31 |
#include <assert.h> |
#include <assert.h> |
32 |
#include <stdio.h> |
#include <stdio.h> |
33 |
#include <stdlib.h> |
#include <stdlib.h> |
34 |
|
#include <string.h> // memcpy |
35 |
|
#include <math.h> // lrint |
36 |
|
|
37 |
#include "../encoder.h" |
#include "../encoder.h" |
38 |
#include "../utils/mbfunctions.h" |
#include "../utils/mbfunctions.h" |
44 |
#include "motion.h" |
#include "motion.h" |
45 |
#include "sad.h" |
#include "sad.h" |
46 |
#include "../utils/emms.h" |
#include "../utils/emms.h" |
47 |
|
#include "../dct/fdct.h" |
48 |
|
|
49 |
|
/***************************************************************************** |
50 |
|
* Modified rounding tables -- declared in motion.h |
51 |
|
* Original tables see ISO spec tables 7-6 -> 7-9 |
52 |
|
****************************************************************************/ |
53 |
|
|
54 |
|
const uint32_t roundtab[16] = |
55 |
|
{0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2 }; |
56 |
|
|
57 |
|
/* K = 4 */ |
58 |
|
const uint32_t roundtab_76[16] = |
59 |
|
{ 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1 }; |
60 |
|
|
61 |
|
/* K = 2 */ |
62 |
|
const uint32_t roundtab_78[8] = |
63 |
|
{ 0, 0, 1, 1, 0, 0, 0, 1 }; |
64 |
|
|
65 |
|
/* K = 1 */ |
66 |
|
const uint32_t roundtab_79[4] = |
67 |
|
{ 0, 1, 0, 0 }; |
68 |
|
|
69 |
#define INITIAL_SKIP_THRESH (10) |
#define INITIAL_SKIP_THRESH (10) |
70 |
#define FINAL_SKIP_THRESH (50) |
#define FINAL_SKIP_THRESH (50) |
71 |
#define MAX_SAD00_FOR_SKIP (20) |
#define MAX_SAD00_FOR_SKIP (20) |
72 |
#define MAX_CHROMA_SAD_FOR_SKIP (22) |
#define MAX_CHROMA_SAD_FOR_SKIP (22) |
|
#define SKIP_THRESH_B (25) |
|
73 |
|
|
74 |
#define CHECK_CANDIDATE(X,Y,D) { \ |
#define CHECK_CANDIDATE(X,Y,D) { \ |
75 |
(*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } |
CheckCandidate((X),(Y), (D), &iDirection, data ); } |
76 |
|
|
77 |
#define iDiamondSize 2 |
/***************************************************************************** |
78 |
|
* Code |
79 |
|
****************************************************************************/ |
80 |
|
|
81 |
|
static __inline uint32_t |
82 |
|
d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv) |
83 |
|
{ |
84 |
|
int bits; |
85 |
|
const int q = (1 << (iFcode - 1)) - 1; |
86 |
|
|
87 |
|
x <<= qpel; |
88 |
|
y <<= qpel; |
89 |
|
if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); } |
90 |
|
|
91 |
|
x -= pred.x; |
92 |
|
bits = (x != 0 ? iFcode:0); |
93 |
|
x = abs(x); |
94 |
|
x += q; |
95 |
|
x >>= (iFcode - 1); |
96 |
|
bits += mvtab[x]; |
97 |
|
|
98 |
static __inline int |
y -= pred.y; |
99 |
d_mv_bits(int x, int y, const uint32_t iFcode) |
bits += (y != 0 ? iFcode:0); |
100 |
|
y = abs(y); |
101 |
|
y += q; |
102 |
|
y >>= (iFcode - 1); |
103 |
|
bits += mvtab[y]; |
104 |
|
|
105 |
|
return bits; |
106 |
|
} |
107 |
|
|
108 |
|
static int32_t ChromaSAD2(const int fx, const int fy, const int bx, const int by, |
109 |
|
const SearchData * const data) |
110 |
{ |
{ |
111 |
int xb, yb; |
int sad; |
112 |
|
const uint32_t stride = data->iEdgedWidth/2; |
113 |
|
uint8_t * f_refu = data->RefQ, |
114 |
|
* f_refv = data->RefQ + 8, |
115 |
|
* b_refu = data->RefQ + 16, |
116 |
|
* b_refv = data->RefQ + 24; |
117 |
|
int offset = (fx>>1) + (fy>>1)*stride; |
118 |
|
|
119 |
if (x == 0) xb = 1; |
switch (((fx & 1) << 1) | (fy & 1)) { |
120 |
else { |
case 0: |
121 |
if (x < 0) x = -x; |
f_refu = (uint8_t*)data->RefP[4] + offset; |
122 |
x += (1 << (iFcode - 1)) - 1; |
f_refv = (uint8_t*)data->RefP[5] + offset; |
123 |
x >>= (iFcode - 1); |
break; |
124 |
if (x > 32) x = 32; |
case 1: |
125 |
xb = mvtab[x] + iFcode; |
interpolate8x8_halfpel_v(f_refu, data->RefP[4] + offset, stride, data->rounding); |
126 |
|
interpolate8x8_halfpel_v(f_refv, data->RefP[5] + offset, stride, data->rounding); |
127 |
|
break; |
128 |
|
case 2: |
129 |
|
interpolate8x8_halfpel_h(f_refu, data->RefP[4] + offset, stride, data->rounding); |
130 |
|
interpolate8x8_halfpel_h(f_refv, data->RefP[5] + offset, stride, data->rounding); |
131 |
|
break; |
132 |
|
default: |
133 |
|
interpolate8x8_halfpel_hv(f_refu, data->RefP[4] + offset, stride, data->rounding); |
134 |
|
interpolate8x8_halfpel_hv(f_refv, data->RefP[5] + offset, stride, data->rounding); |
135 |
|
break; |
136 |
} |
} |
137 |
|
|
138 |
if (y == 0) yb = 1; |
offset = (bx>>1) + (by>>1)*stride; |
139 |
else { |
switch (((bx & 1) << 1) | (by & 1)) { |
140 |
if (y < 0) y = -y; |
case 0: |
141 |
y += (1 << (iFcode - 1)) - 1; |
b_refu = (uint8_t*)data->b_RefP[4] + offset; |
142 |
y >>= (iFcode - 1); |
b_refv = (uint8_t*)data->b_RefP[5] + offset; |
143 |
if (y > 32) y = 32; |
break; |
144 |
yb = mvtab[y] + iFcode; |
case 1: |
145 |
|
interpolate8x8_halfpel_v(b_refu, data->b_RefP[4] + offset, stride, data->rounding); |
146 |
|
interpolate8x8_halfpel_v(b_refv, data->b_RefP[5] + offset, stride, data->rounding); |
147 |
|
break; |
148 |
|
case 2: |
149 |
|
interpolate8x8_halfpel_h(b_refu, data->b_RefP[4] + offset, stride, data->rounding); |
150 |
|
interpolate8x8_halfpel_h(b_refv, data->b_RefP[5] + offset, stride, data->rounding); |
151 |
|
break; |
152 |
|
default: |
153 |
|
interpolate8x8_halfpel_hv(b_refu, data->b_RefP[4] + offset, stride, data->rounding); |
154 |
|
interpolate8x8_halfpel_hv(b_refv, data->b_RefP[5] + offset, stride, data->rounding); |
155 |
|
break; |
156 |
} |
} |
157 |
return xb + yb; |
|
158 |
|
sad = sad8bi(data->CurU, b_refu, f_refu, stride); |
159 |
|
sad += sad8bi(data->CurV, b_refv, f_refv, stride); |
160 |
|
|
161 |
|
return sad; |
162 |
} |
} |
163 |
|
|
164 |
static int32_t |
static int32_t |
165 |
ChromaSAD(int dx, int dy, const SearchData * const data) |
ChromaSAD(const int dx, const int dy, const SearchData * const data) |
166 |
{ |
{ |
167 |
int sad; |
int sad; |
168 |
dx = (dx >> 1) + roundtab_79[dx & 0x3]; |
const uint32_t stride = data->iEdgedWidth/2; |
169 |
dy = (dy >> 1) + roundtab_79[dy & 0x3]; |
int offset = (dx>>1) + (dy>>1)*stride; |
170 |
|
|
171 |
|
if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently |
172 |
|
data->temp[5] = dx; data->temp[6] = dy; // backup |
173 |
|
|
174 |
switch (((dx & 1) << 1) + (dy & 1)) { // ((dx%2)?2:0)+((dy%2)?1:0) |
switch (((dx & 1) << 1) | (dy & 1)) { |
175 |
case 0: |
case 0: |
176 |
sad = sad8(data->CurU, data->RefCU + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); |
sad = sad8(data->CurU, data->RefP[4] + offset, stride); |
177 |
sad += sad8(data->CurV, data->RefCV + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); |
sad += sad8(data->CurV, data->RefP[5] + offset, stride); |
178 |
break; |
break; |
179 |
case 1: |
case 1: |
180 |
dx = dx / 2; dy = (dy - 1) / 2; |
sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + stride, stride); |
181 |
sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); |
sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + stride, stride); |
|
sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); |
|
182 |
break; |
break; |
183 |
case 2: |
case 2: |
184 |
dx = (dx - 1) / 2; dy = dy / 2; |
sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + 1, stride); |
185 |
sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); |
sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + 1, stride); |
|
sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); |
|
186 |
break; |
break; |
187 |
default: |
default: |
188 |
dx = (dx - 1) / 2; dy = (dy - 1) / 2; |
interpolate8x8_halfpel_hv(data->RefQ, data->RefP[4] + offset, stride, data->rounding); |
189 |
interpolate8x8_halfpel_hv(data->RefQ, |
sad = sad8(data->CurU, data->RefQ, stride); |
190 |
data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, |
|
191 |
data->rounding); |
interpolate8x8_halfpel_hv(data->RefQ, data->RefP[5] + offset, stride, data->rounding); |
192 |
sad = sad8(data->CurU, data->RefQ, data->iEdgedWidth/2); |
sad += sad8(data->CurV, data->RefQ, stride); |
|
interpolate8x8_halfpel_hv(data->RefQ, |
|
|
data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, |
|
|
data->rounding); |
|
|
sad += sad8(data->CurV, data->RefQ, data->iEdgedWidth/2); |
|
193 |
break; |
break; |
194 |
} |
} |
195 |
|
data->temp[7] = sad; //backup, part 2 |
196 |
return sad; |
return sad; |
197 |
} |
} |
198 |
|
|
199 |
static __inline const uint8_t * |
static __inline const uint8_t * |
200 |
GetReference(const int x, const int y, const int dir, const SearchData * const data) |
GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data) |
201 |
{ |
{ |
202 |
// dir : 0 = forward, 1 = backward |
// dir : 0 = forward, 1 = backward |
203 |
switch ( (dir << 2) | ((x&1)<<1) | (y&1) ) { |
const uint8_t *const *const direction = ( dir == 0 ? data->RefP : data->b_RefP ); |
204 |
case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth); |
const int picture = ((x&1)<<1) | (y&1); |
205 |
case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); |
const int offset = (x>>1) + (y>>1)*data->iEdgedWidth; |
206 |
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); |
|
|
|
|
207 |
} |
} |
208 |
|
|
209 |
|
// this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate |
210 |
|
static __inline const uint8_t * |
211 |
|
GetReference(const int x, const int y, const SearchData * const data) |
212 |
|
{ |
213 |
|
const int picture = ((x&1)<<1) | (y&1); |
214 |
|
const int offset = (x>>1) + (y>>1)*data->iEdgedWidth; |
215 |
|
return data->RefP[picture] + offset; |
216 |
} |
} |
217 |
|
|
218 |
static uint8_t * |
static uint8_t * |
219 |
Interpolate8x8qpel(const int x, const int y, const int block, const int dir, const SearchData * const data) |
Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data) |
220 |
{ |
{ |
221 |
// create or find a qpel-precision reference picture; return pointer to it |
// create or find a qpel-precision reference picture; return pointer to it |
222 |
uint8_t * Reference = (uint8_t *)data->RefQ; |
uint8_t * Reference = data->RefQ + 16*dir; |
223 |
const int32_t iEdgedWidth = data->iEdgedWidth; |
const uint32_t iEdgedWidth = data->iEdgedWidth; |
224 |
const uint32_t rounding = data->rounding; |
const uint32_t rounding = data->rounding; |
225 |
const int halfpel_x = x/2; |
const int halfpel_x = x/2; |
226 |
const int halfpel_y = y/2; |
const int halfpel_y = y/2; |
227 |
const uint8_t *ref1, *ref2, *ref3, *ref4; |
const uint8_t *ref1, *ref2, *ref3, *ref4; |
228 |
|
|
229 |
ref1 = GetReference(halfpel_x, halfpel_y, dir, data); // this reference is used in all cases |
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
230 |
ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
231 |
switch( ((x&1)<<1) + (y&1) ) { |
switch( ((x&1)<<1) + (y&1) ) { |
232 |
case 0: // pure halfpel position |
case 3: // x and y in qpel resolution - the "corners" (top left/right and |
233 |
Reference = (uint8_t *) GetReference(halfpel_x, halfpel_y, dir, data); |
// bottom left/right) during qpel refinement |
234 |
Reference += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
235 |
|
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
236 |
|
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
237 |
|
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
238 |
|
ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
239 |
|
ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
240 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
241 |
break; |
break; |
242 |
|
|
243 |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
244 |
ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
245 |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
246 |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
247 |
break; |
break; |
248 |
|
|
249 |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
250 |
ref2 = GetReference(x - halfpel_x, halfpel_y, dir, data); |
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
251 |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
252 |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
253 |
break; |
break; |
254 |
|
|
255 |
default: // x and y in qpel resolution - the "corners" (top left/right and |
default: // pure halfpel position |
256 |
// bottom left/right) during qpel refinement |
return (uint8_t *) ref1; |
257 |
ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
|
|
ref3 = GetReference(x - halfpel_x, halfpel_y, dir, data); |
|
|
ref4 = GetReference(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; |
|
258 |
} |
} |
259 |
return Reference; |
return Reference; |
260 |
} |
} |
261 |
|
|
262 |
static uint8_t * |
static uint8_t * |
263 |
Interpolate16x16qpel(const int x, const int y, const int dir, const SearchData * const data) |
Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data) |
264 |
{ |
{ |
265 |
// create or find a qpel-precision reference picture; return pointer to it |
// create or find a qpel-precision reference picture; return pointer to it |
266 |
uint8_t * Reference = (uint8_t *)data->RefQ; |
uint8_t * Reference = data->RefQ + 16*dir; |
267 |
const int32_t iEdgedWidth = data->iEdgedWidth; |
const uint32_t iEdgedWidth = data->iEdgedWidth; |
268 |
const uint32_t rounding = data->rounding; |
const uint32_t rounding = data->rounding; |
269 |
const int halfpel_x = x/2; |
const int halfpel_x = x/2; |
270 |
const int halfpel_y = y/2; |
const int halfpel_y = y/2; |
271 |
const uint8_t *ref1, *ref2, *ref3, *ref4; |
const uint8_t *ref1, *ref2, *ref3, *ref4; |
272 |
|
|
273 |
ref1 = GetReference(halfpel_x, halfpel_y, dir, data); // this reference is used in all cases |
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
274 |
switch( ((x&1)<<1) + (y&1) ) { |
switch( ((x&1)<<1) + (y&1) ) { |
275 |
case 0: // pure halfpel position |
case 3: // x and y in qpel resolution - the "corners" (top left/right and |
276 |
return (uint8_t *) GetReference(halfpel_x, halfpel_y, dir, data); |
// bottom left/right) during qpel refinement |
277 |
|
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
278 |
|
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
279 |
|
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
280 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
281 |
|
interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); |
282 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); |
283 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); |
284 |
|
break; |
285 |
|
|
286 |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
287 |
ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
288 |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
289 |
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, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); |
292 |
break; |
break; |
293 |
|
|
294 |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
295 |
ref2 = GetReference(x - halfpel_x, halfpel_y, dir, data); |
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
296 |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
297 |
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
298 |
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); |
299 |
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); |
300 |
break; |
break; |
301 |
|
|
302 |
default: // x and y in qpel resolution - the "corners" (top left/right and |
default: // pure halfpel position |
303 |
// bottom left/right) during qpel refinement |
return (uint8_t *) ref1; |
|
ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
|
|
ref3 = GetReference(x - halfpel_x, halfpel_y, dir, data); |
|
|
ref4 = GetReference(x - halfpel_x, y - halfpel_y, dir, data); |
|
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
|
|
interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); |
|
|
interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, 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); |
|
|
break; |
|
304 |
} |
} |
305 |
return Reference; |
return Reference; |
306 |
} |
} |
310 |
static void |
static void |
311 |
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 int Direction, int * const dir, const SearchData * const data) |
312 |
{ |
{ |
313 |
int t, xc, yc; |
int xc, yc; |
314 |
const uint8_t * Reference; |
const uint8_t * Reference; |
315 |
VECTOR * current; |
VECTOR * current; |
316 |
|
int32_t sad; uint32_t t; |
317 |
|
|
318 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if (( x > data->max_dx) || ( x < data->min_dx) |
319 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
320 |
|
|
321 |
if (data->qpel_precision) { // x and y are in 1/4 precision |
if (!data->qpel_precision) { |
322 |
|
Reference = GetReference(x, y, data); |
323 |
|
current = data->currentMV; |
324 |
|
xc = x; yc = y; |
325 |
|
} else { // x and y are in 1/4 precision |
326 |
Reference = Interpolate16x16qpel(x, y, 0, data); |
Reference = Interpolate16x16qpel(x, y, 0, data); |
|
t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
|
327 |
xc = x/2; yc = y/2; //for chroma sad |
xc = x/2; yc = y/2; //for chroma sad |
328 |
current = data->currentQMV; |
current = data->currentQMV; |
|
} else { |
|
|
switch ( ((x&1)<<1) + (y&1) ) { |
|
|
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
|
|
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
|
|
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
|
|
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
|
|
} |
|
|
if (data->qpel) t = d_mv_bits(2*x - data->predMV.x, 2*y - data->predMV.y, data->iFcode); |
|
|
else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
|
|
current = data->currentMV; |
|
|
xc = x; yc = y; |
|
329 |
} |
} |
330 |
|
|
331 |
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
332 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
333 |
|
|
334 |
data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; |
sad += (data->lambda16 * t * sad)>>10; |
335 |
