52 |
#define CHECK_CANDIDATE(X,Y,D) { \ |
#define CHECK_CANDIDATE(X,Y,D) { \ |
53 |
(*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } |
(*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } |
54 |
|
|
|
#define GET_REFERENCE(X, Y, REF) { \ |
|
|
switch ( (((X)&1)<<1) + ((Y)&1) ) \ |
|
|
{ \ |
|
|
case 0 : REF = (uint8_t *)data->Ref + (X)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ |
|
|
case 1 : REF = (uint8_t *)data->RefV + (X)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ |
|
|
case 2 : REF = (uint8_t *)data->RefH + ((X)-1)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ |
|
|
default : REF = (uint8_t *)data->RefHV + ((X)-1)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ |
|
|
} \ |
|
|
} |
|
|
|
|
55 |
#define iDiamondSize 2 |
#define iDiamondSize 2 |
56 |
|
|
57 |
static __inline int |
static __inline int |
79 |
return xb + yb; |
return xb + yb; |
80 |
} |
} |
81 |
|
|
82 |
|
static int32_t |
83 |
|
ChromaSAD(int dx, int dy, const SearchData * const data) |
84 |
|
{ |
85 |
|
int sad; |
86 |
|
dx = (dx >> 1) + roundtab_79[dx & 0x3]; |
87 |
|
dy = (dy >> 1) + roundtab_79[dy & 0x3]; |
88 |
|
|
89 |
|
switch (((dx & 1) << 1) + (dy & 1)) { // ((dx%2)?2:0)+((dy%2)?1:0) |
90 |
|
case 0: |
91 |
|
sad = sad8(data->CurU, data->RefCU + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); |
92 |
|
sad += sad8(data->CurV, data->RefCV + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); |
93 |
|
break; |
94 |
|
case 1: |
95 |
|
dx = dx / 2; dy = (dy - 1) / 2; |
96 |
|
sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); |
97 |
|
sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); |
98 |
|
break; |
99 |
|
case 2: |
100 |
|
dx = (dx - 1) / 2; dy = dy / 2; |
101 |
|
sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); |
102 |
|
sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); |
103 |
|
break; |
104 |
|
default: |
105 |
|
dx = (dx - 1) / 2; dy = (dy - 1) / 2; |
106 |
|
interpolate8x8_halfpel_hv(data->RefQ, |
107 |
|
data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, |
108 |
|
data->rounding); |
109 |
|
sad = sad8(data->CurU, data->RefQ, data->iEdgedWidth/2); |
110 |
|
interpolate8x8_halfpel_hv(data->RefQ, |
111 |
|
data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, |
112 |
|
data->rounding); |
113 |
|
sad += sad8(data->CurV, data->RefQ, data->iEdgedWidth/2); |
114 |
|
break; |
115 |
|
} |
116 |
|
return sad; |
117 |
|
} |
118 |
|
|
119 |
/* CHECK_CANDIATE FUNCTIONS START */ |
static __inline const uint8_t * |
120 |
|
GetReference(const int x, const int y, const int dir, const SearchData * const data) |
|
static void |
|
|
CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
121 |
{ |
{ |
122 |
int t; |
// dir : 0 = forward, 1 = backward |
123 |
const uint8_t * Reference; |
switch ( (dir << 2) | ((x&1)<<1) | (y&1) ) { |
124 |
|
case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth); |
125 |
|
case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); |
126 |
|
case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); |
127 |
|
case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); |
128 |
|
case 4 : return data->bRef + x/2 + (y/2)*(data->iEdgedWidth); |
129 |
|
case 5 : return data->bRefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); |
130 |
|
case 6 : return data->bRefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); |
131 |
|
default : return data->bRefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); |
132 |
|
|
133 |
if (( x > data->max_dx) || ( x < data->min_dx) |
} |
134 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
} |
135 |
|
|
136 |
|
static uint8_t * |
137 |
|
Interpolate8x8qpel(const int x, const int y, const int block, const int dir, const SearchData * const data) |
138 |
|
{ |
139 |
|
// create or find a qpel-precision reference picture; return pointer to it |
140 |
|
uint8_t * Reference = (uint8_t *)data->RefQ; |
141 |
|
const int32_t iEdgedWidth = data->iEdgedWidth; |
142 |
|
const uint32_t rounding = data->rounding; |
143 |
|
const int halfpel_x = x/2; |
144 |
|
const int halfpel_y = y/2; |
145 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
146 |
|
|
147 |
|
ref1 = GetReference(halfpel_x, halfpel_y, dir, data); // this reference is used in all cases |
148 |
|
ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
149 |
switch ( ((x&1)<<1) + (y&1) ) { |
switch ( ((x&1)<<1) + (y&1) ) { |
150 |
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
case 0: // pure halfpel position |
151 |
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
Reference = (uint8_t *) GetReference(halfpel_x, halfpel_y, dir, data); |
152 |
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
Reference += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
153 |
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
break; |
|
} |
|
154 |
|
|
155 |
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
156 |
|
ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
157 |
|
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
158 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
159 |
|
break; |
160 |
|
|
161 |
t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
162 |
data->temp[0] += data->lambda16 * t; |
ref2 = GetReference(x - halfpel_x, halfpel_y, dir, data); |
163 |
data->temp[1] += data->lambda8 * t; |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
164 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
165 |
|
break; |
166 |
|
|
167 |
if (data->temp[0] < data->iMinSAD[0]) { |
default: // x and y in qpel resolution - the "corners" (top left/right and |
168 |
data->iMinSAD[0] = data->temp[0]; |
// bottom left/right) during qpel refinement |
169 |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
170 |
*dir = Direction; } |
ref3 = GetReference(x - halfpel_x, halfpel_y, dir, data); |
171 |
|
ref4 = GetReference(x - halfpel_x, y - halfpel_y, dir, data); |
172 |
|
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
173 |
|
ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
174 |
|
ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
175 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
176 |
|
break; |
177 |
|
} |
178 |
|
return Reference; |
179 |
|
} |
180 |
|
|
181 |
if (data->temp[1] < data->iMinSAD[1]) { |
static uint8_t * |
182 |
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
Interpolate16x16qpel(const int x, const int y, const int dir, const SearchData * const data) |
183 |
if (data->temp[2] < data->iMinSAD[2]) { |
{ |
184 |
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
// create or find a qpel-precision reference picture; return pointer to it |
185 |
if (data->temp[3] < data->iMinSAD[3]) { |
uint8_t * Reference = (uint8_t *)data->RefQ; |
186 |
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
const int32_t iEdgedWidth = data->iEdgedWidth; |
187 |
if (data->temp[4] < data->iMinSAD[4]) { |
const uint32_t rounding = data->rounding; |
188 |
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
const int halfpel_x = x/2; |
189 |
|
const int halfpel_y = y/2; |
190 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
191 |
|
|
192 |
|
ref1 = GetReference(halfpel_x, halfpel_y, dir, data); // this reference is used in all cases |
193 |
|
switch( ((x&1)<<1) + (y&1) ) { |
194 |
|
case 0: // pure halfpel position |
195 |
|
return (uint8_t *) GetReference(halfpel_x, halfpel_y, dir, data); |
196 |
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
197 |
|
ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
198 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
199 |
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); |
200 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); |
201 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); |
202 |
|
break; |
203 |
|
|
204 |
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
205 |
|
ref2 = GetReference(x - halfpel_x, halfpel_y, dir, data); |
206 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
207 |
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); |
208 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); |
209 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); |
210 |
|
break; |
211 |
|
|
212 |
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
213 |
|
// bottom left/right) during qpel refinement |
214 |
|
ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
215 |
|
ref3 = GetReference(x - halfpel_x, halfpel_y, dir, data); |
216 |
|
ref4 = GetReference(x - halfpel_x, y - halfpel_y, dir, data); |
217 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
218 |
|
interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); |
219 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); |
220 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); |
221 |
|
break; |
222 |
|
} |
223 |
|
return Reference; |
224 |
} |
} |
225 |
|
|
226 |
|
/* CHECK_CANDIATE FUNCTIONS START */ |
227 |
|
|
228 |
static void |
static void |
229 |
CheckCandidate16Q(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) |
230 |
{ |
{ |
231 |
//this function expects X and Y in halfpel positions, but assumes that qpel positions are our goal |
int t, xc, yc; |
|
int t; |
|
232 |
const uint8_t * Reference; |
const uint8_t * Reference; |
233 |
|
VECTOR * current; |
234 |
|
|
235 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if (( x > data->max_dx) || ( x < data->min_dx) |
236 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
237 |
|
|
238 |
|
if (data->qpel_precision) { // x and y are in 1/4 precision |
239 |
|
Reference = Interpolate16x16qpel(x, y, 0, data); |
240 |
|
t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
241 |
|
xc = x/2; yc = y/2; //for chroma sad |
242 |
|
current = data->currentQMV; |
243 |
|
} else { |
244 |
switch ( ((x&1)<<1) + (y&1) ) { |
switch ( ((x&1)<<1) + (y&1) ) { |
245 |
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
246 |
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
247 |
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
248 |
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
249 |
} |
} |
250 |
|
if (data->qpel) t = d_mv_bits(2*x - data->predMV.x, 2*y - data->predMV.y, data->iFcode); |
251 |
|
else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
252 |
|
current = data->currentMV; |
253 |
|
xc = x; yc = y; |
254 |
|
} |
255 |
|
|
256 |
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
257 |
|
|
258 |
t = d_mv_bits(2 * x - data->predQMV.x, 2 * y - data->predQMV.y, data->iFcode); |
data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; |
259 |
data->temp[0] += data->lambda16 * t; |
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; |
260 |
data->temp[1] += data->lambda8 * t; |
|
261 |
|
if (data->chroma) data->temp[0] += ChromaSAD(xc, yc, data); |
262 |
|
|
263 |
if (data->temp[0] < data->iMinSAD[0]) { |
if (data->temp[0] < data->iMinSAD[0]) { |
