[i+1]; |
/* if we skipped some search steps, we have to assume that SAD would be lower with them */ |
2272 |
} |
MB->sad16 = Data->iMinSAD[i+1] - (simplicity<<7); |
2273 |
} |
} |
2274 |
} |
} |
2275 |
|
|
2276 |
#define INTRA_BIAS 2500 |
#define INTRA_THRESH 2200 |
2277 |
#define INTRA_THRESH 1500 |
#define INTER_THRESH 40 |
2278 |
#define INTER_THRESH 1400 |
#define INTRA_THRESH2 95 |
2279 |
|
|
2280 |
int |
int |
2281 |
MEanalysis( const IMAGE * const pRef, |
MEanalysis( const IMAGE * const pRef, |
2282 |
FRAMEINFO * const Current, |
const FRAMEINFO * const Current, |
2283 |
MBParam * const pParam, |
const MBParam * const pParam, |
2284 |
int maxIntra, //maximum number if non-I frames |
const int maxIntra, //maximum number if non-I frames |
2285 |
int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame |
const int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame |
2286 |
int bCount) // number of B frames in a row |
const int bCount, // number of B frames in a row |
2287 |
|
const int b_thresh) |
2288 |
{ |
{ |
2289 |
uint32_t x, y, intra = 0; |
uint32_t x, y, intra = 0; |
2290 |
int sSAD = 0; |
int sSAD = 0; |
2291 |
MACROBLOCK * const pMBs = Current->mbs; |
MACROBLOCK * const pMBs = Current->mbs; |
2292 |
const IMAGE * const pCurrent = &Current->image; |
const IMAGE * const pCurrent = &Current->image; |
2293 |
int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH; |
int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + b_thresh; |
2294 |
|
int blocks = 0; |
2295 |
|
int complexity = 0; |
2296 |
|
|
2297 |
int32_t iMinSAD[5], temp[5]; |
int32_t iMinSAD[5], temp[5]; |
2298 |
|
uint32_t dir; |
2299 |
VECTOR currentMV[5]; |
VECTOR currentMV[5]; |
2300 |
SearchData Data; |
SearchData Data; |
2301 |
Data.iEdgedWidth = pParam->edged_width; |
Data.iEdgedWidth = pParam->edged_width; |
2302 |
Data.currentMV = currentMV; |
Data.currentMV = currentMV; |
2303 |
Data.iMinSAD = iMinSAD; |
Data.iMinSAD = iMinSAD; |
2304 |
Data.iFcode = Current->fcode; |
Data.iFcode = Current->fcode; |
|
Data.rrv = Current->global_flags & XVID_REDUCED; |
|
2305 |
Data.temp = temp; |
Data.temp = temp; |
2306 |
CheckCandidate = CheckCandidate32I; |
Data.dir = &dir; |
2307 |
|
Data.qpel = (pParam->vol_flags & XVID_VOL_QUARTERPEL)? 1: 0; |
2308 |
if (intraCount != 0 && intraCount < 10) // we're right after an I frame |
Data.qpel_precision = 0; |
2309 |
IntraThresh += 4 * (intraCount - 10) * (intraCount - 10); |
|
2310 |
|
if (intraCount != 0) { |
2311 |
|
if (intraCount < 10) // we're right after an I frame |
2312 |
|
IntraThresh += 15* (intraCount - 10) * (intraCount - 10); |
2313 |
else |
else |
2314 |
if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec |
if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec |
2315 |
IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; |
IntraThresh -= (IntraThresh * (maxIntra - 8*(maxIntra - intraCount)))/maxIntra; |
2316 |
|
} |
2317 |
|
|
2318 |
InterThresh += 400 * (1 - bCount); |
InterThresh -= 20 * bCount; |
2319 |
if (InterThresh < 300) InterThresh = 300; |
if (InterThresh < 10 + b_thresh) InterThresh = 10 + b_thresh; |
2320 |
|
|
2321 |
if (sadInit) (*sadInit) (); |
if (sadInit) (*sadInit) (); |
2322 |
|
|
2323 |
for (y = 1; y < pParam->mb_height-1; y += 2) { |
for (y = 1; y < pParam->mb_height-1; y += 2) { |
2324 |
for (x = 1; x < pParam->mb_width-1; x += 2) { |
for (x = 1; x < pParam->mb_width-1; x += 2) { |
2325 |
int i; |
int i; |
2326 |
|
blocks += 10; |
2327 |
|
|
2328 |
if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; |
if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; |
2329 |
|
else { //extrapolation of the vector found for last frame |
2330 |
|
pMBs[x + y * pParam->mb_width].mvs[0].x = |
2331 |
|
(pMBs[x + y * pParam->mb_width].mvs[0].x * (bCount+1) ) / bCount; |
2332 |
|
pMBs[x + y * pParam->mb_width].mvs[0].y = |
2333 |
|
(pMBs[x + y * pParam->mb_width].mvs[0].y * (bCount+1) ) / bCount; |
2334 |
|
} |
2335 |
|
|
2336 |
MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data); |
MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data); |
2337 |
|
|
2338 |
for (i = 0; i < 4; i++) { |
for (i = 0; i < 4; i++) { |
2339 |
int dev; |
int dev; |
2340 |
MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width]; |
MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width]; |
|
if (pMB->sad16 > IntraThresh) { |
|
2341 |
dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16, |
dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16, |
2342 |
pParam->edged_width); |
pParam->edged_width); |
2343 |
|
|
2344 |
|
complexity += MAX(dev, 300); |
2345 |
if (dev + IntraThresh < pMB->sad16) { |
if (dev + IntraThresh < pMB->sad16) { |
2346 |
pMB->mode = MODE_INTRA; |
pMB->mode = MODE_INTRA; |
2347 |
if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return I_VOP; |
if (++intra > ((pParam->mb_height-2)*(pParam->mb_width-2))/2) return I_VOP; |
|
} |
|
|
} |
|
|
sSAD += pMB->sad16; |
|
|
} |
|
|
} |
|
|
} |
|
|
sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); |
|
|
// if (sSAD > IntraThresh + INTRA_BIAS) return I_VOP; |
|
|
if (sSAD > InterThresh ) return P_VOP; |
|
|
emms(); |
|
|
return B_VOP; |
|
|
|
|
|
} |
|
|
|
|
|
|
|
|
static WARPPOINTS |
|
|
GlobalMotionEst(const MACROBLOCK * const pMBs, |
|
|
const MBParam * const pParam, |
|
|
const FRAMEINFO * const current, |
|
|
const FRAMEINFO * const reference, |
|
|
const IMAGE * const pRefH, |
|
|
const IMAGE * const pRefV, |
|
|
const IMAGE * const pRefHV ) |
|
|
{ |
|
|
|
|
|
const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs |
|
|
const int deltay=8; |
|
|
const int grad=512; // lower bound for deviation in MB |
|
|
|
|
|
WARPPOINTS gmc; |
|
|
|
|
|
uint32_t mx, my; |
|
|
|
|
|
int MBh = pParam->mb_height; |
|
|
int MBw = pParam->mb_width; |
|
|
|
|
|
int *MBmask= calloc(MBh*MBw,sizeof(int)); |
|
|
double DtimesF[4] = { 0.,0., 0., 0. }; |
|
|
double sol[4] = { 0., 0., 0., 0. }; |
|
|
double a,b,c,n,denom; |
|
|
double meanx,meany; |
|
|
int num,oldnum; |
|
|
|
|
|
if (!MBmask) { fprintf(stderr,"Mem error\n"); |
|
|
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; |
|
|
return gmc; } |
|
|
|
|
|
// filter mask of all blocks |
|
|
|
|
|
for (my = 1; my < (uint32_t)MBh-1; my++) |
|
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
|
|
{ |
|
|
const int mbnum = mx + my * MBw; |
|
|
const MACROBLOCK *pMB = &pMBs[mbnum]; |
|
|