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; |
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; |
336 |
|
|
337 |
if (data->chroma) data->temp[0] += ChromaSAD(xc, yc, data); |
if (data->chroma) sad += ChromaSAD( (xc >> 1) + roundtab_79[xc & 0x3], |
338 |
|
(yc >> 1) + roundtab_79[yc & 0x3], data); |
339 |
|
|
340 |
if (data->temp[0] < data->iMinSAD[0]) { |
if (sad < data->iMinSAD[0]) { |
341 |
data->iMinSAD[0] = data->temp[0]; |
data->iMinSAD[0] = sad; |
342 |
current[0].x = x; current[0].y = y; |
current[0].x = x; current[0].y = y; |
343 |
*dir = Direction; } |
*dir = Direction; |
344 |
|
} |
345 |
|
|
346 |
if (data->temp[1] < data->iMinSAD[1]) { |
if (data->temp[1] < data->iMinSAD[1]) { |
347 |
data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y= y; } |
data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y= y; } |
351 |
data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; } |
data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; } |
352 |
if (data->temp[4] < data->iMinSAD[4]) { |
if (data->temp[4] < data->iMinSAD[4]) { |
353 |
data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; } |
data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; } |
354 |
|
} |
355 |
|
|
356 |
|
static void |
357 |
|
CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
358 |
|
{ |
359 |
|
int32_t sad; uint32_t t; |
360 |
|
const uint8_t * Reference; |
361 |
|
VECTOR * current; |
362 |
|
|
363 |
|
if ( (x > data->max_dx) || (x < data->min_dx) |
364 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
365 |
|
|
366 |
|
if (!data->qpel_precision) { |
367 |
|
Reference = GetReference(x, y, data); |
368 |
|
current = data->currentMV; |
369 |
|
} else { // x and y are in 1/4 precision |
370 |
|
Reference = Interpolate8x8qpel(x, y, 0, 0, data); |
371 |
|
current = data->currentQMV; |
372 |
|
} |
373 |
|
|
374 |
|
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
375 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
376 |
|
|
377 |
|
sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10; |
378 |
|
|
379 |
|
if (sad < *(data->iMinSAD)) { |
380 |
|
*(data->iMinSAD) = sad; |
381 |
|
current->x = x; current->y = y; |
382 |
|
*dir = Direction; |
383 |
|
} |
384 |
|
} |
385 |
|
|
386 |
|
static void |
387 |
|
CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
388 |
|
{ |
389 |
|
uint32_t t; |
390 |
|
const uint8_t * Reference; |
391 |
|
|
392 |
|
if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero even value |
393 |
|
(x > data->max_dx) || (x < data->min_dx) |
394 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
395 |
|
|
396 |
|
Reference = GetReference(x, y, data); |
397 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1); |
398 |
|
|
399 |
|
data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
400 |
|
|
401 |
|
data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10; |
402 |
|
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; |
403 |
|
|
404 |
|
if (data->temp[0] < data->iMinSAD[0]) { |
405 |
|
data->iMinSAD[0] = data->temp[0]; |
406 |
|
data->currentMV[0].x = x; data->currentMV[0].y = y; |
407 |
|
*dir = Direction; } |
408 |
|
|
409 |
|
if (data->temp[1] < data->iMinSAD[1]) { |
410 |
|
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
411 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
412 |
|
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
413 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
414 |
|
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
415 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
416 |
|
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
417 |
} |
} |
418 |
|
|
419 |
static void |
static void |
420 |
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 int Direction, int * const dir, const SearchData * const data) |
421 |
{ |
{ |
422 |
int32_t sad; |
int32_t sad, xc, yc; |
423 |
const uint8_t * Reference; |
const uint8_t * Reference; |
424 |
int t; |
uint32_t t; |
425 |
VECTOR * current; |
VECTOR * current; |
426 |
|
|
427 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if (( x > data->max_dx) || ( x < data->min_dx) |
428 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
429 |
|
|
430 |
|
if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; //non-zero even value |
431 |
|
|
432 |
if (data->qpel_precision) { // x and y are in 1/4 precision |
if (data->qpel_precision) { // x and y are in 1/4 precision |
433 |
Reference = Interpolate16x16qpel(x, y, 0, data); |
Reference = Interpolate16x16qpel(x, y, 0, data); |
|
t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
|
434 |
current = data->currentQMV; |
current = data->currentQMV; |
435 |
|
xc = x/2; yc = y/2; |
436 |
} else { |
} else { |
437 |
switch ( ((x&1)<<1) + (y&1) ) { |
Reference = GetReference(x, y, data); |
|
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
|
|
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
|
|
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
|
|
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
|
|
} |
|
|
if (data->qpel) t = d_mv_bits(2*x - data->predMV.x, 2*y - data->predMV.y, data->iFcode); |
|
|
else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
|
438 |
current = data->currentMV; |
current = data->currentMV; |
439 |
|
xc = x; yc = y; |
440 |
} |
} |
441 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, |
442 |
|
data->qpel^data->qpel_precision, data->rrv); |
443 |
|
|
444 |
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
445 |
sad += (data->lambda16 * t * sad)/1000; |
sad += (data->lambda16 * t * sad)>>10; |
446 |
|
|
447 |
|
if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
448 |
|
(yc >> 1) + roundtab_79[yc & 0x3], data); |
449 |
|
|
450 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
451 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
452 |
current->x = x; current->y = y; |
current->x = x; current->y = y; |
453 |
*dir = Direction; } |
*dir = Direction; |
454 |
|
} |
455 |
} |
} |
456 |
|
|
457 |
static void |
static void |
458 |
CheckCandidate16no4vI(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
459 |
{ |
{ |
460 |
// maximum speed - for P/B/I decision |
// maximum speed - for P/B/I decision |
461 |
int32_t sad; |
int32_t sad; |
463 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if (( x > data->max_dx) || ( x < data->min_dx) |
464 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
465 |
|
|
466 |
sad = sad16(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), |
sad = sad32v_c(data->Cur, data->RefP[0] + (x>>1) + (y>>1)*(data->iEdgedWidth), |
467 |
data->iEdgedWidth, 256*4096); |
data->iEdgedWidth, data->temp+1); |
468 |
|
|
469 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
470 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
471 |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
472 |
*dir = Direction; } |
*dir = Direction; |
473 |
} |
} |
474 |
|
if (data->temp[1] < data->iMinSAD[1]) { |
475 |
|
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
476 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
477 |
|
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
478 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
479 |
|
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
480 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
481 |
|
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
482 |
|
|
483 |
|
} |
484 |
|
|
485 |
static void |
static void |
486 |
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 int Direction, int * const dir, const SearchData * const data) |
487 |
{ |
{ |
488 |
int32_t sad; |
int32_t sad, xb, yb, xcf, ycf, xcb, ycb; |
489 |
int xb, yb, t; |
uint32_t t; |
490 |
const uint8_t *ReferenceF, *ReferenceB; |
const uint8_t *ReferenceF, *ReferenceB; |
491 |
VECTOR *current; |
VECTOR *current; |
492 |
|
|
493 |
if (( xf > data->max_dx) || ( xf < data->min_dx) |
if ((xf > data->max_dx) || (xf < data->min_dx) || |
494 |
|| ( yf > data->max_dy) || (yf < data->min_dy)) return; |
(yf > data->max_dy) || (yf < data->min_dy)) |
495 |
|
return; |
496 |
|
|
497 |
if (data->qpel_precision) { |
if (!data->qpel_precision) { |
498 |
|
ReferenceF = GetReference(xf, yf, data); |
499 |
|
xb = data->currentMV[1].x; yb = data->currentMV[1].y; |
500 |
|
ReferenceB = GetReferenceB(xb, yb, 1, data); |
501 |
|
current = data->currentMV; |
502 |
|
xcf = xf; ycf = yf; |
503 |
|
xcb = xb; ycb = yb; |
504 |
|
} else { |
505 |
ReferenceF = Interpolate16x16qpel(xf, yf, 0, data); |
ReferenceF = Interpolate16x16qpel(xf, yf, 0, data); |
506 |
xb = data->currentQMV[1].x; yb = data->currentQMV[1].y; |
xb = data->currentQMV[1].x; yb = data->currentQMV[1].y; |
507 |
current = data->currentQMV; |
current = data->currentQMV; |
508 |
ReferenceB = Interpolate16x16qpel(xb, yb, 1, data); |
ReferenceB = Interpolate16x16qpel(xb, yb, 1, data); |
509 |
t = d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) |
xcf = xf/2; ycf = yf/2; |
510 |
+ d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode); |
xcb = xb/2; ycb = yb/2; |
|
} else { |
|
|
ReferenceF = Interpolate16x16qpel(2*xf, 2*yf, 0, data); |
|
|
xb = data->currentMV[1].x; yb = data->currentMV[1].y; |
|
|
ReferenceB = Interpolate16x16qpel(2*xb, 2*yb, 1, data); |
|
|
current = data->currentMV; |
|
|
if (data->qpel) |
|
|
t = d_mv_bits(2*xf - data->predMV.x, 2*yf - data->predMV.y, data->iFcode) |
|
|
+ d_mv_bits(2*xb - data->bpredMV.x, 2*yb - data->bpredMV.y, data->iFcode); |
|
|
else |
|
|
t = d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) |
|
|
+ d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode); |
|
511 |
} |
} |
512 |
|
|
513 |
|
t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0) |
514 |
|
+ d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
515 |
|
|
516 |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
517 |
sad += (data->lambda16 * t * sad)/1000; |
sad += (data->lambda16 * t * sad)>>10; |
518 |
|
|
519 |
|
if (data->chroma) sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3], |
520 |
|
(ycf >> 1) + roundtab_79[ycf & 0x3], |
521 |
|
(xcb >> 1) + roundtab_79[xcb & 0x3], |
522 |
|
(ycb >> 1) + roundtab_79[ycb & 0x3], data); |
523 |
|
|
524 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
525 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
526 |
current->x = xf; current->y = yf; |
current->x = xf; current->y = yf; |
527 |
*dir = Direction; } |
*dir = Direction; |
528 |
|
} |
529 |
} |
} |
530 |
|
|
531 |
static void |
static void |
532 |
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 int Direction, int * const dir, const SearchData * const data) |
533 |
{ |
{ |
534 |
int32_t sad = 0; |
int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0; |
535 |
int k; |
uint32_t k; |
536 |
const uint8_t *ReferenceF; |
const uint8_t *ReferenceF; |
537 |
const uint8_t *ReferenceB; |
const uint8_t *ReferenceB; |
538 |
VECTOR mvs, b_mvs; |
VECTOR mvs, b_mvs; |
550 |
data->directmvB[k].y |
data->directmvB[k].y |
551 |
: mvs.y - data->referencemv[k].y); |
: mvs.y - data->referencemv[k].y); |
552 |
|
|
553 |
if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
if ((mvs.x > data->max_dx) || (mvs.x < data->min_dx) || |
554 |
|| ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) |
(mvs.y > data->max_dy) || (mvs.y < data->min_dy) || |
555 |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
(b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx) || |
556 |
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
(b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) ) |
557 |
|
return; |
558 |
if (!data->qpel) { |
|
559 |
mvs.x *= 2; mvs.y *= 2; |
if (data->qpel) { |
560 |
b_mvs.x *= 2; b_mvs.y *= 2; //we move to qpel precision anyway |
xcf += mvs.x/2; ycf += mvs.y/2; |
561 |
|
xcb += b_mvs.x/2; ycb += b_mvs.y/2; |
562 |
|
} else { |
563 |
|
xcf += mvs.x; ycf += mvs.y; |
564 |
|
xcb += b_mvs.x; ycb += b_mvs.y; |
565 |
|
mvs.x *= 2; mvs.y *= 2; //we move to qpel precision anyway |
566 |
|
b_mvs.x *= 2; b_mvs.y *= 2; |
567 |
} |
} |
568 |
|
|
569 |
ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); |
ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); |
570 |
ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); |
ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); |
571 |
|
|
572 |
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
573 |
ReferenceF, ReferenceB, |
ReferenceF, ReferenceB, data->iEdgedWidth); |
|
data->iEdgedWidth); |
|
574 |
if (sad > *(data->iMinSAD)) return; |
if (sad > *(data->iMinSAD)) return; |
575 |
} |
} |
576 |
|
|
577 |
sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000; |
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
578 |
|
|
579 |
|
if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
580 |
|
(ycf >> 3) + roundtab_76[ycf & 0xf], |
581 |
|
(xcb >> 3) + roundtab_76[xcb & 0xf], |
582 |
|
(ycb >> 3) + roundtab_76[ycb & 0xf], data); |
583 |
|
|
584 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
585 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
586 |
data->currentMV->x = x; data->currentMV->y = y; |
data->currentMV->x = x; data->currentMV->y = y; |
587 |
*dir = Direction; } |
*dir = Direction; |
588 |
|
} |
589 |
} |
} |
590 |
|
|
591 |
static void |
static void |
592 |
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 int Direction, int * const dir, const SearchData * const data) |
593 |
{ |
{ |
594 |
int32_t sad; |
int32_t sad, xcf, ycf, xcb, ycb; |
595 |
const uint8_t *ReferenceF; |
const uint8_t *ReferenceF; |
596 |
const uint8_t *ReferenceB; |
const uint8_t *ReferenceB; |
597 |
VECTOR mvs, b_mvs; |
VECTOR mvs, b_mvs; |
613 |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
614 |
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
615 |
|
|
616 |
if (!data->qpel) { |
if (data->qpel) { |
617 |
mvs.x *= 2; mvs.y *= 2; |
xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2); |
618 |
b_mvs.x *= 2; b_mvs.y *= 2; //we move to qpel precision anyway |
xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2); |
|
} |
|
619 |
ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data); |
ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data); |
620 |
ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); |
ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); |
621 |
|
} else { |
622 |
|
xcf = 4*mvs.x; ycf = 4*mvs.y; |
623 |
|
xcb = 4*b_mvs.x; ycb = 4*b_mvs.y; |
624 |
|
ReferenceF = GetReference(mvs.x, mvs.y, data); |
625 |
|
ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data); |
626 |
|
} |
627 |
|
|
628 |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
629 |
sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000; |
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
630 |
|
|
631 |
|
if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
632 |
|
(ycf >> 3) + roundtab_76[ycf & 0xf], |
633 |
|
(xcb >> 3) + roundtab_76[xcb & 0xf], |
634 |
|
(ycb >> 3) + roundtab_76[ycb & 0xf], data); |
635 |
|
|
636 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
637 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
638 |
data->currentMV->x = x; data->currentMV->y = y; |
data->currentMV->x = x; data->currentMV->y = y; |
639 |
*dir = Direction; } |
*dir = Direction; |
640 |
} |
} |
641 |
|
} |
642 |
|
|
643 |
|
|
644 |
static void |
static void |
645 |
CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidateBits16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
646 |
{ |
{ |
647 |
int32_t sad; int t; |
|
648 |
const uint8_t * Reference; |
int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; |
649 |
|
int32_t bits = 0; |
650 |
|
VECTOR * current; |
651 |
|
const uint8_t * ptr; |
652 |
|
int i, cbp = 0, t, xc, yc; |
653 |
|
|
654 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if (( x > data->max_dx) || ( x < data->min_dx) |
655 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
656 |
|
|
657 |
if (data->qpel) Reference = Interpolate16x16qpel(x, y, 0, data); |
if (!data->qpel_precision) { |
658 |
else Reference = Interpolate16x16qpel(2*x, 2*y, 0, data); |
ptr = GetReference(x, y, data); |
659 |
|
current = data->currentMV; |
660 |
|
xc = x; yc = y; |
661 |
|
} else { // x and y are in 1/4 precision |
662 |
|
ptr = Interpolate16x16qpel(x, y, 0, data); |
663 |
|
current = data->currentQMV; |
664 |
|
xc = x/2; yc = y/2; |
665 |
|
} |
666 |
|
|
667 |
|
for(i = 0; i < 4; i++) { |
668 |
|
int s = 8*((i&1) + (i>>1)*data->iEdgedWidth); |
669 |
|
transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth); |
670 |
|
bits += data->temp[i] = Block_CalcBits(coeff, in, data->iQuant, data->quant_type, &cbp, i, 0); |
671 |
|
} |
672 |
|
|
673 |
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
bits += t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
|
if (data->qpel) t = d_mv_bits(2 * x - data->predMV.x, 2 * y - data->predMV.y, data->iFcode); |
|
|
else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
|
674 |
|
|
675 |
sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))/100; |
bits += xvid_cbpy_tab[15-(cbp>>2)].len; |
676 |
|
|
677 |
if (sad < *(data->iMinSAD)) { |
if (bits >= data->iMinSAD[0]) return; |
678 |
*(data->iMinSAD) = sad; |
|
679 |
data->currentMV->x = x; data->currentMV->y = y; |
//chroma |
680 |
*dir = Direction; } |
xc = (xc >> 1) + roundtab_79[xc & 0x3]; |
681 |
|
yc = (yc >> 1) + roundtab_79[yc & 0x3]; |
682 |
|
|
683 |
|
//chroma U |
684 |
|
ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefP[4], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); |
685 |
|
transfer_8to16subro(in, ptr, data->CurU, data->iEdgedWidth/2); |
686 |
|
bits += Block_CalcBits(coeff, in, data->iQuant, data->quant_type, &cbp, 4, 0); |
687 |
|
if (bits >= data->iMinSAD[0]) return; |
688 |
|
|
689 |
|
