31 |
#include <assert.h> |
#include <assert.h> |
32 |
#include <stdio.h> |
#include <stdio.h> |
33 |
#include <stdlib.h> |
#include <stdlib.h> |
34 |
|
#include <string.h> // memcpy |
35 |
|
|
36 |
#include "../encoder.h" |
#include "../encoder.h" |
37 |
#include "../utils/mbfunctions.h" |
#include "../utils/mbfunctions.h" |
38 |
#include "../prediction/mbprediction.h" |
#include "../prediction/mbprediction.h" |
39 |
#include "../global.h" |
#include "../global.h" |
40 |
#include "../utils/timer.h" |
#include "../utils/timer.h" |
41 |
|
#include "../image/interpolate8x8.h" |
42 |
#include "motion_est.h" |
#include "motion_est.h" |
43 |
#include "motion.h" |
#include "motion.h" |
44 |
#include "sad.h" |
#include "sad.h" |
48 |
#define FINAL_SKIP_THRESH (50) |
#define FINAL_SKIP_THRESH (50) |
49 |
#define MAX_SAD00_FOR_SKIP (20) |
#define MAX_SAD00_FOR_SKIP (20) |
50 |
#define MAX_CHROMA_SAD_FOR_SKIP (22) |
#define MAX_CHROMA_SAD_FOR_SKIP (22) |
|
#define SKIP_THRESH_B (25) |
|
51 |
|
|
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 iDiamondSize 2 |
|
|
|
|
55 |
static __inline int |
static __inline int |
56 |
d_mv_bits(int x, int y, const uint32_t iFcode) |
d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv) |
57 |
{ |
{ |
58 |
int xb, yb; |
int xb, yb; |
59 |
|
if (qpel) { x *= 2; y *= 2;} |
60 |
|
else if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); } |
61 |
|
x = pred.x - x; |
62 |
|
y = pred.y - y; |
63 |
|
|
64 |
if (x == 0) xb = 1; |
if (x == 0) xb = 1; |
65 |
else { |
else { |
81 |
return xb + yb; |
return xb + yb; |
82 |
} |
} |
83 |
|
|
84 |
|
static int32_t |
85 |
|
ChromaSAD(int dx, int dy, const SearchData * const data) |
86 |
|
{ |
87 |
|
int sad; |
88 |
|
dx = (dx >> 1) + roundtab_79[dx & 0x3]; |
89 |
|
dy = (dy >> 1) + roundtab_79[dy & 0x3]; |
90 |
|
|
91 |
|
if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently |
92 |
|
|
93 |
|
switch (((dx & 1) << 1) | (dy & 1)) { |
94 |
|
case 0: |
95 |
|
sad = sad8(data->CurU, data->RefCU + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); |
96 |
|
sad += sad8(data->CurV, data->RefCV + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); |
97 |
|
break; |
98 |
|
case 1: |
99 |
|
dx = dx / 2; dy = (dy - 1) / 2; |
100 |
|
sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); |
101 |
|
sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); |
102 |
|
break; |
103 |
|
case 2: |
104 |
|
dx = (dx - 1) / 2; dy = dy / 2; |
105 |
|
sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); |
106 |
|
sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); |
107 |
|
break; |
108 |
|
default: |
109 |
|
dx = (dx - 1) / 2; dy = (dy - 1) / 2; |
110 |
|
interpolate8x8_halfpel_hv(data->RefQ, |
111 |
|
data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, |
112 |
|
data->rounding); |
113 |
|
sad = sad8(data->CurU, data->RefQ, data->iEdgedWidth/2); |
114 |
|
interpolate8x8_halfpel_hv(data->RefQ, |
115 |
|
data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, |
116 |
|
data->rounding); |
117 |
|
sad += sad8(data->CurV, data->RefQ, data->iEdgedWidth/2); |
118 |
|
break; |
119 |
|
} |
120 |
|
data->temp[5] = dx; data->temp[6] = dy; data->temp[7] = sad; //backup |
121 |
|
return sad; |
122 |
|
} |
123 |
|
|
124 |
|
static __inline const uint8_t * |
125 |
|
GetReference(const int x, const int y, const int dir, const SearchData * const data) |
126 |
|
{ |
127 |
|
// dir : 0 = forward, 1 = backward |
128 |
|
switch ( (dir << 2) | ((x&1)<<1) | (y&1) ) { |
129 |
|
case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth); |
130 |
|
case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); |
131 |
|
case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); |
132 |
|
case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); |
133 |
|
case 4 : return data->bRef + x/2 + (y/2)*(data->iEdgedWidth); |
134 |
|
case 5 : return data->bRefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); |
135 |
|
case 6 : return data->bRefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); |
136 |
|
default : return data->bRefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); |
137 |
|
} |
138 |
|
} |
139 |
|
|
140 |
|
static uint8_t * |
141 |
|
Interpolate8x8qpel(const int x, const int y, const int block, const int dir, const SearchData * const data) |
142 |
|
{ |
143 |
|
// create or find a qpel-precision reference picture; return pointer to it |
144 |
|
uint8_t * Reference = (uint8_t *)data->RefQ + 16*dir; |
145 |
|
const int32_t iEdgedWidth = data->iEdgedWidth; |
146 |
|
const uint32_t rounding = data->rounding; |
147 |
|
const int halfpel_x = x/2; |
148 |
|
const int halfpel_y = y/2; |
149 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
150 |
|
|
151 |
|
ref1 = GetReference(halfpel_x, halfpel_y, dir, data); // this reference is used in all cases |
152 |
|
ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
153 |
|
switch( ((x&1)<<1) + (y&1) ) { |
154 |
|
case 0: // pure halfpel position |
155 |
|
Reference = (uint8_t *) GetReference(halfpel_x, halfpel_y, dir, data); |
156 |
|
Reference += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
157 |
|
break; |
158 |
|
|
159 |
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
160 |
|
ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
161 |
|
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
162 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
163 |
|
break; |
164 |
|
|
165 |
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
166 |
|
ref2 = GetReference(x - halfpel_x, halfpel_y, dir, data); |
167 |
|
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
168 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
169 |
|
break; |
170 |
|
|
171 |
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
172 |
|
// bottom left/right) during qpel refinement |
173 |
|
ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
174 |
|
ref3 = GetReference(x - halfpel_x, halfpel_y, dir, data); |
175 |
|
ref4 = GetReference(x - halfpel_x, y - halfpel_y, dir, data); |
176 |
|
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
177 |
|
ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
178 |
|
ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
179 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
180 |
|
break; |
181 |
|
} |
182 |
|
return Reference; |
183 |
|
} |
184 |
|
|
185 |
|
static uint8_t * |
186 |
|
Interpolate16x16qpel(const int x, const int y, const int dir, const SearchData * const data) |
187 |
|
{ |
188 |
|
// create or find a qpel-precision reference picture; return pointer to it |
189 |
|
uint8_t * Reference = (uint8_t *)data->RefQ + 16*dir; |
190 |
|
const int32_t iEdgedWidth = data->iEdgedWidth; |
191 |
|
const uint32_t rounding = data->rounding; |
192 |
|
const int halfpel_x = x/2; |
193 |
|
const int halfpel_y = y/2; |
194 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
195 |
|
|
196 |
|
ref1 = GetReference(halfpel_x, halfpel_y, dir, data); // this reference is used in all cases |
197 |
|
switch( ((x&1)<<1) + (y&1) ) { |
198 |
|
case 0: // pure halfpel position |
199 |
|
return (uint8_t *) GetReference(halfpel_x, halfpel_y, dir, data); |
200 |
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
201 |
|
ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
202 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
203 |
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
204 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); |
205 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
206 |
|
break; |
207 |
|
|
208 |
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
209 |
|
ref2 = GetReference(x - halfpel_x, halfpel_y, dir, data); |
210 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
211 |
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
212 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); |
213 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
214 |
|
break; |
215 |
|
|
216 |
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
217 |
|
// bottom left/right) during qpel refinement |
218 |
|
ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