const VECTOR mv = pMB->mvs[0]; |
|
|
|
|
|
if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) |
|
|
continue; |
|
|
|
|
|
if ( ( (ABS(mv.x - (pMB-1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-1)->mvs[0].y) < deltay) ) |
|
|
&& ( (ABS(mv.x - (pMB+1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+1)->mvs[0].y) < deltay) ) |
|
|
&& ( (ABS(mv.x - (pMB-MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-MBw)->mvs[0].y) < deltay) ) |
|
|
&& ( (ABS(mv.x - (pMB+MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) ) |
|
|
MBmask[mbnum]=1; |
|
|
} |
|
|
|
|
|
for (my = 1; my < (uint32_t)MBh-1; my++) |
|
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
|
|
{ |
|
|
const uint8_t *const pCur = current->image.y + 16*my*pParam->edged_width + 16*mx; |
|
|
|
|
|
const int mbnum = mx + my * MBw; |
|
|
if (!MBmask[mbnum]) |
|
|
continue; |
|
|
|
|
|
if (sad16 ( pCur, pCur+1 , pParam->edged_width, 65536) <= (uint32_t)grad ) |
|
|
MBmask[mbnum] = 0; |
|
|
if (sad16 ( pCur, pCur+pParam->edged_width, pParam->edged_width, 65536) <= (uint32_t)grad ) |
|
|
MBmask[mbnum] = 0; |
|
|
|
|
2348 |
} |
} |
2349 |
|
|
2350 |
emms(); |
if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0) |
2351 |
|
if (dev > 1000 && pMB->sad16 < 1000) |
2352 |
do { /* until convergence */ |
sSAD += 1000; |
|
|
|
|
a = b = c = n = 0; |
|
|
DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.; |
|
|
for (my = 0; my < (uint32_t)MBh; my++) |
|
|
for (mx = 0; mx < (uint32_t)MBw; mx++) |
|
|
{ |
|
|
const int mbnum = mx + my * MBw; |
|
|
const MACROBLOCK *pMB = &pMBs[mbnum]; |
|
|
const VECTOR mv = pMB->mvs[0]; |
|
|
|
|
|
if (!MBmask[mbnum]) |
|
|
continue; |
|
|
|
|
|
n++; |
|
|
a += 16*mx+8; |
|
|
b += 16*my+8; |
|
|
c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8); |
|
2353 |
|
|
2354 |
DtimesF[0] += (double)mv.x; |
sSAD += (dev < 4000) ? pMB->sad16 : pMB->sad16/2; /* blocks with big contrast differences usually have large SAD - while they look very good in b-frames */ |
|
DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8); |
|
|
DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8); |
|
|
DtimesF[3] += (double)mv.y; |
|
2355 |
} |
} |
|
|
|
|
denom = a*a+b*b-c*n; |
|
|
|
|
|
/* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */ |
|
|
/* D'*E*F has been calculated in the same loop as matrix */ |
|
|
|
|
|
sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2]; |
|
|
sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3]; |
|
|
sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3]; |
|
|
sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3]; |
|
|
|
|
|
sol[0] /= denom; |
|
|
sol[1] /= denom; |
|
|
sol[2] /= denom; |
|
|
sol[3] /= denom; |
|
|
|
|
|
meanx = meany = 0.; |
|
|
oldnum = 0; |
|
|
for (my = 0; my < (uint32_t)MBh; my++) |
|
|
for (mx = 0; mx < (uint32_t)MBw; mx++) |
|
|
{ |
|
|
const int mbnum = mx + my * MBw; |
|
|
const MACROBLOCK *pMB = &pMBs[mbnum]; |
|
|
const VECTOR mv = pMB->mvs[0]; |
|
|
|
|
|
if (!MBmask[mbnum]) |
|
|
continue; |
|
|
|
|
|
oldnum++; |
|
|
meanx += ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ); |
|
|
meany += ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ); |
|
2356 |
} |
} |
|
|
|
|
if (4*meanx > oldnum) /* better fit than 0.25 is useless */ |
|
|
meanx /= oldnum; |
|
|
else |
|
|
meanx = 0.25; |
|
|
|
|
|
if (4*meany > oldnum) |
|
|
meany /= oldnum; |
|
|
else |
|
|
meany = 0.25; |
|
|
|
|
|
/* fprintf(stderr,"sol = (%8.5f, %8.5f, %8.5f, %8.5f)\n",sol[0],sol[1],sol[2],sol[3]); |
|
|
fprintf(stderr,"meanx = %8.5f meany = %8.5f %d\n",meanx,meany, oldnum); |
|
|
*/ |
|
|
num = 0; |
|
|
for (my = 0; my < (uint32_t)MBh; my++) |
|
|
for (mx = 0; mx < (uint32_t)MBw; mx++) |
|
|
{ |
|
|
const int mbnum = mx + my * MBw; |
|
|
const MACROBLOCK *pMB = &pMBs[mbnum]; |
|
|
const VECTOR mv = pMB->mvs[0]; |
|
|
|
|
|
if (!MBmask[mbnum]) |
|
|
continue; |
|
|
|
|
|
if ( ( ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ) > meanx ) |
|
|
|| ( ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ) > meany ) ) |
|
|
MBmask[mbnum]=0; |
|
|
else |
|
|
num++; |
|
2357 |
} |
} |
2358 |
|
complexity >>= 7; |
2359 |
|
|
2360 |
} while ( (oldnum != num) && (num>=4) ); |
sSAD /= complexity + 4*blocks; |
|
|
|
|
if (num < 4) |
|
|
{ |
|
|
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; |
|
|
} else { |
|
|
|
|
|
gmc.duv[0].x=(int)(sol[0]+0.5); |
|
|
gmc.duv[0].y=(int)(sol[3]+0.5); |
|
|
|
|
|
gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5); |
|
|
gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5); |
|
2361 |
|
|
2362 |
gmc.duv[2].x=0; |
if (intraCount > 80 && sSAD > INTRA_THRESH2 ) return I_VOP; |
2363 |
gmc.duv[2].y=0; |
if (sSAD > InterThresh ) return P_VOP; |
2364 |
|
emms(); |
2365 |
|
return B_VOP; |
2366 |
} |
} |
|
// 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); |
|
|
|
|
|
free(MBmask); |
|
2367 |
|
|
|
return gmc; |
|
|
} |
|
2368 |
|
|
2369 |
// functions which perform BITS-based search/bitcount |
/* functions which perform BITS-based search/bitcount */ |
2370 |
|
|
2371 |
static int |
static int |
2372 |
CountMBBitsInter(SearchData * const Data, |
findRDinter(SearchData * const Data, |
2373 |
const MACROBLOCK * const pMBs, const int x, const int y, |
const MACROBLOCK * const pMBs, const int x, const int y, |
2374 |
const MBParam * const pParam, |
const MBParam * const pParam, |
2375 |
const uint32_t MotionFlags) |
const uint32_t MotionFlags) |
2376 |
{ |
{ |
2377 |
int i, iDirection; |
int i; |
2378 |
int32_t bsad[5]; |
int32_t bsad[5]; |
2379 |
|
|
|
CheckCandidate = CheckCandidateBits16; |
|
|
|
|
2380 |
if (Data->qpel) { |
if (Data->qpel) { |
2381 |
for(i = 0; i < 5; i++) { |
for(i = 0; i < 5; i++) { |
2382 |
Data->currentMV[i].x = Data->currentQMV[i].x/2; |
Data->currentMV[i].x = Data->currentQMV[i].x/2; |
2383 |
Data->currentMV[i].y = Data->currentQMV[i].y/2; |
Data->currentMV[i].y = Data->currentQMV[i].y/2; |
2384 |
} |
} |
2385 |
Data->qpel_precision = 1; |
Data->qpel_precision = 1; |
2386 |
CheckCandidateBits16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data); |
CheckCandidateRD16(Data->currentQMV[0].x, Data->currentQMV[0].y, Data, 255); |
2387 |
|
|
2388 |
//checking if this vector is perfect. if it is, we stop. |
if (MotionFlags & (XVID_ME_HALFPELREFINE16_RD | XVID_ME_EXTSEARCH_RD)) { /* we have to prepare for halfpixel-precision search */ |
|