//chroma V |
690 |
|
ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefP[5], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); |
691 |
|
transfer_8to16subro(in, ptr, data->CurV, data->iEdgedWidth/2); |
692 |
|
bits += Block_CalcBits(coeff, in, data->iQuant, data->quant_type, &cbp, 5, 0); |
693 |
|
|
694 |
|
bits += mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; |
695 |
|
|
696 |
|
if (bits < data->iMinSAD[0]) { |
697 |
|
data->iMinSAD[0] = bits; |
698 |
|
current[0].x = x; current[0].y = y; |
699 |
|
*dir = Direction; |
700 |
|
} |
701 |
|
|
702 |
|
if (data->temp[0] + t < data->iMinSAD[1]) { |
703 |
|
data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; } |
704 |
|
if (data->temp[1] < data->iMinSAD[2]) { |
705 |
|
data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; } |
706 |
|
if (data->temp[2] < data->iMinSAD[3]) { |
707 |
|
data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; } |
708 |
|
if (data->temp[3] < data->iMinSAD[4]) { |
709 |
|
data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; } |
710 |
|
|
711 |
|
} |
712 |
|
static void |
713 |
|
CheckCandidateBits8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
714 |
|
{ |
715 |
|
|
716 |
|
int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; |
717 |
|
int32_t bits; |
718 |
|
VECTOR * current; |
719 |
|
const uint8_t * ptr; |
720 |
|
int cbp = 0; |
721 |
|
|
722 |
|
if ( (x > data->max_dx) || (x < data->min_dx) |
723 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
724 |
|
|
725 |
|
if (!data->qpel_precision) { |
726 |
|
ptr = GetReference(x, y, data); |
727 |
|
current = data->currentMV; |
728 |
|
} else { // x and y are in 1/4 precision |
729 |
|
ptr = Interpolate8x8qpel(x, y, 0, 0, data); |
730 |
|
current = data->currentQMV; |
731 |
|
} |
732 |
|
|
733 |
|
transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth); |
734 |
|
bits = Block_CalcBits(coeff, in, data->iQuant, data->quant_type, &cbp, 5, 0); |
735 |
|
bits += d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
736 |
|
|
737 |
|
if (bits < data->iMinSAD[0]) { |
738 |
|
data->temp[0] = cbp; |
739 |
|
data->iMinSAD[0] = bits; |
740 |
|
current[0].x = x; current[0].y = y; |
741 |
|
*dir = Direction; |
742 |
|
} |
743 |
} |
} |
744 |
|
|
745 |
/* CHECK_CANDIATE FUNCTIONS END */ |
/* CHECK_CANDIATE FUNCTIONS END */ |
754 |
|
|
755 |
int iDirection; |
int iDirection; |
756 |
|
|
757 |
do { |
for(;;) { //forever |
758 |
iDirection = 0; |
iDirection = 0; |
759 |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
760 |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
763 |
|
|
764 |
/* now we're doing diagonal checks near our candidate */ |
/* now we're doing diagonal checks near our candidate */ |
765 |
|
|
766 |
if (iDirection) { //checking if anything found |
if (iDirection) { //if anything found |
767 |
bDirection = iDirection; |
bDirection = iDirection; |
768 |
iDirection = 0; |
iDirection = 0; |
769 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
772 |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
773 |
} else { // what remains here is up or down |
} else { // what remains here is up or down |
774 |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
775 |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); } |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
776 |
|
} |
777 |
|
|
778 |
if (iDirection) { |
if (iDirection) { |
779 |
bDirection += iDirection; |
bDirection += iDirection; |
780 |
x = data->currentMV->x; y = data->currentMV->y; } |
x = data->currentMV->x; y = data->currentMV->y; |
781 |
|
} |
782 |
} else { //about to quit, eh? not so fast.... |
} else { //about to quit, eh? not so fast.... |
783 |
switch (bDirection) { |
switch (bDirection) { |
784 |
case 2: |
case 2: |
829 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
830 |
} |
} |
831 |
} |
} |
|
while (1); //forever |
|
832 |
} |
} |
833 |
|
|
834 |
static void |
static void |
878 |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
879 |
} else { // what remains here is up or down |
} else { // what remains here is up or down |
880 |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
881 |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); } |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
882 |
|
} |
883 |
bDirection += iDirection; |
bDirection += iDirection; |
884 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
885 |
} |
} |
889 |
|
|
890 |
/* MAINSEARCH FUNCTIONS END */ |
/* MAINSEARCH FUNCTIONS END */ |
891 |
|
|
|
/* HALFPELREFINE COULD BE A MAINSEARCH FUNCTION, BUT THERE IS NO NEED FOR IT */ |
|
|
|
|
892 |
static void |
static void |
893 |
SubpelRefine(const SearchData * const data) |
SubpelRefine(const SearchData * const data) |
894 |
{ |
{ |
895 |
/* Do a half-pel or q-pel refinement */ |
/* Do a half-pel or q-pel refinement */ |
896 |
VECTOR backupMV; |
const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV; |
897 |
int iDirection; //not needed |
int iDirection; //only needed because macro expects it |
|
|
|
|
if (data->qpel_precision) |
|
|
backupMV = *(data->currentQMV); |
|
|
else backupMV = *(data->currentMV); |
|
|
|
|
|
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); |
|
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); |
|
|
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); |
|
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0); |
|
898 |
|
|
899 |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); |
CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0); |
900 |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); |
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0); |
901 |
|
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0); |
902 |
CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); |
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0); |
903 |
CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); |
CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0); |
904 |
|
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0); |
905 |
|
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0); |
906 |
|
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0); |
907 |
} |
} |
908 |
|
|
909 |
static __inline int |
static __inline int |
910 |
SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
911 |
const int x, const int y, |
const int x, const int y, |
912 |
const uint32_t iEdgedWidth, const uint32_t iQuant) |
const uint32_t stride, const uint32_t iQuant, int rrv) |
913 |
|
|
914 |
{ |
{ |
915 |
/* keep repeating checks for all b-frames before this P frame, |
int offset = (x + y*stride)*8; |
916 |
to make sure that SKIP is possible (todo) |
if(!rrv) { |
917 |
how: if skip is not possible set sad00 to a very high value */ |
uint32_t sadC = sad8(current->u + offset, |
918 |
|
reference->u + offset, stride); |
|
uint32_t sadC = sad8(current->u + x*8 + y*(iEdgedWidth/2)*8, |
|
|
reference->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); |
|
919 |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
920 |
sadC += sad8(current->v + (x + y*(iEdgedWidth/2))*8, |
sadC += sad8(current->v + offset, |
921 |
reference->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); |
reference->v + offset, stride); |
922 |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
923 |
|
return 1; |
924 |
|
|
925 |
|
} else { |
926 |
|
uint32_t sadC = sad16(current->u + 2*offset, |
927 |
|
reference->u + 2*offset, stride, 256*4096); |
928 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
929 |
|
sadC += sad16(current->v + 2*offset, |
930 |
|
reference->v + 2*offset, stride, 256*4096); |
931 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
932 |
return 1; |
return 1; |
933 |
} |
} |
934 |
|
} |
935 |
|
|
936 |
static __inline void |
static __inline void |
937 |
SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
938 |
{ |
{ |
939 |
pMB->mode = MODE_NOT_CODED; |
pMB->mode = MODE_NOT_CODED; |
940 |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV; |
941 |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; |
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV; |
|
|
|
|
pMB->qmvs[0].x = pMB->qmvs[1].x = pMB->qmvs[2].x = pMB->qmvs[3].x = 0; |
|
|
pMB->qmvs[0].y = pMB->qmvs[1].y = pMB->qmvs[2].y = pMB->qmvs[3].y = 0; |
|
|
|
|
942 |
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; |
943 |
} |
} |
944 |
|
|
945 |
bool |
static __inline void |
946 |
MotionEstimation(MBParam * const pParam, |
ModeDecision(SearchData * const Data, |
947 |
FRAMEINFO * const current, |
MACROBLOCK * const pMB, |
948 |
FRAMEINFO * const reference, |
const MACROBLOCK * const pMBs, |
949 |
const IMAGE * const pRefH, |
const int x, const int y, |
950 |
const IMAGE * const pRefV, |
const MBParam * const pParam, |
951 |
const IMAGE * const pRefHV, |
const uint32_t MotionFlags, |
952 |
const uint32_t iLimit) |
const uint32_t GlobalFlags, |
953 |
|
const IMAGE * const pCurrent, |
954 |
|
const IMAGE * const pRef) |
955 |
{ |
{ |
956 |
MACROBLOCK *const pMBs = current->mbs; |
int mode = MODE_INTER; |
957 |
const IMAGE *const pCurrent = ¤t->image; |
int inter4v = (GlobalFlags & XVID_INTER4V) && (pMB->dquant == NO_CHANGE); |
958 |
const IMAGE *const pRef = &reference->image; |
const uint32_t iQuant = pMB->quant; |
959 |
|
|
960 |
const VECTOR zeroMV = { 0, 0 }; |
const int skip_possible = (!(GlobalFlags & XVID_GMC)) && (pMB->dquant == NO_CHANGE); |
961 |
|
|
962 |
uint32_t x, y; |
if (!(GlobalFlags & XVID_MODEDECISION_BITS)) { //normal, fast, SAD-based mode decision |
963 |
uint32_t iIntra = 0; |
int sad; |
964 |
int32_t InterBias, quant = current->quant, sad00; |
int InterBias = MV16_INTER_BIAS; |
965 |
uint8_t *qimage; |
if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
966 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) { |
967 |
// some pre-initialized thingies for SearchP |
mode = MODE_INTER; |
968 |
int32_t temp[5]; |
sad = Data->iMinSAD[0]; |
969 |
VECTOR currentMV[5]; |
} else { |
970 |
VECTOR currentQMV[5]; |
mode = MODE_INTER4V; |
971 |
int32_t iMinSAD[5]; |
sad = Data->iMinSAD[1] + Data->iMinSAD[2] + |
972 |
SearchData Data; |
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant; |
973 |
Data.iEdgedWidth = pParam->edged_width; |
Data->iMinSAD[0] = sad; |
974 |
Data.currentMV = currentMV; |
} |
|
Data.currentQMV = currentQMV; |
|
|
Data.iMinSAD = iMinSAD; |
|
|
Data.temp = temp; |
|
|
Data.iFcode = current->fcode; |
|
|
Data.rounding = pParam->m_rounding_type; |
|
|
Data.qpel = pParam->m_quarterpel; |
|
|
Data.chroma = current->global_flags & XVID_ME_COLOUR; |
|
975 |
|
|
976 |
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
977 |
return 1; // allocate some mem for qpel interpolated blocks |
if (skip_possible && (pMB->sad16 < (int)iQuant * MAX_SAD00_FOR_SKIP)) |
978 |
// somehow this is dirty since I think we shouldn't use malloc outside |
if ( (100*sad)/(pMB->sad16+1) > FINAL_SKIP_THRESH) |
979 |
// encoder_create() - so please fix me! |
if (Data->chroma || SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant, Data->rrv)) { |
980 |
Data.RefQ = qimage; |
mode = MODE_NOT_CODED; |
981 |
if (sadInit) (*sadInit) (); |
sad = 0; |
982 |
|
} |
983 |
|
|
984 |
for (y = 0; y < pParam->mb_height; y++) { |
/* intra decision */ |
|
for (x = 0; x < pParam->mb_width; x++) { |
|
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
|
985 |
|
|
986 |
pMB->sad16 |
if (iQuant > 8) InterBias += 100 * (iQuant - 8); // to make high quants work |
987 |
= sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, |
if (y != 0) |
988 |
pRef->y + (x + y * pParam->edged_width) * 16, |
if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; |
989 |
pParam->edged_width, pMB->sad8 ); |
if (x != 0) |
990 |
|
if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; |
991 |
|
|
992 |
if (Data.chroma) { |
if (Data->chroma) InterBias += 50; // dev8(chroma) ??? |
993 |
pMB->sad16 += sad8(pCurrent->u + x*8 + y*(pParam->edged_width/2)*8, |
if (Data->rrv) InterBias *= 4; |
994 |
pRef->u + x*8 + y*(pParam->edged_width/2)*8, pParam->edged_width/2); |
|
995 |
|
if (InterBias < pMB->sad16) { |
996 |
|
int32_t deviation; |
997 |
|
if (!Data->rrv) deviation = dev16(Data->Cur, Data->iEdgedWidth); |
998 |
|
else deviation = dev16(Data->Cur, Data->iEdgedWidth) + |
999 |
|
dev16(Data->Cur+16, Data->iEdgedWidth) + |
1000 |
|
dev16(Data->Cur + 16*Data->iEdgedWidth, Data->iEdgedWidth) + |
1001 |
|
dev16(Data->Cur+16+16*Data->iEdgedWidth, Data->iEdgedWidth); |
1002 |
|
|
1003 |
pMB->sad16 += sad8(pCurrent->v + (x + y*(pParam->edged_width/2))*8, |
if (deviation < (sad - InterBias)) mode = MODE_INTRA; |
|
pRef->v + (x + y*(pParam->edged_width/2))*8, pParam->edged_width/2); |
|
1004 |
} |
} |
1005 |
|
|
1006 |
sad00 = pMB->sad16; //if no gmc; else sad00 = (..) |
} else { // BITS |
1007 |
|
|
1008 |
if (!(current->global_flags & XVID_LUMIMASKING)) { |
int bits, intra, i; |
1009 |
pMB->dquant = NO_CHANGE; |
VECTOR backup[5], *v; |
1010 |
pMB->quant = current->quant; |
Data->iQuant = iQuant; |
1011 |
|
|
1012 |
|
v = Data->qpel ? Data->currentQMV : Data->currentMV; |
1013 |
|
for (i = 0; i < 5; i++) { |
1014 |
|
Data->iMinSAD[i] = 256*4096; |
1015 |
|
backup[i] = v[i]; |
1016 |
|
} |
1017 |
|
|
1018 |
|
bits = CountMBBitsInter(Data, pMBs, x, y, pParam, MotionFlags); |
1019 |
|
if (bits == 0) |
1020 |
|
mode = MODE_INTER; // quick stop |
1021 |
|
else { |
1022 |
|
if (inter4v) { |
1023 |
|
int bits_inter4v = CountMBBitsInter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup); |
1024 |
|
if (bits_inter4v < bits) { Data->iMinSAD[0] = bits = bits_inter4v; mode = MODE_INTER4V; } |
1025 |
|
} |
1026 |
|
|
1027 |
|
intra = CountMBBitsIntra(Data); |
1028 |
|
|
1029 |
|
if (intra < bits) { *Data->iMinSAD = bits = intra; mode = MODE_INTRA; } |
1030 |
|
} |
1031 |
|
} |
1032 |
|
|
1033 |
|
if (Data->rrv) { |
1034 |
|
Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); |
1035 |
|
Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); |
1036 |
|
} |
1037 |
|
|
1038 |
|
if (mode == MODE_INTER) { |
1039 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
1040 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
1041 |
|
|
1042 |
|
if(Data->qpel) { |
1043 |
|
pMB->qmvs[0] = pMB->qmvs[1] |
1044 |
|
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
1045 |
|
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; |
1046 |
|
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; |
1047 |
} else { |
} else { |
1048 |
|
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
1049 |
|
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
1050 |
|
} |
1051 |
|
|
1052 |
|
} else if (mode == MODE_INTER4V) |
1053 |
|
pMB->sad16 = Data->iMinSAD[0]; |
1054 |
|
else // INTRA, NOT_CODED |
1055 |
|
SkipMacroblockP(pMB, 0); |
1056 |
|
|
1057 |
|
pMB->mode = mode; |
1058 |
|
} |
1059 |
|
|
1060 |
|
bool |
1061 |
|
MotionEstimation(MBParam * const pParam, |
1062 |
|
FRAMEINFO * const current, |
1063 |
|
FRAMEINFO * const reference, |
1064 |
|
const IMAGE * const pRefH, |
1065 |
|
const IMAGE * const pRefV, |
1066 |
|
const IMAGE * const pRefHV, |
1067 |
|
const uint32_t iLimit) |
1068 |
|
{ |
1069 |
|
MACROBLOCK *const pMBs = current->mbs; |
1070 |
|
const IMAGE *const pCurrent = ¤t->image; |
1071 |
|
const IMAGE *const pRef = &reference->image; |
1072 |
|
|
1073 |
|
uint32_t mb_width = pParam->mb_width; |
1074 |
|
uint32_t mb_height = pParam->mb_height; |
1075 |
|
const uint32_t iEdgedWidth = pParam->edged_width; |
1076 |
|
const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->global_flags); |
1077 |
|
|
1078 |
|
uint32_t x, y; |
1079 |
|
uint32_t iIntra = 0; |
1080 |
|
int32_t quant = current->quant, sad00; |
1081 |
|
int skip_thresh = INITIAL_SKIP_THRESH * |
1082 |
|
(current->global_flags & XVID_REDUCED ? 4:1) * |
1083 |
|
(current->global_flags & XVID_MODEDECISION_BITS ? 2:1); |
1084 |
|
|
1085 |
|
// some pre-initialized thingies for SearchP |
1086 |
|
int32_t temp[8]; |
1087 |
|
VECTOR currentMV[5]; |
1088 |
|
VECTOR currentQMV[5]; |
1089 |
|
int32_t iMinSAD[5]; |
1090 |
|
DECLARE_ALIGNED_MATRIX(dct_space, 2, 64, int16_t, CACHE_LINE); |
1091 |
|
SearchData Data; |
1092 |
|
memset(&Data, 0, sizeof(SearchData)); |
1093 |
|
Data.iEdgedWidth = iEdgedWidth; |
1094 |
|
Data.currentMV = currentMV; |
1095 |
|
Data.currentQMV = currentQMV; |
1096 |
|
Data.iMinSAD = iMinSAD; |
1097 |
|
Data.temp = temp; |
1098 |
|
Data.iFcode = current->fcode; |
1099 |
|
Data.rounding = pParam->m_rounding_type; |
1100 |
|
Data.qpel = pParam->m_quarterpel; |
1101 |
|
Data.chroma = MotionFlags & PMV_CHROMA16; |
1102 |
|
Data.rrv = current->global_flags & XVID_REDUCED; |
1103 |
|
Data.dctSpace = dct_space; |
1104 |
|
Data.quant_type = pParam->m_quant_type; |
1105 |
|
|
1106 |
|
if ((current->global_flags & XVID_REDUCED)) { |
1107 |
|
mb_width = (pParam->width + 31) / 32; |
1108 |
|
mb_height = (pParam->height + 31) / 32; |
1109 |
|
Data.qpel = 0; |
1110 |
|
} |
1111 |
|
|
1112 |
|
Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose) |
1113 |
|
if (sadInit) (*sadInit) (); |
1114 |
|
|
1115 |
|
for (y = 0; y < mb_height; y++) { |
1116 |
|
for (x = 0; x < mb_width; x++) { |
1117 |
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
1118 |
|
|
1119 |
|
if (!Data.rrv) pMB->sad16 = |
1120 |
|
sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16, |
1121 |
|
pRef->y + (x + y * iEdgedWidth) * 16, |
1122 |
|
pParam->edged_width, pMB->sad8 ); |
1123 |
|
|
1124 |
|
else pMB->sad16 = |
1125 |
|
sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32, |
1126 |
|
pRef->y + (x + y * iEdgedWidth) * 32, |
1127 |
|
pParam->edged_width, pMB->sad8 ); |
1128 |
|
|
1129 |
|
if (Data.chroma) { |
1130 |
|
Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8, |