219 |
|
ref3 = GetReference(x - halfpel_x, halfpel_y, dir, data); |
220 |
|
ref4 = GetReference(x - halfpel_x, y - halfpel_y, dir, data); |
221 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
222 |
|
interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); |
223 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); |
224 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); |
225 |
|
break; |
226 |
|
} |
227 |
|
return Reference; |
228 |
|
} |
229 |
|
|
230 |
/* CHECK_CANDIATE FUNCTIONS START */ |
/* CHECK_CANDIATE FUNCTIONS START */ |
231 |
|
|
232 |
static void |
static void |
233 |
CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
234 |
{ |
{ |
235 |
int32_t * const sad = data->temp; |
int t, xc, yc; |
|
int t; |
|
236 |
const uint8_t * Reference; |
const uint8_t * Reference; |
237 |
|
VECTOR * current; |
238 |
|
|
239 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if (( x > data->max_dx) || ( x < data->min_dx) |
240 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
241 |
|
|
242 |
switch ( ((x&1)<<1) + (y&1) ) { |
if (data->qpel_precision) { // x and y are in 1/4 precision |
243 |
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
Reference = Interpolate16x16qpel(x, y, 0, data); |
244 |
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
xc = x/2; yc = y/2; //for chroma sad |
245 |
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
current = data->currentQMV; |
246 |
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
} else { |
247 |
|
Reference = GetReference(x, y, 0, data); |
248 |
|
current = data->currentMV; |
249 |
|
xc = x; yc = y; |
250 |
} |
} |
251 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0); |
252 |
|
|
253 |
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, sad+1); |
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
254 |
|
|
255 |
t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; |
256 |
data->temp[0] += lambda_vec16[data->iQuant] * t; |
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; |
257 |
data->temp[1] += lambda_vec8[data->iQuant] * t; |
|
258 |
|
if (data->chroma) data->temp[0] += ChromaSAD(xc, yc, data); |
259 |
|
|
260 |
|
if (data->temp[0] < data->iMinSAD[0]) { |
261 |
|
data->iMinSAD[0] = data->temp[0]; |
262 |
|
current[0].x = x; current[0].y = y; |
263 |
|
*dir = Direction; } |
264 |
|
|
265 |
|
if (data->temp[1] < data->iMinSAD[1]) { |
266 |
|
data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y= y; } |
267 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
268 |
|
data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; } |
269 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
270 |
|
data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; } |
271 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
272 |
|
data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; } |
273 |
|
|
274 |
|
} |
275 |
|
|
276 |
|
static void |
277 |
|
CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
278 |
|
{ |
279 |
|
int t; |
280 |
|
const uint8_t * Reference; |
281 |
|
|
282 |
|
if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero integer value |
283 |
|
( x > data->max_dx) || ( x < data->min_dx) |
284 |
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
285 |
|
|
286 |
|
Reference = GetReference(x, y, 0, data); |
287 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1); |
288 |
|
|
289 |
|
data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
290 |
|
|
291 |
|
data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; |
292 |
|
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; |
293 |
|
|
294 |
if (data->temp[0] < data->iMinSAD[0]) { |
if (data->temp[0] < data->iMinSAD[0]) { |
295 |
data->iMinSAD[0] = data->temp[0]; |
data->iMinSAD[0] = data->temp[0]; |
304 |
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
305 |
if (data->temp[4] < data->iMinSAD[4]) { |
if (data->temp[4] < data->iMinSAD[4]) { |
306 |
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
|
|
|
307 |
} |
} |
308 |
|
|
309 |
static void |
static void |
311 |
{ |
{ |
312 |
int32_t sad; |
int32_t sad; |
313 |
const uint8_t * Reference; |
const uint8_t * Reference; |
314 |
|
int t; |
315 |
|
VECTOR * current; |
316 |
|
|
317 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if (( x > data->max_dx) || ( x < data->min_dx) |
318 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
319 |
|
|
320 |
switch ( ((x&1)<<1) + (y&1) ) |
if (data->rrv) |
321 |
{ |
if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) ) return; //non-zero integer value |
322 |
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
|
323 |
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
if (data->qpel_precision) { // x and y are in 1/4 precision |
324 |
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
Reference = Interpolate16x16qpel(x, y, 0, data); |
325 |
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
current = data->currentQMV; |
326 |
|
} else { |
327 |
|
Reference = GetReference(x, y, 0, data); |
328 |
|
current = data->currentMV; |
329 |
} |
} |
330 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, |
331 |
|
data->qpel && !data->qpel_precision && !data->rrv, data->rrv); |
332 |
|
|
333 |
sad = lambda_vec16[data->iQuant] * |
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
334 |
d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
sad += (data->lambda16 * t * sad)/1000; |
|
sad += sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
|
335 |
|
|
336 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
337 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
338 |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
current->x = x; current->y = y; |
339 |
*dir = Direction; } |
*dir = Direction; } |
340 |
} |
} |
341 |
|
|
342 |
static void |
static void |
343 |
CheckCandidate16no4vI(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
344 |
{ |
{ |
345 |
int32_t sad; |
// maximum speed - for P/B/I decision |
346 |
|
|
347 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if (( x > data->max_dx) || ( x < data->min_dx) |
348 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
349 |
|
|
350 |
sad = lambda_vec16[data->iQuant] * |
data->temp[0] = sad32v_c(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), |
351 |
d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
data->iEdgedWidth, data->temp+1); |
352 |
|
if (data->temp[0] < *(data->iMinSAD)) { |
353 |
sad += sad16(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), |
*(data->iMinSAD) = data->temp[0]; |
|
data->iEdgedWidth, 256*4096); |
|
|
|
|
|
if (sad < *(data->iMinSAD)) { |
|
|
*(data->iMinSAD) = sad; |
|
354 |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
355 |
*dir = Direction; } |
*dir = Direction; } |
356 |
|
if (data->temp[1] < data->iMinSAD[1]) { |
357 |
|
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
358 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
359 |
|
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
360 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
361 |
|
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
362 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
363 |
|
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
364 |
|
|
365 |
} |
} |
366 |
|
|
367 |
|
|
369 |
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) |
370 |
{ |
{ |
371 |
int32_t sad; |
int32_t sad; |
372 |
const int xb = data->currentMV[1].x; |
int xb, yb, t; |
|
const int yb = data->currentMV[1].y; |
|
373 |
const uint8_t *ReferenceF, *ReferenceB; |
const uint8_t *ReferenceF, *ReferenceB; |
374 |
|
VECTOR *current; |
375 |
|
|
376 |
if (( xf > data->max_dx) || ( xf < data->min_dx) |
if (( xf > data->max_dx) || ( xf < data->min_dx) |
377 |
|| ( yf > data->max_dy) || (yf < data->min_dy)) return; |
|| ( yf > data->max_dy) || (yf < data->min_dy)) return; |
378 |
|
|
379 |
switch ( ((xf&1)<<1) + (yf&1) ) { |
if (data->qpel_precision) { |
380 |
case 0 : ReferenceF = data->Ref + xf/2 + (yf/2)*(data->iEdgedWidth); break; |
ReferenceF = Interpolate16x16qpel(xf, yf, 0, data); |
381 |
case 1 : ReferenceF = data->RefV + xf/2 + ((yf-1)/2)*(data->iEdgedWidth); break; |