264 |
data->iMinSAD[0] = data->temp[0]; |
data->iMinSAD[0] = data->temp[0]; |
265 |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
current[0].x = x; current[0].y = y; |
266 |
*dir = Direction; } |
*dir = Direction; } |
267 |
|
|
268 |
if (data->temp[1] < data->iMinSAD[1]) { |
if (data->temp[1] < data->iMinSAD[1]) { |
269 |
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y= y; } |
270 |
if (data->temp[2] < data->iMinSAD[2]) { |
if (data->temp[2] < data->iMinSAD[2]) { |
271 |
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; } |
272 |
if (data->temp[3] < data->iMinSAD[3]) { |
if (data->temp[3] < data->iMinSAD[3]) { |
273 |
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; } |
274 |
if (data->temp[4] < data->iMinSAD[4]) { |
if (data->temp[4] < data->iMinSAD[4]) { |
275 |
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; } |
276 |
|
|
277 |
} |
} |
278 |
|
|
281 |
{ |
{ |
282 |
int32_t sad; |
int32_t sad; |
283 |
const uint8_t * Reference; |
const uint8_t * Reference; |
284 |
|
int t; |
285 |
|
VECTOR * current; |
286 |
|
|
287 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if (( x > data->max_dx) || ( x < data->min_dx) |
288 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
289 |
|
|
290 |
switch ( ((x&1)<<1) + (y&1) ) |
if (data->qpel_precision) { // x and y are in 1/4 precision |
291 |
{ |
Reference = Interpolate16x16qpel(x, y, 0, data); |
292 |
|
t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
293 |
|
current = data->currentQMV; |
294 |
|
} else { |
295 |
|
switch ( ((x&1)<<1) + (y&1) ) { |
296 |
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
297 |
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
298 |
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
299 |
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
300 |
} |
} |
301 |
|
if (data->qpel) t = d_mv_bits(2*x - data->predMV.x, 2*y - data->predMV.y, data->iFcode); |
302 |
|
else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
303 |
|
current = data->currentMV; |
304 |
|
} |
305 |
|
|
306 |
sad = data->lambda16 * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
307 |
sad += sad16(data->Cur, Reference, data->iEdgedWidth, MV_MAX_ERROR); |
sad += (data->lambda16 * t * sad)/1000; |
308 |
|
|
309 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
310 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
311 |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
current->x = x; current->y = y; |
312 |
*dir = Direction; } |
*dir = Direction; } |
313 |
} |
} |
314 |
|
|
315 |
static void |
static void |
|
CheckCandidate16_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
|
|
|
|
// CheckCandidate16 variant which expects x and y in quarter pixel resolution |
|
|
// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!) |
|
|
// around currentMV! |
|
|
{ |
|
|
int t; |
|
|
uint8_t * Reference = (uint8_t *)data->RefQ; |
|
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
|
|
VECTOR halfpelMV = *(data->currentMV); |
|
|
|
|
|
int32_t iEdgedWidth = data->iEdgedWidth; |
|
|
uint32_t rounding = data->rounding; |
|
|
|
|
|
if (( x > data->max_dx) || ( x < data->min_dx) |
|
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|
|
|
|
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this refenrence is used in all cases |
|
|
switch( ((x&1)<<1) + (y&1) ) |
|
|
{ |
|
|
case 0: // pure halfpel position - shouldn't happen during a refinement step |
|
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, Reference); |
|
|
break; |
|
|
|
|
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
|
|
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
|
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
|
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); |
|
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); |
|
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); |
|
|
break; |
|
|
|
|
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
|
|
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); |
|
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
|
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); |
|
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); |
|
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); |
|
|
break; |
|
|
|
|
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
|
|
// bottom left/right) during qpel refinement |
|
|
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
|
|
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); |
|
|
GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); |
|
|
|
|
|
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; |
|
|
} |
|
|
|
|
|
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp+1); |
|
|
|
|
|
t = d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode); |
|
|
data->temp[0] += data->lambda16 * t; |
|
|
data->temp[1] += data->lambda8 * t; |
|
|
|
|
|
if (data->temp[0] < data->iMinSAD[0]) { |
|
|
data->iMinSAD[0] = data->temp[0]; |
|
|
data->currentQMV[0].x = x; data->currentQMV[0].y = y; |
|
|
/* *dir = Direction;*/ } |
|
|
|
|
|
if (data->temp[1] < data->iMinSAD[1]) { |
|
|
data->iMinSAD[1] = data->temp[1]; data->currentQMV[1].x = x; data->currentQMV[1].y = y; } |
|
|
if (data->temp[2] < data->iMinSAD[2]) { |
|
|
data->iMinSAD[2] = data->temp[2]; data->currentQMV[2].x = x; data->currentQMV[2].y = y; } |
|
|
if (data->temp[3] < data->iMinSAD[3]) { |
|
|
data->iMinSAD[3] = data->temp[3]; data->currentQMV[3].x = x; data->currentQMV[3].y = y; } |
|
|
if (data->temp[4] < data->iMinSAD[4]) { |
|
|
data->iMinSAD[4] = data->temp[4]; data->currentQMV[4].x = x; data->currentQMV[4].y = y; } |
|
|
} |
|
|
|
|
|
static void |
|
316 |
CheckCandidate16no4vI(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidate16no4vI(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
317 |
{ |
{ |
318 |
|
// maximum speed - for P/B/I decision |
319 |
int32_t sad; |
int32_t sad; |
320 |
|
|
321 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if (( x > data->max_dx) || ( x < data->min_dx) |
335 |
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) |
336 |
{ |
{ |
337 |
int32_t sad; |
int32_t sad; |
338 |
const int xb = data->currentMV[1].x; |
int xb, yb, t; |
|
const int yb = data->currentMV[1].y; |
|
339 |
const uint8_t *ReferenceF, *ReferenceB; |
const uint8_t *ReferenceF, *ReferenceB; |
340 |
|
VECTOR *current; |
341 |
|
|
342 |
if (( xf > data->max_dx) || ( xf < data->min_dx) |
if (( xf > data->max_dx) || ( xf < data->min_dx) |
343 |
|| ( yf > data->max_dy) || (yf < data->min_dy)) return; |
|| ( yf > data->max_dy) || (yf < data->min_dy)) return; |
344 |
|
|
345 |
switch ( ((xf&1)<<1) + (yf&1) ) { |
if (data->qpel_precision) { |
346 |
case 0 : ReferenceF = data->Ref + xf/2 + (yf/2)*(data->iEdgedWidth); break; |
ReferenceF = Interpolate16x16qpel(xf, yf, 0, data); |
347 |
case 1 : ReferenceF = data->RefV + xf/2 + ((yf-1)/2)*(data->iEdgedWidth); break; |
xb = data->currentQMV[1].x; yb = data->currentQMV[1].y; |
348 |
case 2 : ReferenceF = data->RefH + (xf-1)/2 + (yf/2)*(data->iEdgedWidth); break; |
current = data->currentQMV; |
349 |
default : ReferenceF = data->RefHV + (xf-1)/2 + ((yf-1)/2)*(data->iEdgedWidth); break; |
ReferenceB = Interpolate16x16qpel(xb, yb, 1, data); |
350 |
} |
t = d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) |
351 |
|
+ d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode); |
352 |
switch ( ((xb&1)<<1) + (yb&1) ) { |
} else { |
353 |
case 0 : ReferenceB = data->bRef + xb/2 + (yb/2)*(data->iEdgedWidth); break; |
ReferenceF = Interpolate16x16qpel(2*xf, 2*yf, 0, data); |
354 |
case 1 : ReferenceB = data->bRefV + xb/2 + ((yb-1)/2)*(data->iEdgedWidth); break; |
xb = data->currentMV[1].x; yb = data->currentMV[1].y; |
355 |
case 2 : ReferenceB = data->bRefH + (xb-1)/2 + (yb/2)*(data->iEdgedWidth); break; |
ReferenceB = Interpolate16x16qpel(2*xb, 2*yb, 1, data); |
356 |
default : ReferenceB = data->bRefHV + (xb-1)/2 + ((yb-1)/2)*(data->iEdgedWidth); break; |
current = data->currentMV; |
357 |
|
if (data->qpel) |
358 |
|
t = d_mv_bits(2*xf - data->predMV.x, 2*yf - data->predMV.y, data->iFcode) |
359 |
|
+ d_mv_bits(2*xb - data->bpredMV.x, 2*yb - data->bpredMV.y, data->iFcode); |
360 |
|
else |
361 |
|
t = d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) |
362 |
|
+ d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode); |
363 |
} |
} |
364 |
|
|
365 |
sad = data->lambda16 * |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
366 |
( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) + |
sad += (data->lambda16 * t * sad)/1000; |
|
d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode) ); |
|
|
|
|
|
sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
|
367 |
|
|
368 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
369 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
370 |
data->currentMV->x = xf; data->currentMV->y = yf; |
current->x = xf; current->y = yf; |
371 |
*dir = Direction; } |
*dir = Direction; } |
372 |
} |
} |
373 |
|
|
374 |
static void |
static void |
375 |
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) |
376 |
{ |
{ |
377 |
int32_t sad; |
int32_t sad = 0; |
378 |
int k; |
int k; |
379 |
const uint8_t *ReferenceF; |
const uint8_t *ReferenceF; |
380 |
const uint8_t *ReferenceB; |
const uint8_t *ReferenceB; |
382 |
|
|
383 |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
384 |
|
|
|
sad = data->lambda16 * d_mv_bits(x, y, 1); |
|
|
|
|
385 |
for (k = 0; k < 4; k++) { |
for (k = 0; k < 4; k++) { |
386 |
mvs.x = data->directmvF[k].x + x; |
mvs.x = data->directmvF[k].x + x; |
387 |
b_mvs.x = ((x == 0) ? |
b_mvs.x = ((x == 0) ? |
398 |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
399 |
|| ( 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; |
400 |
|
|
401 |
switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
if (!data->qpel) { |
402 |
case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
mvs.x *= 2; mvs.y *= 2; |
403 |
case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
b_mvs.x *= 2; b_mvs.y *= 2; //we move to qpel precision anyway |
|
case 2 : ReferenceF = data->RefH + (mvs.x-1)/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
|
|
default : ReferenceF = data->RefHV + (mvs.x-1)/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
|
|
} |
|
|
|
|
|
switch ( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { |
|
|
case 0 : ReferenceB = data->bRef + b_mvs.x/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
|
|
case 1 : ReferenceB = data->bRefV + b_mvs.x/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
|
|
case 2 : ReferenceB = data->bRefH + (b_mvs.x-1)/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
|
|
default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
|
404 |
} |
} |
405 |
|
ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); |
406 |
|
ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); |