1131 |
|
pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2) |
1132 |
|
+ sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8, |
1133 |
|
pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); |
1134 |
|
pMB->sad16 += Data.temp[7]; |
1135 |
|
} |
1136 |
|
|
1137 |
|
sad00 = pMB->sad16; |
1138 |
|
|
1139 |
|
if (!(current->global_flags & XVID_LUMIMASKING)) |
1140 |
|
pMB->dquant = NO_CHANGE; |
1141 |
|
else { |
1142 |
if (pMB->dquant != NO_CHANGE) { |
if (pMB->dquant != NO_CHANGE) { |
1143 |
quant += DQtab[pMB->dquant]; |
quant += DQtab[pMB->dquant]; |
1144 |
if (quant > 31) quant = 31; |
if (quant > 31) quant = 31; |
1145 |
else if (quant < 1) quant = 1; |
else if (quant < 1) quant = 1; |
1146 |
} |
} |
|
pMB->quant = quant; |
|
1147 |
} |
} |
1148 |
|
pMB->quant = quant; |
1149 |
|
|
1150 |
//initial skip decision |
//initial skip decision |
1151 |
/* no early skip for GMC (global vector = skip vector is unknown!) */ |
/* no early skip for GMC (global vector = skip vector is unknown!) */ |
1152 |
if (current->coding_type == P_VOP) { /* no fast SKIP for S(GMC)-VOPs */ |
if (!(current->global_flags & XVID_GMC)) { /* no fast SKIP for S(GMC)-VOPs */ |
1153 |
if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH) |
if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * skip_thresh) |
1154 |
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) { |
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) { |
1155 |
SkipMacroblockP(pMB, sad00); |
SkipMacroblockP(pMB, sad00); |
1156 |
continue; |
continue; |
1157 |
} |
} |
1158 |
} |
} |
1159 |
|
|
1160 |
SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
1161 |
y, current->motion_flags, pMB->quant, |
y, MotionFlags, current->global_flags, |
1162 |
&Data, pParam, pMBs, reference->mbs, |
&Data, pParam, pMBs, reference->mbs, pMB); |
|
current->global_flags & XVID_INTER4V, pMB); |
|
|
|
|
|
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
|
|
if (current->coding_type == P_VOP) { |
|
|
if ( (pMB->dquant == NO_CHANGE) && (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
|
|
&& ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) ) |
|
|
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) { |
|
|
SkipMacroblockP(pMB, sad00); |
|
|
continue; |
|
|
} |
|
|
} |
|
|
|
|
|
/* finally, intra decision */ |
|
|
|
|
|
InterBias = MV16_INTER_BIAS; |
|
|
if (pMB->quant > 8) InterBias += 100 * (pMB->quant - 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; |
|
1163 |
|
|
1164 |
if (Data.chroma) InterBias += 50; // to compensate bigger SAD |
ModeDecision(&Data, pMB, pMBs, x, y, pParam, |
1165 |
|
MotionFlags, current->global_flags, |
1166 |
|
pCurrent, pRef); |
1167 |
|
|
1168 |
if (InterBias < pMB->sad16) { |
if (pMB->mode == MODE_INTRA) |
1169 |
const int32_t deviation = |
if (++iIntra > iLimit) return 1; |
|
dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
|
|
pParam->edged_width); |
|
|
|
|
|
if (deviation < (pMB->sad16 - InterBias)) { |
|
|
if (++iIntra >= iLimit) { free(qimage); return 1; } |
|
|
pMB->mode = MODE_INTRA; |
|
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = |
|
|
pMB->mvs[3] = zeroMV; |
|
|
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = |
|
|
pMB->qmvs[3] = zeroMV; |
|
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = |
|
|
pMB->sad8[3] = 0; |
|
1170 |
} |
} |
1171 |
} |
} |
|
} |
|
|
} |
|
|
free(qimage); |
|
|
|
|
|
if (current->coding_type == S_VOP) /* first GMC step only for S(GMC)-VOPs */ |
|
|
current->GMC_MV = GlobalMotionEst( pMBs, pParam, current->fcode ); |
|
|
else |
|
|
current->GMC_MV = zeroMV; |
|
1172 |
|
|
1173 |
|
if (current->global_flags & XVID_GMC ) /* GMC only for S(GMC)-VOPs */ |
1174 |
|
{ |
1175 |
|
current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV); |
1176 |
|
} |
1177 |
return 0; |
return 0; |
1178 |
} |
} |
1179 |
|
|
1180 |
|
|
|
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
|
|
|
|
1181 |
static __inline int |
static __inline int |
1182 |
make_mask(const VECTOR * const pmv, const int i) |
make_mask(const VECTOR * const pmv, const int i) |
1183 |
{ |
{ |
1185 |
for (j = 0; j < i; j++) { |
for (j = 0; j < i; j++) { |
1186 |
if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already |
if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already |
1187 |
if (pmv[i].x == pmv[j].x) { |
if (pmv[i].x == pmv[j].x) { |
1188 |
if (pmv[i].y == pmv[j].y + iDiamondSize) { mask &= ~4; continue; } |
if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4; |
1189 |
if (pmv[i].y == pmv[j].y - iDiamondSize) { mask &= ~8; continue; } |
else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8; |
1190 |
} else |
} else |
1191 |
if (pmv[i].y == pmv[j].y) { |
if (pmv[i].y == pmv[j].y) { |
1192 |
if (pmv[i].x == pmv[j].x + iDiamondSize) { mask &= ~1; continue; } |
if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1; |
1193 |
if (pmv[i].x == pmv[j].x - iDiamondSize) { mask &= ~2; continue; } |
else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2; |
1194 |
} |
} |
1195 |
} |
} |
1196 |
return mask; |
return mask; |
1197 |
} |
} |
1198 |
|
|
1199 |
static __inline void |
static __inline void |
1200 |
PreparePredictionsP(VECTOR * const pmv, int x, int y, const int iWcount, |
PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, |
1201 |
const int iHcount, const MACROBLOCK * const prevMB) |
int iHcount, const MACROBLOCK * const prevMB, int rrv) |
1202 |
{ |
{ |
1203 |
|
|
1204 |
//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself |
//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself |
1205 |
|
if (rrv) { iWcount /= 2; iHcount /= 2; } |
1206 |
|
|
1207 |
if ( (y != 0) && (x != (iWcount-1)) ) { // [5] top-right neighbour |
if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour |
1208 |
pmv[5].x = EVEN(pmv[3].x); |
pmv[5].x = EVEN(pmv[3].x); |
1209 |
pmv[5].y = EVEN(pmv[3].y); |
pmv[5].y = EVEN(pmv[3].y); |
1210 |
} else pmv[5].x = pmv[5].y = 0; |
} else pmv[5].x = pmv[5].y = 0; |
1223 |
pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
1224 |
pmv[2].y = EVEN(prevMB->mvs[0].y); |
pmv[2].y = EVEN(prevMB->mvs[0].y); |
1225 |
|
|
1226 |
if ((x != iWcount-1) && (y != iHcount-1)) { |
if ((x < iWcount-1) && (y < iHcount-1)) { |
1227 |
pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame |
pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame |
1228 |
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
1229 |
} else pmv[6].x = pmv[6].y = 0; |
} else pmv[6].x = pmv[6].y = 0; |
1230 |
|
|
1231 |
|
if (rrv) { |
1232 |
|
int i; |
1233 |
|
for (i = 0; i < 7; i++) { |
1234 |
|
pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); |
1235 |
|
pmv[i].y = RRV_MV_SCALEUP(pmv[i].y); |
1236 |
|
} |
1237 |
|
} |
1238 |
} |
} |
1239 |
|
|
1240 |
static void |
static void |
1246 |
const int x, |
const int x, |
1247 |
const int y, |
const int y, |
1248 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
1249 |
const uint32_t iQuant, |
const uint32_t GlobalFlags, |
1250 |
SearchData * const Data, |
SearchData * const Data, |
1251 |
const MBParam * const pParam, |
const MBParam * const pParam, |
1252 |
const MACROBLOCK * const pMBs, |
const MACROBLOCK * const pMBs, |
1253 |
const MACROBLOCK * const prevMBs, |
const MACROBLOCK * const prevMBs, |
|
int inter4v, |
|
1254 |
MACROBLOCK * const pMB) |
MACROBLOCK * const pMB) |
1255 |
{ |
{ |
1256 |
|
|
1257 |
int i, iDirection = 255, mask, threshA; |
int i, iDirection = 255, mask, threshA; |
1258 |
VECTOR pmv[7]; |
VECTOR pmv[7]; |
1259 |
|
int inter4v = (GlobalFlags & XVID_INTER4V) && (pMB->dquant == NO_CHANGE); |
1260 |
|
|
|
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); //has to be changed to get_pmv(2)() |
|
1261 |
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, 16, |
1262 |
pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
1263 |
|
|
1264 |
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16; |
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); |
|
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
|
|
Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
|
|
|
|
|
Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
|
|
Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
|
1265 |
|
|
1266 |
Data->lambda16 = lambda_vec16[iQuant]; |
Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache |
1267 |
Data->lambda8 = lambda_vec8[iQuant]; |
i = Data->rrv ? 2 : 1; |
1268 |
Data->qpel_precision = 0; |
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i; |
1269 |
|
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1270 |
if (!(MotionFlags & PMV_HALFPEL16)) { |
Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1271 |
Data->min_dx = EVEN(Data->min_dx); |
|
1272 |
Data->max_dx = EVEN(Data->max_dx); |
Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16*i; |
1273 |
Data->min_dy = EVEN(Data->min_dy); |
Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16*i; |
1274 |
Data->max_dy = EVEN(Data->max_dy); } |
Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16*i; |
1275 |
|
Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16*i; |
1276 |
|
Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1277 |
|
Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1278 |
|
|
1279 |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
Data->lambda16 = lambda_vec16[pMB->quant]; |
1280 |
|
Data->lambda8 = lambda_vec8[pMB->quant]; |
1281 |
|
Data->qpel_precision = 0; |
1282 |
|
|
1283 |
for(i = 0; i < 5; i++) |
memset(Data->currentMV, 0, 5*sizeof(VECTOR)); |
|
Data->currentMV[i].x = Data->currentMV[i].y = 0; |
|
1284 |
|
|
1285 |
if (pParam->m_quarterpel) 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); |
1286 |
else Data->predMV = pmv[0]; |
else Data->predMV = pmv[0]; |
1287 |
|
|
1288 |
i = d_mv_bits(Data->predMV.x, Data->predMV.y, Data->iFcode); |
i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0); |
1289 |
Data->iMinSAD[0] = pMB->sad16 + (Data->lambda16 * i * pMB->sad16)/1000; |
Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10); |
1290 |
Data->iMinSAD[1] = pMB->sad8[0] + (Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS))/100; |
Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10); |
1291 |
Data->iMinSAD[2] = pMB->sad8[1]; |
Data->iMinSAD[2] = pMB->sad8[1]; |
1292 |
Data->iMinSAD[3] = pMB->sad8[2]; |
Data->iMinSAD[3] = pMB->sad8[2]; |
1293 |
Data->iMinSAD[4] = pMB->sad8[3]; |
Data->iMinSAD[4] = pMB->sad8[3]; |
1294 |
|
|
1295 |
if ((x == 0) && (y == 0)) threshA = 512; |
if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) || (x | y)) { |
1296 |
else { |
threshA = Data->temp[0]; // that's where we keep this SAD atm |
|
threshA = Data->temp[0]; // that's when we keep this SAD atm |
|
1297 |
if (threshA < 512) threshA = 512; |
if (threshA < 512) threshA = 512; |
1298 |
if (threshA > 1024) threshA = 1024; } |
else if (threshA > 1024) threshA = 1024; |
1299 |
|
} else |
1300 |
|
threshA = 512; |
1301 |
|
|
1302 |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
1303 |
prevMBs + x + y * pParam->mb_width); |
prevMBs + x + y * pParam->mb_width, Data->rrv); |
1304 |
|
|
1305 |
if (inter4v || Data->chroma) CheckCandidate = CheckCandidate16; |
if (!Data->rrv) { |
1306 |
|
if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16; |
1307 |
else CheckCandidate = CheckCandidate16no4v; //for extra speed |
else CheckCandidate = CheckCandidate16no4v; //for extra speed |
1308 |
|
} else CheckCandidate = CheckCandidate32; |
1309 |
|
|
1310 |
/* main loop. checking all predictions */ |
/* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/ |
1311 |
|
|
1312 |
for (i = 1; i < 7; i++) { |
for (i = 1; i < 7; i++) { |
1313 |
if (!(mask = make_mask(pmv, i)) ) continue; |
if (!(mask = make_mask(pmv, i)) ) continue; |
1314 |
(*CheckCandidate)(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1315 |
if (Data->iMinSAD[0] <= threshA) break; |
if (Data->iMinSAD[0] <= threshA) break; |
1316 |
} |
} |
1317 |
|
|
1318 |
if ((Data->iMinSAD[0] <= threshA) || |
if ((Data->iMinSAD[0] <= threshA) || |
1319 |
(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
1320 |
(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { |
(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) |
1321 |
inter4v = 0; |
inter4v = 0; |
1322 |
} else { |
else { |
1323 |
|
|
1324 |
MainSearchFunc * MainSearchPtr; |
MainSearchFunc * MainSearchPtr; |
1325 |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1326 |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1327 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1328 |
|
|
1329 |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
1330 |
|
|
1331 |
/* extended search, diamond starting in 0,0 and in prediction. |
/* extended search, diamond starting in 0,0 and in prediction. |
1332 |
note that this search is/might be done in halfpel positions, |
note that this search is/might be done in halfpel positions, |
1335 |
if (MotionFlags & PMV_EXTSEARCH16) { |
if (MotionFlags & PMV_EXTSEARCH16) { |
1336 |
int32_t bSAD; |
int32_t bSAD; |
1337 |
VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; |
VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; |
1338 |
if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? |
if (Data->rrv) { |
1339 |
startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); |
startMV.x = RRV_MV_SCALEUP(startMV.x); |
1340 |
|
startMV.y = RRV_MV_SCALEUP(startMV.y); |
1341 |
|
} |
1342 |
if (!(MVequal(startMV, backupMV))) { |
if (!(MVequal(startMV, backupMV))) { |
1343 |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1344 |
|
|
1345 |
(*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); |
CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data); |
1346 |
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
MainSearchPtr(startMV.x, startMV.y, Data, 255); |
1347 |
if (bSAD < Data->iMinSAD[0]) { |
if (bSAD < Data->iMinSAD[0]) { |
1348 |
Data->currentMV[0] = backupMV; |
Data->currentMV[0] = backupMV; |
1349 |
Data->iMinSAD[0] = bSAD; } |
Data->iMinSAD[0] = bSAD; } |
1350 |
} |
} |
1351 |
|
|
1352 |
backupMV = Data->currentMV[0]; |
backupMV = Data->currentMV[0]; |
1353 |
if (MotionFlags & PMV_HALFPELREFINE16) startMV.x = startMV.y = 1; |
startMV.x = startMV.y = 1; |
|
else startMV.x = startMV.y = 0; |
|
1354 |
if (!(MVequal(startMV, backupMV))) { |
if (!(MVequal(startMV, backupMV))) { |
1355 |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1356 |
|
|
1357 |
(*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); |
CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data); |
1358 |
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
MainSearchPtr(startMV.x, startMV.y, Data, 255); |
1359 |
if (bSAD < Data->iMinSAD[0]) { |
if (bSAD < Data->iMinSAD[0]) { |
1360 |
Data->currentMV[0] = backupMV; |
Data->currentMV[0] = backupMV; |
1361 |
Data->iMinSAD[0] = bSAD; } |
Data->iMinSAD[0] = bSAD; } |
1363 |
} |
} |
1364 |
} |
} |
1365 |
|
|
1366 |
if (MotionFlags & PMV_HALFPELREFINE16) SubpelRefine(Data); |
if (MotionFlags & PMV_HALFPELREFINE16) |
1367 |
|
SubpelRefine(Data); |
1368 |
|
|
1369 |
for(i = 0; i < 5; i++) { |
for(i = 0; i < 5; i++) { |
1370 |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors |
1371 |
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
1372 |
} |
} |
1373 |
|
|
1374 |
if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { |
if (MotionFlags & PMV_QUARTERPELREFINE16) { |
1375 |
|
|
|
Data->qpel_precision = 1; |
|
1376 |
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, 16, |
1377 |
pParam->width, pParam->height, Data->iFcode, 0); |
pParam->width, pParam->height, Data->iFcode, 1, 0); |
1378 |
|
Data->qpel_precision = 1; |
1379 |
SubpelRefine(Data); |
SubpelRefine(Data); |
1380 |
} |
} |
1381 |
|
|
1382 |
if (Data->iMinSAD[0] < (int32_t)iQuant * 30 ) inter4v = 0; |
if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) && (Data->iMinSAD[0] < (int32_t)pMB->quant * 30)) |
1383 |
|
inter4v = 0; |
1384 |
|
|
1385 |
if (inter4v) { |
if (inter4v) { |
1386 |
SearchData Data8; |
SearchData Data8; |
1387 |
Data8.iFcode = Data->iFcode; |
memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data |
1388 |
Data8.lambda8 = Data->lambda8; |
|
|
Data8.iEdgedWidth = Data->iEdgedWidth; |
|
|
Data8.RefQ = Data->RefQ; |
|
|
Data8.qpel = Data->qpel; |
|
1389 |
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
1390 |
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); |
1391 |
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); |
1392 |
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); |
1393 |
|
|
1394 |
if (Data->chroma) { |
if ((Data->chroma) && (!(GlobalFlags & XVID_MODEDECISION_BITS))) { |
1395 |
int sumx, sumy, dx, dy; |
// chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, it will not be used |
1396 |
|
int sumx = 0, sumy = 0; |
1397 |
if(pParam->m_quarterpel) { |
|
1398 |
sumx= pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2; |
if (Data->qpel) |
1399 |
sumy = pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2; |
for (i = 1; i < 5; i++) { |
1400 |
} else { |
sumx += Data->currentQMV[i].x/2; |
1401 |
sumx = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
sumy += Data->currentQMV[i].y/2; |
|
sumy = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
|
|
} |
|
|
dx = (sumx >> 3) + roundtab_76[sumx & 0xf]; |
|
|
dy = (sumy >> 3) + roundtab_76[sumy & 0xf]; |
|
|
|
|
|
Data->iMinSAD[1] += ChromaSAD(dx, dy, Data); |
|
1402 |
} |
} |
1403 |
|
else |
1404 |
|
for (i = 1; i < 5; i++) { |
1405 |
|
sumx += Data->currentMV[i].x; |
1406 |
|
sumy += Data->currentMV[i].y; |
1407 |
} |
} |
1408 |
|
|
1409 |
if (!(inter4v) || |
Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf], |
1410 |
(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