if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0) |
|
|
return 0; //quick stop |
|
|
|
|
|
if (MotionFlags & (HALFPELREFINE16_BITS | EXTSEARCH_BITS)) { //we have to prepare for halfpixel-precision search |
|
2389 |
for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i]; |
for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i]; |
2390 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
2391 |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, Data->rrv); |
2392 |
Data->qpel_precision = 0; |
Data->qpel_precision = 0; |
2393 |
if (Data->currentQMV->x & 1 || Data->currentQMV->y & 1) |
if (Data->currentQMV->x & 1 || Data->currentQMV->y & 1) |
2394 |
CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); |
CheckCandidateRD16(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255); |
2395 |
} |
} |
2396 |
|
|
2397 |
} else { // not qpel |
} else { /* not qpel */ |
2398 |
|
|
2399 |
CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); |
CheckCandidateRD16(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255); |
|
//checking if this vector is perfect. if it is, we stop. |
|
|
if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0) { |
|
|
return 0; //inter |
|
|
} |
|
2400 |
} |
} |
2401 |
|
|
2402 |
if (MotionFlags&EXTSEARCH_BITS) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
if (MotionFlags&XVID_ME_EXTSEARCH_RD) |
2403 |
|
SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, 255, CheckCandidateRD16); |
2404 |
|
|
2405 |
if (MotionFlags&HALFPELREFINE16_BITS) SubpelRefine(Data); |
if (MotionFlags&XVID_ME_HALFPELREFINE16_RD) |
2406 |
|
SubpelRefine(Data, CheckCandidateRD16); |
2407 |
|
|
2408 |
if (Data->qpel) { |
if (Data->qpel) { |
2409 |
if (MotionFlags&(EXTSEARCH_BITS | HALFPELREFINE16_BITS)) { // there was halfpel-precision search |
if (MotionFlags&(XVID_ME_EXTSEARCH_RD | XVID_ME_HALFPELREFINE16_RD)) { /* there was halfpel-precision search */ |
2410 |
for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) { |
for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) { |
2411 |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // we have found a better match |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* we have found a better match */ |
2412 |
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
2413 |
} |
} |
2414 |
|
|
2415 |
// preparing for qpel-precision search |
/* preparing for qpel-precision search */ |
2416 |
Data->qpel_precision = 1; |
Data->qpel_precision = 1; |
2417 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
2418 |
pParam->width, pParam->height, Data->iFcode, 1, 0); |
pParam->width, pParam->height, Data->iFcode, 2, 0); |
2419 |
} |
} |
2420 |
if (MotionFlags&QUARTERPELREFINE16_BITS) SubpelRefine(Data); |
if (MotionFlags&XVID_ME_QUARTERPELREFINE16_RD) |
2421 |
|
SubpelRefine(Data, CheckCandidateRD16); |
2422 |
} |
} |
2423 |
|
|
2424 |
if (MotionFlags&CHECKPREDICTION_BITS) { //let's check vector equal to prediction |
if (MotionFlags&XVID_ME_CHECKPREDICTION_RD) { /* let's check vector equal to prediction */ |
2425 |
VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV; |
VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV; |
2426 |
if (!(Data->predMV.x == v->x && Data->predMV.y == v->y)) |
if (!(Data->predMV.x == v->x && Data->predMV.y == v->y)) |
2427 |
CheckCandidateBits16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data); |
CheckCandidateRD16(Data->predMV.x, Data->predMV.y, Data, 255); |
2428 |
} |
} |
2429 |
return Data->iMinSAD[0]; |
return Data->iMinSAD[0]; |
2430 |
} |
} |
2431 |
|
|
|
|
|
2432 |
static int |
static int |
2433 |
CountMBBitsInter4v(const SearchData * const Data, |
findRDinter4v(const SearchData * const Data, |
2434 |
MACROBLOCK * const pMB, const MACROBLOCK * const pMBs, |
MACROBLOCK * const pMB, const MACROBLOCK * const pMBs, |
2435 |
const int x, const int y, |
const int x, const int y, |
2436 |
const MBParam * const pParam, const uint32_t MotionFlags, |
const MBParam * const pParam, const uint32_t MotionFlags, |
2437 |
const VECTOR * const backup) |
const VECTOR * const backup) |
2438 |
{ |
{ |
2439 |
|
|
2440 |
int cbp = 0, bits = 0, t = 0, i, iDirection; |
int cbp = 0, bits = 0, t = 0, i; |
2441 |
SearchData Data2, *Data8 = &Data2; |
SearchData Data2, *Data8 = &Data2; |
2442 |
int sumx = 0, sumy = 0; |
int sumx = 0, sumy = 0; |
2443 |
int16_t in[64], coeff[64]; |
int16_t *in = Data->dctSpace, *coeff = Data->dctSpace + 64; |
2444 |
|
uint8_t * ptr; |
2445 |
|
|
2446 |
memcpy(Data8, Data, sizeof(SearchData)); |
memcpy(Data8, Data, sizeof(SearchData)); |
|
CheckCandidate = CheckCandidateBits8; |
|
2447 |
|
|
2448 |
for (i = 0; i < 4; i++) { |
for (i = 0; i < 4; i++) { /* for all luma blocks */ |
2449 |
|
|
2450 |
Data8->iMinSAD = Data->iMinSAD + i + 1; |
Data8->iMinSAD = Data->iMinSAD + i + 1; |
2451 |
Data8->currentMV = Data->currentMV + i + 1; |
Data8->currentMV = Data->currentMV + i + 1; |
2452 |
Data8->currentQMV = Data->currentQMV + i + 1; |
Data8->currentQMV = Data->currentQMV + i + 1; |
2453 |
Data8->Cur = Data->Cur + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
Data8->Cur = Data->Cur + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2454 |
Data8->Ref = Data->Ref + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
Data8->RefP[0] = Data->RefP[0] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2455 |
Data8->RefH = Data->RefH + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
Data8->RefP[2] = Data->RefP[2] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2456 |
Data8->RefV = Data->RefV + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
Data8->RefP[1] = Data->RefP[1] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2457 |
Data8->RefHV = Data->RefHV + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
Data8->RefP[3] = Data->RefP[3] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2458 |
|
*Data8->cbp = (Data->cbp[1] & (1<<(5-i))) ? 1:0; // copy corresponding cbp bit |
2459 |
|
|
2460 |
if(Data->qpel) { |
if(Data->qpel) { |
2461 |
Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i); |
Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i); |
2467 |
Data8->predMV, Data8->iFcode, 0, 0); |
Data8->predMV, Data8->iFcode, 0, 0); |
2468 |
} |
} |
2469 |
|
|
2470 |
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, |
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 3, |
2471 |
pParam->width, pParam->height, Data8->iFcode, Data8->qpel, 0); |