xb = data->currentQMV[1].x; yb = data->currentQMV[1].y; |
382 |
case 2 : ReferenceF = data->RefH + (xf-1)/2 + (yf/2)*(data->iEdgedWidth); break; |
current = data->currentQMV; |
383 |
default : ReferenceF = data->RefHV + (xf-1)/2 + ((yf-1)/2)*(data->iEdgedWidth); break; |
ReferenceB = Interpolate16x16qpel(xb, yb, 1, data); |
384 |
} |
} else { |
385 |
|
ReferenceF = GetReference(xf, yf, 0, data); |
386 |
switch ( ((xb&1)<<1) + (yb&1) ) { |
xb = data->currentMV[1].x; yb = data->currentMV[1].y; |
387 |
case 0 : ReferenceB = data->bRef + xb/2 + (yb/2)*(data->iEdgedWidth); break; |
ReferenceB = GetReference(xb, yb, 1, data); |
388 |
case 1 : ReferenceB = data->bRefV + xb/2 + ((yb-1)/2)*(data->iEdgedWidth); break; |
current = data->currentMV; |
|
case 2 : ReferenceB = data->bRefH + (xb-1)/2 + (yb/2)*(data->iEdgedWidth); break; |
|
|
default : ReferenceB = data->bRefHV + (xb-1)/2 + ((yb-1)/2)*(data->iEdgedWidth); break; |
|
389 |
} |
} |
390 |
|
|
391 |
sad = lambda_vec16[data->iQuant] * |
t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0) |
392 |
( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) + |
+ d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel && !data->qpel_precision, 0); |
|
d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode) ); |
|
393 |
|
|
394 |
sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
395 |
|
sad += (data->lambda16 * t * sad)/1000; |
396 |
|
|
397 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
398 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
399 |
data->currentMV->x = xf; data->currentMV->y = yf; |
current->x = xf; current->y = yf; |
400 |
*dir = Direction; } |
*dir = Direction; } |
401 |
} |
} |
402 |
|
|
403 |
static void |
static void |
404 |
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) |
405 |
{ |
{ |
406 |
int32_t sad; |
int32_t sad = 0; |
407 |
int k; |
int k; |
408 |
const uint8_t *ReferenceF; |
const uint8_t *ReferenceF; |
409 |
const uint8_t *ReferenceB; |
const uint8_t *ReferenceB; |
410 |
VECTOR mvs, b_mvs; |
VECTOR mvs, b_mvs; |
411 |
|
const VECTOR zeroMV={0,0}; |
412 |
|
|
413 |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
414 |
|
|
|
sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1); |
|
|
|
|
415 |
for (k = 0; k < 4; k++) { |
for (k = 0; k < 4; k++) { |
416 |
mvs.x = data->directmvF[k].x + x; |
mvs.x = data->directmvF[k].x + x; |
417 |
b_mvs.x = ((x == 0) ? |
b_mvs.x = ((x == 0) ? |
428 |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
429 |
|| ( 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; |
430 |
|
|
431 |
switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
if (!data->qpel) { |
432 |
case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
mvs.x *= 2; mvs.y *= 2; |
433 |
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; |
|
434 |
} |
} |
435 |
|
ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); |
436 |
|
ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); |
437 |
|
|
438 |
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
439 |
ReferenceF + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
ReferenceF, ReferenceB, |
|
ReferenceB + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
|
440 |
data->iEdgedWidth); |
data->iEdgedWidth); |
441 |
if (sad > *(data->iMinSAD)) return; |
if (sad > *(data->iMinSAD)) return; |
442 |
} |
} |
443 |
|
|
444 |
|
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)/1000; |
445 |
|
|
446 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
447 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
448 |
data->currentMV->x = x; data->currentMV->y = y; |
data->currentMV->x = x; data->currentMV->y = y; |
456 |
const uint8_t *ReferenceF; |
const uint8_t *ReferenceF; |
457 |
const uint8_t *ReferenceB; |
const uint8_t *ReferenceB; |
458 |
VECTOR mvs, b_mvs; |
VECTOR mvs, b_mvs; |
459 |
|
const VECTOR zeroMV = {0,0}; |
460 |
|
|
461 |
if (( x > 31) || ( x < -31) || ( y > 31) || (y < -31)) return; |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
|
|
|
|
sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1); |
|
462 |
|
|
463 |
mvs.x = data->directmvF[0].x + x; |
mvs.x = data->directmvF[0].x + x; |
464 |
b_mvs.x = ((x == 0) ? |
b_mvs.x = ((x == 0) ? |
475 |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
476 |
|| ( 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; |
477 |
|
|
478 |
switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
if (!data->qpel) { |
479 |
case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
mvs.x *= 2; mvs.y *= 2; |
480 |
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; |
|
481 |
} |
} |
482 |
|
ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data); |
483 |
|
ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); |
484 |
|
|
485 |
sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
486 |
|
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)/1000; |
487 |
|
|
488 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
489 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
494 |
static void |
static void |
495 |
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) |
496 |
{ |
{ |
497 |
int32_t sad; |
int32_t sad; int t; |
498 |
const uint8_t * Reference; |
const uint8_t * Reference; |
499 |
|
|
500 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if (( x > data->max_dx) || ( x < data->min_dx) |
501 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
502 |
|
|
503 |
switch ( ((x&1)<<1) + (y&1) ) |
if (data->qpel) Reference = Interpolate16x16qpel(x, y, 0, data); |
504 |
{ |
else Reference = GetReference(x, 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; |
|
|
} |
|
505 |
|
|
506 |
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
507 |
sad += lambda_vec8[data->iQuant] * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0); |
508 |
|
|
509 |
|
sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))/100; |
510 |
|
|
511 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
512 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
526 |
|
|
527 |
int iDirection; |
int iDirection; |
528 |
|
|
529 |
do { |
for(;;) { //forever |
530 |
iDirection = 0; |
iDirection = 0; |
531 |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
532 |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
544 |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
545 |
} else { // what remains here is up or down |
} else { // what remains here is up or down |
546 |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
547 |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); } |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
548 |
|
} |
549 |
|
|
550 |
if (iDirection) { |
if (iDirection) { |
551 |
bDirection += iDirection; |
bDirection += iDirection; |
552 |
x = data->currentMV->x; y = data->currentMV->y; } |
x = data->currentMV->x; y = data->currentMV->y; |
553 |
|
} |
554 |
} else { //about to quit, eh? not so fast.... |
} else { //about to quit, eh? not so fast.... |
555 |
switch (bDirection) { |
switch (bDirection) { |
556 |
case 2: |
case 2: |
601 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
602 |
} |
} |
603 |
} |
} |
|
while (1); //forever |
|
604 |
} |
} |
605 |
|
|
606 |
static void |
static void |
650 |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
651 |
} else { // what remains here is up or down |
} else { // what remains here is up or down |
652 |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
653 |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); } |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
654 |
|
} |
655 |
bDirection += iDirection; |
bDirection += iDirection; |
656 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
657 |
} |
} |
664 |
/* 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 */ |
665 |
|
|
666 |
static void |
static void |
667 |
HalfpelRefine(const SearchData * const data) |
SubpelRefine(const SearchData * const data) |
668 |
{ |
{ |
669 |
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
/* Do a half-pel or q-pel refinement */ |
670 |
|
VECTOR backupMV; |
|