407 |
|
|
408 |
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
409 |
ReferenceF + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
ReferenceF, ReferenceB, |
|
ReferenceB + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
|
410 |
data->iEdgedWidth); |
data->iEdgedWidth); |
411 |
if (sad > *(data->iMinSAD)) return; |
if (sad > *(data->iMinSAD)) return; |
412 |
} |
} |
413 |
|
|
414 |
|
sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000; |
415 |
|
|
416 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
417 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
418 |
data->currentMV->x = x; data->currentMV->y = y; |
data->currentMV->x = x; data->currentMV->y = y; |
429 |
|
|
430 |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
431 |
|
|
|
sad = data->lambda16 * d_mv_bits(x, y, 1); |
|
|
|
|
432 |
mvs.x = data->directmvF[0].x + x; |
mvs.x = data->directmvF[0].x + x; |
433 |
b_mvs.x = ((x == 0) ? |
b_mvs.x = ((x == 0) ? |
434 |
data->directmvB[0].x |
data->directmvB[0].x |
444 |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
445 |
|| ( 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; |
446 |
|
|
447 |
switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
if (!data->qpel) { |
448 |
case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
mvs.x *= 2; mvs.y *= 2; |
449 |
case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
b_mvs.x *= 2; b_mvs.y *= 2; //we move to qpel precision anyway |
|
case 2 : ReferenceF = data->RefH + (mvs.x-1)/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
|
|
default : ReferenceF = data->RefHV + (mvs.x-1)/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
|
450 |
} |
} |
451 |
|
ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data); |
452 |
|
ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); |
453 |
|
|
454 |
switch ( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
455 |
case 0 : ReferenceB = data->bRef + b_mvs.x/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000; |
|
case 1 : ReferenceB = data->bRefV + b_mvs.x/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
|
|
case 2 : ReferenceB = data->bRefH + (b_mvs.x-1)/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
|
|
default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
|
|
} |
|
|
|
|
|
sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
|
456 |
|
|
457 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
458 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
463 |
static void |
static void |
464 |
CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
465 |
{ |
{ |
466 |
int32_t sad; |
int32_t sad; int t; |
467 |
const uint8_t * Reference; |
const uint8_t * Reference; |
468 |
|
|
469 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if (( x > data->max_dx) || ( x < data->min_dx) |
470 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
471 |
|
|
472 |
switch ( ((x&1)<<1) + (y&1) ) |
if (data->qpel) Reference = Interpolate16x16qpel(x, y, 0, data); |
473 |
{ |
else Reference = Interpolate16x16qpel(2*x, 2*y, 0, 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; |
|
|
} |
|
474 |
|
|
475 |
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
476 |
sad += data->lambda8 * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
if (data->qpel) t = d_mv_bits(2 * x - data->predMV.x, 2 * y - data->predMV.y, data->iFcode); |
477 |
|
else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
|
if (sad < *(data->iMinSAD)) { |
|
|
*(data->iMinSAD) = sad; |
|
|
data->currentMV->x = x; data->currentMV->y = y; |
|
|
*dir = Direction; } |
|
|
} |
|
|
|
|
|
static void |
|
|
CheckCandidate8Q(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
|
{ |
|
|
int32_t sad; |
|
|
const uint8_t * Reference; |
|
|
|
|
|
if (( x > data->max_dx) || ( x < data->min_dx) |
|
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|
478 |
|
|
479 |
switch ( ((x&1)<<1) + (y&1) ) |
sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))/100; |
|
{ |
|
|
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; |
|
|
} |
|
|
|
|
|
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
|
|
sad += data->lambda8 * d_mv_bits(2 * x - data->predQMV.x, 2 * y - data->predQMV.y, data->iFcode); |
|
480 |
|
|
481 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
482 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
484 |
*dir = Direction; } |
*dir = Direction; } |
485 |
} |
} |
486 |
|
|
|
static void |
|
|
CheckCandidate8_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
|
// CheckCandidate16no4v variant which expects x and y in quarter pixel resolution |
|
|
// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!) |
|
|
// around currentMV! |
|
|
|
|
|
{ |
|
|
int32_t sad; |
|
|
uint8_t *Reference = (uint8_t *) data->RefQ; |
|
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
|
|
VECTOR halfpelMV = *(data->currentMV); |
|
|
|
|
|
int32_t iEdgedWidth = data->iEdgedWidth; |
|
|
uint32_t rounding = data->rounding; |
|
|
|
|
|
if (( x > data->max_dx) || ( x < data->min_dx) |
|
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|
|
|
|
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); |
|
|
switch( ((x&1)<<1) + (y&1) ) |
|
|
{ |
|
|
case 0: // pure halfpel position - shouldn't happen during a refinement step |
|
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, Reference); |
|
|
break; |
|
|
|
|
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
|
|
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
|
|
|
|
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
|
|
break; |
|
|
|
|
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
|
|
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); |
|
|
|
|
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
|
|
break; |
|
|
|
|
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
|
|
// bottom left/right) during qpel refinement |
|
|
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
|
|
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); |
|
|
GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); |
|
|
|
|
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
|
|
break; |
|
|
} |
|
|
|
|
|
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
|
|
sad += data->lambda8 * d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode); |
|
|
|
|
|
if (sad < *(data->iMinSAD)) { |
|
|
*(data->iMinSAD) = sad; |
|
|
data->currentQMV->x = x; data->currentQMV->y = y; |
|
|
*dir = Direction; } |
|
|
} |
|
|
|
|
487 |
/* CHECK_CANDIATE FUNCTIONS END */ |
/* CHECK_CANDIATE FUNCTIONS END */ |
488 |
|
|
489 |
/* MAINSEARCH FUNCTIONS START */ |
/* MAINSEARCH FUNCTIONS START */ |
633 |
/* HALFPELREFINE COULD BE A MAINSEARCH FUNCTION, BUT THERE IS NO NEED FOR IT */ |
/* HALFPELREFINE COULD BE A MAINSEARCH FUNCTION, BUT THERE IS NO NEED FOR IT */ |
634 |
|
|
635 |
static void |
static void |
636 |
HalfpelRefine(const SearchData * const data) |
SubpelRefine(const SearchData * const data) |
637 |
{ |
{ |
638 |
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
/* Do a half-pel or q-pel refinement */ |
639 |
|
VECTOR backupMV; |
|
VECTOR backupMV = *(data->currentMV); |
|
640 |
int iDirection; //not needed |
int iDirection; //not needed |
641 |
|
|
642 |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); |
if (data->qpel_precision) |
643 |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); |
backupMV = *(data->currentQMV); |
644 |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); |
else backupMV = *(data->currentMV); |
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0); |
|
|
|
|
|
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); |
|
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); |
|
|
|
|
|
CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); |
|
|
CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); |
|
|
} |
|
|
|
|
|
|
|
|
static void |
|
|
QuarterpelRefine(const SearchData * const data) |
|
|
{ |
|
|
/* Perform quarter pixel refinement*/ |
|
|
|
|
|
VECTOR backupMV = *(data->currentQMV); |
|
|
int iDirection; //not needed |
|
645 |
|
|
646 |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); |
647 |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); |
653 |
|
|
654 |
CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); |
CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); |
655 |
CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); |
CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); |
|
|
|
656 |
} |
} |
657 |
|
|
658 |
static __inline int |
static __inline int |
705 |
|
|
706 |
uint32_t x, y; |
uint32_t x, y; |
707 |
uint32_t iIntra = 0; |
uint32_t iIntra = 0; |
708 |
int32_t InterBias, quant = current->quant; |
int32_t InterBias, quant = current->quant, sad00; |
709 |
uint8_t *qimage; |
uint8_t *qimage; |
710 |
|
|
711 |
// some pre-initialized thingies for SearchP |
// some pre-initialized thingies for SearchP |
721 |
Data.temp = temp; |
Data.temp = temp; |
722 |
Data.iFcode = current->fcode; |
Data.iFcode = current->fcode; |
723 |
Data.rounding = pParam->m_rounding_type; |
Data.rounding = pParam->m_rounding_type; |
724 |
|
Data.qpel = pParam->m_quarterpel; |
725 |
|
Data.chroma = current->global_flags & XVID_ME_COLOUR; |
726 |
|
|
727 |
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
728 |
return 1; // allocate some mem for qpel interpolated blocks |
return 1; // allocate some mem for qpel interpolated blocks |
729 |
// somehow this is dirty since I think we shouldn't use malloc outside |
// somehow this is dirty since I think we shouldn't use malloc outside |
730 |
// encoder_create() - so please fix me! |
// encoder_create() - so please fix me! |
731 |
|
Data.RefQ = qimage; |
732 |
if (sadInit) (*sadInit) (); |
if (sadInit) (*sadInit) (); |
733 |
|
|
734 |
for (y = 0; y < pParam->mb_height; y++) { |
for (y = 0; y < pParam->mb_height; y++) { |
735 |
for (x = 0; x < pParam->mb_width; x++) { |
for (x = 0; x < pParam->mb_width; x++) { |
|
|
|
736 |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
737 |
int32_t sad00 = pMB->sad16 |
|
738 |
|
pMB->sad16 |
739 |
= sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, |
= sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, |
740 |
pRef->y + (x + y * pParam->edged_width) * 16, |
pRef->y + (x + y * pParam->edged_width) * 16, |
741 |
pParam->edged_width, pMB->sad8 ); |
pParam->edged_width, pMB->sad8 ); |
742 |
|
|
743 |
|
if (Data.chroma) { |
744 |
|
pMB->sad16 += sad8(pCurrent->u + x*8 + y*(pParam->edged_width/2)*8, |
745 |
|
pRef->u + x*8 + y*(pParam->edged_width/2)*8, pParam->edged_width/2); |
746 |
|
|
747 |
|
pMB->sad16 += sad8(pCurrent->v + (x + y*(pParam->edged_width/2))*8, |
748 |
|
pRef->v + (x + y*(pParam->edged_width/2))*8, pParam->edged_width/2); |
749 |
|
} |
750 |
|
|
751 |
|
sad00 = pMB->sad16; //if no gmc; else sad00 = (..) |
752 |
|
|
753 |