pParam->width, pParam->height, Data8->iFcode, Data8->qpel+1, 0); |
2472 |
|
|
2473 |
*Data8->iMinSAD += t; |
*Data8->iMinSAD += BITS_MULT*t; |
2474 |
|
|
2475 |
Data8->qpel_precision = Data8->qpel; |
Data8->qpel_precision = Data8->qpel; |
2476 |
// checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far) |
/* checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far) */ |
2477 |
if (Data8->qpel) { |
{ |
2478 |
if (!(Data8->currentQMV->x == backup[i+1].x && Data8->currentQMV->y == backup[i+1].y)) |
VECTOR *v = Data8->qpel ? Data8->currentQMV : Data8->currentMV; |
2479 |
CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8); |
if (!MVequal (*v, backup[i+1]) ) |
2480 |
} else { |
CheckCandidateRD8(backup[i+1].x, backup[i+1].y, Data8, 255); |
|
if (!(Data8->currentMV->x == backup[i+1].x && Data8->currentMV->y == backup[i+1].y)) |
|
|
CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8); |
|
2481 |
} |
} |
2482 |
|
|
2483 |
if (Data8->qpel) { |
if (Data8->qpel) { |
2484 |
if (MotionFlags&HALFPELREFINE8_BITS || (MotionFlags&PMV_EXTSEARCH8 && MotionFlags&EXTSEARCH_BITS)) { // halfpixel motion search follows |
if (MotionFlags&XVID_ME_HALFPELREFINE8_RD || (MotionFlags&XVID_ME_EXTSEARCH8 && MotionFlags&XVID_ME_EXTSEARCH_RD)) { /* halfpixel motion search follows */ |
2485 |
int32_t s = *Data8->iMinSAD; |
int32_t s = *Data8->iMinSAD; |
2486 |
Data8->currentMV->x = Data8->currentQMV->x/2; |
Data8->currentMV->x = Data8->currentQMV->x/2; |
2487 |
Data8->currentMV->y = Data8->currentQMV->y/2; |
Data8->currentMV->y = Data8->currentQMV->y/2; |
2488 |
Data8->qpel_precision = 0; |
Data8->qpel_precision = 0; |
2489 |
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, |
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 3, |
2490 |
pParam->width, pParam->height, Data8->iFcode - 1, 0, 0); |
pParam->width, pParam->height, Data8->iFcode - 1, 1, 0); |
2491 |
|
|
2492 |
if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1) |
if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1) |
2493 |
CheckCandidateBits8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8); |
CheckCandidateRD8(Data8->currentMV->x, Data8->currentMV->y, Data8, 255); |
2494 |
|
|
2495 |
if (MotionFlags & PMV_EXTSEARCH8 && MotionFlags & EXTSEARCH_BITS) |
if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_RD) |
2496 |
SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255); |
SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255, CheckCandidateRD8); |
2497 |
|
|
2498 |
if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8); |
if (MotionFlags & XVID_ME_HALFPELREFINE8_RD) |
2499 |
|
SubpelRefine(Data8, CheckCandidateRD8); |
2500 |
|
|
2501 |
if(s > *Data8->iMinSAD) { //we have found a better match |
if(s > *Data8->iMinSAD) { /* we have found a better match */ |
2502 |
Data8->currentQMV->x = 2*Data8->currentMV->x; |
Data8->currentQMV->x = 2*Data8->currentMV->x; |
2503 |
Data8->currentQMV->y = 2*Data8->currentMV->y; |
Data8->currentQMV->y = 2*Data8->currentMV->y; |
2504 |
} |
} |
2505 |
|
|
2506 |
Data8->qpel_precision = 1; |
Data8->qpel_precision = 1; |
2507 |
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, |
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 3, |
2508 |
pParam->width, pParam->height, Data8->iFcode, 1, 0); |
pParam->width, pParam->height, Data8->iFcode, 2, 0); |
2509 |
|
|
2510 |
} |
} |
2511 |
if (MotionFlags & QUARTERPELREFINE8_BITS) SubpelRefine(Data8); |
if (MotionFlags & XVID_ME_QUARTERPELREFINE8_RD) |
2512 |
|
SubpelRefine(Data8, CheckCandidateRD8); |
2513 |
|
|
2514 |
} else // not qpel |
} else { /* not qpel */ |
|
if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8); //halfpel mode, halfpel refinement |
|
2515 |
|
|
2516 |
//checking vector equal to predicion |
if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_RD) /* extsearch */ |
2517 |
if (i != 0 && MotionFlags & CHECKPREDICTION_BITS) { |
SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255, CheckCandidateRD8); |
2518 |
|
|
2519 |
|
if (MotionFlags & XVID_ME_HALFPELREFINE8_RD) |
2520 |
|
SubpelRefine(Data8, CheckCandidateRD8); /* halfpel refinement */ |
2521 |
|
} |
2522 |
|
|
2523 |
|
/* checking vector equal to predicion */ |
2524 |
|
if (i != 0 && MotionFlags & XVID_ME_CHECKPREDICTION_RD) { |
2525 |
const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV; |
const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV; |
2526 |
if (!(Data8->predMV.x == v->x && Data8->predMV.y == v->y)) |
if (!MVequal(*v, Data8->predMV)) |
2527 |
CheckCandidateBits8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8); |
CheckCandidateRD8(Data8->predMV.x, Data8->predMV.y, Data8, 255); |
2528 |
} |
} |
2529 |
|
|
2530 |
bits += *Data8->iMinSAD; |
bits += *Data8->iMinSAD; |
2531 |
if (bits >= Data->iMinSAD[0]) break; // no chances for INTER4V |
if (bits >= Data->iMinSAD[0]) return bits; /* no chances for INTER4V */ |
2532 |
|
|
2533 |
// MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else |
/* MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else */ |
2534 |
if(Data->qpel) { |
if(Data->qpel) { |
2535 |
pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x; |
pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x; |
2536 |
pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y; |
pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y; |
2545 |
} |
} |
2546 |
pMB->mvs[i] = *Data8->currentMV; |
pMB->mvs[i] = *Data8->currentMV; |
2547 |
pMB->sad8[i] = 4 * *Data8->iMinSAD; |
pMB->sad8[i] = 4 * *Data8->iMinSAD; |
2548 |
if (Data8->temp[0]) cbp |= 1 << (5 - i); |
if (Data8->cbp[0]) cbp |= 1 << (5 - i); |
|
} |
|
2549 |
|
|
2550 |
if (bits < *Data->iMinSAD) { // there is still a chance for inter4v mode. let's check chroma |
} /* end - for all luma blocks */ |
2551 |
const uint8_t * ptr; |
|
2552 |
|
bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; |
2553 |
|
|
2554 |
|
/* let's check chroma */ |
2555 |
sumx = (sumx >> 3) + roundtab_76[sumx & 0xf]; |
sumx = (sumx >> 3) + roundtab_76[sumx & 0xf]; |
2556 |
sumy = (sumy >> 3) + roundtab_76[sumy & 0xf]; |
sumy = (sumy >> 3) + roundtab_76[sumy & 0xf]; |
2557 |
|
|
2558 |
//chroma U |
/* chroma U */ |
2559 |
ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefCU, 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[4], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
2560 |
transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2); |
transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2); |
2561 |
fdct(in); |
bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4); |
|
if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16); |
|
|
else i = quant4_inter(coeff, in, Data->lambda16); |
|
|
if (i > 0) { |
|
|
bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); |
|
|
cbp |= 1 << (5 - 4); |
|
|
} |
|
2562 |
|
|
2563 |
if (bits < *Data->iMinSAD) { // still possible |
if (bits >= *Data->iMinSAD) return bits; |
2564 |
//chroma V |
|
2565 |
ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefCV, 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
/* chroma V */ |
2566 |
|
ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[5], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
2567 |
transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2); |
transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2); |
2568 |
fdct(in); |
bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5); |
2569 |
if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16); |
|
2570 |
else i = quant4_inter(coeff, in, Data->lambda16); |
bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len; |
|
if (i > 0) { |
|
|
bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); |
|
|
cbp |= 1 << (5 - 5); |
|
|
} |
|
|
bits += cbpy_tab[15-(cbp>>2)].len; |
|
|
bits += mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len; |
|
|
} |
|
|
} |
|
2571 |
|
|
2572 |
|
*Data->cbp = cbp; |
2573 |
return bits; |
return bits; |
2574 |
} |
} |
2575 |
|
|
|
|
|
2576 |
static int |
static int |
2577 |
CountMBBitsIntra(const SearchData * const Data) |
findRDintra(const SearchData * const Data) |
2578 |
{ |
{ |
2579 |
int bits = 1; //this one is ac/dc prediction flag. always 1. |
int bits = BITS_MULT*1; /* this one is ac/dc prediction flag bit */ |
2580 |
int cbp = 0, i, t, dc = 0, b_dc = 1024; |
int cbp = 0, i, dc = 0; |
2581 |
const uint32_t iQuant = Data->lambda16; |
int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64; |
|
int16_t in[64], coeff[64]; |
|
2582 |
|
|
2583 |
for(i = 0; i < 4; i++) { |
for(i = 0; i < 4; i++) { |
|
uint32_t iDcScaler = get_dc_scaler(iQuant, 1); |
|
|
|
|
2584 |
int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2585 |
transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth); |
transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth); |
2586 |
fdct(in); |
bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i, &dc); |
2587 |
b_dc = dc; |
|
2588 |
dc = in[0]; |
if (bits >= Data->iMinSAD[0]) return bits; |
2589 |
in[0] -= b_dc; |
} |
2590 |
if (Data->lambda8 == 0) quant_intra_c(coeff, in, iQuant, iDcScaler); |
|
2591 |
else quant4_intra_c(coeff, in, iQuant, iDcScaler); |
bits += BITS_MULT*xvid_cbpy_tab[cbp>>2].len; |
2592 |
|
|
2593 |
b_dc = dc; |
/*chroma U */ |
|
dc = coeff[0]; |
|
|
if (i != 0) coeff[0] -= b_dc; |
|
|
|
|
|
bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcy_tab[coeff[0] + 255].len;; |
|
|
Data->temp[i] = t; |
|
|
if (t != 0) cbp |= 1 << (5 - i); |
|
|
if (bits >= Data->iMinSAD[0]) break; |
|
|
} |
|
|
|
|
|
if (bits < Data->iMinSAD[0]) { // INTRA still looks good, let's add chroma |
|
|
uint32_t iDcScaler = get_dc_scaler(iQuant, 0); |
|
|
//chroma U |
|
2594 |
transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2); |
transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2); |
2595 |
fdct(in); |
bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4, &dc); |
2596 |
in[0] -= 1024; |
|
2597 |
if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler); |
if (bits >= Data->iMinSAD[0]) return bits; |
|
else quant4_intra(coeff, in, iQuant, iDcScaler); |
|
|
|
|
|
bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len; |
|
|
if (t != 0) cbp |= 1 << (5 - 4); |
|
|
Data->temp[4] = t; |
|
2598 |
|
|
2599 |
if (bits < Data->iMinSAD[0]) { |
/* chroma V */ |
|
//chroma V |
|
2600 |
transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2); |
transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2); |
2601 |
fdct(in); |
bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5, &dc); |
2602 |
in[0] -= 1024; |
|
2603 |
if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler); |
bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len; |
2604 |
else quant4_intra(coeff, in, iQuant, iDcScaler); |
|
2605 |
|
return bits; |
2606 |
|
} |
2607 |
|
|
2608 |
|
static int |
2609 |
|
findRDgmc(const SearchData * const Data, const IMAGE * const vGMC, const int x, const int y) |
2610 |
|
{ |
2611 |
|
int bits = BITS_MULT*1; /* this one is mcsel */ |
2612 |
|
int cbp = 0, i; |
2613 |
|
int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64; |
2614 |
|
|
2615 |
|
for(i = 0; i < 4; i++) { |
2616 |
|
int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2617 |
|
transfer_8to16subro(in, Data->Cur + s, vGMC->y + s + 16*(x+y*Data->iEdgedWidth), Data->iEdgedWidth); |
2618 |
|
bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i); |
2619 |
|
if (bits >= Data->iMinSAD[0]) return bits; |
2620 |
|
} |
2621 |
|
|
2622 |
|
bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; |
2623 |
|
|
2624 |
|
/*chroma U */ |
2625 |
|
transfer_8to16subro(in, Data->CurU, vGMC->u + 8*(x+y*(Data->iEdgedWidth/2)), Data->iEdgedWidth/2); |
2626 |
|
bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4); |
2627 |
|
|
2628 |
|
if (bits >= Data->iMinSAD[0]) return bits; |
2629 |
|
|
2630 |
|
/* chroma V */ |
2631 |
|
transfer_8to16subro(in, Data->CurV , vGMC->v + 8*(x+y*(Data->iEdgedWidth/2)), Data->iEdgedWidth/2); |
2632 |
|
bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5); |
2633 |
|
|
2634 |
|
bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; |
2635 |
|
|
2636 |
|
*Data->cbp = cbp; |
2637 |
|
|
2638 |
|
return bits; |
2639 |
|
} |
2640 |
|
|
2641 |
|
|
2642 |
|
|
2643 |
|
|
2644 |
|
static __inline void |
2645 |
|
GMEanalyzeMB ( const uint8_t * const pCur, |
2646 |
|
const uint8_t * const pRef, |
2647 |
|
const uint8_t * const pRefH, |
2648 |
|
const uint8_t * const pRefV, |
2649 |
|
const uint8_t * const pRefHV, |
2650 |
|
const int x, |
2651 |
|
const int y, |
2652 |
|
const MBParam * const pParam, |
2653 |
|
MACROBLOCK * const pMBs, |
2654 |
|
SearchData * const Data) |
2655 |
|
{ |
2656 |
|
|
2657 |
|
int i=0; |
2658 |
|
MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width]; |
2659 |
|
|
2660 |
|
Data->iMinSAD[0] = MV_MAX_ERROR; |
2661 |
|
|
2662 |