VECTOR backupMV = *(data->currentMV); |
|
671 |
int iDirection; //not needed |
int iDirection; //not needed |
672 |
|
|
673 |
|
if (data->qpel_precision) |
674 |
|
backupMV = *(data->currentQMV); |
675 |
|
else backupMV = *(data->currentMV); |
676 |
|
|
677 |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); |
678 |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); |
679 |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); |
689 |
static __inline int |
static __inline int |
690 |
SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
691 |
const int x, const int y, |
const int x, const int y, |
692 |
const uint32_t iEdgedWidth, const uint32_t iQuant) |
const uint32_t iEdgedWidth, const uint32_t iQuant, int rrv) |
693 |
|
|
694 |
{ |
{ |
695 |
/* keep repeating checks for all b-frames before this P frame, |
/* keep repeating checks for all b-frames before this P frame, |
696 |
to make sure that SKIP is possible (todo) |
to make sure that SKIP is possible (todo) |
697 |
how: if skip is not possible set sad00 to a very high value */ |
how: if skip is not possible set sad00 to a very high value */ |
698 |
|
if(rrv) { |
699 |
|
uint32_t sadC = sad16(current->u + x*16 + y*(iEdgedWidth/2)*16, |
700 |
|
reference->u + x*16 + y*(iEdgedWidth/2)*16, iEdgedWidth/2, 256*4096); |
701 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
702 |
|
sadC += sad16(current->v + (x + y*(iEdgedWidth/2))*16, |
703 |
|
reference->v + (x + y*(iEdgedWidth/2))*16, iEdgedWidth/2, 256*4096); |
704 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
705 |
|
return 1; |
706 |
|
} else { |
707 |
uint32_t sadC = sad8(current->u + x*8 + y*(iEdgedWidth/2)*8, |
uint32_t sadC = sad8(current->u + x*8 + y*(iEdgedWidth/2)*8, |
708 |
reference->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); |
reference->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); |
709 |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
710 |
sadC += sad8(current->v + (x + y*(iEdgedWidth/2))*8, |
sadC += sad8(current->v + (x + y*(iEdgedWidth/2))*8, |
711 |
reference->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); |
reference->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); |
712 |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
|
|
|
713 |
return 1; |
return 1; |
714 |
} |
} |
715 |
|
} |
716 |
|
|
717 |
static __inline void |
static __inline void |
718 |
SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
720 |
pMB->mode = MODE_NOT_CODED; |
pMB->mode = MODE_NOT_CODED; |
721 |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
722 |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; |
723 |
|
|
724 |
|
pMB->qmvs[0].x = pMB->qmvs[1].x = pMB->qmvs[2].x = pMB->qmvs[3].x = 0; |
725 |
|
pMB->qmvs[0].y = pMB->qmvs[1].y = pMB->qmvs[2].y = pMB->qmvs[3].y = 0; |
726 |
|
|
727 |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
728 |
} |
} |
729 |
|
|
742 |
|
|
743 |
const VECTOR zeroMV = { 0, 0 }; |
const VECTOR zeroMV = { 0, 0 }; |
744 |
|
|
745 |
|
uint32_t mb_width = pParam->mb_width; |
746 |
|
uint32_t mb_height = pParam->mb_height; |
747 |
|
|
748 |
uint32_t x, y; |
uint32_t x, y; |
749 |
uint32_t iIntra = 0; |
uint32_t iIntra = 0; |
750 |
int32_t InterBias; |
int32_t InterBias, quant = current->quant, sad00; |
751 |
|
uint8_t *qimage; |
752 |
|
|
753 |
// some pre-initialized thingies for SearchP |
// some pre-initialized thingies for SearchP |
754 |
int32_t temp[5]; |
int32_t temp[8]; |
755 |
VECTOR currentMV[5]; |
VECTOR currentMV[5]; |
756 |
|
VECTOR currentQMV[5]; |
757 |
int32_t iMinSAD[5]; |
int32_t iMinSAD[5]; |
758 |
SearchData Data; |
SearchData Data; |
759 |
|
memset(&Data, 0, sizeof(SearchData)); |
760 |
Data.iEdgedWidth = pParam->edged_width; |
Data.iEdgedWidth = pParam->edged_width; |
761 |
Data.currentMV = currentMV; |
Data.currentMV = currentMV; |
762 |
|
Data.currentQMV = currentQMV; |
763 |
Data.iMinSAD = iMinSAD; |
Data.iMinSAD = iMinSAD; |
764 |
Data.temp = temp; |
Data.temp = temp; |
765 |
Data.iFcode = current->fcode; |
Data.iFcode = current->fcode; |
766 |
|
Data.rounding = pParam->m_rounding_type; |
767 |
|
Data.qpel = pParam->m_quarterpel; |
768 |
|
Data.chroma = current->global_flags & XVID_ME_COLOUR; |
769 |
|
Data.rrv = current->global_flags & XVID_REDUCED; |
770 |
|
|
771 |
|
if ((current->global_flags & XVID_REDUCED)) { |
772 |
|
mb_width = (pParam->width + 31) / 32; |
773 |
|
mb_height = (pParam->height + 31) / 32; |
774 |
|
Data.qpel = Data.chroma = 0; |
775 |
|
} |
776 |
|
|
777 |
|
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
778 |
|
return 1; // allocate some mem for qpel interpolated blocks |
779 |
|
// somehow this is dirty since I think we shouldn't use malloc outside |
780 |
|
// encoder_create() - so please fix me! |
781 |
|
Data.RefQ = qimage; |
782 |
if (sadInit) (*sadInit) (); |
if (sadInit) (*sadInit) (); |
783 |
|
|
784 |
for (y = 0; y < pParam->mb_height; y++) { |
for (y = 0; y < mb_height; y++) { |
785 |
for (x = 0; x < pParam->mb_width; x++) { |
for (x = 0; x < mb_width; x++) { |
|
|
|
786 |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
787 |
int32_t sad00 = pMB->sad16 |
|
788 |
= sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, |
if (Data.rrv) pMB->sad16 = |
789 |
|
sad32v_c(pCurrent->y + (x + y * pParam->edged_width) * 32, |
790 |
|
pRef->y + (x + y * pParam->edged_width) * 32, |
791 |
|
pParam->edged_width, pMB->sad8 ); |
792 |
|
|
793 |
|
else pMB->sad16 = |
794 |
|
sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, |
795 |
pRef->y + (x + y * pParam->edged_width) * 16, |
pRef->y + (x + y * pParam->edged_width) * 16, |
796 |
pParam->edged_width, pMB->sad8 ); |
pParam->edged_width, pMB->sad8 ); |
797 |
|
|
798 |
|
if (Data.chroma) { |
799 |
|
pMB->sad16 += sad8(pCurrent->u + x*8 + y*(pParam->edged_width/2)*8, |
800 |
|
pRef->u + x*8 + y*(pParam->edged_width/2)*8, pParam->edged_width/2); |
801 |
|
|
802 |
|
pMB->sad16 += sad8(pCurrent->v + (x + y*(pParam->edged_width/2))*8, |
803 |
|
pRef->v + (x + y*(pParam->edged_width/2))*8, pParam->edged_width/2); |
804 |
|
} |
805 |
|
|
806 |
|
sad00 = pMB->sad16; //if no gmc; else sad00 = (..) |
807 |
|
|
808 |
if (!(current->global_flags & XVID_LUMIMASKING)) { |
if (!(current->global_flags & XVID_LUMIMASKING)) { |
809 |
pMB->dquant = NO_CHANGE; |
pMB->dquant = NO_CHANGE; |
810 |
pMB->quant = current->quant; } |
pMB->quant = current->quant; |
811 |
|
} else { |
812 |
|
if (pMB->dquant != NO_CHANGE) { |
813 |
|
quant += DQtab[pMB->dquant]; |
814 |
|
if (quant > 31) quant = 31; |
815 |
|
else if (quant < 1) quant = 1; |
816 |
|
} |
817 |
|
pMB->quant = quant; |
818 |
|
} |
819 |
|
|
820 |
//initial skip decision |
//initial skip decision |
821 |
|
/* no early skip for GMC (global vector = skip vector is unknown!) */ |
822 |
if ((pMB->dquant == NO_CHANGE) && (sad00 <= MAX_SAD00_FOR_SKIP * pMB->quant) |
if (current->coding_type == P_VOP) { /* no fast SKIP for S(GMC)-VOPs */ |
823 |
&& (SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) ) { |
if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH * (Data.rrv ? 4:1) ) |
824 |
if (pMB->sad16 < pMB->quant * INITIAL_SKIP_THRESH) { |
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant, Data.rrv)) { |
825 |
SkipMacroblockP(pMB, sad00); |
SkipMacroblockP(pMB, sad00); |
826 |
continue; |
continue; |
827 |
} |
} |
828 |
} else sad00 = 256*4096; // skip not allowed - for final skip decision |
} |
829 |
|
|
830 |
SearchP(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
831 |
y, current->motion_flags, pMB->quant, |
y, current->motion_flags, pMB->quant, |
832 |
&Data, pParam, pMBs, reference->mbs, |
&Data, pParam, pMBs, reference->mbs, |
833 |
current->global_flags & XVID_INTER4V, pMB); |
current->global_flags & XVID_INTER4V, pMB); |
834 |
|
|
835 |
/* 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?" */ |
836 |
if (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
if (current->coding_type == P_VOP) { |
837 |
if ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) |
if ( (pMB->dquant == NO_CHANGE) && (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
838 |