if (!(current->global_flags & XVID_LUMIMASKING)) { |
if (!(current->global_flags & XVID_LUMIMASKING)) { |
754 |
pMB->dquant = NO_CHANGE; |
pMB->dquant = NO_CHANGE; |
755 |
pMB->quant = current->quant; |
pMB->quant = current->quant; |
763 |
} |
} |
764 |
|
|
765 |
//initial skip decision |
//initial skip decision |
766 |
|
/* no early skip for GMC (global vector = skip vector is unknown!) */ |
767 |
|
if (current->coding_type == P_VOP) { /* no fast SKIP for S(GMC)-VOPs */ |
768 |
if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH) |
if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH) |
769 |
if (SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) { |
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) { |
770 |
SkipMacroblockP(pMB, sad00); |
SkipMacroblockP(pMB, sad00); |
771 |
continue; |
continue; |
772 |
} |
} |
773 |
|
} |
774 |
|
|
775 |
SearchP(pRef->y, pRefH->y, pRefV->y, pRefHV->y, qimage, pCurrent, x, |
SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
776 |
y, current->motion_flags, pMB->quant, |
y, current->motion_flags, pMB->quant, |
777 |
&Data, pParam, pMBs, reference->mbs, |
&Data, pParam, pMBs, reference->mbs, |
778 |
current->global_flags & XVID_INTER4V, pMB); |
current->global_flags & XVID_INTER4V, pMB); |
779 |
|
|
780 |
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
781 |
if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * MAX_SAD00_FOR_SKIP |
if (current->coding_type == P_VOP) { |
782 |
|
if ( (pMB->dquant == NO_CHANGE) && (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
783 |
&& ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) ) |
&& ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) ) |
784 |
if (SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) { |
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) { |
785 |
SkipMacroblockP(pMB, sad00); |
SkipMacroblockP(pMB, sad00); |
786 |
continue; |
continue; |
787 |
} |
} |
788 |
|
} |
789 |
|
|
790 |
/* finally, intra decision */ |
/* finally, intra decision */ |
791 |
|
|
792 |
InterBias = MV16_INTER_BIAS; |
InterBias = MV16_INTER_BIAS; |
793 |
if (pMB->quant > 8) InterBias += 50 * (pMB->quant - 8); // to make high quants work |
if (pMB->quant > 8) InterBias += 100 * (pMB->quant - 8); // to make high quants work |
794 |
if (y != 0) |
if (y != 0) |
795 |
if ((pMB - pParam->mb_width)->mode == MODE_INTER ) InterBias -= 50; |
if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; |
796 |
if (x != 0) |
if (x != 0) |
797 |
if ((pMB - 1)->mode == MODE_INTER ) InterBias -= 50; |
if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; |
798 |
|
|
799 |
|
if (Data.chroma) InterBias += 50; // to compensate bigger SAD |
800 |
|
|
801 |
if (InterBias < pMB->sad16) { |
if (InterBias < pMB->sad16) { |
802 |
const int32_t deviation = |
const int32_t deviation = |
817 |
} |
} |
818 |
} |
} |
819 |
free(qimage); |
free(qimage); |
820 |
|
|
821 |
|
if (current->coding_type == S_VOP) /* first GMC step only for S(GMC)-VOPs */ |
822 |
|
current->GMC_MV = GlobalMotionEst( pMBs, pParam, current->fcode ); |
823 |
|
else |
824 |
|
current->GMC_MV = zeroMV; |
825 |
|
|
826 |
return 0; |
return 0; |
827 |
} |
} |
828 |
|
|
880 |
} |
} |
881 |
|
|
882 |
static void |
static void |
883 |
SearchP(const uint8_t * const pRef, |
SearchP(const IMAGE * const pRef, |
884 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
885 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
886 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
|
const uint8_t * const pRefQ, |
|
887 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
888 |
const int x, |
const int x, |
889 |
const int y, |
const int y, |
904 |
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, |
905 |
pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); |
pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); |
906 |
|
|
|
Data->predMV = pmv[0]; |
|
|
|
|
907 |
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16; |
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16; |
908 |
Data->Ref = pRef + (x + Data->iEdgedWidth*y)*16; |
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
909 |
|
Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
910 |
|
|
911 |
|
Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16; |
912 |
Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16; |
Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16; |
913 |
Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16; |
Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16; |
914 |
Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16; |
Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16; |
915 |
Data->RefQ = pRefQ; |
Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
916 |
|
Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
917 |
|
|
918 |
Data->lambda16 = lambda_vec16[iQuant]; |
Data->lambda16 = lambda_vec16[iQuant]; |
919 |
Data->lambda8 = lambda_vec8[iQuant]; |
Data->lambda8 = lambda_vec8[iQuant]; |
920 |
|
Data->qpel_precision = 0; |
921 |
|
|
922 |
if (!(MotionFlags & PMV_HALFPEL16)) { |
if (!(MotionFlags & PMV_HALFPEL16)) { |
923 |
Data->min_dx = EVEN(Data->min_dx); |
Data->min_dx = EVEN(Data->min_dx); |
930 |
for(i = 0; i < 5; i++) |
for(i = 0; i < 5; i++) |
931 |
Data->currentMV[i].x = Data->currentMV[i].y = 0; |
Data->currentMV[i].x = Data->currentMV[i].y = 0; |
932 |
|
|
933 |
if (pParam->m_quarterpel) { |
if (pParam->m_quarterpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
934 |
Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
else Data->predMV = pmv[0]; |
|
i = d_mv_bits(Data->predQMV.x, Data->predQMV.y, Data->iFcode); |
|
|
} else i = d_mv_bits(Data->predMV.x, Data->predMV.y, Data->iFcode); |
|
935 |
|
|
936 |
Data->iMinSAD[0] = pMB->sad16 + lambda_vec16[iQuant] * i; |
i = d_mv_bits(Data->predMV.x, Data->predMV.y, Data->iFcode); |
937 |
Data->iMinSAD[1] = pMB->sad8[0] + lambda_vec8[iQuant] * i; |
Data->iMinSAD[0] = pMB->sad16 + (Data->lambda16 * i * pMB->sad16)/1000; |
938 |
|
Data->iMinSAD[1] = pMB->sad8[0] + (Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS))/100; |
939 |
Data->iMinSAD[2] = pMB->sad8[1]; |
Data->iMinSAD[2] = pMB->sad8[1]; |
940 |
Data->iMinSAD[3] = pMB->sad8[2]; |
Data->iMinSAD[3] = pMB->sad8[2]; |
941 |
Data->iMinSAD[4] = pMB->sad8[3]; |
Data->iMinSAD[4] = pMB->sad8[3]; |
949 |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
950 |
prevMBs + x + y * pParam->mb_width); |
prevMBs + x + y * pParam->mb_width); |
951 |
|
|
952 |
if (pParam->m_quarterpel) CheckCandidate = CheckCandidate16Q; |
if (inter4v || Data->chroma) CheckCandidate = CheckCandidate16; |
953 |
else |
else CheckCandidate = CheckCandidate16no4v; //for extra speed |
|
if (inter4v) CheckCandidate = CheckCandidate16; |
|
|
else CheckCandidate = CheckCandidate16no4v; |
|
954 |
|
|
955 |
/* main loop. checking all predictions */ |
/* main loop. checking all predictions */ |
956 |
|
|
1007 |
} |
} |
1008 |
} |
} |
1009 |
|
|
1010 |
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data); |
if (MotionFlags & PMV_HALFPELREFINE16) SubpelRefine(Data); |
1011 |
|
|
1012 |
for(i = 0; i < 5; i++) { |
for(i = 0; i < 5; i++) { |
1013 |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors |
1016 |
|
|
1017 |
if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { |
if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { |
1018 |
|
|
1019 |
CheckCandidate = CheckCandidate16_qpel; |
Data->qpel_precision = 1; |
1020 |
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, |
1021 |
pParam->width, pParam->height, Data->iFcode, 0); |
pParam->width, pParam->height, Data->iFcode, 0); |
1022 |
|
|
1023 |
QuarterpelRefine(Data); |
SubpelRefine(Data); |
1024 |
} |
} |
1025 |
|
|
1026 |
|
if (Data->iMinSAD[0] < (int32_t)iQuant * 30 ) inter4v = 0; |
1027 |
if (inter4v) { |
if (inter4v) { |
1028 |
SearchData Data8; |
SearchData Data8; |
1029 |
Data8.iFcode = Data->iFcode; |
Data8.iFcode = Data->iFcode; |
1030 |
Data8.lambda8 = Data->lambda8; |
Data8.lambda8 = Data->lambda8; |
1031 |
Data8.iEdgedWidth = Data->iEdgedWidth; |
Data8.iEdgedWidth = Data->iEdgedWidth; |
1032 |
Data8.RefQ = Data->RefQ; |
Data8.RefQ = Data->RefQ; |
1033 |
|
Data8.qpel = Data->qpel; |
1034 |
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
1035 |
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); |
1036 |
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); |
1037 |
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); |
1038 |
|
|
1039 |
|
if (Data->chroma) { |
1040 |
|
int sumx, sumy, dx, dy; |
1041 |
|
|
1042 |
|
if(pParam->m_quarterpel) { |
1043 |
|
sumx= pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2; |
1044 |
|
sumy = pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2; |
1045 |
|
} else { |
1046 |
|
sumx = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
1047 |
|
sumy = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
1048 |
|
} |
1049 |
|
dx = (sumx >> 3) + roundtab_76[sumx & 0xf]; |
1050 |
|
dy = (sumy >> 3) + roundtab_76[sumy & 0xf]; |
1051 |
|
|
1052 |
|
Data->iMinSAD[1] += ChromaSAD(dx, dy, Data); |
1053 |
|
} |
1054 |
} |
} |
1055 |
|
|
1056 |
if (!(inter4v) || |
if (!(inter4v) || |
1061 |
pMB->mvs[0] = pMB->mvs[1] |
pMB->mvs[0] = pMB->mvs[1] |
1062 |
= pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
= pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
1063 |
|
|
|
pMB->qmvs[0] = pMB->qmvs[1] |
|
|
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
|
|
|
|
1064 |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
1065 |
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
1066 |
|
|
1067 |
if(pParam->m_quarterpel) { |
if(pParam->m_quarterpel) { |
1068 |
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predQMV.x; |
pMB->qmvs[0] = pMB->qmvs[1] |
1069 |
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predQMV.y; |
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
1070 |
} |
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; |
1071 |
else { |
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; |
1072 |
|
} else { |
1073 |
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
1074 |
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
1075 |
} |
} |
1096 |
Data->currentQMV = OldData->currentQMV + 1 + block; |
Data->currentQMV = OldData->currentQMV + 1 + block; |
1097 |
|
|
1098 |
if(pParam->m_quarterpel) { |
if(pParam->m_quarterpel) { |
1099 |
Data->predQMV = 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); |
1100 |
if (block != 0) *(Data->iMinSAD) += Data->lambda8 * |
if (block != 0) *(Data->iMinSAD) += (Data->lambda8 * |
1101 |
d_mv_bits( Data->currentQMV->x - Data->predQMV.x, |
d_mv_bits( Data->currentQMV->x - Data->predMV.x, |
1102 |
Data->currentQMV->y - Data->predQMV.y, |
Data->currentQMV->y - Data->predMV.y, |
1103 |
Data->iFcode); |
Data->iFcode) * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100; |