|
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
2663 |
|
|
2664 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
2665 |
|
pParam->width, pParam->height, 16, 1, 0); |
2666 |
|
|
2667 |
bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len; |
Data->Cur = pCur + 16*(x + y * pParam->edged_width); |
2668 |
if (t != 0) cbp |= 1 << (5 - 5); |
Data->RefP[0] = pRef + 16*(x + y * pParam->edged_width); |
2669 |
|
Data->RefP[1] = pRefV + 16*(x + y * pParam->edged_width); |
2670 |
|
Data->RefP[2] = pRefH + 16*(x + y * pParam->edged_width); |
2671 |
|
Data->RefP[3] = pRefHV + 16*(x + y * pParam->edged_width); |
2672 |
|
|
2673 |
Data->temp[5] = t; |
Data->currentMV[0].x = Data->currentMV[0].y = 0; |
2674 |
|
CheckCandidate16I(0, 0, Data, 255); |
2675 |
|
|
2676 |
bits += t = cbpy_tab[cbp>>2].len; |
if ( (Data->predMV.x !=0) || (Data->predMV.y != 0) ) |
2677 |
Data->temp[6] = t; |
CheckCandidate16I(Data->predMV.x, Data->predMV.y, Data, 255); |
2678 |
|
|
2679 |
bits += t = mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len; |
DiamondSearch(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255, CheckCandidate16I); |
|
Data->temp[7] = t; |
|
2680 |
|
|
2681 |
|
SubpelRefine(Data, CheckCandidate16I); |
2682 |
|
|
2683 |
|
|
2684 |
|
/* for QPel halfpel positions are worse than in halfpel mode :( */ |
2685 |
|
/* if (Data->qpel) { |
2686 |
|
Data->currentQMV->x = 2*Data->currentMV->x; |
2687 |
|
Data->currentQMV->y = 2*Data->currentMV->y; |
2688 |
|
Data->qpel_precision = 1; |
2689 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
2690 |
|
pParam->width, pParam->height, iFcode, 2, 0); |
2691 |
|
SubpelRefine(Data); |
2692 |
} |
} |
2693 |
|
*/ |
2694 |
|
|
2695 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
2696 |
|
pMB->sad16 = Data->iMinSAD[0]; |
2697 |
|
pMB->mode = MODE_INTER; |
2698 |
|
pMB->sad16 += 10*d_mv_bits(pMB->mvs[0].x, pMB->mvs[0].y, Data->predMV, Data->iFcode, 0, 0); |
2699 |
|
return; |
2700 |
} |
} |
2701 |
|
|
2702 |
return bits; |
void |
2703 |
|
GMEanalysis(const MBParam * const pParam, |
2704 |
|
const FRAMEINFO * const current, |
2705 |
|
const FRAMEINFO * const reference, |
2706 |
|
const IMAGE * const pRefH, |
2707 |
|
const IMAGE * const pRefV, |
2708 |
|
const IMAGE * const pRefHV) |
2709 |
|
{ |
2710 |
|
uint32_t x, y; |
2711 |
|
MACROBLOCK * const pMBs = current->mbs; |
2712 |
|
const IMAGE * const pCurrent = ¤t->image; |
2713 |
|
const IMAGE * const pReference = &reference->image; |
2714 |
|
|
2715 |
|
int32_t iMinSAD[5], temp[5]; |
2716 |
|
VECTOR currentMV[5]; |
2717 |
|
uint32_t dir; |
2718 |
|
SearchData Data; |
2719 |
|
memset(&Data, 0, sizeof(SearchData)); |
2720 |
|
|
2721 |
|
Data.iEdgedWidth = pParam->edged_width; |
2722 |
|
Data.rounding = pParam->m_rounding_type; |
2723 |
|
|
2724 |
|
Data.currentMV = ¤tMV[0]; |
2725 |
|
Data.iMinSAD = &iMinSAD[0]; |
2726 |
|
Data.iFcode = current->fcode; |
2727 |
|
Data.temp = temp; |
2728 |
|
Data.dir = &dir; |
2729 |
|
|
2730 |
|
if (sadInit) (*sadInit) (); |
2731 |
|
|
2732 |
|
for (y = 0; y < pParam->mb_height; y ++) { |
2733 |
|
for (x = 0; x < pParam->mb_width; x ++) { |
2734 |
|
GMEanalyzeMB(pCurrent->y, pReference->y, pRefH->y, pRefV->y, pRefHV->y, x, y, pParam, pMBs, &Data); |
2735 |
|
} |
2736 |
|
} |
2737 |
|
return; |
2738 |
|
} |
2739 |
|
|
2740 |
|
|
2741 |
|
WARPPOINTS |
2742 |
|
GlobalMotionEst(MACROBLOCK * const pMBs, |
2743 |
|
const MBParam * const pParam, |
2744 |
|
const FRAMEINFO * const current, |
2745 |
|
const FRAMEINFO * const reference, |
2746 |
|
const IMAGE * const pRefH, |
2747 |
|
const IMAGE * const pRefV, |
2748 |
|
const IMAGE * const pRefHV) |
2749 |
|
{ |
2750 |
|
|
2751 |
|
const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs |
2752 |
|
const int deltay=8; |
2753 |
|
const unsigned int gradx=512; // lower bound for gradient in MB (ignore "flat" blocks) |
2754 |
|
const unsigned int grady=512; |
2755 |
|
|
2756 |
|
double sol[4] = { 0., 0., 0., 0. }; |
2757 |
|
|
2758 |
|
WARPPOINTS gmc; |
2759 |
|
|
2760 |
|
uint32_t mx, my; |
2761 |
|
|
2762 |
|
int MBh = pParam->mb_height; |
2763 |
|
int MBw = pParam->mb_width; |
2764 |
|
const int minblocks = 9; //MBh*MBw/32+3; /* just some reasonable number 3% + 3 */ |
2765 |
|
const int maxblocks = MBh*MBw/4; /* just some reasonable number 3% + 3 */ |
2766 |
|
|
2767 |
|
int num=0; |
2768 |
|
int oldnum; |
2769 |
|
|
2770 |
|
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; |
2771 |
|
|
2772 |
|
GMEanalysis(pParam,current, reference, pRefH, pRefV, pRefHV); |
2773 |
|
|
2774 |
|
/* block based ME isn't done, yet, so do a quick presearch */ |
2775 |
|
|
2776 |
|
// filter mask of all blocks |
2777 |
|
|
2778 |
|
for (my = 0; my < (uint32_t)MBh; my++) |
2779 |
|
for (mx = 0; mx < (uint32_t)MBw; mx++) |
2780 |
|
{ |
2781 |
|
const int mbnum = mx + my * MBw; |
2782 |
|
pMBs[mbnum].mcsel = 0; |
2783 |
|
} |
2784 |
|
|
2785 |
|
|
2786 |
|
for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */ |
2787 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */ |
2788 |
|
{ |
2789 |
|
const int mbnum = mx + my * MBw; |
2790 |
|
MACROBLOCK *const pMB = &pMBs[mbnum]; |
2791 |
|
const VECTOR mv = pMB->mvs[0]; |
2792 |
|
|
2793 |
|
/* don't use object boundaries */ |
2794 |
|
if ( (abs(mv.x - (pMB-1)->mvs[0].x) < deltax) |
2795 |
|
&& (abs(mv.y - (pMB-1)->mvs[0].y) < deltay) |
2796 |
|
&& (abs(mv.x - (pMB+1)->mvs[0].x) < deltax) |
2797 |
|
&& (abs(mv.y - (pMB+1)->mvs[0].y) < deltay) |
2798 |
|
&& (abs(mv.x - (pMB-MBw)->mvs[0].x) < deltax) |
2799 |
|
&& (abs(mv.y - (pMB-MBw)->mvs[0].y) < deltay) |
2800 |
|
&& (abs(mv.x - (pMB+MBw)->mvs[0].x) < deltax) |
2801 |
|
&& (abs(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) |
2802 |
|
{ const int iEdgedWidth = pParam->edged_width; |
2803 |
|
const uint8_t *const pCur = current->image.y + 16*(my*iEdgedWidth + mx); |
2804 |
|
if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx ) |
2805 |
|
&& (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) ) |
2806 |
|
{ pMB->mcsel = 1; |
2807 |
|
num++; |
2808 |
|
} |
2809 |
|
|
2810 |
|
/* only use "structured" blocks */ |
2811 |
|
} |
2812 |
|
} |
2813 |
|
emms(); |
2814 |
|
|
2815 |
|
/* further filtering would be possible, but during iteration, remaining |
2816 |
|
outliers usually are removed, too */ |
2817 |
|
|
2818 |
|
if (num>= minblocks) |
2819 |
|
do { /* until convergence */ |
2820 |
|
double DtimesF[4]; |