{ SkipMacroblockP(pMB, sad00); continue; } |
&& ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1)) ) |
839 |
|
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant, Data.rrv)) { |
840 |
|
SkipMacroblockP(pMB, sad00); |
841 |
|
continue; |
842 |
|
} |
843 |
|
} |
844 |
|
|
845 |
/* finally, intra decision */ |
/* finally, intra decision */ |
846 |
|
|
847 |
InterBias = MV16_INTER_BIAS; |
InterBias = MV16_INTER_BIAS; |
848 |
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 |
849 |
if (y != 0) |
if (y != 0) |
850 |
if ((pMB - pParam->mb_width)->mode == MODE_INTER ) InterBias -= 50; |
if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; |
851 |
if (x != 0) |
if (x != 0) |
852 |
if ((pMB - 1)->mode == MODE_INTER ) InterBias -= 50; |
if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; |
853 |
|
|
854 |
|
if (Data.chroma) InterBias += 50; // to compensate bigger SAD |
855 |
|
if (Data.rrv) InterBias *= 4; //?? |
856 |
|
|
857 |
if (InterBias < pMB->sad16) { |
if (InterBias < pMB->sad16) { |
858 |
const int32_t deviation = |
int32_t deviation; |
859 |
dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
if (Data.rrv) { |
860 |
|
deviation = dev16(pCurrent->y + (x + y * pParam->edged_width) * 32, |
861 |
|
pParam->edged_width) |
862 |
|
+ dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16, |
863 |
|
pParam->edged_width) |
864 |
|
+ dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16 * pParam->edged_width, |
865 |
|
pParam->edged_width) |
866 |
|
+ dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16 * (pParam->edged_width+1), |
867 |
|
pParam->edged_width); |
868 |
|
} else |
869 |
|
deviation = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
870 |
pParam->edged_width); |
pParam->edged_width); |
871 |
|
|
872 |
if (deviation < (pMB->sad16 - InterBias)) { |
if (deviation < (pMB->sad16 - InterBias)) { |
873 |
if (++iIntra >= iLimit) return 1; |
if (++iIntra >= iLimit) { free(qimage); return 1; } |
874 |
pMB->mode = MODE_INTRA; |
pMB->mode = MODE_INTRA; |
875 |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = |
876 |
pMB->mvs[3] = zeroMV; |
pMB->mvs[3] = zeroMV; |
877 |
|
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = |
878 |
|
pMB->qmvs[3] = zeroMV; |
879 |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = |
880 |
pMB->sad8[3] = 0; |
pMB->sad8[3] = 0; |
881 |
} |
} |
882 |
} |
} |
883 |
} |
} |
884 |
} |
} |
885 |
|
free(qimage); |
886 |
|
|
887 |
|
if (current->coding_type == S_VOP) /* first GMC step only for S(GMC)-VOPs */ |
888 |
|
current->GMC_MV = GlobalMotionEst( pMBs, pParam, current->fcode ); |
889 |
|
else |
890 |
|
current->GMC_MV = zeroMV; |
891 |
|
|
892 |
return 0; |
return 0; |
893 |
} |
} |
894 |
|
|
914 |
} |
} |
915 |
|
|
916 |
static __inline void |
static __inline void |
917 |
PreparePredictionsP(VECTOR * const pmv, int x, int y, const int iWcount, |
PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, |
918 |
const int iHcount, const MACROBLOCK * const prevMB) |
int iHcount, const MACROBLOCK * const prevMB, int rrv) |
919 |
{ |
{ |
920 |
|
|
921 |
//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself |
//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself |
922 |
|
if (rrv) { iWcount /= 2; iHcount /= 2; } |
923 |
|
|
924 |
if ( (y != 0) && (x != (iWcount-1)) ) { // [5] top-right neighbour |
if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour |
925 |
pmv[5].x = EVEN(pmv[3].x); |
pmv[5].x = EVEN(pmv[3].x); |
926 |
pmv[5].y = EVEN(pmv[3].y); |
pmv[5].y = EVEN(pmv[3].y); |
927 |
} else pmv[5].x = pmv[5].y = 0; |
} else pmv[5].x = pmv[5].y = 0; |
933 |
else pmv[4].x = pmv[4].y = 0; |
else pmv[4].x = pmv[4].y = 0; |
934 |
|
|
935 |
// [1] median prediction |
// [1] median prediction |
936 |
pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); |
if (rrv) { //median is in halfzero-precision |
937 |
|
pmv[1].x = RRV_MV_SCALEUP(pmv[0].x); |
938 |
|
pmv[1].y = RRV_MV_SCALEUP(pmv[0].y); |
939 |
|
} else { pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); } |
940 |
|
|
941 |
pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask |
pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask |
942 |
|
|
943 |
pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
944 |
pmv[2].y = EVEN(prevMB->mvs[0].y); |
pmv[2].y = EVEN(prevMB->mvs[0].y); |
945 |
|
|
946 |
if ((x != iWcount-1) && (y != iHcount-1)) { |
if ((x < iWcount-1) && (y < iHcount-1)) { |
947 |
pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame |
pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame |
948 |
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
949 |
} else pmv[6].x = pmv[6].y = 0; |
} else pmv[6].x = pmv[6].y = 0; |
950 |
|
|
951 |
|
if (rrv) { |
952 |
|
int i; |
953 |
|
for (i = 0; i < 7; i++) { |
954 |
|
pmv[i].x = RRV_MV_SCALEDOWN(pmv[i].x); |
955 |
|
pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); // a trick |
956 |
|
} |
957 |
|
} |
958 |
} |
} |
959 |
|
|
960 |
static void |
static void |
961 |
SearchP(const uint8_t * const pRef, |
SearchP(const IMAGE * const pRef, |
962 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
963 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
964 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
978 |
int i, iDirection = 255, mask, threshA; |
int i, iDirection = 255, mask, threshA; |
979 |
VECTOR pmv[7]; |
VECTOR pmv[7]; |
980 |
|
|
|
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); //has to be changed to get_pmv(2)() |
|
981 |
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, |
982 |
pParam->width, pParam->height, Data->iFcode); |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
983 |
|
|
984 |
|
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); //has to be changed to get_pmv(2)() |
985 |
|
|
986 |
Data->predMV = pmv[0]; |
Data->temp[5] = Data->temp[7] = 256*4096; // to reset chroma-sad cache |
987 |
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16; |
if (Data->rrv) i = 2; else i = 1; |
988 |
Data->Ref = pRef + (x + Data->iEdgedWidth*y)*16; |
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i; |
989 |
Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16; |
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
990 |
Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16; |
Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
991 |
Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16; |
|
992 |
|
Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16*i; |
993 |
Data->iQuant = iQuant; |
Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16*i; |
994 |
|
Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16*i; |
995 |
if (!(MotionFlags & PMV_HALFPEL16)) { |
Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16*i; |
996 |
Data->min_dx = EVEN(Data->min_dx); |
Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
997 |
Data->max_dx = EVEN(Data->max_dx); |
Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
998 |
Data->min_dy = EVEN(Data->min_dy); |
|
999 |
Data->max_dy = EVEN(Data->max_dy); } |
Data->lambda16 = lambda_vec16[iQuant]; |
1000 |
|
Data->lambda8 = lambda_vec8[iQuant]; |
1001 |
|
Data->qpel_precision = 0; |
1002 |
|
|
1003 |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
1004 |
|
|
1005 |
if (inter4v) CheckCandidate = CheckCandidate16; |
for(i = 0; i < 5; i++) |
1006 |
else CheckCandidate = CheckCandidate16no4v; |
Data->currentMV[i].x = Data->currentMV[i].y = 0; |
1007 |
|
|
1008 |
for(i = 0; i < 5; i++) Data->currentMV[i].x = Data->currentMV[i].y = 0; |
if (pParam->m_quarterpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
1009 |
|
else Data->predMV = pmv[0]; |
1010 |
|
|
1011 |
i = d_mv_bits(Data->predMV.x, Data->predMV.y, Data->iFcode); |
i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0); |
1012 |
Data->iMinSAD[0] = pMB->sad16 + lambda_vec16[iQuant] * i; |
Data->iMinSAD[0] = pMB->sad16 + (Data->lambda16 * i * pMB->sad16)/1000; |
1013 |
Data->iMinSAD[1] = pMB->sad8[0] + lambda_vec8[iQuant] * i; |
Data->iMinSAD[1] = pMB->sad8[0] + (Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS))/100; |
1014 |
Data->iMinSAD[2] = pMB->sad8[1]; |