|
CheckCandidate = CheckCandidate8Q; |
|
1104 |
} else { |
} else { |
1105 |
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); |
1106 |
if (block != 0) *(Data->iMinSAD) += Data->lambda8 * |
if (block != 0) *(Data->iMinSAD) += (Data->lambda8 * |
1107 |
d_mv_bits( Data->currentMV->x - Data->predMV.x, |
d_mv_bits( Data->currentMV->x - Data->predMV.x, |
1108 |
Data->currentMV->y - Data->predMV.y, |
Data->currentMV->y - Data->predMV.y, |
1109 |
Data->iFcode); |
Data->iFcode) * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100; |
|
CheckCandidate = CheckCandidate8; |
|
1110 |
} |
} |
1111 |
|
|
1112 |
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { |
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { |
1117 |
Data->RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1118 |
|
|
1119 |
Data->Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1120 |
|
Data->qpel_precision = 0; |
1121 |
|
|
1122 |
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, |
1123 |
pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel); |
pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel); |
1124 |
|
CheckCandidate = CheckCandidate8; |
1125 |
|
|
1126 |
if (MotionFlags & PMV_EXTSEARCH8) { |
if (MotionFlags & PMV_EXTSEARCH8) { |
1127 |
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
1142 |
if (MotionFlags & PMV_HALFPELREFINE8) { |
if (MotionFlags & PMV_HALFPELREFINE8) { |
1143 |
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
1144 |
|
|
1145 |
HalfpelRefine(Data); // perform halfpel refine of current best vector |
SubpelRefine(Data); // perform halfpel refine of current best vector |
1146 |
|
|
1147 |
if(*(Data->iMinSAD) < temp_sad) { // we have found a better match |
if(*(Data->iMinSAD) < temp_sad) { // we have found a better match |
1148 |
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
1153 |
if(pParam->m_quarterpel) { |
if(pParam->m_quarterpel) { |
1154 |
if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && |
if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && |
1155 |
(MotionFlags & PMV_QUARTERPELREFINE8)) { |
(MotionFlags & PMV_QUARTERPELREFINE8)) { |
1156 |
CheckCandidate = CheckCandidate8_qpel; |
Data->qpel_precision = 1; |
1157 |
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, |
1158 |
pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel); |
pParam->width, pParam->height, OldData->iFcode, 0); |
1159 |
QuarterpelRefine(Data); |
SubpelRefine(Data); |
1160 |
} |
} |
1161 |
} |
} |
1162 |
} |
} |
1163 |
|
|
1164 |
if(pParam->m_quarterpel) { |
if(pParam->m_quarterpel) { |
1165 |
pMB->pmvs[block].x = Data->currentQMV->x - Data->predQMV.x; |
pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x; |
1166 |
pMB->pmvs[block].y = Data->currentQMV->y - Data->predQMV.y; |
pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y; |
1167 |
|
pMB->qmvs[block] = *(Data->currentQMV); |
1168 |
} |
} |
1169 |
else { |
else { |
1170 |
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
1172 |
} |
} |
1173 |
|
|
1174 |
pMB->mvs[block] = *(Data->currentMV); |
pMB->mvs[block] = *(Data->currentMV); |
|
pMB->qmvs[block] = *(Data->currentQMV); |
|
|
|
|
1175 |
pMB->sad8[block] = 4 * (*Data->iMinSAD); |
pMB->sad8[block] = 4 * (*Data->iMinSAD); |
1176 |
} |
} |
1177 |
|
|
1249 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1250 |
*Data->iMinSAD = MV_MAX_ERROR; |
*Data->iMinSAD = MV_MAX_ERROR; |
1251 |
Data->iFcode = iFcode; |
Data->iFcode = iFcode; |
1252 |
|
Data->qpel_precision = 0; |
1253 |
|
|
1254 |
Data->Ref = pRef + (x + y * iEdgedWidth) * 16; |
Data->Ref = pRef + (x + y * iEdgedWidth) * 16; |
1255 |
Data->RefH = pRefH + (x + y * iEdgedWidth) * 16; |
Data->RefH = pRefH + (x + y * iEdgedWidth) * 16; |
1262 |
pParam->width, pParam->height, iFcode, pParam->m_quarterpel); |
pParam->width, pParam->height, iFcode, pParam->m_quarterpel); |
1263 |
|
|
1264 |
pmv[0] = Data->predMV; |
pmv[0] = Data->predMV; |
1265 |
|
if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } |
1266 |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
1267 |
|
|
1268 |
Data->currentMV->x = Data->currentMV->y = 0; |
Data->currentMV->x = Data->currentMV->y = 0; |
|
|
|
1269 |
CheckCandidate = CheckCandidate16no4v; |
CheckCandidate = CheckCandidate16no4v; |
1270 |
|
|
1271 |
// main loop. checking all predictions |
// main loop. checking all predictions |
1282 |
|
|
1283 |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1284 |
|
|
1285 |
HalfpelRefine(Data); |
SubpelRefine(Data); |
1286 |
|
|
1287 |
|
if (Data->qpel) { |
1288 |
|
Data->currentQMV->x = 2*Data->currentMV->x; |
1289 |
|
Data->currentQMV->y = 2*Data->currentMV->y; |
1290 |
|
Data->qpel_precision = 1; |
1291 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1292 |
|
pParam->width, pParam->height, iFcode, 0); |
1293 |
|
SubpelRefine(Data); |
1294 |
|
} |
1295 |
|
|
1296 |
// three bits are needed to code backward mode. four for forward |
// three bits are needed to code backward mode. four for forward |
1297 |
// we treat the bits just like they were vector's |
// we treat the bits just like they were vector's |
1298 |
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16 * 2; |
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
1299 |
else *Data->iMinSAD += 3 * Data->lambda16 * 2; |
else *Data->iMinSAD += 3 * Data->lambda16; |
1300 |
|
|
1301 |
if (*Data->iMinSAD < *best_sad) { |
if (*Data->iMinSAD < *best_sad) { |
1302 |
*best_sad = *Data->iMinSAD; |
*best_sad = *Data->iMinSAD; |
1303 |
pMB->mode = mode_current; |
pMB->mode = mode_current; |
1304 |
|
if (Data->qpel) { |
1305 |
|
pMB->pmvs[0].x = Data->currentQMV->x - predMV->x; |
1306 |
|
pMB->pmvs[0].y = Data->currentQMV->y - predMV->y; |
1307 |
|
if (mode_current == MODE_FORWARD) |
1308 |
|
pMB->qmvs[0] = *Data->currentQMV; |
1309 |
|
else |
1310 |
|
pMB->b_qmvs[0] = *Data->currentQMV; |
1311 |
|
} else { |
1312 |
pMB->pmvs[0].x = Data->currentMV->x - predMV->x; |
pMB->pmvs[0].x = Data->currentMV->x - predMV->x; |
1313 |
pMB->pmvs[0].y = Data->currentMV->y - predMV->y; |
pMB->pmvs[0].y = Data->currentMV->y - predMV->y; |
1314 |
if (mode_current == MODE_FORWARD) pMB->mvs[0] = *(Data->currentMV+2) = *Data->currentMV; |
} |
1315 |
else pMB->b_mvs[0] = *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search |
if (mode_current == MODE_FORWARD) |
1316 |
|
pMB->mvs[0] = *(Data->currentMV+2) = *Data->currentMV; |
1317 |
|
else |
1318 |
|
pMB->b_mvs[0] = *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search |
1319 |
|
|
1320 |
} |
} |
1321 |
|
|
1322 |
} |
} |
1347 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1348 |
|
|
1349 |
*Data->iMinSAD = 256*4096; |
*Data->iMinSAD = 256*4096; |
|
Data->referencemv = b_mb->mvs; |
|
1350 |
|
|
1351 |
Data->Ref = f_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
Data->Ref = f_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
1352 |
Data->RefH = f_RefH + (x + Data->iEdgedWidth*y) * 16; |
Data->RefH = f_RefH + (x + Data->iEdgedWidth*y) * 16; |
1361 |
Data->max_dy = 2 * pParam->height - 2 * (y) * 16; |
Data->max_dy = 2 * pParam->height - 2 * (y) * 16; |
1362 |
Data->min_dx = -(2 * 16 + 2 * (x) * 16); |
Data->min_dx = -(2 * 16 + 2 * (x) * 16); |
1363 |
Data->min_dy = -(2 * 16 + 2 * (y) * 16); |
Data->min_dy = -(2 * 16 + 2 * (y) * 16); |
1364 |
|
if (Data->qpel) { //we measure in qpixels |
1365 |
|
Data->max_dx *= 2; |
1366 |
|
Data->max_dy *= 2; |
1367 |
|
Data->min_dx *= 2; |
1368 |
|
Data->min_dy *= 2; |
1369 |
|
Data->referencemv = b_mb->qmvs; |
1370 |
|
} else Data->referencemv = b_mb->mvs; |
1371 |
|
Data->qpel_precision = 0; // it'm a trick. it's 1 not 0, but we need 0 here |
1372 |
|
|
1373 |
for (k = 0; k < 4; k++) { |
for (k = 0; k < 4; k++) { |
1374 |
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); |
1393 |
} |
} |
1394 |
} |
} |
1395 |
|
|
1396 |
if (b_mb->mode == MODE_INTER4V) |
|
1397 |
CheckCandidate = CheckCandidateDirect; |
if (b_mb->mode == MODE_INTER4V) CheckCandidate = CheckCandidateDirect; |
1398 |
else CheckCandidate = CheckCandidateDirectno4v; |
else CheckCandidate = CheckCandidateDirectno4v; |
1399 |
|
|
1400 |
(*CheckCandidate)(0, 0, 255, &k, Data); |
(*CheckCandidate)(0, 0, 255, &k, Data); |
1401 |
|
|
1402 |
// skip decision |
// skip decision |
1403 |
if (*Data->iMinSAD - 2 * Data->lambda16 < (uint32_t)pMB->quant * SKIP_THRESH_B) { |
if (*Data->iMinSAD < pMB->quant * SKIP_THRESH_B) { |
1404 |
//possible skip - checking chroma. everything copied from MC |
//possible skip - checking chroma. everything copied from MC |
1405 |
//this is not full chroma compensation, only it's fullpel approximation. should work though |
//this is not full chroma compensation, only it's fullpel approximation. should work though |
1406 |
int sum, dx, dy, b_dx, b_dy; |
int sum, dx, dy, b_dx, b_dy; |
1407 |
|
|
1408 |
|
if (Data->qpel) { |
1409 |
|
sum = pMB->mvs[0].y/2 + pMB->mvs[1].y/2 + pMB->mvs[2].y/2 + pMB->mvs[3].y/2; |
1410 |
|
dy = (sum >> 3) + roundtab_76[sum & 0xf]; |
1411 |
|
sum = pMB->mvs[0].x/2 + pMB->mvs[1].x/2 + pMB->mvs[2].x/2 + pMB->mvs[3].x/2; |
1412 |
|
dx = (sum >> 3) + roundtab_76[sum & 0xf]; |
1413 |
|
|
1414 |
|
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; |
1415 |
|
b_dy = (sum >> 3) + roundtab_76[sum & 0xf]; |
1416 |
|
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; |
1417 |
|
b_dx = (sum >> 3) + roundtab_76[sum & 0xf]; |
1418 |
|
|
1419 |
|
} else { |
1420 |
sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
1421 |
dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
|
|
1422 |
sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
1423 |
dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
1424 |
|
|
1425 |
sum = pMB->b_mvs[0].x + pMB->b_mvs[1].x + pMB->b_mvs[2].x + pMB->b_mvs[3].x; |
sum = pMB->b_mvs[0].x + pMB->b_mvs[1].x + pMB->b_mvs[2].x + pMB->b_mvs[3].x; |
1426 |
b_dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
b_dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
|
|
1427 |
sum = pMB->b_mvs[0].y + pMB->b_mvs[1].y + pMB->b_mvs[2].y + pMB->b_mvs[3].y; |
sum = pMB->b_mvs[0].y + pMB->b_mvs[1].y + pMB->b_mvs[2].y + pMB->b_mvs[3].y; |
1428 |
b_dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
b_dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
1429 |
|
} |
1430 |
sum = sad8bi(pCur->u + 8*x + 8*y*(Data->iEdgedWidth/2), |
sum = sad8bi(pCur->u + 8*x + 8*y*(Data->iEdgedWidth/2), |
1431 |
f_Ref->u + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
f_Ref->u + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
1432 |
b_Ref->u + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
b_Ref->u + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
1453 |
|
|
1454 |
(*MainSearchPtr)(0, 0, Data, 255); |
(*MainSearchPtr)(0, 0, Data, 255); |
1455 |
|
|
1456 |
HalfpelRefine(Data); |
SubpelRefine(Data); |
1457 |
|
|
1458 |