(sumy >> 3) + roundtab_76[sumy & 0xf], Data); |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
|
|
// INTER MODE |
|
|
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(pParam->m_quarterpel) { |
|
|
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 { |
|
|
// INTER4V MODE; all other things are already set in Search8 |
|
|
pMB->mode = MODE_INTER4V; |
|
|
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + |
|
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant; |
|
1411 |
} |
} |
1412 |
|
} else Data->iMinSAD[1] = 4096*256; |
1413 |
} |
} |
1414 |
|
|
1415 |
static void |
static void |
1422 |
const int block, |
const int block, |
1423 |
SearchData * const Data) |
SearchData * const Data) |
1424 |
{ |
{ |
1425 |
|
int i = 0; |
1426 |
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
1427 |
Data->currentMV = OldData->currentMV + 1 + block; |
Data->currentMV = OldData->currentMV + 1 + block; |
1428 |
Data->currentQMV = OldData->currentQMV + 1 + block; |
Data->currentQMV = OldData->currentQMV + 1 + block; |
1429 |
|
|
1430 |
if(pParam->m_quarterpel) { |
if(Data->qpel) { |
1431 |
Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
1432 |
if (block != 0) *(Data->iMinSAD) += (Data->lambda8 * |
if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y, |
1433 |
d_mv_bits( Data->currentQMV->x - Data->predMV.x, |
Data->predMV, Data->iFcode, 0, 0); |
|
Data->currentQMV->y - Data->predMV.y, |
|
|
Data->iFcode) * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100; |
|
1434 |
} else { |
} else { |
1435 |
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
1436 |
if (block != 0) *(Data->iMinSAD) += (Data->lambda8 * |
if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y, |
1437 |
d_mv_bits( Data->currentMV->x - Data->predMV.x, |
Data->predMV, Data->iFcode, 0, Data->rrv); |
|
Data->currentMV->y - Data->predMV.y, |
|
|
Data->iFcode) * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100; |
|
1438 |
} |
} |
1439 |
|
|
1440 |
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { |
*(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10; |
1441 |
|
|
1442 |
Data->Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8|PMV_QUARTERPELREFINE8)) { |
|
Data->RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
|
Data->RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
|
Data->RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
1443 |
|
|
1444 |
Data->Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
if (Data->rrv) i = 16; else i = 8; |
1445 |
|
|
1446 |
|
Data->RefP[0] = OldData->RefP[0] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1447 |
|
Data->RefP[1] = OldData->RefP[1] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1448 |
|
Data->RefP[2] = OldData->RefP[2] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1449 |
|
Data->RefP[3] = OldData->RefP[3] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1450 |
|
|
1451 |
|
Data->Cur = OldData->Cur + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1452 |
Data->qpel_precision = 0; |
Data->qpel_precision = 0; |
1453 |
|
|
1454 |
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, 8, |
1455 |
pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel); |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
1456 |
CheckCandidate = CheckCandidate8; |
|
1457 |
|
if (!Data->rrv) CheckCandidate = CheckCandidate8; |
1458 |
|
else CheckCandidate = CheckCandidate16no4v; |
1459 |
|
|
1460 |
if (MotionFlags & PMV_EXTSEARCH8) { |
if (MotionFlags & PMV_EXTSEARCH8 && (!(MotionFlags & EXTSEARCH_BITS))) { |
1461 |
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
1462 |
|
|
1463 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1465 |
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
1466 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1467 |
|
|
1468 |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1469 |
|
|
1470 |
if(*(Data->iMinSAD) < temp_sad) { |
if(*(Data->iMinSAD) < temp_sad) { |
1471 |
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
1484 |
} |
} |
1485 |
} |
} |
1486 |
|
|
1487 |
if(pParam->m_quarterpel) { |
if (Data->qpel && MotionFlags & PMV_QUARTERPELREFINE8) { |
|
if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && |
|
|
(MotionFlags & PMV_QUARTERPELREFINE8)) { |
|
1488 |
Data->qpel_precision = 1; |
Data->qpel_precision = 1; |
1489 |
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, 8, |
1490 |
pParam->width, pParam->height, OldData->iFcode, 0); |
pParam->width, pParam->height, Data->iFcode, 1, 0); |
1491 |
SubpelRefine(Data); |
SubpelRefine(Data); |
1492 |
} |
} |
1493 |
} |
} |
1494 |
|
|
1495 |
|
if (Data->rrv) { |
1496 |
|
Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x); |
1497 |
|
Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y); |
1498 |
} |
} |
1499 |
|
|
1500 |
if(pParam->m_quarterpel) { |
if(Data->qpel) { |
1501 |
pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x; |
pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x; |
1502 |
pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y; |
pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y; |
1503 |
pMB->qmvs[block] = *(Data->currentQMV); |
pMB->qmvs[block] = *Data->currentQMV; |
1504 |
} |
} else { |
|
else { |
|
1505 |
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
1506 |
pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; |
pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; |
1507 |
} |
} |
1508 |
|
|
1509 |
pMB->mvs[block] = *(Data->currentMV); |
pMB->mvs[block] = *Data->currentMV; |
1510 |
pMB->sad8[block] = 4 * (*Data->iMinSAD); |
pMB->sad8[block] = 4 * *Data->iMinSAD; |
1511 |
} |
} |
1512 |
|
|
1513 |
/* B-frames code starts here */ |
/* motion estimation for B-frames */ |
1514 |
|
|
1515 |
static __inline VECTOR |
static __inline VECTOR |
1516 |
ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) |
ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) |
1517 |
{ |
{ |
1518 |
/* the stupidiest function ever */ |
/* the stupidiest function ever */ |
1519 |
if (mode == MODE_FORWARD) return pMB->mvs[0]; |
return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]); |
|
else return pMB->b_mvs[0]; |
|
1520 |
} |
} |
1521 |
|
|
1522 |
static void __inline |
static void __inline |
1549 |
pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
1550 |
} else pmv[5].x = pmv[5].y = 0; |
} else pmv[5].x = pmv[5].y = 0; |
1551 |
|
|
1552 |
if ((x != 0)&&(y != 0)) { |
if (x != 0 && y != 0) { |
1553 |
pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); |
pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); |
1554 |
pmv[6].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y); |
1555 |
} else pmv[6].x = pmv[6].y = 0; |
} else pmv[6].x = pmv[6].y = 0; |
|
|
|
|
// more? |
|
1556 |
} |
} |
1557 |
|
|
1558 |
|
|
1559 |
/* search backward or forward, for b-frames */ |
/* search backward or forward */ |
1560 |
static void |
static void |
1561 |
SearchBF( const uint8_t * const pRef, |
SearchBF( const IMAGE * const pRef, |
1562 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1563 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
1564 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
1574 |
SearchData * const Data) |
SearchData * const Data) |
1575 |
{ |
{ |
1576 |
|
|
1577 |
const int32_t iEdgedWidth = pParam->edged_width; |
int i, iDirection = 255, mask; |
|
|
|
|
int i, iDirection, mask; |
|
1578 |
VECTOR pmv[7]; |
VECTOR pmv[7]; |
1579 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1580 |
*Data->iMinSAD = MV_MAX_ERROR; |
*Data->iMinSAD = MV_MAX_ERROR; |
1581 |
Data->iFcode = iFcode; |
Data->iFcode = iFcode; |
1582 |
Data->qpel_precision = 0; |
Data->qpel_precision = 0; |
1583 |
|
Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; // reset chroma-sad cache |
1584 |
|
|
1585 |
Data->Ref = pRef + (x + y * iEdgedWidth) * 16; |
Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16; |
1586 |
Data->RefH = pRefH + (x + y * iEdgedWidth) * 16; |
Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16; |
1587 |
Data->RefV = pRefV + (x + y * iEdgedWidth) * 16; |
Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16; |
1588 |
Data->RefHV = pRefHV + (x + y * iEdgedWidth) * 16; |
Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16; |
1589 |
|
Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
1590 |
|
Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
1591 |
|
|
1592 |
Data->predMV = *predMV; |
Data->predMV = *predMV; |
1593 |
|
|
1594 |
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, 16, |
1595 |
pParam->width, pParam->height, iFcode, pParam->m_quarterpel); |
pParam->width, pParam->height, iFcode - Data->qpel, 0, 0); |
1596 |
|
|
1597 |
pmv[0] = Data->predMV; |
pmv[0] = Data->predMV; |
1598 |
if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } |
if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } |
1599 |
|
|
1600 |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
1601 |
|
|
1602 |
Data->currentMV->x = Data->currentMV->y = 0; |
Data->currentMV->x = Data->currentMV->y = 0; |
1603 |
CheckCandidate = CheckCandidate16no4v; |
CheckCandidate = CheckCandidate16no4v; |
1604 |
|
|
1605 |
// main loop. checking all predictions |
// main loop. checking all predictions |
1606 |
for (i = 0; i < 8; i++) { |
for (i = 0; i < 7; i++) { |
1607 |
if (!(mask = make_mask(pmv, i)) ) continue; |
if (!(mask = make_mask(pmv, i)) ) continue; |
1608 |
CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1609 |
} |
} |
1610 |
|
|
1611 |
if (MotionFlags & PMV_USESQUARES16) |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1612 |
MainSearchPtr = SquareSearch; |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
|
|
MainSearchPtr = AdvDiamondSearch; |
|
1613 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1614 |
|
|
1615 |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
1616 |
|
|
1617 |
SubpelRefine(Data); |
SubpelRefine(Data); |
1618 |
|
|
1619 |
if (Data->qpel) { |
if (Data->qpel && *Data->iMinSAD < *best_sad + 300) { |
1620 |
Data->currentQMV->x = 2*Data->currentMV->x; |
Data->currentQMV->x = 2*Data->currentMV->x; |
1621 |
Data->currentQMV->y = 2*Data->currentMV->y; |
Data->currentQMV->y = 2*Data->currentMV->y; |
1622 |
Data->qpel_precision = 1; |
Data->qpel_precision = 1; |
1623 |
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, 16, |
1624 |
pParam->width, pParam->height, iFcode, 0); |
pParam->width, pParam->height, iFcode, 1, 0); |
1625 |
SubpelRefine(Data); |
SubpelRefine(Data); |
1626 |
} |
} |
1627 |
|
|
1628 |
// three bits are needed to code backward mode. four for forward |
// three bits are needed to code backward mode. four for forward |
1629 |
// we treat the bits just like they were vector's |
|
1630 |
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
1631 |
else *Data->iMinSAD += 3 * Data->lambda16; |
else *Data->iMinSAD += 3 * Data->lambda16; |
1632 |
|
|
1644 |
pMB->pmvs[0].x = Data->currentMV->x - predMV->x; |
pMB->pmvs[0].x = Data->currentMV->x - predMV->x; |
1645 |
pMB->pmvs[0].y = Data->currentMV->y - predMV->y; |
pMB->pmvs[0].y = Data->currentMV->y - predMV->y; |
1646 |
} |
} |
1647 |
if (mode_current == MODE_FORWARD) |
if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; |
1648 |
pMB->mvs[0] = *(Data->currentMV+2) = *Data->currentMV; |
else pMB->b_mvs[0] = *Data->currentMV; |
1649 |
else |
} |
|
pMB->b_mvs[0] = *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search |
|
1650 |
|
|
1651 |
|
if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; |
1652 |
|
else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search |
1653 |
} |
} |
1654 |
|
|
1655 |
|
static void |
1656 |
|
SkipDecisionB(const IMAGE * const pCur, |
1657 |
|
const IMAGE * const f_Ref, |
1658 |
|
const IMAGE * const b_Ref, |
1659 |
|
MACROBLOCK * const pMB, |
1660 |
|
const uint32_t x, const uint32_t y, |
1661 |
|
const SearchData * const Data) |
1662 |
|
{ |
1663 |
|
int dx = 0, dy = 0, b_dx = 0, b_dy = 0; |
1664 |
|
int32_t sum; |
1665 |
|
const int div = 1 + Data->qpel; |
1666 |
|
int k; |
1667 |
|
const uint32_t stride = Data->iEdgedWidth/2; |
1668 |
|
//this is not full chroma compensation, only it's fullpel approximation. should work though |
1669 |
|
|
1670 |
|
for (k = 0; k < 4; k++) { |
1671 |
|
dy += Data->directmvF[k].y / div; |
1672 |
|
dx += Data->directmvF[k].x / div; |
1673 |
|
b_dy += Data->directmvB[k].y / div; |
1674 |
|
b_dx += Data->directmvB[k].x / div; |
1675 |
|
} |
1676 |
|
|
1677 |
|
dy = (dy >> 3) + roundtab_76[dy & 0xf]; |
1678 |
|
dx = (dx >> 3) + roundtab_76[dx & 0xf]; |
1679 |
|
b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf]; |
1680 |
|
b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf]; |
1681 |
|
|
1682 |
|
sum = sad8bi(pCur->u + 8 * x + 8 * y * stride, |
1683 |
|
f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2, |
1684 |
|
b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, |
1685 |
|
stride); |
1686 |
|
|
1687 |
|
if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; //no skip |
1688 |
|
|
1689 |
|
sum += sad8bi(pCur->v + 8*x + 8 * y * stride, |
1690 |
|
f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2, |
1691 |
|
b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, |
1692 |
|
stride); |
1693 |
|
|
1694 |
|
if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) { |
1695 |
|
pMB->mode = MODE_DIRECT_NONE_MV; //skipped |
1696 |
|
for (k = 0; k < 4; k++) { |
1697 |
|
pMB->qmvs[k] = pMB->mvs[k]; |
1698 |
|
pMB->b_qmvs[k] = pMB->b_mvs[k]; |
1699 |
|
} |
1700 |
|
} |
1701 |
} |
} |
1702 |
|
|
1703 |
static int32_t |
static __inline uint32_t |
1704 |
SearchDirect(const IMAGE * const f_Ref, |
SearchDirect(const IMAGE * const f_Ref, |
1705 |
const uint8_t * const f_RefH, |
const uint8_t * const f_RefH, |
1706 |
const uint8_t * const f_RefV, |
const uint8_t * const f_RefV, |
1721 |
|
|
1722 |
{ |
{ |
1723 |
int32_t skip_sad; |
int32_t skip_sad; |
1724 |
int k; |
int k = (x + Data->iEdgedWidth*y) * 16; |
|
|
|
1725 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1726 |
|
|
1727 |
*Data->iMinSAD = 256*4096; |
*Data->iMinSAD = 256*4096; |
1728 |
|
Data->RefP[0] = f_Ref->y + k; |
1729 |
|
Data->RefP[2] = f_RefH + k; |
1730 |
|
Data->RefP[1] = f_RefV + k; |
1731 |
|
Data->RefP[3] = f_RefHV + k; |
1732 |
|
Data->b_RefP[0] = b_Ref->y + k; |
1733 |
|
Data->b_RefP[2] = b_RefH + k; |
1734 |
|
Data->b_RefP[1] = b_RefV + k; |
1735 |
|
Data->b_RefP[3] = b_RefHV + k; |
1736 |
|
Data->RefP[4] = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; |
1737 |
|
Data->RefP[5] = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; |
1738 |
|
Data->b_RefP[4] = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; |
1739 |
|
Data->b_RefP[5] = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; |
1740 |
|
|
1741 |
|
k = Data->qpel ? 4 : 2; |
1742 |
|
Data->max_dx = k * (pParam->width - x * 16); |
1743 |
|
Data->max_dy = k * (pParam->height - y * 16); |
1744 |
|
Data->min_dx = -k * (16 + x * 16); |
1745 |
|
Data->min_dy = -k * (16 + y * 16); |
1746 |
|
|
1747 |
Data->Ref = f_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs; |
1748 |
Data->RefH = f_RefH + (x + Data->iEdgedWidth*y) * 16; |
Data->qpel_precision = 0; |
|
Data->RefV = f_RefV + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefHV = f_RefHV + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->bRef = b_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->bRefH = b_RefH + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->bRefV = b_RefV + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->bRefHV = b_RefHV + (x + Data->iEdgedWidth*y) * 16; |
|
|
|
|
|
Data->max_dx = 2 * pParam->width - 2 * (x) * 16; |
|
|
Data->max_dy = 2 * pParam->height - 2 * (y) * 16; |
|
|
Data->min_dx = -(2 * 16 + 2 * (x) * 16); |
|
|
Data->min_dy = -(2 * 16 + 2 * (y) * 16); |
|
|
if (Data->qpel) { //we measure in qpixels |
|
|
Data->max_dx *= 2; |
|
|
Data->max_dy *= 2; |
|
|
Data->min_dx *= 2; |
|
|
Data->min_dy *= 2; |
|
|
Data->referencemv = b_mb->qmvs; |
|
|
} else Data->referencemv = b_mb->mvs; |
|
|
Data->qpel_precision = 0; // it'm a trick. it's 1 not 0, but we need 0 here |
|
1749 |
|
|
1750 |
for (k = 0; k < 4; k++) { |
for (k = 0; k < 4; k++) { |
1751 |
pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); |
pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); |
1753 |
pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); |
pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); |
1754 |
pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; |
pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; |
1755 |
|
|
1756 |
if ( ( pMB->b_mvs[k].x > Data->max_dx ) || ( pMB->b_mvs[k].x < Data->min_dx ) |
if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx) |
1757 |
|| ( pMB->b_mvs[k].y > Data->max_dy ) || ( pMB->b_mvs[k].y < Data->min_dy )) { |
| (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) { |
1758 |
|
|
1759 |
*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 |
1760 |
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" |
1761 |
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
1762 |
return 0; |
return 256*4096; |
1763 |
} |
} |
1764 |
if (b_mb->mode != MODE_INTER4V) { |
if (b_mb->mode != MODE_INTER4V) { |
1765 |
pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; |
pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; |
1770 |
} |
} |
1771 |
} |
} |
1772 |
|
|
1773 |
|
CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v; |
1774 |
|
|
1775 |
if (b_mb->mode == MODE_INTER4V) CheckCandidate = CheckCandidateDirect; |
CheckCandidate(0, 0, 255, &k, Data); |
|
else CheckCandidate = CheckCandidateDirectno4v; |
|
1776 |
|
|
1777 |
(*CheckCandidate)(0, 0, 255, &k, Data); |
// initial (fast) skip decision |
1778 |
|