Data->iMinSAD[2] = pMB->sad8[1]; |
1015 |
Data->iMinSAD[3] = pMB->sad8[2]; |
Data->iMinSAD[3] = pMB->sad8[2]; |
1016 |
Data->iMinSAD[4] = pMB->sad8[3]; |
Data->iMinSAD[4] = pMB->sad8[3]; |
1022 |
if (threshA > 1024) threshA = 1024; } |
if (threshA > 1024) threshA = 1024; } |
1023 |
|
|
1024 |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
1025 |
prevMBs + x + y * pParam->mb_width); |
prevMBs + x + y * pParam->mb_width, Data->rrv); |
|
|
|
|
if (inter4v) CheckCandidate = CheckCandidate16; |
|
|
else CheckCandidate = CheckCandidate16no4v; |
|
1026 |
|
|
1027 |
|
if (Data->rrv) CheckCandidate = CheckCandidate32; |
1028 |
|
else if (inter4v || Data->chroma) CheckCandidate = CheckCandidate16; |
1029 |
|
else CheckCandidate = CheckCandidate16no4v; //for extra speed |
1030 |
|
|
1031 |
/* main loop. checking all predictions */ |
/* main loop. checking all predictions */ |
1032 |
|
|
1056 |
if (MotionFlags & PMV_EXTSEARCH16) { |
if (MotionFlags & PMV_EXTSEARCH16) { |
1057 |
int32_t bSAD; |
int32_t bSAD; |
1058 |
VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; |
VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; |
1059 |
|
if (Data->rrv) { |
1060 |
|
startMV.x = RRV_MV_SCALEUP(startMV.x); |
1061 |
|
startMV.y = RRV_MV_SCALEUP(startMV.y); |
1062 |
|
} else |
1063 |
if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? |
if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? |
1064 |
startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); |
startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); |
1065 |
if (!(MVequal(startMV, backupMV))) { |
if (!(MVequal(startMV, backupMV))) { |
1066 |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1067 |
|
|
1068 |
CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, Data); |
(*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); |
1069 |
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
1070 |
if (bSAD < Data->iMinSAD[0]) { |
if (bSAD < Data->iMinSAD[0]) { |
1071 |
Data->currentMV[0] = backupMV; |
Data->currentMV[0] = backupMV; |
1073 |
} |
} |
1074 |
|
|
1075 |
backupMV = Data->currentMV[0]; |
backupMV = Data->currentMV[0]; |
1076 |
if (MotionFlags & PMV_HALFPELREFINE16) startMV.x = startMV.y = 1; |
if (!MotionFlags & PMV_HALFPELREFINE16 || Data->rrv) startMV.x = startMV.y = 0; |
1077 |
else startMV.x = startMV.y = 0; |
else startMV.x = startMV.y = 1; |
1078 |
if (!(MVequal(startMV, backupMV))) { |
if (!(MVequal(startMV, backupMV))) { |
1079 |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1080 |
|
|
1081 |
CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, Data); |
(*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); |
1082 |
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
1083 |
if (bSAD < Data->iMinSAD[0]) { |
if (bSAD < Data->iMinSAD[0]) { |
1084 |
Data->currentMV[0] = backupMV; |
Data->currentMV[0] = backupMV; |
1087 |
} |
} |
1088 |
} |
} |
1089 |
|
|
1090 |
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data); |
if (MotionFlags & PMV_HALFPELREFINE16) SubpelRefine(Data); |
1091 |
|
|
1092 |
|
for(i = 0; i < 5; i++) { |
1093 |
|
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors |
1094 |
|
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
1095 |
|
} |
1096 |
|
|
1097 |
|
if((!Data->rrv) && (pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { |
1098 |
|
|
1099 |
|
Data->qpel_precision = 1; |
1100 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1101 |
|
pParam->width, pParam->height, Data->iFcode, 1, 0); |
1102 |
|
|
1103 |
|
SubpelRefine(Data); |
1104 |
|
} |
1105 |
|
|
1106 |
|
if (Data->iMinSAD[0] < (int32_t)iQuant * 30 ) inter4v = 0; |
1107 |
if (inter4v) { |
if (inter4v) { |
1108 |
SearchData Data8; |
SearchData Data8; |
1109 |
Data8.iFcode = Data->iFcode; |
Data8.iFcode = Data->iFcode; |
1110 |
Data8.iQuant = Data->iQuant; |
Data8.lambda8 = Data->lambda8; |
1111 |
Data8.iEdgedWidth = Data->iEdgedWidth; |
Data8.iEdgedWidth = Data->iEdgedWidth; |
1112 |
|
Data8.RefQ = Data->RefQ; |
1113 |
|
Data8.qpel = Data->qpel; |
1114 |
|
Data8.rrv = Data->rrv; |
1115 |
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
1116 |
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); |
1117 |
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); |
1118 |
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); |
1119 |
|
|
1120 |
|
if (Data->chroma) { |
1121 |
|
int sumx, sumy, dx, dy; |
1122 |
|
|
1123 |
|
if(pParam->m_quarterpel) { |
1124 |
|
sumx= pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2; |
1125 |
|
sumy = pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2; |
1126 |
|
} else { |
1127 |
|
sumx = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
1128 |
|
sumy = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
1129 |
} |
} |
1130 |
|
dx = (sumx >> 3) + roundtab_76[sumx & 0xf]; |
1131 |
|
dy = (sumy >> 3) + roundtab_76[sumy & 0xf]; |
1132 |
|
|
1133 |
|
Data->iMinSAD[1] += ChromaSAD(dx, dy, Data); |
1134 |
|
} |
1135 |
|
} |
1136 |
|
|
1137 |
|
if (Data->rrv) { |
1138 |
|
Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); |
1139 |
|
Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); |
1140 |
|
} |
1141 |
if (!(inter4v) || |
if (!(inter4v) || |
1142 |
(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
1143 |
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
1149 |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
1150 |
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
1151 |
|
|
1152 |
|
if(pParam->m_quarterpel) { |
1153 |
|
pMB->qmvs[0] = pMB->qmvs[1] |
1154 |
|
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
1155 |
|
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; |
1156 |
|
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; |
1157 |
|
} else { |
1158 |
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
1159 |
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
1160 |
|
} |
1161 |
} else { |
} else { |
1162 |
// INTER4V MODE; all other things are already set in Search8 |
// INTER4V MODE; all other things are already set in Search8 |
1163 |
pMB->mode = MODE_INTER4V; |
pMB->mode = MODE_INTER4V; |
1164 |
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + |
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + |
1165 |
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant; |
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant; |
1166 |
} |
} |
|
|
|
1167 |
} |
} |
1168 |
|
|
1169 |
static void |
static void |
1176 |
const int block, |
const int block, |
1177 |
SearchData * const Data) |
SearchData * const Data) |
1178 |
{ |
{ |
1179 |
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
int i = 0; |
1180 |
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
1181 |
Data->currentMV = OldData->currentMV + 1 + block; |
Data->currentMV = OldData->currentMV + 1 + block; |
1182 |
|
Data->currentQMV = OldData->currentQMV + 1 + block; |
1183 |
|
|
1184 |
|
if(pParam->m_quarterpel) { |
1185 |
|
Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
1186 |
|
if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y, |
1187 |
|
Data->predMV, Data->iFcode, 0, 0); |
1188 |
|
|
1189 |
|
} else { |
1190 |
|
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
1191 |
|
if (block != 0) { |
1192 |
|
if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y, |
1193 |
|
Data->predMV, Data->iFcode, 0, Data->rrv); |
1194 |
|
} |
1195 |
|
} |
1196 |
|
|
1197 |
if (block != 0) |
*(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100; |
|
*(Data->iMinSAD) += lambda_vec8[Data->iQuant] * |
|
|
d_mv_bits( Data->currentMV->x - Data->predMV.x, |
|
|
Data->currentMV->y - Data->predMV.y, |
|
|
Data->iFcode); |
|
1198 |
|
|
1199 |
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { |
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { |
1200 |
|
if (Data->rrv) i = 2; else i = 1; |
1201 |
|
|
1202 |
Data->Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->Ref = OldData->Ref + i*8 * ((block&1) + pParam->edged_width*(block>>1)); |
1203 |
Data->RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->RefH = OldData->RefH + i*8 * ((block&1) + pParam->edged_width*(block>>1)); |