*Data->iMinSAD += 1 * Data->lambda16 * 2; // one bit is needed to code direct mode |
*Data->iMinSAD += 1 * Data->lambda16; // one bit is needed to code direct mode |
1459 |
*best_sad = *Data->iMinSAD; |
*best_sad = *Data->iMinSAD; |
1460 |
|
|
1461 |
if (b_mb->mode == MODE_INTER4V) |
// if (b_mb->mode == MODE_INTER4V) |
1462 |
pMB->mode = MODE_DIRECT; |
pMB->mode = MODE_DIRECT; |
1463 |
else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
// else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
1464 |
|
|
1465 |
pMB->pmvs[3] = *Data->currentMV; |
pMB->pmvs[3] = *Data->currentMV; |
1466 |
|
|
1473 |
pMB->b_mvs[k].y = ((Data->currentMV->y == 0) |
pMB->b_mvs[k].y = ((Data->currentMV->y == 0) |
1474 |
? Data->directmvB[k].y |
? Data->directmvB[k].y |
1475 |
: pMB->mvs[k].y - Data->referencemv[k].y); |
: pMB->mvs[k].y - Data->referencemv[k].y); |
1476 |
|
if (Data->qpel) { |
1477 |
|
pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2; |
1478 |
|
pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2; |
1479 |
|
pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2; |
1480 |
|
pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2; |
1481 |
|
} |
1482 |
|
|
1483 |
if (b_mb->mode != MODE_INTER4V) { |
if (b_mb->mode != MODE_INTER4V) { |
1484 |
pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; |
pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; |
1485 |
pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; |
pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; |
1486 |
|
pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0]; |
1487 |
|
pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0]; |
1488 |
break; |
break; |
1489 |
} |
} |
1490 |
} |
} |
1516 |
{ |
{ |
1517 |
|
|
1518 |
const int32_t iEdgedWidth = pParam->edged_width; |
const int32_t iEdgedWidth = pParam->edged_width; |
|
|
|
1519 |
int iDirection, i, j; |
int iDirection, i, j; |
1520 |
SearchData bData; |
SearchData bData; |
1521 |
|
|
1522 |
*(bData.iMinSAD = fData->iMinSAD) = 4096*256; |
*(bData.iMinSAD = fData->iMinSAD) = 4096*256; |
1523 |
bData.Cur = fData->Cur; |
bData.Cur = fData->Cur; |
1524 |
fData->iEdgedWidth = bData.iEdgedWidth = iEdgedWidth; |
fData->iEdgedWidth = bData.iEdgedWidth = iEdgedWidth; |
1525 |
bData.currentMV = fData->currentMV + 1; |
bData.currentMV = fData->currentMV + 1; bData.currentQMV = fData->currentQMV + 1; |
1526 |
bData.lambda16 = fData->lambda16; |
bData.lambda16 = fData->lambda16; |
1527 |
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
1528 |
|
|
1534 |
bData.RefH = fData->bRefH = b_RefH + (x + y * iEdgedWidth) * 16; |
bData.RefH = fData->bRefH = b_RefH + (x + y * iEdgedWidth) * 16; |
1535 |
bData.RefV = fData->bRefV = b_RefV + (x + y * iEdgedWidth) * 16; |
bData.RefV = fData->bRefV = b_RefV + (x + y * iEdgedWidth) * 16; |
1536 |
bData.RefHV = fData->bRefHV = b_RefHV + (x + y * iEdgedWidth) * 16; |
bData.RefHV = fData->bRefHV = b_RefHV + (x + y * iEdgedWidth) * 16; |
1537 |
|
bData.RefQ = fData->RefQ; |
1538 |
|
fData->qpel_precision = bData.qpel_precision = 0; |
1539 |
|
bData.rounding = 0; |
1540 |
|
|
1541 |
bData.bpredMV = fData->predMV = *f_predMV; |
bData.bpredMV = fData->predMV = *f_predMV; |
1542 |
fData->bpredMV = bData.predMV = *b_predMV; |
fData->bpredMV = bData.predMV = *b_predMV; |
1543 |
|
|
1544 |
fData->currentMV[0] = fData->currentMV[3]; //forward search stored it's vector here. backward stored it in the place it's needed |
fData->currentMV[0] = fData->currentMV[2]; |
1545 |
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); |
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); |
1546 |
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(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, pParam->m_quarterpel); |
1547 |
|
|
1572 |
// backward MV moves |
// backward MV moves |
1573 |
i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
1574 |
fData->currentMV[2] = fData->currentMV[0]; |
fData->currentMV[2] = fData->currentMV[0]; |
|
|
|
1575 |
CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); |
CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); |
1576 |
CheckCandidateInt(i, j + 1, 0, &iDirection, &bData); |
CheckCandidateInt(i, j + 1, 0, &iDirection, &bData); |
1577 |
CheckCandidateInt(i - 1, j, 0, &iDirection, &bData); |
CheckCandidateInt(i - 1, j, 0, &iDirection, &bData); |
1579 |
|
|
1580 |
} while (!(iDirection)); |
} while (!(iDirection)); |
1581 |
|
|
1582 |
// two bits are needed to code interpolate mode. we treat the bits just like they were vector's |
*fData->iMinSAD += 2 * fData->lambda16; // two bits are needed to code interpolate mode. |
1583 |
*fData->iMinSAD += 2 * fData->lambda16 * 2; |
|
1584 |
|
if (fData->qpel) { |
1585 |
|
fData->qpel_precision = bData.qpel_precision = 1; |
1586 |
|
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 0); |
1587 |
|
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 0); |
1588 |
|
fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x; |
1589 |
|
fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y; |
1590 |
|
fData->currentQMV[1].x = 2 * fData->currentMV[1].x; |
1591 |
|
fData->currentQMV[1].y = 2 * fData->currentMV[1].y; |
1592 |
|
SubpelRefine(fData); |
1593 |
|
fData->currentQMV[2] = fData->currentQMV[0]; |
1594 |
|
SubpelRefine(&bData); |
1595 |
|
} |
1596 |
|
|
1597 |
if (*fData->iMinSAD < *best_sad) { |
if (*fData->iMinSAD < *best_sad) { |
1598 |
*best_sad = *fData->iMinSAD; |
*best_sad = *fData->iMinSAD; |
1599 |
pMB->mvs[0] = fData->currentMV[0]; |
pMB->mvs[0] = fData->currentMV[0]; |
1600 |
pMB->b_mvs[0] = fData->currentMV[1]; |
pMB->b_mvs[0] = fData->currentMV[1]; |
1601 |
pMB->mode = MODE_INTERPOLATE; |
pMB->mode = MODE_INTERPOLATE; |
1602 |
|
if (fData->qpel) { |
1603 |
|
pMB->qmvs[0] = fData->currentQMV[0]; |
1604 |
|
pMB->b_qmvs[0] = fData->currentQMV[1]; |
1605 |
|
pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x; |
1606 |
|
pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y; |
1607 |
|
pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x; |
1608 |
|
pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y; |
1609 |
|
} else { |
1610 |
pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; |
pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; |
1611 |
pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; |
pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; |
1612 |
pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; |
pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; |
1613 |
pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; |
pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; |
1614 |
} |
} |
1615 |
} |
} |
1616 |
|
} |
1617 |
|
|
1618 |
void |
void |
1619 |
MotionEstimationBVOP(MBParam * const pParam, |
MotionEstimationBVOP(MBParam * const pParam, |
1627 |
const IMAGE * const f_refV, |
const IMAGE * const f_refV, |
1628 |
const IMAGE * const f_refHV, |
const IMAGE * const f_refHV, |
1629 |
// backward (future) reference |
// backward (future) reference |
1630 |
const MACROBLOCK * const b_mbs, |
const FRAMEINFO * const b_reference, |
1631 |
const IMAGE * const b_ref, |
const IMAGE * const b_ref, |
1632 |
const IMAGE * const b_refH, |
const IMAGE * const b_refH, |
1633 |
const IMAGE * const b_refV, |
const IMAGE * const b_refV, |
1637 |
int32_t best_sad, skip_sad; |
int32_t best_sad, skip_sad; |
1638 |
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; |
1639 |
static const VECTOR zeroMV={0,0}; |
static const VECTOR zeroMV={0,0}; |
1640 |
|
const MACROBLOCK * const b_mbs = b_reference->mbs; |
1641 |
|
|
1642 |
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
1643 |
|
|
1644 |
const int32_t TRB = time_pp - time_bp; |
const int32_t TRB = time_pp - time_bp; |
1645 |
const int32_t TRD = time_pp; |
const int32_t TRD = time_pp; |
1646 |
|
uint8_t * qimage; |
1647 |
|
|
1648 |
// some pre-inintialized data for the rest of the search |
// some pre-inintialized data for the rest of the search |
1649 |
|
|
1650 |
SearchData Data; |
SearchData Data; |
1651 |
int32_t iMinSAD; |
int32_t iMinSAD; |
1652 |
VECTOR currentMV[3]; |
VECTOR currentMV[3]; |
1653 |
|
VECTOR currentQMV[3]; |
1654 |
Data.iEdgedWidth = pParam->edged_width; |
Data.iEdgedWidth = pParam->edged_width; |
1655 |
Data.currentMV = currentMV; |
Data.currentMV = currentMV; Data.currentQMV = currentQMV; |
1656 |
Data.iMinSAD = &iMinSAD; |
Data.iMinSAD = &iMinSAD; |
1657 |
Data.lambda16 = lambda_vec16[frame->quant]; |
Data.lambda16 = lambda_vec16[frame->quant]; |
1658 |
|
Data.qpel = pParam->m_quarterpel; |
1659 |
|
Data.rounding = 0; |
1660 |
|
|
1661 |
// note: i==horizontal, j==vertical |
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
1662 |
|
return; // allocate some mem for qpel interpolated blocks |
1663 |
|
// somehow this is dirty since I think we shouldn't use malloc outside |
1664 |
|
// encoder_create() - so please fix me! |
1665 |
|
Data.RefQ = qimage; |
1666 |
|
|
1667 |
|
// note: i==horizontal, j==vertical |
1668 |
for (j = 0; j < pParam->mb_height; j++) { |
for (j = 0; j < pParam->mb_height; j++) { |
1669 |
|
|
1670 |
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
1673 |
MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; |
MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; |
1674 |
const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; |
const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; |
1675 |
|
|
1676 |
/* special case, if collocated block is SKIPed: encoding is forward (0,0), cpb=0 without further ado */ |
/* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */ |
1677 |
|
if (b_reference->coding_type != S_VOP) |
1678 |
if (b_mb->mode == MODE_NOT_CODED) { |
if (b_mb->mode == MODE_NOT_CODED) { |
1679 |
pMB->mode = MODE_NOT_CODED; |
pMB->mode = MODE_NOT_CODED; |
1680 |
continue; |
continue; |
1682 |
|
|
1683 |
Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
1684 |
pMB->quant = frame->quant; |
pMB->quant = frame->quant; |
1685 |
|
|
1686 |
/* direct search comes first, because it (1) checks for SKIP-mode |
/* direct search comes first, because it (1) checks for SKIP-mode |
1687 |
and (2) sets very good predictions for forward and backward search */ |
and (2) sets very good predictions for forward and backward search */ |
|
|
|
1688 |
skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
1689 |
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
1690 |
&frame->image, |
&frame->image, |
1698 |
|
|
1699 |
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
1700 |
|
|
|
// best_sad = 256*4096; //uncomment to disable Directsearch. |
|
|
// To disable any other mode, just comment the function call |
|
|
|
|
1701 |
// forward search |
// forward search |
1702 |
SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
1703 |
&frame->image, i, j, |
&frame->image, i, j, |
1730 |
switch (pMB->mode) { |
switch (pMB->mode) { |
1731 |
case MODE_FORWARD: |