2821 |
|
double a,b,c,n,invdenom; |
2822 |
|
double meanx,meany; |
2823 |
|
|
2824 |
|
a = b = c = n = 0; |
2825 |
|
DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.; |
2826 |
|
for (my = 1; my < (uint32_t)MBh-1; my++) |
2827 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
2828 |
|
{ |
2829 |
|
const int mbnum = mx + my * MBw; |
2830 |
|
const VECTOR mv = pMBs[mbnum].mvs[0]; |
2831 |
|
|
2832 |
|
if (!pMBs[mbnum].mcsel) |
2833 |
|
continue; |
2834 |
|
|
2835 |
|
n++; |
2836 |
|
a += 16*mx+8; |
2837 |
|
b += 16*my+8; |
2838 |
|
c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8); |
2839 |
|
|
2840 |
|
DtimesF[0] += (double)mv.x; |
2841 |
|
DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8); |
2842 |
|
DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8); |
2843 |
|
DtimesF[3] += (double)mv.y; |
2844 |
|
} |
2845 |
|
|
2846 |
|
invdenom = a*a+b*b-c*n; |
2847 |
|
|
2848 |
|
/* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */ |
2849 |
|
/* D'*E*F has been calculated in the same loop as matrix */ |
2850 |
|
|
2851 |
|
sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2]; |
2852 |
|
sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3]; |
2853 |
|
sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3]; |
2854 |
|
sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3]; |
2855 |
|
|
2856 |
|
sol[0] /= invdenom; |
2857 |
|
sol[1] /= invdenom; |
2858 |
|
sol[2] /= invdenom; |
2859 |
|
sol[3] /= invdenom; |
2860 |
|
|
2861 |
|
meanx = meany = 0.; |
2862 |
|
oldnum = 0; |
2863 |
|
for (my = 1; my < (uint32_t)MBh-1; my++) |
2864 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
2865 |
|
{ |
2866 |
|
const int mbnum = mx + my * MBw; |
2867 |
|
const VECTOR mv = pMBs[mbnum].mvs[0]; |
2868 |
|
|
2869 |
|
if (!pMBs[mbnum].mcsel) |
2870 |
|
continue; |
2871 |
|
|
2872 |
|
oldnum++; |
2873 |
|
meanx += fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - (double)mv.x ); |
2874 |
|
meany += fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - (double)mv.y ); |
2875 |
|
} |
2876 |
|
|
2877 |
|
if (4*meanx > oldnum) /* better fit than 0.25 (=1/4pel) is useless */ |
2878 |
|
meanx /= oldnum; |
2879 |
|
else |
2880 |
|
meanx = 0.25; |
2881 |
|
|
2882 |
|
if (4*meany > oldnum) |
2883 |
|
meany /= oldnum; |
2884 |
|
else |
2885 |
|
meany = 0.25; |
2886 |
|
|
2887 |
|
num = 0; |
2888 |
|
for (my = 0; my < (uint32_t)MBh; my++) |
2889 |
|
for (mx = 0; mx < (uint32_t)MBw; mx++) |
2890 |
|
{ |
2891 |
|
const int mbnum = mx + my * MBw; |
2892 |
|
const VECTOR mv = pMBs[mbnum].mvs[0]; |
2893 |
|
|
2894 |
|
if (!pMBs[mbnum].mcsel) |
2895 |
|
continue; |
2896 |
|
|
2897 |
|
if ( ( fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - (double)mv.x ) > meanx ) |
2898 |
|
|| ( fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - (double)mv.y ) > meany ) ) |
2899 |
|
pMBs[mbnum].mcsel=0; |
2900 |
|
else |
2901 |
|
num++; |
2902 |
|
} |
2903 |
|
|
2904 |
|
} while ( (oldnum != num) && (num>= minblocks) ); |
2905 |
|
|
2906 |
|
if (num < minblocks) |
2907 |
|
{ |
2908 |
|
const int iEdgedWidth = pParam->edged_width; |
2909 |
|
num = 0; |
2910 |
|
|
2911 |
|
/* fprintf(stderr,"Warning! Unreliable GME (%d/%d blocks), falling back to translation.\n",num,MBh*MBw); |
2912 |
|
*/ |
2913 |
|
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; |
2914 |
|
|
2915 |
|
if (!(current->motion_flags & XVID_ME_GME_REFINE)) |
2916 |
|
return gmc; |
2917 |
|
|
2918 |
|
for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */ |
2919 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */ |
2920 |
|
{ |
2921 |
|
const int mbnum = mx + my * MBw; |
2922 |
|
MACROBLOCK *const pMB = &pMBs[mbnum]; |
2923 |
|
const uint8_t *const pCur = current->image.y + 16*(my*iEdgedWidth + mx); |
2924 |
|
if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx ) |
2925 |
|
&& (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) ) |
2926 |
|
{ pMB->mcsel = 1; |
2927 |
|
gmc.duv[0].x += pMB->mvs[0].x; |
2928 |
|
gmc.duv[0].y += pMB->mvs[0].y; |
2929 |
|
num++; |
2930 |
|
} |
2931 |
|
} |
2932 |
|
|
2933 |
|
if (gmc.duv[0].x) |
2934 |
|
gmc.duv[0].x /= num; |
2935 |
|
if (gmc.duv[0].y) |
2936 |
|
gmc.duv[0].y /= num; |
2937 |
|
} else { |
2938 |
|
|
2939 |
|
gmc.duv[0].x=(int)(sol[0]+0.5); |
2940 |
|
gmc.duv[0].y=(int)(sol[3]+0.5); |
2941 |
|
|
2942 |
|
gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5); |
2943 |
|
gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5); |
2944 |
|
|
2945 |
|
gmc.duv[2].x=-gmc.duv[1].y; /* two warp points only */ |
2946 |
|
gmc.duv[2].y=gmc.duv[1].x; |
2947 |
|
} |
2948 |
|
if (num>maxblocks) |
2949 |
|
{ for (my = 1; my < (uint32_t)MBh-1; my++) |
2950 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
2951 |
|
{ |
2952 |
|
const int mbnum = mx + my * MBw; |
2953 |
|
if (pMBs[mbnum-1].mcsel) |
2954 |
|
pMBs[mbnum].mcsel=0; |
2955 |
|
else |
2956 |
|
if (pMBs[mbnum-MBw].mcsel) |
2957 |
|
pMBs[mbnum].mcsel=0; |
2958 |
|
} |
2959 |
|
} |
2960 |
|
return gmc; |
2961 |
|
} |
2962 |
|
|
2963 |
|
int |
2964 |
|
GlobalMotionEstRefine( |
2965 |
|
WARPPOINTS *const startwp, |
2966 |
|
MACROBLOCK * const pMBs, |
2967 |
|
const MBParam * const pParam, |
2968 |
|
const FRAMEINFO * const current, |
2969 |
|
const FRAMEINFO * const reference, |
2970 |
|
const IMAGE * const pCurr, |
2971 |
|
const IMAGE * const pRef, |
2972 |
|
const IMAGE * const pRefH, |
2973 |
|
const IMAGE * const pRefV, |
2974 |
|
const IMAGE * const pRefHV) |
2975 |
|
{ |
2976 |
|
uint8_t* GMCblock = (uint8_t*)malloc(16*pParam->edged_width); |
2977 |
|
WARPPOINTS bestwp=*startwp; |
2978 |
|
WARPPOINTS centerwp,currwp; |
2979 |
|
int gmcminSAD=0; |
2980 |
|
int gmcSAD=0; |
2981 |
|
int direction; |
2982 |
|
// int mx,my; |
2983 |
|
|
2984 |
|
/* use many blocks... */ |
2985 |
|
/* for (my = 0; my < (uint32_t)pParam->mb_height; my++) |
2986 |
|
for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++) |
2987 |
|
{ |
2988 |
|
const int mbnum = mx + my * pParam->mb_width; |
2989 |
|
pMBs[mbnum].mcsel=1; |
2990 |
|
} |
2991 |
|
*/ |
2992 |
|
|
2993 |
|
/* or rather don't use too many blocks... */ |
2994 |
|
/* |
2995 |
|
for (my = 1; my < (uint32_t)MBh-1; my++) |
2996 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
2997 |
|
{ |
2998 |
|
const int mbnum = mx + my * MBw; |
2999 |
|
if (MBmask[mbnum-1]) |
3000 |
|
MBmask[mbnum-1]=0; |
3001 |
|
else |
3002 |
|
if (MBmask[mbnum-MBw]) |