1204 |
Data->RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->RefV = OldData->RefV + i*8 * ((block&1) + pParam->edged_width*(block>>1)); |
1205 |
Data->RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->RefHV = OldData->RefHV + i*8 * ((block&1) + pParam->edged_width*(block>>1)); |
1206 |
|
|
1207 |
Data->Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->Cur = OldData->Cur + i*8 * ((block&1) + pParam->edged_width*(block>>1)); |
1208 |
|
Data->qpel_precision = 0; |
1209 |
|
|
1210 |
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, |
1211 |
pParam->width, pParam->height, OldData->iFcode); |
pParam->width, pParam->height, OldData->iFcode - Data->qpel, 0, Data->rrv); |
1212 |
|
|
1213 |
CheckCandidate = CheckCandidate8; |
if (Data->rrv) CheckCandidate = CheckCandidate16no4v; |
1214 |
|
else CheckCandidate = CheckCandidate8; |
1215 |
|
|
1216 |
if (MotionFlags & PMV_EXTSEARCH8) { |
if (MotionFlags & PMV_EXTSEARCH8) { |
1217 |
|
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
1218 |
|
|
1219 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1220 |
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
1221 |
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
1222 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1223 |
|
|
1224 |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); } |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1225 |
|
|
1226 |
|
if(*(Data->iMinSAD) < temp_sad) { |
1227 |
|
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
1228 |
|
Data->currentQMV->y = 2 * Data->currentMV->y; |
1229 |
|
} |
1230 |
|
} |
1231 |
|
|
1232 |
|
if (MotionFlags & PMV_HALFPELREFINE8) { |
1233 |
|
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
1234 |
|
|
1235 |
|
SubpelRefine(Data); // perform halfpel refine of current best vector |
1236 |
|
|
1237 |
|
if(*(Data->iMinSAD) < temp_sad) { // we have found a better match |
1238 |
|
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
1239 |
|
Data->currentQMV->y = 2 * Data->currentMV->y; |
1240 |
|
} |
1241 |
|
} |
1242 |
|
|
1243 |
|
if(!Data->rrv && Data->qpel) { |
1244 |
|
if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && |
1245 |
|
(MotionFlags & PMV_QUARTERPELREFINE8)) { |
1246 |
|
Data->qpel_precision = 1; |
1247 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
1248 |
|
pParam->width, pParam->height, OldData->iFcode, 1, 0); |
1249 |
|
SubpelRefine(Data); |
1250 |
|
} |
1251 |
|
} |
1252 |
|
} |
1253 |
|
|
1254 |
if (MotionFlags & PMV_HALFPELREFINE8) HalfpelRefine(Data); |
if (Data->rrv) { |
1255 |
|
Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x); |
1256 |
|
Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y); |
1257 |
} |
} |
1258 |
|
|
1259 |
|
if(Data->qpel) { |
1260 |
|
pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x; |
1261 |
|
pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y; |
1262 |
|
pMB->qmvs[block] = *(Data->currentQMV); |
1263 |
|
} else { |
1264 |
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
1265 |
pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; |
pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; |
1266 |
|
} |
1267 |
|
|
1268 |
pMB->mvs[block] = *(Data->currentMV); |
pMB->mvs[block] = *(Data->currentMV); |
1269 |
pMB->sad8[block] = 4 * (*Data->iMinSAD); |
pMB->sad8[block] = 4 * (*Data->iMinSAD); |
|
|
|
1270 |
} |
} |
1271 |
|
|
1272 |
/* B-frames code starts here */ |
/* B-frames code starts here */ |
1343 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1344 |
*Data->iMinSAD = MV_MAX_ERROR; |
*Data->iMinSAD = MV_MAX_ERROR; |
1345 |
Data->iFcode = iFcode; |
Data->iFcode = iFcode; |
1346 |
|
Data->qpel_precision = 0; |
1347 |
|
|
1348 |
Data->Ref = pRef + (x + y * iEdgedWidth) * 16; |
Data->Ref = pRef + (x + y * iEdgedWidth) * 16; |
1349 |
Data->RefH = pRefH + (x + y * iEdgedWidth) * 16; |
Data->RefH = pRefH + (x + y * iEdgedWidth) * 16; |
1353 |
Data->predMV = *predMV; |
Data->predMV = *predMV; |
1354 |
|
|
1355 |
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, |
1356 |
pParam->width, pParam->height, iFcode); |
pParam->width, pParam->height, iFcode - Data->qpel, 0, 0); |
1357 |
|
|
1358 |
pmv[0] = Data->predMV; |
pmv[0] = Data->predMV; |
1359 |
|
if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } |
1360 |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
1361 |
|
|
1362 |
Data->currentMV->x = Data->currentMV->y = 0; |
Data->currentMV->x = Data->currentMV->y = 0; |
|
|
|
1363 |
CheckCandidate = CheckCandidate16no4v; |
CheckCandidate = CheckCandidate16no4v; |
1364 |
|
|
1365 |
// main loop. checking all predictions |
// main loop. checking all predictions |
1376 |
|
|
1377 |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1378 |
|
|
1379 |
HalfpelRefine(Data); |
SubpelRefine(Data); |
1380 |
|
|
1381 |
|
if (Data->qpel) { |
1382 |
|
Data->currentQMV->x = 2*Data->currentMV->x; |
1383 |
|
Data->currentQMV->y = 2*Data->currentMV->y; |
1384 |
|
Data->qpel_precision = 1; |
1385 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1386 |
|
pParam->width, pParam->height, iFcode, 1, 0); |
1387 |
|
SubpelRefine(Data); |
1388 |
|
} |
1389 |
|
|
1390 |
// three bits are needed to code backward mode. four for forward |
// three bits are needed to code backward mode. four for forward |
1391 |
// we treat the bits just like they were vector's |
// we treat the bits just like they were vector's |
1392 |
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * lambda_vec16[Data->iQuant]; |
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
1393 |
else *Data->iMinSAD += 3 * lambda_vec16[Data->iQuant]; |
else *Data->iMinSAD += 3 * Data->lambda16; |
|
|
|
1394 |
|
|
1395 |
if (*Data->iMinSAD < *best_sad) { |
if (*Data->iMinSAD < *best_sad) { |
1396 |
*best_sad = *Data->iMinSAD; |
*best_sad = *Data->iMinSAD; |
1397 |
pMB->mode = mode_current; |
pMB->mode = mode_current; |
1398 |
|
if (Data->qpel) { |
1399 |
|
pMB->pmvs[0].x = Data->currentQMV->x - predMV->x; |
1400 |
|
pMB->pmvs[0].y = Data->currentQMV->y - predMV->y; |
1401 |
|
if (mode_current == MODE_FORWARD) |
1402 |
|
pMB->qmvs[0] = *Data->currentQMV; |
1403 |
|
else |
1404 |
|
pMB->b_qmvs[0] = *Data->currentQMV; |
1405 |
|
} else { |
1406 |
pMB->pmvs[0].x = Data->currentMV->x - predMV->x; |
pMB->pmvs[0].x = Data->currentMV->x - predMV->x; |
1407 |
pMB->pmvs[0].y = Data->currentMV->y - predMV->y; |
pMB->pmvs[0].y = Data->currentMV->y - predMV->y; |
1408 |
|
} |
1409 |
if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; |
if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; |
1410 |
else pMB->b_mvs[0] = *Data->currentMV; |
else pMB->b_mvs[0] = *Data->currentMV; |
1411 |
} |
} |
1412 |
|
if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; |
1413 |
|
else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search |
1414 |
|
|
1415 |
} |
} |
1416 |
|
|
1417 |
static int32_t |
static void |
1418 |
|
SkipDecisionB(const IMAGE * const pCur, |
1419 |
|
const IMAGE * const f_Ref, |
1420 |
|
const IMAGE * const b_Ref, |
1421 |
|
MACROBLOCK * const pMB, |
1422 |
|
const uint32_t quant, |
1423 |
|
const uint32_t x, const uint32_t y, |
1424 |
|
const SearchData * const Data) |
1425 |
|
{ |
1426 |
|
int dx, dy, b_dx, b_dy; |
1427 |
|
uint32_t sum; |
1428 |
|
//this is not full chroma compensation, only it's fullpel approximation. should work though |
1429 |
|
if (Data->qpel) { |
1430 |
|
dy = Data->directmvF[0].y/2 + Data->directmvF[1].y/2 + |
1431 |
|
Data->directmvF[2].y/2 + Data->directmvF[3].y/2; |
1432 |
|
|
1433 |
|
dx = Data->directmvF[0].x/2 + Data->directmvF[1].x/2 + |
1434 |
|
Data->directmvF[2].x/2 + Data->directmvF[3].x/2; |
1435 |
|
|
1436 |
|
b_dy = Data->directmvB[0].y/2 + Data->directmvB[1].y/2 + |
1437 |
|
Data->directmvB[2].y/2 + Data->directmvB[3].y/2; |
1438 |
|
|
1439 |
|
b_dx = Data->directmvB[0].x/2 + Data->directmvB[1].x/2 + |
1440 |
|
Data->directmvB[2].x/2 + Data->directmvB[3].x/2; |
1441 |
|
|
1442 |
|
} else { |
1443 |
|
dy = Data->directmvF[0].y + Data->directmvF[1].y + |
1444 |
|
Data->directmvF[2].y + Data->directmvF[3].y; |
1445 |
|
|
1446 |
|