case MODE_FORWARD: |
1732 |
f_count++; |
f_count++; |
1733 |
f_predMV = pMB->mvs[0]; |
if (pParam->m_quarterpel) f_predMV = pMB->qmvs[0]; |
1734 |
|
else f_predMV = pMB->mvs[0]; |
1735 |
break; |
break; |
1736 |
case MODE_BACKWARD: |
case MODE_BACKWARD: |
1737 |
b_count++; |
b_count++; |
1738 |
b_predMV = pMB->b_mvs[0]; |
if (pParam->m_quarterpel) b_predMV = pMB->b_qmvs[0]; |
1739 |
|
else b_predMV = pMB->b_mvs[0]; |
1740 |
break; |
break; |
1741 |
case MODE_INTERPOLATE: |
case MODE_INTERPOLATE: |
1742 |
i_count++; |
i_count++; |
1743 |
|
if (pParam->m_quarterpel) { |
1744 |
|
f_predMV = pMB->qmvs[0]; |
1745 |
|
b_predMV = pMB->b_qmvs[0]; |
1746 |
|
} else { |
1747 |
f_predMV = pMB->mvs[0]; |
f_predMV = pMB->mvs[0]; |
1748 |
b_predMV = pMB->b_mvs[0]; |
b_predMV = pMB->b_mvs[0]; |
1749 |
|
} |
1750 |
break; |
break; |
1751 |
case MODE_DIRECT: |
case MODE_DIRECT: |
1752 |
case MODE_DIRECT_NO4V: |
case MODE_DIRECT_NO4V: |
1757 |
} |
} |
1758 |
} |
} |
1759 |
} |
} |
1760 |
|
free(qimage); |
|
// fprintf(debug,"B-Stat: F: %04d B: %04d I: %04d D: %04d, N: %04d\n", |
|
|
// f_count,b_count,i_count,d_count,n_count); |
|
|
|
|
1761 |
} |
} |
1762 |
|
|
1763 |
/* Hinted ME starts here */ |
/* Hinted ME starts here */ |
1764 |
|
|
1765 |
static void |
static void |
1766 |
Search8hinted(const SearchData * const OldData, |
SearchPhinted ( const IMAGE * const pRef, |
|
const int x, const int y, |
|
|
const uint32_t MotionFlags, |
|
|
const MBParam * const pParam, |
|
|
MACROBLOCK * const pMB, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const int block, |
|
|
SearchData * const Data) |
|
|
{ |
|
|
int32_t temp_sad; |
|
|
MainSearchFunc *MainSearchPtr; |
|
|
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
|
|
Data->currentMV = OldData->currentMV + 1 + block; |
|
|
Data->currentQMV = OldData->currentQMV + 1 + block; |
|
|
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
|
|
|
|
|
if(pParam->m_quarterpel) { |
|
|
Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
|
|
if (block != 0) *(Data->iMinSAD) += Data->lambda8 * |
|
|
d_mv_bits( Data->currentQMV->x - Data->predQMV.x, |
|
|
Data->currentQMV->y - Data->predQMV.y, |
|
|
Data->iFcode); |
|
|
CheckCandidate = CheckCandidate8Q; |
|
|
} else { |
|
|
if (block != 0) *(Data->iMinSAD) += Data->lambda8 * |
|
|
d_mv_bits( Data->currentMV->x - Data->predMV.x, |
|
|
Data->currentMV->y - Data->predMV.y, |
|
|
Data->iFcode); |
|
|
CheckCandidate = CheckCandidate8; |
|
|
} |
|
|
|
|
|
Data->Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
|
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)); |
|
|
|
|
|
Data->Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
|
|
|
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
|
|
pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel); |
|
|
|
|
|
temp_sad = *(Data->iMinSAD); // store current MinSAD |
|
|
|
|
|
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
|
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
|
|
else MainSearchPtr = DiamondSearch; |
|
|
|
|
|
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
|
|
|
|
|
if(*(Data->iMinSAD) < temp_sad) { |
|
|
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
|
|
Data->currentQMV->y = 2 * Data->currentMV->y; |
|
|
} |
|
|
|
|
|
if (MotionFlags & PMV_HALFPELREFINE8) { |
|
|
temp_sad = *(Data->iMinSAD); // store current MinSAD |
|
|
|
|
|
HalfpelRefine(Data); // perform halfpel refine of current best vector |
|
|
|
|
|
if(*(Data->iMinSAD) < temp_sad) { // we have found a better match |
|
|
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
|
|
Data->currentQMV->y = 2 * Data->currentMV->y; |
|
|
} |
|
|
} |
|
|
|
|
|
if(pParam->m_quarterpel) { |
|
|
if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && |
|
|
(MotionFlags & PMV_QUARTERPELREFINE8)) { |
|
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
|
|
pParam->width, pParam->height, Data->iFcode, 0); |
|
|
CheckCandidate = CheckCandidate8_qpel; |
|
|
QuarterpelRefine(Data); |
|
|
} |
|
|
pMB->pmvs[block].x = Data->currentQMV->x - Data->predQMV.x; |
|
|
pMB->pmvs[block].y = Data->currentQMV->y - Data->predQMV.y; |
|
|
} else { |
|
|
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
|
|
pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; |
|
|
} |
|
|
|
|
|
pMB->mvs[block] = *(Data->currentMV); |
|
|
pMB->qmvs[block] = *(Data->currentQMV); |
|
|
|
|
|
pMB->sad8[block] = 4 * (*Data->iMinSAD); |
|
|
} |
|
|
|
|
|
static void |
|
|
SearchPhinted ( const uint8_t * const pRef, |
|
1767 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1768 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
1769 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
1779 |
SearchData * const Data) |
SearchData * const Data) |
1780 |
{ |
{ |
1781 |
|
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
|
|
|
1782 |
int i, t; |
int i, t; |
1783 |
MainSearchFunc * MainSearchPtr; |
MainSearchFunc * MainSearchPtr; |
1784 |
|
|
|
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
|
1785 |
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, |
1786 |
pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); |
pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); |
1787 |
|
|
1788 |
Data->Cur = pCur->y + (x + y * iEdgedWidth) * 16; |
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16; |
1789 |
Data->Ref = pRef + (x + iEdgedWidth*y)*16; |
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
1790 |
Data->RefH = pRefH + (x + iEdgedWidth*y) * 16; |
Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
1791 |
Data->RefV = pRefV + (x + iEdgedWidth*y) * 16; |
|
1792 |
Data->RefHV = pRefHV + (x + iEdgedWidth*y) * 16; |
Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16; |
1793 |
Data->lambda16 = lambda_vec16[iQuant]; |
Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16; |
1794 |
Data->lambda8 = lambda_vec8[iQuant]; |
Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16; |
1795 |
|
Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16; |
1796 |
|
Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
1797 |
|
Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
1798 |
|
Data->qpel_precision = 0; |
1799 |
|
|
1800 |
if (!(MotionFlags & PMV_HALFPEL16)) { |
if (!(MotionFlags & PMV_HALFPEL16)) { |
1801 |
Data->min_dx = EVEN(Data->min_dx); |
Data->min_dx = EVEN(Data->min_dx); |
1803 |
Data->min_dy = EVEN(Data->min_dy); |
Data->min_dy = EVEN(Data->min_dy); |
1804 |
Data->max_dy = EVEN(Data->max_dy); |
Data->max_dy = EVEN(Data->max_dy); |
1805 |
} |
} |
1806 |
|
if (pParam->m_quarterpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
1807 |
|
else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
1808 |
|
|
1809 |
for(i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
for(i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
1810 |
|
|
1811 |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
1812 |
|
|
1813 |
if(pParam->m_quarterpel) CheckCandidate = CheckCandidate16Q; |
if (inter4v || Data->chroma) CheckCandidate = CheckCandidate16; |
|
else |
|
|
if (inter4v) CheckCandidate = CheckCandidate16; |
|
1814 |
else CheckCandidate = CheckCandidate16no4v; |
else CheckCandidate = CheckCandidate16no4v; |
1815 |
|
|
1816 |
pMB->mvs[0].x = EVEN(pMB->mvs[0].x); |
pMB->mvs[0].x = EVEN(pMB->mvs[0].x); |
1838 |
|
|
1839 |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1840 |
|
|
1841 |
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data); |
if (MotionFlags & PMV_HALFPELREFINE16) SubpelRefine(Data); |
1842 |
|
|
1843 |
for(i = 0; i < 5; i++) { |
for(i = 0; i < 5; i++) { |
1844 |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors |
1848 |
if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { |
if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { |
1849 |
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, |
1850 |
pParam->width, pParam->height, Data->iFcode, 0); |
pParam->width, pParam->height, Data->iFcode, 0); |
1851 |
CheckCandidate = CheckCandidate16_qpel; |
Data->qpel_precision = 1; |
1852 |
QuarterpelRefine(Data); |
SubpelRefine(Data); |
1853 |
} |
} |
1854 |
|
|
1855 |
if (inter4v) { |
if (inter4v) { |
1858 |
Data8.lambda8 = Data->lambda8; |
Data8.lambda8 = Data->lambda8; |
1859 |
Data8.iEdgedWidth = Data->iEdgedWidth; |
Data8.iEdgedWidth = Data->iEdgedWidth; |
1860 |
Data8.RefQ = Data->RefQ; |
Data8.RefQ = Data->RefQ; |
1861 |
Search8hinted(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
Data8.qpel = Data->qpel; |
1862 |
Search8hinted(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); |
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
1863 |
Search8hinted(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); |
Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); |
1864 |
Search8hinted(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); |
Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); |
1865 |
|
Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); |
1866 |
|
|
1867 |
|
if (Data->chroma) { |
1868 |
|
int sumx, sumy, dx, dy; |
1869 |
|
|
1870 |
|
if(pParam->m_quarterpel) { |
1871 |
|
sumx= pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2; |
1872 |
|
sumy = pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2; |
1873 |
|
} else { |
1874 |
|
sumx = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
1875 |
|
sumy = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
1876 |
|
} |
1877 |
|
dx = (sumx >> 3) + roundtab_76[sumx & 0xf]; |
1878 |
|
dy = (sumy >> 3) + roundtab_76[sumy & 0xf]; |
1879 |
|
|
1880 |
|
Data->iMinSAD[1] += ChromaSAD(dx, dy, Data); |
1881 |
|
} |
1882 |
} |
} |
1883 |
|
|
1884 |
if (!(inter4v) || |
if (!(inter4v) || |
1896 |
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
1897 |
|
|
1898 |
if(pParam->m_quarterpel) { |
if(pParam->m_quarterpel) { |
1899 |
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predQMV.x; |
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; |
1900 |
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predQMV.y; |
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; |
1901 |
} |
} else { |
|
else { |
|
1902 |
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
1903 |
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
1904 |
} |
} |
1905 |
} else { |
} else { |
1906 |
// INTER4V MODE; all other things are already set in Search8hinted |
// INTER4V MODE; all other things are already set in Search8 |
1907 |
pMB->mode = MODE_INTER4V; |
pMB->mode = MODE_INTER4V; |
1908 |
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] |
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] |
1909 |
+ Data->iMinSAD[4] + IMV16X16 * iQuant; |
+ Data->iMinSAD[4] + IMV16X16 * iQuant; |
1936 |
Data.temp = temp; |
Data.temp = temp; |
1937 |
Data.iFcode = current->fcode; |
Data.iFcode = current->fcode; |
1938 |
Data.rounding = pParam->m_rounding_type; |
Data.rounding = pParam->m_rounding_type; |