if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (2 + Data->chroma?1:0)) { |
1779 |
// skip decision |
//possible skip |
1780 |
if (*Data->iMinSAD < pMB->quant * SKIP_THRESH_B) { |
if (Data->chroma) { |
|
//possible skip - checking chroma. everything copied from MC |
|
|
//this is not full chroma compensation, only it's fullpel approximation. should work though |
|
|
int sum, dx, dy, b_dx, b_dy; |
|
|
|
|
|
if (Data->qpel) { |
|
|
sum = pMB->mvs[0].y/2 + pMB->mvs[1].y/2 + pMB->mvs[2].y/2 + pMB->mvs[3].y/2; |
|
|
dy = (sum >> 3) + roundtab_76[sum & 0xf]; |
|
|
sum = pMB->mvs[0].x/2 + pMB->mvs[1].x/2 + pMB->mvs[2].x/2 + pMB->mvs[3].x/2; |
|
|
dx = (sum >> 3) + roundtab_76[sum & 0xf]; |
|
|
|
|
|
sum = pMB->b_mvs[0].y/2 + pMB->b_mvs[1].y/2 + pMB->b_mvs[2].y/2 + pMB->b_mvs[3].y/2; |
|
|
b_dy = (sum >> 3) + roundtab_76[sum & 0xf]; |
|
|
sum = pMB->b_mvs[0].x/2 + pMB->b_mvs[1].x/2 + pMB->b_mvs[2].x/2 + pMB->b_mvs[3].x/2; |
|
|
b_dx = (sum >> 3) + roundtab_76[sum & 0xf]; |
|
|
|
|
|
} else { |
|
|
sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
|
|
dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
|
sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
|
|
dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
|
|
|
|
sum = pMB->b_mvs[0].x + pMB->b_mvs[1].x + pMB->b_mvs[2].x + pMB->b_mvs[3].x; |
|
|
b_dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
|
sum = pMB->b_mvs[0].y + pMB->b_mvs[1].y + pMB->b_mvs[2].y + pMB->b_mvs[3].y; |
|
|
b_dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
|
} |
|
|
sum = sad8bi(pCur->u + 8*x + 8*y*(Data->iEdgedWidth/2), |
|
|
f_Ref->u + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
|
|
b_Ref->u + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
|
|
Data->iEdgedWidth/2); |
|
|
sum += sad8bi(pCur->v + 8*x + 8*y*(Data->iEdgedWidth/2), |
|
|
f_Ref->v + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
|
|
b_Ref->v + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
|
|
Data->iEdgedWidth/2); |
|
|
|
|
|
if (sum < MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) { |
|
1781 |
pMB->mode = MODE_DIRECT_NONE_MV; |
pMB->mode = MODE_DIRECT_NONE_MV; |
1782 |
return *Data->iMinSAD; |
return *Data->iMinSAD; // skip. |
1783 |
|
} else { |
1784 |
|
SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data); |
1785 |
|
if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip. |
1786 |
} |
} |
1787 |
} |
} |
1788 |
|
|
1789 |
|
*Data->iMinSAD += Data->lambda16; |
1790 |
skip_sad = *Data->iMinSAD; |
skip_sad = *Data->iMinSAD; |
1791 |
|
|
1792 |
// DIRECT MODE DELTA VECTOR SEARCH. |
// DIRECT MODE DELTA VECTOR SEARCH. |
1796 |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1797 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1798 |
|
|
1799 |
(*MainSearchPtr)(0, 0, Data, 255); |
MainSearchPtr(0, 0, Data, 255); |
1800 |
|
|
1801 |
SubpelRefine(Data); |
SubpelRefine(Data); |
1802 |
|
|
|
*Data->iMinSAD += 1 * Data->lambda16; // one bit is needed to code direct mode |
|
1803 |
*best_sad = *Data->iMinSAD; |
*best_sad = *Data->iMinSAD; |
1804 |
|
|
1805 |
// if (b_mb->mode == MODE_INTER4V) |
if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT; |
1806 |
pMB->mode = MODE_DIRECT; |
else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
|
// else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
|
1807 |
|
|
1808 |
pMB->pmvs[3] = *Data->currentMV; |
pMB->pmvs[3] = *Data->currentMV; |
1809 |
|
|
1834 |
return skip_sad; |
return skip_sad; |
1835 |
} |
} |
1836 |
|
|
1837 |
|
static void |
1838 |
static __inline void |
SearchInterpolate(const IMAGE * const f_Ref, |
|
SearchInterpolate(const uint8_t * const f_Ref, |
|
1839 |
const uint8_t * const f_RefH, |
const uint8_t * const f_RefH, |
1840 |
const uint8_t * const f_RefV, |
const uint8_t * const f_RefV, |
1841 |
const uint8_t * const f_RefHV, |
const uint8_t * const f_RefHV, |
1842 |
const uint8_t * const b_Ref, |
const IMAGE * const b_Ref, |
1843 |
const uint8_t * const b_RefH, |
const uint8_t * const b_RefH, |
1844 |
const uint8_t * const b_RefV, |
const uint8_t * const b_RefV, |
1845 |
const uint8_t * const b_RefHV, |
const uint8_t * const b_RefHV, |
1857 |
|
|
1858 |
{ |
{ |
1859 |
|
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
1860 |
int iDirection, i, j; |
int iDirection, i, j; |
1861 |
SearchData bData; |
SearchData bData; |
1862 |
|
|
1863 |
*(bData.iMinSAD = fData->iMinSAD) = 4096*256; |
fData->qpel_precision = 0; |
1864 |
bData.Cur = fData->Cur; |
memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data |
1865 |
fData->iEdgedWidth = bData.iEdgedWidth = iEdgedWidth; |
*fData->iMinSAD = 4096*256; |
1866 |
bData.currentMV = fData->currentMV + 1; bData.currentQMV = fData->currentQMV + 1; |
bData.currentMV++; bData.currentQMV++; |
|
bData.lambda16 = fData->lambda16; |
|
1867 |
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
1868 |
|
|
1869 |
bData.bRef = fData->Ref = f_Ref + (x + y * iEdgedWidth) * 16; |
i = (x + y * fData->iEdgedWidth) * 16; |
1870 |
bData.bRefH = fData->RefH = f_RefH + (x + y * iEdgedWidth) * 16; |
|
1871 |
bData.bRefV = fData->RefV = f_RefV + (x + y * iEdgedWidth) * 16; |
bData.b_RefP[0] = fData->RefP[0] = f_Ref->y + i; |
1872 |
bData.bRefHV = fData->RefHV = f_RefHV + (x + y * iEdgedWidth) * 16; |
bData.b_RefP[2] = fData->RefP[2] = f_RefH + i; |
1873 |
bData.Ref = fData->bRef = b_Ref + (x + y * iEdgedWidth) * 16; |
bData.b_RefP[1] = fData->RefP[1] = f_RefV + i; |
1874 |
bData.RefH = fData->bRefH = b_RefH + (x + y * iEdgedWidth) * 16; |
bData.b_RefP[3] = fData->RefP[3] = f_RefHV + i; |
1875 |
bData.RefV = fData->bRefV = b_RefV + (x + y * iEdgedWidth) * 16; |
bData.RefP[0] = fData->b_RefP[0] = b_Ref->y + i; |
1876 |
bData.RefHV = fData->bRefHV = b_RefHV + (x + y * iEdgedWidth) * 16; |
bData.RefP[2] = fData->b_RefP[2] = b_RefH + i; |
1877 |
bData.RefQ = fData->RefQ; |
bData.RefP[1] = fData->b_RefP[1] = b_RefV + i; |
1878 |
fData->qpel_precision = bData.qpel_precision = 0; |
bData.RefP[3] = fData->b_RefP[3] = b_RefHV + i; |
1879 |
bData.rounding = 0; |
bData.b_RefP[4] = fData->RefP[4] = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; |
1880 |
|
bData.b_RefP[5] = fData->RefP[5] = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; |
1881 |
|
bData.RefP[4] = fData->b_RefP[4] = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; |
1882 |
|
bData.RefP[5] = fData->b_RefP[5] = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; |
1883 |
|
|
1884 |
bData.bpredMV = fData->predMV = *f_predMV; |
bData.bpredMV = fData->predMV = *f_predMV; |
1885 |
fData->bpredMV = bData.predMV = *b_predMV; |
fData->bpredMV = bData.predMV = *b_predMV; |
|
|
|
1886 |
fData->currentMV[0] = fData->currentMV[2]; |
fData->currentMV[0] = fData->currentMV[2]; |
1887 |
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, pParam->m_quarterpel); |
|
1888 |
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, pParam->m_quarterpel); |
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode - fData->qpel, 0, 0); |
1889 |
|
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode - fData->qpel, 0, 0); |
1890 |
|
|
1891 |
if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; |
if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; |
1892 |
if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dy; |
if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx; |
1893 |
if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dx; |
if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy; |
1894 |
if (fData->currentMV[0].y > fData->min_dy) fData->currentMV[0].y = fData->min_dy; |
if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy; |
1895 |
|
|
1896 |
if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx; |
if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx; |
1897 |
if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dy; |
if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx; |
1898 |
if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dx; |
if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy; |
1899 |
if (fData->currentMV[1].y > bData.min_dy) fData->currentMV[1].y = bData.min_dy; |
if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy; |
1900 |
|
|
1901 |
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
1902 |
|
|
1903 |
//diamond. I wish we could use normal mainsearch functions (square, advdiamond) |
//diamond |
|
|
|
1904 |
do { |
do { |
1905 |
iDirection = 255; |
iDirection = 255; |
1906 |
// forward MV moves |
// forward MV moves |
1921 |
|
|
1922 |
} while (!(iDirection)); |
} while (!(iDirection)); |
1923 |
|
|
1924 |
*fData->iMinSAD += 2 * fData->lambda16; // two bits are needed to code interpolate mode. |
//qpel refinement |
|
|
|
1925 |
if (fData->qpel) { |
if (fData->qpel) { |
1926 |
|
if (*fData->iMinSAD > *best_sad + 500) return; |
1927 |
|
CheckCandidate = CheckCandidateInt; |
1928 |
fData->qpel_precision = bData.qpel_precision = 1; |
fData->qpel_precision = bData.qpel_precision = 1; |
1929 |
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 0); |
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0); |
1930 |
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 0); |
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0); |
1931 |
fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x; |
fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x; |
1932 |
fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y; |
fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y; |
1933 |
fData->currentQMV[1].x = 2 * fData->currentMV[1].x; |
fData->currentQMV[1].x = 2 * fData->currentMV[1].x; |
1934 |
fData->currentQMV[1].y = 2 * fData->currentMV[1].y; |
fData->currentQMV[1].y = 2 * fData->currentMV[1].y; |
1935 |
SubpelRefine(fData); |
SubpelRefine(fData); |
1936 |
|
if (*fData->iMinSAD > *best_sad + 300) return; |
1937 |
fData->currentQMV[2] = fData->currentQMV[0]; |
fData->currentQMV[2] = fData->currentQMV[0]; |
1938 |
SubpelRefine(&bData); |
SubpelRefine(&bData); |
1939 |
} |
} |
1940 |
|
|
1941 |
|
*fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode. |
1942 |
|
|
1943 |
if (*fData->iMinSAD < *best_sad) { |
if (*fData->iMinSAD < *best_sad) { |
1944 |
*best_sad = *fData->iMinSAD; |
*best_sad = *fData->iMinSAD; |
1945 |
pMB->mvs[0] = fData->currentMV[0]; |
pMB->mvs[0] = fData->currentMV[0]; |
1980 |
const IMAGE * const b_refHV) |
const IMAGE * const b_refHV) |
1981 |
{ |
{ |
1982 |
uint32_t i, j; |
uint32_t i, j; |
1983 |
int32_t best_sad, skip_sad; |
int32_t best_sad; |
1984 |
|
uint32_t skip_sad; |
1985 |
int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; |
int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; |
|
static const VECTOR zeroMV={0,0}; |
|
1986 |
const MACROBLOCK * const b_mbs = b_reference->mbs; |
const MACROBLOCK * const b_mbs = b_reference->mbs; |
1987 |
|
|
1988 |
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
1989 |
|
|
1990 |
const int32_t TRB = time_pp - time_bp; |
const int32_t TRB = time_pp - time_bp; |
1991 |
const int32_t TRD = time_pp; |
const int32_t TRD = time_pp; |
|
uint8_t * qimage; |
|
1992 |
|
|
1993 |
// some pre-inintialized data for the rest of the search |
// some pre-inintialized data for the rest of the search |
1994 |
|
|
1996 |
int32_t iMinSAD; |
int32_t iMinSAD; |
1997 |
VECTOR currentMV[3]; |
VECTOR currentMV[3]; |
1998 |
VECTOR currentQMV[3]; |
VECTOR currentQMV[3]; |
1999 |
|
int32_t temp[8]; |
2000 |
|
memset(&Data, 0, sizeof(SearchData)); |
2001 |
Data.iEdgedWidth = pParam->edged_width; |
Data.iEdgedWidth = pParam->edged_width; |
2002 |
Data.currentMV = currentMV; Data.currentQMV = currentQMV; |
Data.currentMV = currentMV; Data.currentQMV = currentQMV; |
2003 |
Data.iMinSAD = &iMinSAD; |
Data.iMinSAD = &iMinSAD; |
2004 |
Data.lambda16 = lambda_vec16[frame->quant]; |
Data.lambda16 = lambda_vec16[frame->quant]; |
2005 |
Data.qpel = pParam->m_quarterpel; |
Data.qpel = pParam->m_quarterpel; |
2006 |
Data.rounding = 0; |
Data.rounding = 0; |
2007 |
|
Data.chroma = frame->motion_flags & PMV_CHROMA8; |
2008 |
|
Data.temp = temp; |
2009 |
|
|
2010 |
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose) |
|
return; // allocate some mem for qpel interpolated blocks |
|
|
// somehow this is dirty since I think we shouldn't use malloc outside |
|
|
// encoder_create() - so please fix me! |
|
|
Data.RefQ = qimage; |
|
|
|
|
2011 |
// note: i==horizontal, j==vertical |
// note: i==horizontal, j==vertical |
2012 |
for (j = 0; j < pParam->mb_height; j++) { |
for (j = 0; j < pParam->mb_height; j++) { |
2013 |
|
|
2025 |
} |
} |
2026 |
|
|
2027 |
Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
2028 |
|
Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8; |
2029 |
|
Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8; |
2030 |
pMB->quant = frame->quant; |
pMB->quant = frame->quant; |
2031 |
|
|
2032 |
/* direct search comes first, because it (1) checks for SKIP-mode |
/* direct search comes first, because it (1) checks for SKIP-mode |
2045 |
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
2046 |
|
|
2047 |
// forward search |
// forward search |
2048 |
SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
2049 |
&frame->image, i, j, |
&frame->image, i, j, |
2050 |
frame->motion_flags, |
frame->motion_flags, |
2051 |
frame->fcode, pParam, |
frame->fcode, pParam, |
2053 |
MODE_FORWARD, &Data); |
MODE_FORWARD, &Data); |
2054 |
|
|
2055 |
// backward search |
// backward search |
2056 |
SearchBF(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y, |
2057 |
&frame->image, i, j, |
&frame->image, i, j, |
2058 |
frame->motion_flags, |
frame->motion_flags, |
2059 |
frame->bcode, pParam, |
frame->bcode, pParam, |
2061 |
MODE_BACKWARD, &Data); |
MODE_BACKWARD, &Data); |
2062 |
|
|
2063 |
// 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 |
2064 |
|
SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
2065 |
SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
|
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
2066 |
&frame->image, |
&frame->image, |
2067 |
i, j, |
i, j, |
2068 |
frame->fcode, frame->bcode, |
frame->fcode, frame->bcode, |
2072 |
pMB, &best_sad, |
pMB, &best_sad, |
2073 |
&Data); |
&Data); |
2074 |
|
|
2075 |
|
// final skip decision |
2076 |
|
if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2) |
2077 |
|
&& ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) |
2078 |
|
SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data); |
2079 |
|
|
2080 |
switch (pMB->mode) { |
switch (pMB->mode) { |
2081 |
case MODE_FORWARD: |
case MODE_FORWARD: |
2082 |
f_count++; |
f_count++; |
2083 |
if (pParam->m_quarterpel) f_predMV = pMB->qmvs[0]; |
f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; |
|
else f_predMV = pMB->mvs[0]; |
|
2084 |
break; |
break; |
2085 |
case MODE_BACKWARD: |
case MODE_BACKWARD: |
2086 |
b_count++; |
b_count++; |
2087 |
if (pParam->m_quarterpel) b_predMV = pMB->b_qmvs[0]; |
b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; |
|
else b_predMV = pMB->b_mvs[0]; |
|
2088 |
break; |
break; |
2089 |
case MODE_INTERPOLATE: |
case MODE_INTERPOLATE: |
2090 |
i_count++; |
i_count++; |
2091 |
if (pParam->m_quarterpel) { |
f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; |
2092 |
f_predMV = pMB->qmvs[0]; |
b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; |
|
b_predMV = pMB->b_qmvs[0]; |
|
|
} else { |
|
|
f_predMV = pMB->mvs[0]; |
|
|
b_predMV = pMB->b_mvs[0]; |
|
|
} |
|
2093 |
break; |
break; |
2094 |
case MODE_DIRECT: |
case MODE_DIRECT: |
2095 |
case MODE_DIRECT_NO4V: |
case MODE_DIRECT_NO4V: |
2096 |
d_count++; |
d_count++; |
|
break; |
|
2097 |
default: |
default: |
2098 |
break; |
break; |
2099 |
} |
} |
2100 |
} |
} |
2101 |
} |
} |
|
free(qimage); |
|
2102 |
} |
} |
2103 |
|
|
2104 |
/* Hinted ME starts here */ |
static __inline void |
2105 |
|
MEanalyzeMB ( const uint8_t * const pRef, |
2106 |
static void |
const uint8_t * const pCur, |
|
SearchPhinted ( const IMAGE * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const IMAGE * const pCur, |
|
2107 |
const int x, |
const int x, |
2108 |
const int y, |
const int y, |
|
const uint32_t MotionFlags, |
|
|
const uint32_t iQuant, |
|
2109 |
const MBParam * const pParam, |
const MBParam * const pParam, |
2110 |
const MACROBLOCK * const pMBs, |
MACROBLOCK * const pMBs, |
|
int inter4v, |
|
|
MACROBLOCK * const pMB, |
|
2111 |
SearchData * const Data) |
SearchData * const Data) |
2112 |
{ |
{ |
2113 |
|
|
2114 |
int i, t; |
int i, mask; |
2115 |
MainSearchFunc * MainSearchPtr; |
VECTOR pmv[3]; |
2116 |
|
MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width]; |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
|
|
pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); |
|
2117 |
|
|
2118 |
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16; |
for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
|
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
|
|
Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
|
|
|
|
|
Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
|
|
Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
|
|
Data->qpel_precision = 0; |
|
2119 |
|
|
2120 |
if (!(MotionFlags & PMV_HALFPEL16)) { |
//median is only used as prediction. it doesn't have to be real |
2121 |
Data->min_dx = EVEN(Data->min_dx); |
if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; |
2122 |
Data->max_dx = EVEN(Data->max_dx); |
else |
2123 |
Data->min_dy = EVEN(Data->min_dy); |
if (x == 1) //left macroblock does not have any vector now |
2124 |
Data->max_dy = EVEN(Data->max_dy); |
Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median |
2125 |
} |
else if (y == 1) // top macroblock doesn't have it's vector |
2126 |