3003 |
|
MBmask[mbnum-1]=0; |
3004 |
|
|
3005 |
|
} |
3006 |
|
*/ |
3007 |
|
gmcminSAD = globalSAD(&bestwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3008 |
|
|
3009 |
|
if ( (reference->coding_type == S_VOP) |
3010 |
|
&& ( (reference->warp.duv[1].x != bestwp.duv[1].x) |
3011 |
|
|| (reference->warp.duv[1].y != bestwp.duv[1].y) |
3012 |
|
|| (reference->warp.duv[0].x != bestwp.duv[0].x) |
3013 |
|
|| (reference->warp.duv[0].y != bestwp.duv[0].y) |
3014 |
|
|| (reference->warp.duv[2].x != bestwp.duv[2].x) |
3015 |
|
|| (reference->warp.duv[2].y != bestwp.duv[2].y) ) ) |
3016 |
|
{ |
3017 |
|
gmcSAD = globalSAD(&reference->warp, pParam, pMBs, |
3018 |
|
current, pRef, pCurr, GMCblock); |
3019 |
|
|
3020 |
|
if (gmcSAD < gmcminSAD) |
3021 |
|
{ bestwp = reference->warp; |
3022 |
|
gmcminSAD = gmcSAD; |
3023 |
|
} |
3024 |
|
} |
3025 |
|
|
3026 |
|
do { |
3027 |
|
direction = 0; |
3028 |
|
centerwp = bestwp; |
3029 |
|
|
3030 |
|
currwp = centerwp; |
3031 |
|
|
3032 |
|
currwp.duv[0].x--; |
3033 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3034 |
|
if (gmcSAD < gmcminSAD) |
3035 |
|
{ bestwp = currwp; |
3036 |
|
gmcminSAD = gmcSAD; |
3037 |
|
direction = 1; |
3038 |
} |
} |
3039 |
|
else |
3040 |
|
{ |
3041 |
|
currwp = centerwp; currwp.duv[0].x++; |
3042 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3043 |
|
if (gmcSAD < gmcminSAD) |
3044 |
|
{ bestwp = currwp; |
3045 |
|
gmcminSAD = gmcSAD; |
3046 |
|
direction = 2; |
3047 |
|
} |
3048 |
|
} |
3049 |
|
if (direction) continue; |
3050 |
|
|
3051 |
|
currwp = centerwp; currwp.duv[0].y--; |
3052 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3053 |
|
if (gmcSAD < gmcminSAD) |
3054 |
|
{ bestwp = currwp; |
3055 |
|
gmcminSAD = gmcSAD; |
3056 |
|
direction = 4; |
3057 |
|
} |
3058 |
|
else |
3059 |
|
{ |
3060 |
|
currwp = centerwp; currwp.duv[0].y++; |
3061 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3062 |
|
if (gmcSAD < gmcminSAD) |
3063 |
|
{ bestwp = currwp; |
3064 |
|
gmcminSAD = gmcSAD; |
3065 |
|
direction = 8; |
3066 |
|
} |
3067 |
|
} |
3068 |
|
if (direction) continue; |
3069 |
|
|
3070 |
|
currwp = centerwp; currwp.duv[1].x++; |
3071 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3072 |
|
if (gmcSAD < gmcminSAD) |
3073 |
|
{ bestwp = currwp; |
3074 |
|
gmcminSAD = gmcSAD; |
3075 |
|
direction = 32; |
3076 |
|
} |
3077 |
|
currwp.duv[2].y++; |
3078 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3079 |
|
if (gmcSAD < gmcminSAD) |
3080 |
|
{ bestwp = currwp; |
3081 |
|
gmcminSAD = gmcSAD; |
3082 |
|
direction = 1024; |
3083 |
|
} |
3084 |
|
|
3085 |
|
currwp = centerwp; currwp.duv[1].x--; |
3086 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3087 |
|
if (gmcSAD < gmcminSAD) |
3088 |
|
{ bestwp = currwp; |
3089 |
|
gmcminSAD = gmcSAD; |
3090 |
|
direction = 16; |
3091 |
|
} |
3092 |
|
else |
3093 |
|
{ |
3094 |
|
currwp = centerwp; currwp.duv[1].x++; |
3095 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3096 |
|
if (gmcSAD < gmcminSAD) |
3097 |
|
{ bestwp = currwp; |
3098 |
|
gmcminSAD = gmcSAD; |
3099 |
|
direction = 32; |
3100 |
|
} |
3101 |
|
} |
3102 |
|
if (direction) continue; |
3103 |
|
|
3104 |
|
|
3105 |
|
currwp = centerwp; currwp.duv[1].y--; |
3106 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3107 |
|
if (gmcSAD < gmcminSAD) |
3108 |
|
{ bestwp = currwp; |
3109 |
|
gmcminSAD = gmcSAD; |
3110 |
|
direction = 64; |
3111 |
|
} |
3112 |
|
else |
3113 |
|
{ |
3114 |
|
currwp = centerwp; currwp.duv[1].y++; |
3115 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3116 |
|
if (gmcSAD < gmcminSAD) |
3117 |
|
{ bestwp = currwp; |
3118 |
|
gmcminSAD = gmcSAD; |
3119 |
|
direction = 128; |
3120 |
|
} |
3121 |
|
} |
3122 |
|
if (direction) continue; |
3123 |
|
|
3124 |
|
currwp = centerwp; currwp.duv[2].x--; |
3125 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3126 |
|
if (gmcSAD < gmcminSAD) |
3127 |
|
{ bestwp = currwp; |
3128 |
|
gmcminSAD = gmcSAD; |
3129 |
|
direction = 256; |
3130 |
|
} |
3131 |
|
else |
3132 |
|
{ |
3133 |
|
currwp = centerwp; currwp.duv[2].x++; |
3134 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3135 |
|
if (gmcSAD < gmcminSAD) |
3136 |
|
{ bestwp = currwp; |
3137 |
|
gmcminSAD = gmcSAD; |
3138 |
|
direction = 512; |
3139 |
|
} |
3140 |
|
} |
3141 |
|
if (direction) continue; |
3142 |
|
|
3143 |
|
currwp = centerwp; currwp.duv[2].y--; |
3144 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3145 |
|
if (gmcSAD < gmcminSAD) |
3146 |
|
{ bestwp = currwp; |
3147 |
|
gmcminSAD = gmcSAD; |
3148 |
|
direction = 1024; |
3149 |
|
} |
3150 |
|
else |
3151 |
|
{ |
3152 |
|
currwp = centerwp; currwp.duv[2].y++; |
3153 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3154 |
|
if (gmcSAD < gmcminSAD) |
3155 |
|
{ bestwp = currwp; |
3156 |
|
gmcminSAD = gmcSAD; |
3157 |
|
direction = 2048; |
3158 |
|
} |
3159 |
|
} |
3160 |
|
} while (direction); |
3161 |
|
free(GMCblock); |
3162 |
|
|
3163 |
|
*startwp = bestwp; |
3164 |
|
|
3165 |
|
return gmcminSAD; |
3166 |
|
} |
3167 |
|
|
3168 |
|
int |
3169 |
|
globalSAD(const WARPPOINTS *const wp, |
3170 |
|
const MBParam * const pParam, |
3171 |
|
const MACROBLOCK * const pMBs, |
3172 |
|
const FRAMEINFO * const current, |
3173 |
|
const IMAGE * const pRef, |
3174 |
|
const IMAGE * const pCurr, |
3175 |
|
uint8_t *const GMCblock) |
3176 |
|
{ |
3177 |
|
NEW_GMC_DATA gmc_data; |
3178 |
|
int iSAD, gmcSAD=0; |
3179 |
|
int num=0; |
3180 |
|
unsigned int mx, my; |
3181 |
|
|
3182 |
|
generate_GMCparameters( 3, 3, wp, pParam->width, pParam->height, &gmc_data); |
3183 |
|
|
3184 |
|
for (my = 0; my < (uint32_t)pParam->mb_height; my++) |
3185 |
|
for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++) { |
3186 |
|
|
3187 |
|
const int mbnum = mx + my * pParam->mb_width; |
3188 |
|
const int iEdgedWidth = pParam->edged_width; |
3189 |
|
|
3190 |
|
if (!pMBs[mbnum].mcsel) |
3191 |
|
continue; |
3192 |
|
|
3193 |
|
gmc_data.predict_16x16(&gmc_data, GMCblock, |
3194 |
|
pRef->y, |
3195 |
|
iEdgedWidth, |
3196 |
|
iEdgedWidth, |
3197 |
|
mx, my, |
3198 |
|
pParam->m_rounding_type); |
3199 |
|
|
3200 |
|
iSAD = sad16 ( pCurr->y + 16*(my*iEdgedWidth + mx), |
3201 |
|
GMCblock , iEdgedWidth, 65536); |
3202 |
|
iSAD -= pMBs[mbnum].sad16; |
3203 |
|
|
3204 |
|
if (iSAD<0) |
3205 |
|
gmcSAD += iSAD; |
3206 |
|
num++; |
3207 |
|
} |
3208 |
|
return gmcSAD; |
3209 |
|
} |
3210 |
|
|