1939 |
|
Data.qpel = pParam->m_quarterpel; |
1940 |
|
Data.chroma = current->global_flags & XVID_ME_COLOUR; |
1941 |
|
|
1942 |
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
1943 |
return; // allocate some mem for qpel interpolated blocks |
return; // allocate some mem for qpel interpolated blocks |
1956 |
//intra mode is copied from the first pass. At least for the time being |
//intra mode is copied from the first pass. At least for the time being |
1957 |
if ((pMB->mode == MODE_INTRA) || (pMB->mode == MODE_NOT_CODED) ) continue; |
if ((pMB->mode == MODE_INTRA) || (pMB->mode == MODE_NOT_CODED) ) continue; |
1958 |
|
|
|
|
|
1959 |
if (!(current->global_flags & XVID_LUMIMASKING)) { |
if (!(current->global_flags & XVID_LUMIMASKING)) { |
1960 |
pMB->dquant = NO_CHANGE; |
pMB->dquant = NO_CHANGE; |
1961 |
pMB->quant = current->quant; } |
pMB->quant = current->quant; } |
1968 |
pMB->quant = quant; |
pMB->quant = quant; |
1969 |
} |
} |
1970 |
|
|
1971 |
SearchPhinted(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
SearchPhinted(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
1972 |
y, current->motion_flags, pMB->quant, |
y, current->motion_flags, pMB->quant, |
1973 |
pParam, pMBs, current->global_flags & XVID_INTER4V, pMB, |
pParam, pMBs, current->global_flags & XVID_INTER4V, pMB, |
1974 |
&Data); |
&Data); |
1991 |
|
|
1992 |
int i = 255, mask; |
int i = 255, mask; |
1993 |
VECTOR pmv[3]; |
VECTOR pmv[3]; |
|
|
|
1994 |
*(Data->iMinSAD) = MV_MAX_ERROR; |
*(Data->iMinSAD) = MV_MAX_ERROR; |
1995 |
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
|
1996 |
|
//median is only used as prediction. it doesn't have to be real |
1997 |
|
if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; |
1998 |
|
else |
1999 |
|
if (x == 1) //left macroblock does not have any vector now |
2000 |
|
Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median |
2001 |
|
else if (y == 1) // top macroblock don't have it's vector |
2002 |
|
Data->predMV = (pMB - 1)->mvs[0]; // left instead of median |
2003 |
|
else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median |
2004 |
|
|
2005 |
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, |
2006 |
pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); |
pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); |
2007 |
|
|
2008 |
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
2009 |
Data->Ref = pRef + (x + y * pParam->edged_width) * 16; |
Data->Ref = pRef + (x + y * pParam->edged_width) * 16; |
2010 |
|
|
|
CheckCandidate = CheckCandidate16no4vI; |
|
|
|
|
2011 |
pmv[1].x = EVEN(pMB->mvs[0].x); |
pmv[1].x = EVEN(pMB->mvs[0].x); |
2012 |
pmv[1].y = EVEN(pMB->mvs[0].y); |
pmv[1].y = EVEN(pMB->mvs[0].y); |
2013 |
pmv[0].x = EVEN(Data->predMV.x); |
pmv[2].x = EVEN(Data->predMV.x); |
2014 |
pmv[0].y = EVEN(Data->predMV.y); |
pmv[2].y = EVEN(Data->predMV.y); |
2015 |
pmv[2].x = pmv[2].y = 0; |
pmv[0].x = pmv[0].y = 0; |
2016 |
|
|
2017 |
|
(*CheckCandidate)(0, 0, 255, &i, Data); |
2018 |
|
|
2019 |
|
//early skip for 0,0 |
2020 |
|
if (*Data->iMinSAD < MAX_SAD00_FOR_SKIP * 4) { |
2021 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
2022 |
|
pMB->mode = MODE_NOT_CODED; |
2023 |
|
return 0; |
2024 |
|
} |
2025 |
|
|
|
CheckCandidate16no4vI(pmv[0].x, pmv[0].y, 255, &i, Data); |
|
2026 |
if (!(mask = make_mask(pmv, 1))) |
if (!(mask = make_mask(pmv, 1))) |
2027 |
CheckCandidate16no4vI(pmv[1].x, pmv[1].y, mask, &i, Data); |
(*CheckCandidate)(pmv[1].x, pmv[1].y, mask, &i, Data); |
2028 |
if (!(mask = make_mask(pmv, 2))) |
if (!(mask = make_mask(pmv, 2))) |
2029 |
CheckCandidate16no4vI(0, 0, mask, &i, Data); |
(*CheckCandidate)(pmv[2].x, pmv[2].y, mask, &i, Data); |
2030 |
|
|
2031 |
|
if (*Data->iMinSAD > MAX_SAD00_FOR_SKIP * 4) // diamond only if needed |
2032 |
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
2033 |
|
|
2034 |
pMB->mvs[0] = *Data->currentMV; |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
2035 |
|
pMB->mode = MODE_INTER; |
2036 |
return *(Data->iMinSAD); |
return *(Data->iMinSAD); |
2037 |
} |
} |
2038 |
|
|
2039 |
#define INTRA_THRESH 1350 |
#define INTRA_THRESH 1350 |
2040 |
#define INTER_THRESH 900 |
#define INTER_THRESH 900 |
2041 |
|
|
2042 |
|
|
2043 |
int |
int |
2044 |
MEanalysis( const IMAGE * const pRef, |
MEanalysis( const IMAGE * const pRef, |
2045 |
const IMAGE * const pCurrent, |
FRAMEINFO * const Current, |
2046 |
MBParam * const pParam, |
MBParam * const pParam, |
2047 |
MACROBLOCK * const pMBs, |
int maxIntra, //maximum number if non-I frames |
2048 |
const uint32_t iFcode) |
int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame |
2049 |
|
int bCount) // number if B frames in a row |
2050 |
{ |
{ |
2051 |
uint32_t x, y, intra = 0; |
uint32_t x, y, intra = 0; |
2052 |
int sSAD = 0; |
int sSAD = 0; |
2053 |
|
MACROBLOCK * const pMBs = Current->mbs; |
2054 |
|
const IMAGE * const pCurrent = &Current->image; |
2055 |
|
int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH; |
2056 |
|
|
2057 |
VECTOR currentMV; |
VECTOR currentMV; |
2058 |
int32_t iMinSAD; |
int32_t iMinSAD; |
2060 |
Data.iEdgedWidth = pParam->edged_width; |
Data.iEdgedWidth = pParam->edged_width; |
2061 |
Data.currentMV = ¤tMV; |
Data.currentMV = ¤tMV; |
2062 |
Data.iMinSAD = &iMinSAD; |
Data.iMinSAD = &iMinSAD; |
2063 |
Data.iFcode = iFcode; |
Data.iFcode = Current->fcode; |
2064 |
|
CheckCandidate = CheckCandidate16no4vI; |
2065 |
|
|
2066 |
|
if (intraCount < 10) // we're right after an I frame |
2067 |
|
IntraThresh += 4 * (intraCount - 10) * (intraCount - 10); |
2068 |
|
else |
2069 |
|
if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec |
2070 |
|
IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; |
2071 |
|
|
2072 |
|
|
2073 |
|
InterThresh += 300 * (1 - bCount); |
2074 |
|
if (InterThresh < 200) InterThresh = 200; |
2075 |
|
|
2076 |
if (sadInit) (*sadInit) (); |
if (sadInit) (*sadInit) (); |
2077 |
|
|
2083 |
sad = MEanalyzeMB(pRef->y, pCurrent->y, x, y, |
sad = MEanalyzeMB(pRef->y, pCurrent->y, x, y, |
2084 |
pParam, pMBs, pMB, &Data); |
pParam, pMBs, pMB, &Data); |
2085 |
|
|
2086 |
if (sad > INTRA_THRESH) { |
if (sad > IntraThresh) { |
2087 |
dev = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
dev = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
2088 |
pParam->edged_width); |
pParam->edged_width); |
2089 |
if (dev + INTRA_THRESH < sad) intra++; |
if (dev + IntraThresh < sad) { |
2090 |
if (intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return 2; // I frame |
pMB->mode = MODE_INTRA; |
2091 |
|
if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return 2; // I frame |
2092 |
|
} |
2093 |
} |
} |
2094 |
sSAD += sad; |
sSAD += sad; |
2095 |
} |
} |
2096 |
} |
} |
2097 |
sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); |
sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); |
2098 |
if (sSAD > INTER_THRESH ) return 1; //P frame |
if (sSAD > InterThresh ) return 1; //P frame |
2099 |
emms(); |
emms(); |
2100 |
return 0; // B frame |
return 0; // B frame |
2101 |
|
|
2130 |
for (i = 1; (max > 32 << (i - 1)); i++); |
for (i = 1; (max > 32 << (i - 1)); i++); |
2131 |
return i; |
return i; |
2132 |
} |
} |
2133 |
|
|
2134 |
|
static void |
2135 |
|
CheckGMC(int x, int y, const int dir, int * iDirection, |
2136 |
|
const MACROBLOCK * const pMBs, uint32_t * bestcount, VECTOR * GMC, |
2137 |
|
const MBParam * const pParam) |
2138 |
|
{ |
2139 |
|
uint32_t mx, my, a, count = 0; |
2140 |
|
|
2141 |
|
for (my = 1; my < pParam->mb_height-1; my++) |
2142 |
|
for (mx = 1; mx < pParam->mb_width-1; mx++) { |
2143 |
|
VECTOR mv; |
2144 |
|
const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; |
2145 |
|
if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) continue; |
2146 |
|
mv = pMB->mvs[0]; |
2147 |
|
a = ABS(mv.x - x) + ABS(mv.y - y); |
2148 |
|
if (a < 6) count += 6 - a; |
2149 |
|
} |
2150 |
|
|
2151 |
|
if (count > *bestcount) { |
2152 |
|
*bestcount = count; |
2153 |
|
*iDirection = dir; |
2154 |
|
GMC->x = x; GMC->y = y; |
2155 |
|
} |
2156 |
|
} |
2157 |
|
|
2158 |
|
|
2159 |
|
static VECTOR |
2160 |
|
GlobalMotionEst(const MACROBLOCK * const pMBs, const MBParam * const pParam, const uint32_t iFcode) |
2161 |
|
{ |
2162 |
|
|
2163 |
|
uint32_t count, bestcount = 0; |
2164 |
|
int x, y; |
2165 |
|
VECTOR gmc = {0,0}; |
2166 |
|
int step, min_x, max_x, min_y, max_y; |
2167 |
|
uint32_t mx, my; |
2168 |
|
int iDirection, bDirection; |
2169 |
|
|
2170 |
|
min_x = min_y = -32<<iFcode; |
2171 |
|
max_x = max_y = 32<<iFcode; |
2172 |
|
|
2173 |
|
//step1: let's find a rough camera panning |
2174 |
|
for (step = 32; step >= 2; step /= 2) { |
2175 |
|
bestcount = 0; |
2176 |
|
for (y = min_y; y <= max_y; y += step) |
2177 |
|
for (x = min_x ; x <= max_x; x += step) { |
2178 |
|
count = 0; |
2179 |
|
//for all macroblocks |
2180 |
|
for (my = 1; my < pParam->mb_height-1; my++) |
2181 |
|
for (mx = 1; mx < pParam->mb_width-1; mx++) { |
2182 |
|
const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; |
2183 |
|
VECTOR mv; |
2184 |
|
|
2185 |
|
if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) |
2186 |
|
continue; |
2187 |
|
|
2188 |
|
mv = pMB->mvs[0]; |
2189 |
|
if ( ABS(mv.x - x) <= step && ABS(mv.y - y) <= step ) /* GMC translation is always halfpel-res */ |
2190 |
|
count++; |
2191 |
|
} |
2192 |
|
if (count >= bestcount) { bestcount = count; gmc.x = x; gmc.y = y; } |
2193 |
|
} |
2194 |
|
min_x = gmc.x - step; |
2195 |
|
max_x = gmc.x + step; |
2196 |
|
min_y = gmc.y - step; |
2197 |
|
max_y = gmc.y + step; |
2198 |
|
|
2199 |
|
} |
2200 |
|
|
2201 |
|
if (bestcount < (pParam->mb_height-2)*(pParam->mb_width-2)/10) |
2202 |
|
gmc.x = gmc.y = 0; //no camara pan, no GMC |
2203 |
|
|
2204 |
|
// step2: let's refine camera panning using gradiend-descent approach. |
2205 |
|
// TODO: more warping points may be evaluated here (like in interpolate mode search - two vectors in one diamond) |
2206 |
|
bestcount = 0; |
2207 |
|
CheckGMC(gmc.x, gmc.y, 255, &iDirection, pMBs, &bestcount, &gmc, pParam); |
2208 |
|
do { |
2209 |
|
x = gmc.x; y = gmc.y; |
2210 |
|
bDirection = iDirection; iDirection = 0; |
2211 |
|
if (bDirection & 1) CheckGMC(x - 1, y, 1+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); |
2212 |
|
if (bDirection & 2) CheckGMC(x + 1, y, 2+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); |
2213 |
|
if (bDirection & 4) CheckGMC(x, y - 1, 1+2+4, &iDirection, pMBs, &bestcount, &gmc, pParam); |
2214 |
|
if (bDirection & 8) CheckGMC(x, y + 1, 1+2+8, &iDirection, pMBs, &bestcount, &gmc, pParam); |
2215 |
|
|
2216 |
|
} while (iDirection); |
2217 |
|
|
2218 |
|
if (pParam->m_quarterpel) { |
2219 |
|
gmc.x *= 2; |
2220 |
|
gmc.y *= 2; /* we store the halfpel value as pseudo-qpel to make comparison easier */ |
2221 |
|
} |
2222 |
|
|
2223 |
|
return gmc; |
2224 |
|
} |