if (pParam->m_quarterpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
Data->predMV = (pMB - 1)->mvs[0]; // left instead of median |
2127 |
else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median |
2128 |
|
|
2129 |
for(i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2130 |
|
pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, 0); |
2131 |
|
|
2132 |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
2133 |
|
Data->RefP[0] = pRef + (x + y * pParam->edged_width) * 16; |
2134 |
|
|
2135 |
if (inter4v || Data->chroma) CheckCandidate = CheckCandidate16; |
pmv[1].x = EVEN(pMB->mvs[0].x); |
2136 |
else CheckCandidate = CheckCandidate16no4v; |
pmv[1].y = EVEN(pMB->mvs[0].y); |
2137 |
|
pmv[2].x = EVEN(Data->predMV.x); |
2138 |
|
pmv[2].y = EVEN(Data->predMV.y); |
2139 |
|
pmv[0].x = pmv[0].y = 0; |
2140 |
|
|
2141 |
pMB->mvs[0].x = EVEN(pMB->mvs[0].x); |
CheckCandidate32I(0, 0, 255, &i, Data); |
|
pMB->mvs[0].y = EVEN(pMB->mvs[0].y); |
|
|
if (pMB->mvs[0].x > Data->max_dx) pMB->mvs[0].x = Data->max_dx; // this is in case iFcode changed |
|
|
if (pMB->mvs[0].x < Data->min_dx) pMB->mvs[0].x = Data->min_dx; |
|
|
if (pMB->mvs[0].y > Data->max_dy) pMB->mvs[0].y = Data->max_dy; |
|
|
if (pMB->mvs[0].y < Data->min_dy) pMB->mvs[0].y = Data->min_dy; |
|
|
|
|
|
(*CheckCandidate)(pMB->mvs[0].x, pMB->mvs[0].y, 0, &t, Data); |
|
|
|
|
|
if (pMB->mode == MODE_INTER4V) |
|
|
for (i = 1; i < 4; i++) { // all four vectors will be used as four predictions for 16x16 search |
|
|
pMB->mvs[i].x = EVEN(pMB->mvs[i].x); |
|
|
pMB->mvs[i].y = EVEN(pMB->mvs[i].y); |
|
|
if (!(make_mask(pMB->mvs, i))) |
|
|
(*CheckCandidate)(pMB->mvs[i].x, pMB->mvs[i].y, 0, &t, Data); |
|
|
} |
|
|
|
|
|
if (MotionFlags & PMV_USESQUARES16) |
|
|
MainSearchPtr = SquareSearch; |
|
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
|
|
MainSearchPtr = AdvDiamondSearch; |
|
|
else MainSearchPtr = DiamondSearch; |
|
2142 |
|
|
2143 |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) { |
2144 |
|
|
2145 |
if (MotionFlags & PMV_HALFPELREFINE16) SubpelRefine(Data); |
if (!(mask = make_mask(pmv, 1))) |
2146 |
|
CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data); |
2147 |
|
if (!(mask = make_mask(pmv, 2))) |
2148 |
|
CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data); |
2149 |
|
|
2150 |
for(i = 0; i < 5; i++) { |
if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) // diamond only if needed |
2151 |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors |
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
|
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
|
2152 |
} |
} |
2153 |
|
|
2154 |
if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { |
for (i = 0; i < 4; i++) { |
2155 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width]; |
2156 |
pParam->width, pParam->height, Data->iFcode, 0); |
MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i]; |
2157 |
Data->qpel_precision = 1; |
MB->mode = MODE_INTER; |
2158 |
SubpelRefine(Data); |
MB->sad16 = Data->iMinSAD[i+1]; |
2159 |
|
} |
2160 |
} |
} |
2161 |
|
|
2162 |
if (inter4v) { |
#define INTRA_THRESH 1800 |
2163 |
SearchData Data8; |
#define INTER_THRESH 1200 |
|
Data8.iFcode = Data->iFcode; |
|
|
Data8.lambda8 = Data->lambda8; |
|
|
Data8.iEdgedWidth = Data->iEdgedWidth; |
|
|
Data8.RefQ = Data->RefQ; |
|
|
Data8.qpel = Data->qpel; |
|
|
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
|
|
Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); |
|
|
Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); |
|
|
Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); |
|
2164 |
|
|
2165 |
if (Data->chroma) { |
int |
2166 |
int sumx, sumy, dx, dy; |
MEanalysis( const IMAGE * const pRef, |
2167 |
|
const FRAMEINFO * const Current, |
2168 |
|
const MBParam * const pParam, |
2169 |
|
const int maxIntra, //maximum number if non-I frames |
2170 |
|
const int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame |
2171 |
|
const int bCount, // number of B frames in a row |
2172 |
|
const int b_thresh) |
2173 |
|
{ |
2174 |
|
uint32_t x, y, intra = 0; |
2175 |
|
int sSAD = 0; |
2176 |
|
MACROBLOCK * const pMBs = Current->mbs; |
2177 |
|
const IMAGE * const pCurrent = &Current->image; |
2178 |
|
int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + 10*b_thresh; |
2179 |
|
int s = 0, blocks = 0; |
2180 |
|
|
2181 |
if(pParam->m_quarterpel) { |
int32_t iMinSAD[5], temp[5]; |
2182 |
sumx= pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2; |
VECTOR currentMV[5]; |
2183 |
sumy = pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2; |
SearchData Data; |
2184 |
} else { |
Data.iEdgedWidth = pParam->edged_width; |
2185 |
sumx = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
Data.currentMV = currentMV; |
2186 |
sumy = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
Data.iMinSAD = iMinSAD; |
2187 |
} |
Data.iFcode = Current->fcode; |
2188 |
dx = (sumx >> 3) + roundtab_76[sumx & 0xf]; |
Data.temp = temp; |
2189 |
dy = (sumy >> 3) + roundtab_76[sumy & 0xf]; |
CheckCandidate = CheckCandidate32I; |
2190 |
|
|
2191 |
Data->iMinSAD[1] += ChromaSAD(dx, dy, Data); |
if (intraCount != 0 && intraCount < 10) // we're right after an I frame |
2192 |
} |
IntraThresh += 15 * (intraCount - 10) * (intraCount - 10); |
2193 |
} |
else |
2194 |
|
if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec |
2195 |
|
IntraThresh -= (IntraThresh * (maxIntra - 8*(maxIntra - intraCount)))/maxIntra; |
2196 |
|
|
2197 |
if (!(inter4v) || |
InterThresh -= (350 - 8*b_thresh) * bCount; |
2198 |
(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] + |
if (InterThresh < 300 + 5*b_thresh) InterThresh = 300 + 5*b_thresh; |
|
Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
|
|
// INTER MODE |
|
|
pMB->mode = MODE_INTER; |
|
|
pMB->mvs[0] = pMB->mvs[1] |
|
|
= pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
|
2199 |
|
|
2200 |
pMB->qmvs[0] = pMB->qmvs[1] |
if (sadInit) (*sadInit) (); |
|
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
|
2201 |
|
|
2202 |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
for (y = 1; y < pParam->mb_height-1; y += 2) { |
2203 |
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
for (x = 1; x < pParam->mb_width-1; x += 2) { |
2204 |
|
int i; |
2205 |
|
blocks += 4; |
2206 |
|
|
2207 |
|
if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; |
2208 |
|
else { //extrapolation of the vector found for last frame |
2209 |
|
pMBs[x + y * pParam->mb_width].mvs[0].x = |
2210 |
|
(pMBs[x + y * pParam->mb_width].mvs[0].x * (bCount+1) ) / bCount; |
2211 |
|
pMBs[x + y * pParam->mb_width].mvs[0].y = |
2212 |
|
(pMBs[x + y * pParam->mb_width].mvs[0].y * (bCount+1) ) / bCount; |
2213 |
|
} |
2214 |
|
|
2215 |
|
MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data); |
2216 |
|
|
2217 |
|
for (i = 0; i < 4; i++) { |
2218 |
|
int dev; |
2219 |
|
MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width]; |
2220 |
|
if (pMB->sad16 > IntraThresh) { |
2221 |
|
dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16, |
2222 |
|
pParam->edged_width); |
2223 |
|
if (dev + IntraThresh < pMB->sad16) { |
2224 |
|
pMB->mode = MODE_INTRA; |
2225 |
|
if (++intra > ((pParam->mb_height-2)*(pParam->mb_width-2))/2) return I_VOP; |
2226 |
|
} |
2227 |
|
} |
2228 |
|
if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0) s++; |
2229 |
|
|
2230 |
if(pParam->m_quarterpel) { |
sSAD += pMB->sad16; |
|
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; |
|
2231 |
} |
} |
|
} else { |
|
|
// INTER4V MODE; all other things are already set in Search8 |
|
|
pMB->mode = MODE_INTER4V; |
|
|
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] |
|
|
+ Data->iMinSAD[4] + IMV16X16 * iQuant; |
|
2232 |
} |
} |
2233 |
|
} |
2234 |
|
|
2235 |
|
sSAD /= blocks; |
2236 |
|
s = (10*s) / blocks; |
2237 |
|
|
2238 |
|
if (s > 4) sSAD += (s - 2) * (60 - 2*b_thresh); //static block - looks bad when in bframe... |
2239 |
|
|
2240 |
|
if (sSAD > InterThresh ) return P_VOP; |
2241 |
|
emms(); |
2242 |
|
return B_VOP; |
2243 |
} |
} |
2244 |
|
|
2245 |
void |
|
2246 |
MotionEstimationHinted( MBParam * const pParam, |
static WARPPOINTS |
2247 |
FRAMEINFO * const current, |
GlobalMotionEst(const MACROBLOCK * const pMBs, |
2248 |
FRAMEINFO * const reference, |
const MBParam * const pParam, |
2249 |
|
const FRAMEINFO * const current, |
2250 |
|
const FRAMEINFO * const reference, |
2251 |
const IMAGE * const pRefH, |
const IMAGE * const pRefH, |
2252 |
const IMAGE * const pRefV, |
const IMAGE * const pRefV, |
2253 |
const IMAGE * const pRefHV) |
const IMAGE * const pRefHV) |
2254 |
{ |
{ |
|
MACROBLOCK *const pMBs = current->mbs; |
|
|
const IMAGE *const pCurrent = ¤t->image; |
|
|
const IMAGE *const pRef = &reference->image; |
|
2255 |
|
|
2256 |
uint32_t x, y; |
const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs |
2257 |
uint8_t * qimage; |
const int deltay=8; |
2258 |
int32_t temp[5], quant = current->quant; |
const int grad=512; // lower bound for deviation in MB |
|
int32_t iMinSAD[5]; |
|
|
VECTOR currentMV[5], currentQMV[5]; |
|
|
SearchData Data; |
|
|
Data.iEdgedWidth = pParam->edged_width; |
|
|
Data.currentMV = currentMV; |
|
|
Data.currentQMV = currentQMV; |
|
|
Data.iMinSAD = iMinSAD; |
|
|
Data.temp = temp; |
|
|
Data.iFcode = current->fcode; |
|
|
Data.rounding = pParam->m_rounding_type; |
|
|
Data.qpel = pParam->m_quarterpel; |
|
|
Data.chroma = current->global_flags & XVID_ME_COLOUR; |
|
|
|
|
|
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
|
|
return; // allocate some mem for qpel interpolated blocks |
|
|
// somehow this is dirty since I think we shouldn't use malloc outside |
|
|
// encoder_create() - so please fix me! |
|
2259 |
|
|
2260 |
Data.RefQ = qimage; |
WARPPOINTS gmc; |
2261 |
|
|
2262 |
if (sadInit) (*sadInit) (); |
uint32_t mx, my; |
2263 |
|
|
2264 |
for (y = 0; y < pParam->mb_height; y++) { |
int MBh = pParam->mb_height; |
2265 |
for (x = 0; x < pParam->mb_width; x++) { |
int MBw = pParam->mb_width; |
2266 |
|
|
2267 |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
int *MBmask= calloc(MBh*MBw,sizeof(int)); |
2268 |
|
double DtimesF[4] = { 0.,0., 0., 0. }; |
2269 |
|
double sol[4] = { 0., 0., 0., 0. }; |
2270 |
|
double a,b,c,n,denom; |
2271 |
|
double meanx,meany; |
2272 |
|
int num,oldnum; |
2273 |
|
|
2274 |
|
if (!MBmask) { fprintf(stderr,"Mem error\n"); |
2275 |
|
gmc.duv[0].x= gmc.duv[0].y = |
2276 |
|
gmc.duv[1].x= gmc.duv[1].y = |
2277 |
|
gmc.duv[2].x= gmc.duv[2].y = 0; |
2278 |
|
return gmc; } |
2279 |
|
|
2280 |
|
// filter mask of all blocks |
2281 |
|
|
2282 |
|
for (my = 1; my < (uint32_t)MBh-1; my++) |
2283 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
2284 |
|
{ |
2285 |
|
const int mbnum = mx + my * MBw; |
2286 |
|
const MACROBLOCK *pMB = &pMBs[mbnum]; |
2287 |
|
const VECTOR mv = pMB->mvs[0]; |
2288 |
|
|
2289 |
//intra mode is copied from the first pass. At least for the time being |
if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) |
2290 |
if ((pMB->mode == MODE_INTRA) || (pMB->mode == MODE_NOT_CODED) ) continue; |
continue; |
2291 |
|
|
2292 |
if (!(current->global_flags & XVID_LUMIMASKING)) { |
if ( ( (abs(mv.x - (pMB-1)->mvs[0].x) < deltax) && (abs(mv.y - (pMB-1)->mvs[0].y) < deltay) ) |
2293 |
pMB->dquant = NO_CHANGE; |
&& ( (abs(mv.x - (pMB+1)->mvs[0].x) < deltax) && (abs(mv.y - (pMB+1)->mvs[0].y) < deltay) ) |
2294 |
pMB->quant = current->quant; } |
&& ( (abs(mv.x - (pMB-MBw)->mvs[0].x) < deltax) && (abs(mv.y - (pMB-MBw)->mvs[0].y) < deltay) ) |
2295 |
else { |
&& ( (abs(mv.x - (pMB+MBw)->mvs[0].x) < deltax) && (abs(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) ) |
2296 |
if (pMB->dquant != NO_CHANGE) { |
MBmask[mbnum]=1; |
|
quant += DQtab[pMB->dquant]; |
|
|
if (quant > 31) quant = 31; |
|
|
else if (quant < 1) quant = 1; |
|
|
} |
|
|
pMB->quant = quant; |
|
2297 |
} |
} |
2298 |
|
|
2299 |
SearchPhinted(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
for (my = 1; my < (uint32_t)MBh-1; my++) |
2300 |
y, current->motion_flags, pMB->quant, |
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
2301 |
pParam, pMBs, current->global_flags & XVID_INTER4V, pMB, |
{ |
2302 |
&Data); |
const uint8_t *const pCur = current->image.y + 16*my*pParam->edged_width + 16*mx; |
2303 |
|
|
2304 |
|
const int mbnum = mx + my * MBw; |
2305 |
|
if (!MBmask[mbnum]) |
2306 |
|
continue; |
2307 |
|
|
2308 |
|
if (sad16 ( pCur, pCur+1 , pParam->edged_width, 65536) <= (uint32_t)grad ) |
2309 |
|
MBmask[mbnum] = 0; |
2310 |
|
if (sad16 ( pCur, pCur+pParam->edged_width, pParam->edged_width, 65536) <= (uint32_t)grad ) |
2311 |
|
MBmask[mbnum] = 0; |
2312 |
|
|
|
} |
|
|
} |
|
|
free(qimage); |
|
2313 |
} |
} |
2314 |
|
|
2315 |
static __inline int |
emms(); |
|
MEanalyzeMB ( const uint8_t * const pRef, |
|
|
const uint8_t * const pCur, |
|
|
const int x, |
|
|
const int y, |
|
|
const MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
MACROBLOCK * const pMB, |
|
|
SearchData * const Data) |
|
|
{ |
|
2316 |
|
|
2317 |
int i = 255, mask; |
do { /* until convergence */ |
|
VECTOR pmv[3]; |
|
|
*(Data->iMinSAD) = MV_MAX_ERROR; |
|
2318 |
|
|
2319 |
//median is only used as prediction. it doesn't have to be real |
a = b = c = n = 0; |
2320 |
if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; |
DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.; |
2321 |
else |
for (my = 0; my < (uint32_t)MBh; my++) |
2322 |
if (x == 1) //left macroblock does not have any vector now |
for (mx = 0; mx < (uint32_t)MBw; mx++) |
2323 |
Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median |
{ |
2324 |
else if (y == 1) // top macroblock don't have it's vector |
const int mbnum = mx + my * MBw; |
2325 |
Data->predMV = (pMB - 1)->mvs[0]; // left instead of median |
const MACROBLOCK *pMB = &pMBs[mbnum]; |
2326 |
else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median |
const VECTOR mv = pMB->mvs[0]; |
2327 |
|
|
2328 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
if (!MBmask[mbnum]) |
2329 |
pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); |
continue; |
2330 |
|
|
2331 |
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
n++; |
2332 |
Data->Ref = pRef + (x + y * pParam->edged_width) * 16; |
a += 16*mx+8; |
2333 |
|
b += 16*my+8; |
2334 |
|
c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8); |
2335 |
|
|
2336 |
|
DtimesF[0] += (double)mv.x; |
2337 |
|
DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8); |
2338 |
|
DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8); |
2339 |
|
DtimesF[3] += (double)mv.y; |
2340 |
|
} |
2341 |
|
|
2342 |
|
denom = a*a+b*b-c*n; |
2343 |
|
|
2344 |
|
/* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */ |
2345 |
|
/* D'*E*F has been calculated in the same loop as matrix */ |
2346 |
|
|
2347 |
|
sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2]; |
2348 |
|
sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3]; |
2349 |
|
sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3]; |
2350 |
|
sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3]; |
2351 |
|
|
2352 |
|
sol[0] /= denom; |
2353 |
|
sol[1] /= denom; |
2354 |
|
sol[2] /= denom; |
2355 |
|
sol[3] /= denom; |
2356 |
|
|
2357 |
|
meanx = meany = 0.; |
2358 |
|
oldnum = 0; |
2359 |
|
for (my = 0; my < (uint32_t)MBh; my++) |
2360 |
|
for (mx = 0; mx < (uint32_t)MBw; mx++) |
2361 |
|
{ |
2362 |
|
const int mbnum = mx + my * MBw; |
2363 |
|
const MACROBLOCK *pMB = &pMBs[mbnum]; |
2364 |
|
const VECTOR mv = pMB->mvs[0]; |
2365 |
|
|
2366 |
pmv[1].x = EVEN(pMB->mvs[0].x); |
if (!MBmask[mbnum]) |
2367 |
pmv[1].y = EVEN(pMB->mvs[0].y); |
continue; |
|
pmv[2].x = EVEN(Data->predMV.x); |
|
|
pmv[2].y = EVEN(Data->predMV.y); |
|
|
pmv[0].x = pmv[0].y = 0; |
|
2368 |
|
|
2369 |
(*CheckCandidate)(0, 0, 255, &i, Data); |
oldnum++; |
2370 |
|
meanx += fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ); |
2371 |
|
meany += fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ); |
2372 |
|
} |
2373 |
|
|
2374 |
//early skip for 0,0 |
if (4*meanx > oldnum) /* better fit than 0.25 is useless */ |
2375 |
if (*Data->iMinSAD < MAX_SAD00_FOR_SKIP * 4) { |
meanx /= oldnum; |
2376 |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
else |
2377 |
pMB->mode = MODE_NOT_CODED; |
meanx = 0.25; |
2378 |
return 0; |
|
2379 |
|
if (4*meany > oldnum) |
2380 |
|
meany /= oldnum; |
2381 |
|
else |
2382 |
|
meany = 0.25; |
2383 |
|
|
2384 |
|
/* fprintf(stderr,"sol = (%8.5f, %8.5f, %8.5f, %8.5f)\n",sol[0],sol[1],sol[2],sol[3]); |
2385 |
|
fprintf(stderr,"meanx = %8.5f meany = %8.5f %d\n",meanx,meany, oldnum); |
2386 |
|
*/ |
2387 |
|
num = 0; |
2388 |
|
for (my = 0; my < (uint32_t)MBh; my++) |
2389 |
|
for (mx = 0; mx < (uint32_t)MBw; mx++) |
2390 |
|
{ |
2391 |
|
const int mbnum = mx + my * MBw; |
2392 |
|
const MACROBLOCK *pMB = &pMBs[mbnum]; |
2393 |
|
const VECTOR mv = pMB->mvs[0]; |
2394 |
|
|
2395 |
|
if (!MBmask[mbnum]) |
2396 |
|
continue; |
2397 |
|
|
2398 |
|
if ( ( fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ) > meanx ) |
2399 |
|
|| ( fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ) > meany ) ) |
2400 |
|
MBmask[mbnum]=0; |
2401 |
|
else |
2402 |
|
num++; |
2403 |
} |
} |
2404 |
|
|
2405 |
if (!(mask = make_mask(pmv, 1))) |
} while ( (oldnum != num) && (num>=4) ); |
|
(*CheckCandidate)(pmv[1].x, pmv[1].y, mask, &i, Data); |
|
|
if (!(mask = make_mask(pmv, 2))) |
|
|
(*CheckCandidate)(pmv[2].x, pmv[2].y, mask, &i, Data); |
|
2406 |
|
|
2407 |
if (*Data->iMinSAD > MAX_SAD00_FOR_SKIP * 4) // diamond only if needed |
if (num < 4) |
2408 |
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
{ |
2409 |
|
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; |
2410 |
|
} else { |
2411 |
|
|
2412 |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
gmc.duv[0].x=(int)(sol[0]+0.5); |
2413 |
pMB->mode = MODE_INTER; |
gmc.duv[0].y=(int)(sol[3]+0.5); |
2414 |
return *(Data->iMinSAD); |
|
2415 |
|
gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5); |
2416 |
|
gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5); |
2417 |
|
|
2418 |
|
gmc.duv[2].x=0; |
2419 |
|
gmc.duv[2].y=0; |
2420 |
} |
} |
2421 |
|
// 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); |
2422 |
|
|
2423 |
#define INTRA_THRESH 1350 |
free(MBmask); |
|
#define INTER_THRESH 900 |
|
2424 |
|
|
2425 |
|
return gmc; |
2426 |
|
} |
2427 |
|
|
2428 |
int |
// functions which perform BITS-based search/bitcount |
2429 |
MEanalysis( const IMAGE * const pRef, |
|
2430 |
FRAMEINFO * const Current, |
static int |
2431 |
MBParam * const pParam, |
CountMBBitsInter(SearchData * const Data, |
2432 |
int maxIntra, //maximum number if non-I frames |
const MACROBLOCK * const pMBs, const int x, const int y, |
2433 |
int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame |
const MBParam * const pParam, |
2434 |
int bCount) // number if B frames in a row |
const uint32_t MotionFlags) |
2435 |
{ |
{ |
2436 |
uint32_t x, y, intra = 0; |
int i, iDirection; |
2437 |
int sSAD = 0; |
int32_t bsad[5]; |
|
MACROBLOCK * const pMBs = Current->mbs; |
|
|
const IMAGE * const pCurrent = &Current->image; |
|
|
int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH; |
|
2438 |
|
|
2439 |
VECTOR currentMV; |
CheckCandidate = CheckCandidateBits16; |
|
int32_t iMinSAD; |
|
|
SearchData Data; |
|
|
Data.iEdgedWidth = pParam->edged_width; |
|
|
Data.currentMV = ¤tMV; |
|
|
Data.iMinSAD = &iMinSAD; |
|
|
Data.iFcode = Current->fcode; |
|
|
CheckCandidate = CheckCandidate16no4vI; |
|
2440 |
|
|
2441 |
if (intraCount < 10) // we're right after an I frame |
if (Data->qpel) { |
2442 |
IntraThresh += 4 * (intraCount - 10) * (intraCount - 10); |
for(i = 0; i < 5; i++) { |
2443 |
else |
Data->currentMV[i].x = Data->currentQMV[i].x/2; |
2444 |
if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec |
Data->currentMV[i].y = Data->currentQMV[i].y/2; |
2445 |
IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; |
} |
2446 |
|
Data->qpel_precision = 1; |
2447 |
|
CheckCandidateBits16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data); |
2448 |
|
|
2449 |
|
//checking if this vector is perfect. if it is, we stop. |
2450 |
|
if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0) |
2451 |
|
return 0; //quick stop |
2452 |
|
|
2453 |
|
if (MotionFlags & (HALFPELREFINE16_BITS | EXTSEARCH_BITS)) { //we have to prepare for halfpixel-precision search |
2454 |
|
for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i]; |
2455 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2456 |
|
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
2457 |
|
Data->qpel_precision = 0; |
2458 |
|
if (Data->currentQMV->x & 1 || Data->currentQMV->y & 1) |
2459 |
|
CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); |
2460 |
|
} |
2461 |
|
|
2462 |
InterThresh += 300 * (1 - bCount); |
} else { // not qpel |
|
if (InterThresh < 200) InterThresh = 200; |
|
2463 |
|
|
2464 |
if (sadInit) (*sadInit) (); |
CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); |
2465 |
|
//checking if this vector is perfect. if it is, we stop. |
2466 |
|
if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0) { |
2467 |
|
return 0; //inter |
2468 |
|
} |
2469 |
|
} |
2470 |
|
|
2471 |
for (y = 1; y < pParam->mb_height-1; y++) { |
if (MotionFlags&EXTSEARCH_BITS) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
|
for (x = 1; x < pParam->mb_width-1; x++) { |
|
|
int sad, dev; |
|
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
|
2472 |
|
|
2473 |
sad = MEanalyzeMB(pRef->y, pCurrent->y, x, y, |
if (MotionFlags&HALFPELREFINE16_BITS) SubpelRefine(Data); |
|
pParam, pMBs, pMB, &Data); |
|
2474 |
|
|
2475 |
if (sad > IntraThresh) { |
if (Data->qpel) { |
2476 |
dev = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
if (MotionFlags&(EXTSEARCH_BITS | HALFPELREFINE16_BITS)) { // there was halfpel-precision search |
2477 |
pParam->edged_width); |
for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) { |
2478 |
if (dev + IntraThresh < sad) { |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // we have found a better match |
2479 |
pMB->mode = MODE_INTRA; |
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
|
if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return 2; // I frame |
|
|
} |
|
2480 |
} |
} |
2481 |
sSAD += sad; |
|
2482 |
|
// preparing for qpel-precision search |
2483 |
|
Data->qpel_precision = 1; |
2484 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2485 |
|
pParam->width, pParam->height, Data->iFcode, 1, 0); |
2486 |
} |
} |
2487 |
|
if (MotionFlags&QUARTERPELREFINE16_BITS) SubpelRefine(Data); |
2488 |
} |
} |
|
sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); |
|
|
if (sSAD > InterThresh ) return 1; //P frame |
|
|
emms(); |
|
|
return 0; // B frame |
|
2489 |
|
|
2490 |
|
if (MotionFlags&CHECKPREDICTION_BITS) { //let's check vector equal to prediction |
2491 |
|
VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV; |
2492 |
|
if (!(Data->predMV.x == v->x && Data->predMV.y == v->y)) |
2493 |
|
CheckCandidateBits16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data); |
2494 |
|
} |
2495 |
|
return Data->iMinSAD[0]; |
2496 |
} |
} |
2497 |
|
|
|
int |
|
|
FindFcode( const MBParam * const pParam, |
|
|
const FRAMEINFO * const current) |
|
|
{ |
|
|
uint32_t x, y; |
|
|
int max = 0, min = 0, i; |
|
2498 |
|
|
2499 |
for (y = 0; y < pParam->mb_height; y++) { |
static int |
2500 |
for (x = 0; x < pParam->mb_width; x++) { |
CountMBBitsInter4v(const SearchData * const Data, |
2501 |
|
MACROBLOCK * const pMB, const MACROBLOCK * const pMBs, |
2502 |
|
const int x, const int y, |
2503 |
|
const MBParam * const pParam, const uint32_t MotionFlags, |
2504 |
|
const VECTOR * const backup) |
2505 |
|
{ |
2506 |
|
|
2507 |
MACROBLOCK *pMB = ¤t->mbs[x + y * pParam->mb_width]; |
int cbp = 0, bits = 0, t = 0, i, iDirection; |
2508 |
for(i = 0; i < (pMB->mode == MODE_INTER4V ? 4:1); i++) { |
SearchData Data2, *Data8 = &Data2; |
2509 |
if (pMB->mvs[i].x > max) max = pMB->mvs[i].x; |
int sumx = 0, sumy = 0; |
2510 |
if (pMB->mvs[i].y > max) max = pMB->mvs[i].y; |
int16_t *in = Data->dctSpace, *coeff = Data->dctSpace + 64; |
2511 |
|
uint8_t * ptr; |
2512 |
|
|
2513 |
|
memcpy(Data8, Data, sizeof(SearchData)); |
2514 |
|
CheckCandidate = CheckCandidateBits8; |
2515 |
|
|
2516 |
|
for (i = 0; i < 4; i++) { //for all luma blocks |
2517 |
|
|
2518 |
|
Data8->iMinSAD = Data->iMinSAD + i + 1; |
2519 |
|
Data8->currentMV = Data->currentMV + i + 1; |
2520 |
|
Data8->currentQMV = Data->currentQMV + i + 1; |
2521 |
|
Data8->Cur = Data->Cur + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2522 |
|
Data8->RefP[0] = Data->RefP[0] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2523 |
|
Data8->RefP[2] = Data->RefP[2] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2524 |
|
Data8->RefP[1] = Data->RefP[1] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2525 |
|
Data8->RefP[3] = Data->RefP[3] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2526 |
|
|
2527 |
if (pMB->mvs[i].x < min) min = pMB->mvs[i].x; |
if(Data->qpel) { |
2528 |
if (pMB->mvs[i].y < min) min = pMB->mvs[i].y; |
Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i); |
2529 |
|
if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y, |
2530 |
|
Data8->predMV, Data8->iFcode, 0, 0); |
2531 |
|
} else { |
2532 |
|
Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i); |
2533 |
|
if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y, |
2534 |
|
Data8->predMV, Data8->iFcode, 0, 0); |
2535 |
} |
} |
2536 |
|
|
2537 |
|
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, |
2538 |
|
pParam->width, pParam->height, Data8->iFcode, Data8->qpel, 0); |
2539 |
|
|
2540 |
|
*Data8->iMinSAD += t; |
2541 |
|
|
2542 |
|
Data8->qpel_precision = Data8->qpel; |
2543 |
|
// checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far) |
2544 |
|
{ |
2545 |
|
VECTOR *v = Data8->qpel ? Data8->currentQMV : Data8->currentMV; |
2546 |
|
if (!MVequal (*v, backup[i+1]) ) |
2547 |
|
CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8); |
2548 |
} |
} |
2549 |
|
|
2550 |
|
if (Data8->qpel) { |
2551 |
|
if (MotionFlags&HALFPELREFINE8_BITS || (MotionFlags&PMV_EXTSEARCH8 && MotionFlags&EXTSEARCH_BITS)) { // halfpixel motion search follows |
2552 |
|
int32_t s = *Data8->iMinSAD; |
2553 |
|
Data8->currentMV->x = Data8->currentQMV->x/2; |
2554 |
|
Data8->currentMV->y = Data8->currentQMV->y/2; |
2555 |
|
Data8->qpel_precision = 0; |
2556 |
|
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, |
2557 |
|
pParam->width, pParam->height, Data8->iFcode - 1, 0, 0); |
2558 |
|
|
2559 |
|
if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1) |
2560 |
|
CheckCandidateBits8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8); |
2561 |
|
|
2562 |
|
if (MotionFlags & PMV_EXTSEARCH8 && MotionFlags & EXTSEARCH_BITS) |
2563 |
|
SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255); |
2564 |
|
|
2565 |
|
if (MotionFlags & HALFPELREFINE8_BITS) |
2566 |
|
SubpelRefine(Data8); |
2567 |
|
|
2568 |
|
if (s > *Data8->iMinSAD) { //we have found a better match |
2569 |
|
Data8->currentQMV->x = 2*Data8->currentMV->x; |
2570 |
|
Data8->currentQMV->y = 2*Data8->currentMV->y; |
2571 |
} |
} |
2572 |
|
|
2573 |
min = -min; |
Data8->qpel_precision = 1; |
2574 |
max += 1; |
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, |
2575 |
if (min > max) max = min; |
pParam->width, pParam->height, Data8->iFcode, 1, 0); |
|
if (pParam->m_quarterpel) max *= 2; |
|
2576 |
|
|
|
for (i = 1; (max > 32 << (i - 1)); i++); |
|
|
return i; |
|
2577 |
} |
} |
2578 |
|
if (MotionFlags & QUARTERPELREFINE8_BITS) SubpelRefine(Data8); |
2579 |
|
|
2580 |
static void |
} else { // not qpel |
|
CheckGMC(int x, int y, const int dir, int * iDirection, |
|
|
const MACROBLOCK * const pMBs, uint32_t * bestcount, VECTOR * GMC, |
|
|
const MBParam * const pParam) |
|
|
{ |
|
|
uint32_t mx, my, a, count = 0; |
|
2581 |
|
|
2582 |
for (my = 1; my < pParam->mb_height-1; my++) |
if (MotionFlags & PMV_EXTSEARCH8 && MotionFlags & EXTSEARCH_BITS) //extsearch |
2583 |
for (mx = 1; mx < pParam->mb_width-1; mx++) { |
SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255); |
2584 |
VECTOR mv; |
|
2585 |
const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; |
if (MotionFlags & HALFPELREFINE8_BITS) |
2586 |
if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) continue; |
SubpelRefine(Data8); //halfpel refinement |
|
mv = pMB->mvs[0]; |
|
|
a = ABS(mv.x - x) + ABS(mv.y - y); |
|
|
if (a < 6) count += 6 - a; |
|
2587 |
} |
} |
2588 |
|
|
2589 |
if (count > *bestcount) { |
//checking vector equal to predicion |
2590 |
*bestcount = count; |
if (i != 0 && MotionFlags & CHECKPREDICTION_BITS) { |
2591 |
*iDirection = dir; |
const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV; |
2592 |
GMC->x = x; GMC->y = y; |
if (!MVequal(*v, Data8->predMV)) |
2593 |
|
CheckCandidateBits8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8); |
2594 |
} |
} |
2595 |
|
|
2596 |
|
bits += *Data8->iMinSAD; |
2597 |
|
if (bits >= Data->iMinSAD[0]) return bits; // no chances for INTER4V |
2598 |
|
|
2599 |
|
// MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else |
2600 |
|
if(Data->qpel) { |
2601 |
|
pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x; |
2602 |
|
pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y; |
2603 |
|
pMB->qmvs[i] = *Data8->currentQMV; |
2604 |
|
sumx += Data8->currentQMV->x/2; |
2605 |
|
sumy += Data8->currentQMV->y/2; |
2606 |
|
} else { |
2607 |
|
pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x; |
2608 |
|
pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y; |
2609 |
|
sumx += Data8->currentMV->x; |
2610 |
|
sumy += Data8->currentMV->y; |
2611 |
} |
} |
2612 |
|
pMB->mvs[i] = *Data8->currentMV; |
2613 |
|
pMB->sad8[i] = 4 * *Data8->iMinSAD; |
2614 |
|
if (Data8->temp[0]) cbp |= 1 << (5 - i); |
2615 |
|
|
2616 |
|
} // /for all luma blocks |
2617 |
|
|
2618 |
static VECTOR |
bits += xvid_cbpy_tab[15-(cbp>>2)].len; |
|
GlobalMotionEst(const MACROBLOCK * const pMBs, const MBParam * const pParam, const uint32_t iFcode) |
|
|
{ |
|
2619 |
|
|
2620 |
uint32_t count, bestcount = 0; |
// let's check chroma |
2621 |
int x, y; |
sumx = (sumx >> 3) + roundtab_76[sumx & 0xf]; |
2622 |
VECTOR gmc = {0,0}; |
sumy = (sumy >> 3) + roundtab_76[sumy & 0xf]; |
|
int step, min_x, max_x, min_y, max_y; |
|
|
uint32_t mx, my; |
|
|
int iDirection, bDirection; |
|
2623 |
|
|
2624 |
min_x = min_y = -32<<iFcode; |
//chroma U |
2625 |
max_x = max_y = 32<<iFcode; |
ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[4], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
2626 |
|
transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2); |
2627 |
|
bits += Block_CalcBits(coeff, in, Data->iQuant, Data->quant_type, &cbp, 4, 0); |
2628 |
|
|
2629 |
//step1: let's find a rough camera panning |
if (bits >= *Data->iMinSAD) return bits; |
|
for (step = 32; step >= 2; step /= 2) { |
|
|
bestcount = 0; |
|
|
for (y = min_y; y <= max_y; y += step) |
|
|
for (x = min_x ; x <= max_x; x += step) { |
|
|
count = 0; |
|
|
//for all macroblocks |
|
|
for (my = 1; my < pParam->mb_height-1; my++) |
|
|
for (mx = 1; mx < pParam->mb_width-1; mx++) { |
|
|
const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; |
|
|
VECTOR mv; |
|
2630 |
|
|
2631 |
if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) |
//chroma V |
2632 |
continue; |
ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[5], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
2633 |
|
transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2); |
2634 |
|
bits += Block_CalcBits(coeff, in, Data->iQuant, Data->quant_type, &cbp, 5, 0); |
2635 |
|
|
2636 |
mv = pMB->mvs[0]; |
bits += mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len; |
|
if ( ABS(mv.x - x) <= step && ABS(mv.y - y) <= step ) /* GMC translation is always halfpel-res */ |
|
|
count++; |
|
|
} |
|
|
if (count >= bestcount) { bestcount = count; gmc.x = x; gmc.y = y; } |
|
|
} |
|
|
min_x = gmc.x - step; |
|
|
max_x = gmc.x + step; |
|
|
min_y = gmc.y - step; |
|
|
max_y = gmc.y + step; |
|
2637 |
|
|
2638 |
|
return bits; |
2639 |
} |
} |
2640 |
|
|
|
if (bestcount < (pParam->mb_height-2)*(pParam->mb_width-2)/10) |
|
|
gmc.x = gmc.y = 0; //no camara pan, no GMC |
|
2641 |
|
|
2642 |
// step2: let's refine camera panning using gradiend-descent approach. |
static int |
2643 |
// TODO: more warping points may be evaluated here (like in interpolate mode search - two vectors in one diamond) |
CountMBBitsIntra(const SearchData * const Data) |
2644 |
bestcount = 0; |
{ |
2645 |
CheckGMC(gmc.x, gmc.y, 255, &iDirection, pMBs, &bestcount, &gmc, pParam); |
int bits = 1; //this one is ac/dc prediction flag. always 1. |
2646 |
do { |
int cbp = 0, i, t, dc = 1024, b_dc; |
2647 |
x = gmc.x; y = gmc.y; |
int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64; |
2648 |
bDirection = iDirection; iDirection = 0; |
uint32_t iDcScaler = get_dc_scaler(Data->iQuant, 1); |
|
if (bDirection & 1) CheckGMC(x - 1, y, 1+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); |
|
|
if (bDirection & 2) CheckGMC(x + 1, y, 2+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); |
|
|
if (bDirection & 4) CheckGMC(x, y - 1, 1+2+4, &iDirection, pMBs, &bestcount, &gmc, pParam); |
|
|
if (bDirection & 8) CheckGMC(x, y + 1, 1+2+8, &iDirection, pMBs, &bestcount, &gmc, pParam); |
|
2649 |
|
|
2650 |
} while (iDirection); |
for(i = 0; i < 4; i++) { |
2651 |
|
int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2652 |
|
transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth); |
2653 |
|
fdct(in); |
2654 |
|
b_dc = in[0]; |
2655 |
|
in[0] -= dc; |
2656 |
|
dc = b_dc; |
2657 |
|
if (Data->quant_type == 0) quant_intra(coeff, in, Data->iQuant, iDcScaler); |
2658 |
|
else quant4_intra(coeff, in, Data->iQuant, iDcScaler); |
2659 |
|
|
2660 |
if (pParam->m_quarterpel) { |
bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcy_tab[coeff[0] + 255].len; |
2661 |
gmc.x *= 2; |
Data->temp[i] = t; |
2662 |
gmc.y *= 2; /* we store the halfpel value as pseudo-qpel to make comparison easier */ |
if (t != 0) cbp |= 1 << (5 - i); |
2663 |
|
if (bits >= Data->iMinSAD[0]) return bits; |
2664 |
} |
} |
2665 |
|
|
2666 |
return gmc; |
bits += xvid_cbpy_tab[cbp>>2].len; |
2667 |
|
|
2668 |
|
iDcScaler = get_dc_scaler(Data->iQuant, 0); |
2669 |
|
|
2670 |
|
//chroma U |
2671 |
|
transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2); |
2672 |
|
fdct(in); |
2673 |
|
in[0] -= 1024; |
2674 |
|
if (Data->quant_type == 0) quant_intra(coeff, in, Data->iQuant, iDcScaler); |
2675 |
|
else quant4_intra(coeff, in, Data->iQuant, iDcScaler); |
2676 |
|
|
2677 |
|
bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len; |
2678 |
|
if (t != 0) cbp |= 1 << (5 - 4); |
2679 |
|
|
2680 |
|
if (bits >= Data->iMinSAD[0]) return bits; |
2681 |
|
|
2682 |
|
//chroma V |
2683 |
|
transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2); |
2684 |
|
fdct(in); |
2685 |
|
in[0] -= 1024; |
2686 |
|
if (Data->quant_type == 0) quant_intra(coeff, in, Data->iQuant, iDcScaler); |
2687 |
|
else quant4_intra(coeff, in, Data->iQuant, iDcScaler); |
2688 |
|
|
2689 |
|
bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len; |
2690 |
|
if (t != 0) cbp |= 1 << (5 - 5); |
2691 |
|
|
2692 |
|
bits += mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len; |
2693 |
|
|
2694 |
|
return bits; |
2695 |
} |
} |