dx = Data->directmvF[0].x + Data->directmvF[1].x + |
1447 |
|
Data->directmvF[2].x + Data->directmvF[3].x; |
1448 |
|
|
1449 |
|
b_dy = Data->directmvB[0].y + Data->directmvB[1].y + |
1450 |
|
Data->directmvB[2].y + Data->directmvB[3].y; |
1451 |
|
|
1452 |
|
b_dx = Data->directmvB[0].x + Data->directmvB[1].x + |
1453 |
|
Data->directmvB[2].x + Data->directmvB[3].x; |
1454 |
|
} |
1455 |
|
|
1456 |
|
|
1457 |
|
dy = (dy >> 3) + roundtab_76[dy & 0xf]; |
1458 |
|
dx = (dx >> 3) + roundtab_76[dx & 0xf]; |
1459 |
|
b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf]; |
1460 |
|
b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf]; |
1461 |
|
|
1462 |
|
sum = sad8bi(pCur->u + 8*x + 8*y*(Data->iEdgedWidth/2), |
1463 |
|
f_Ref->u + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
1464 |
|
b_Ref->u + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
1465 |
|
Data->iEdgedWidth/2); |
1466 |
|
sum += sad8bi(pCur->v + 8*x + 8*y*(Data->iEdgedWidth/2), |
1467 |
|
f_Ref->v + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
1468 |
|
b_Ref->v + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
1469 |
|
Data->iEdgedWidth/2); |
1470 |
|
|
1471 |
|
if (sum < 2*MAX_CHROMA_SAD_FOR_SKIP * quant) pMB->mode = MODE_DIRECT_NONE_MV; //skipped |
1472 |
|
} |
1473 |
|
|
1474 |
|
|
1475 |
|
|
1476 |
|
static __inline uint32_t |
1477 |
SearchDirect(const IMAGE * const f_Ref, |
SearchDirect(const IMAGE * const f_Ref, |
1478 |
const uint8_t * const f_RefH, |
const uint8_t * const f_RefH, |
1479 |
const uint8_t * const f_RefV, |
const uint8_t * const f_RefV, |
1499 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1500 |
|
|
1501 |
*Data->iMinSAD = 256*4096; |
*Data->iMinSAD = 256*4096; |
|
Data->referencemv = b_mb->mvs; |
|
1502 |
|
|
1503 |
Data->Ref = f_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
Data->Ref = f_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
1504 |
Data->RefH = f_RefH + (x + Data->iEdgedWidth*y) * 16; |
Data->RefH = f_RefH + (x + Data->iEdgedWidth*y) * 16; |
1513 |
Data->max_dy = 2 * pParam->height - 2 * (y) * 16; |
Data->max_dy = 2 * pParam->height - 2 * (y) * 16; |
1514 |
Data->min_dx = -(2 * 16 + 2 * (x) * 16); |
Data->min_dx = -(2 * 16 + 2 * (x) * 16); |
1515 |
Data->min_dy = -(2 * 16 + 2 * (y) * 16); |
Data->min_dy = -(2 * 16 + 2 * (y) * 16); |
1516 |
|
if (Data->qpel) { //we measure in qpixels |
1517 |
|
Data->max_dx *= 2; |
1518 |
|
Data->max_dy *= 2; |
1519 |
|
Data->min_dx *= 2; |
1520 |
|
Data->min_dy *= 2; |
1521 |
|
Data->referencemv = b_mb->qmvs; |
1522 |
|
} else Data->referencemv = b_mb->mvs; |
1523 |
|
Data->qpel_precision = 0; // it's a trick. it's 1 not 0, but we need 0 here |
1524 |
|
|
1525 |
for (k = 0; k < 4; k++) { |
for (k = 0; k < 4; k++) { |
1526 |
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); |
1534 |
*best_sad = 256*4096; // in that case, we won't use direct mode |
*best_sad = 256*4096; // in that case, we won't use direct mode |
1535 |
pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV" |
pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV" |
1536 |
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
1537 |
return 0; |
return 256*4096; |
1538 |
} |
} |
1539 |
if (b_mb->mode != MODE_INTER4V) { |
if (b_mb->mode != MODE_INTER4V) { |
1540 |
pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; |
pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; |
1545 |
} |
} |
1546 |
} |
} |
1547 |
|
|
1548 |
if (b_mb->mode == MODE_INTER4V) |
if (b_mb->mode == MODE_INTER4V) CheckCandidate = CheckCandidateDirect; |
|
CheckCandidate = CheckCandidateDirect; |
|
1549 |
else CheckCandidate = CheckCandidateDirectno4v; |
else CheckCandidate = CheckCandidateDirectno4v; |
1550 |
|
|
1551 |
(*CheckCandidate)(0, 0, 255, &k, Data); |
(*CheckCandidate)(0, 0, 255, &k, Data); |
1552 |
|
|
1553 |
// skip decision |
// initial (fast) skip decision |
1554 |
if (*Data->iMinSAD - 2 * lambda_vec16[Data->iQuant] < (int32_t)Data->iQuant * SKIP_THRESH_B) { |
if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH*2) { |
1555 |
//checking chroma. everything copied from MC |
SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data->chroma, Data); //possible skip - checking chroma |
1556 |
//this is not full chroma compensation, only it's fullpel approximation. should work though |
if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip. |
|
int sum, dx, dy, b_dx, b_dy; |
|
|
|
|
|
sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
|
|
dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
|
|
|
|
sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
|
|
dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
|
|
|
|
sum = pMB->b_mvs[0].x + pMB->b_mvs[1].x + pMB->b_mvs[2].x + pMB->b_mvs[3].x; |
|
|
b_dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
|
|
|
|
sum = pMB->b_mvs[0].y + pMB->b_mvs[1].y + pMB->b_mvs[2].y + pMB->b_mvs[3].y; |
|
|
b_dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
|
|
|
|
sum = sad8bi(pCur->u + 8*x + 8*y*(Data->iEdgedWidth/2), |
|
|
f_Ref->u + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
|
|
b_Ref->u + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
|
|
Data->iEdgedWidth/2); |
|
|
sum += sad8bi(pCur->v + 8*x + 8*y*(Data->iEdgedWidth/2), |
|
|
f_Ref->v + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
|
|
b_Ref->v + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
|
|
Data->iEdgedWidth/2); |
|
|
|
|
|
if ((uint32_t) sum < MAX_CHROMA_SAD_FOR_SKIP * Data->iQuant) { |
|
|
pMB->mode = MODE_DIRECT_NONE_MV; |
|
|
return *Data->iMinSAD; |
|
|
} |
|
1557 |
} |
} |
1558 |
|
|
1559 |
skip_sad = *Data->iMinSAD; |
skip_sad = *Data->iMinSAD; |
1567 |
|
|
1568 |
(*MainSearchPtr)(0, 0, Data, 255); |
(*MainSearchPtr)(0, 0, Data, 255); |
1569 |
|
|
1570 |
HalfpelRefine(Data); |
SubpelRefine(Data); |
1571 |
|
|
|
*Data->iMinSAD += 1 * lambda_vec16[Data->iQuant]; // one bit is needed to code direct mode. we treat this bit just like it was vector's |
|
1572 |
*best_sad = *Data->iMinSAD; |
*best_sad = *Data->iMinSAD; |
1573 |
|
|
1574 |
if (b_mb->mode == MODE_INTER4V) |
if (b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT; |
|
pMB->mode = MODE_DIRECT; |
|
1575 |
else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
1576 |
|
|
1577 |
pMB->pmvs[3] = *Data->currentMV; |
pMB->pmvs[3] = *Data->currentMV; |
1585 |
pMB->b_mvs[k].y = ((Data->currentMV->y == 0) |
pMB->b_mvs[k].y = ((Data->currentMV->y == 0) |
1586 |
? Data->directmvB[k].y |
? Data->directmvB[k].y |
1587 |
: pMB->mvs[k].y - Data->referencemv[k].y); |
: pMB->mvs[k].y - Data->referencemv[k].y); |
1588 |
|
if (Data->qpel) { |
1589 |
|
pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2; |
1590 |
|
pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2; |
1591 |
|
pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2; |
1592 |
|
pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2; |
1593 |
|
} |
1594 |
|
|
1595 |
if (b_mb->mode != MODE_INTER4V) { |
if (b_mb->mode != MODE_INTER4V) { |
1596 |
pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; |
pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; |
1597 |
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]; |
1598 |
|
pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0]; |
1599 |
|
pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0]; |
1600 |
break; |
break; |
1601 |
} |
} |
1602 |
} |
} |
1604 |
} |
} |
1605 |
|
|
1606 |
|
|
1607 |
static __inline void |
static void |
1608 |
SearchInterpolate(const uint8_t * const f_Ref, |
SearchInterpolate(const uint8_t * const f_Ref, |
1609 |
const uint8_t * const f_RefH, |
const uint8_t * const f_RefH, |
1610 |
const uint8_t * const f_RefV, |
const uint8_t * const f_RefV, |
1627 |
|
|
1628 |
{ |
{ |
1629 |
|
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
|
|
|
1630 |
int iDirection, i, j; |
int iDirection, i, j; |
1631 |
SearchData bData; |
SearchData bData; |
1632 |
|
|
1633 |
bData.iMinSAD = fData->iMinSAD; |
fData->qpel_precision = 0; |
1634 |
*bData.iMinSAD = 4096*256; |
memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data |
1635 |
bData.Cur = fData->Cur; |
*fData->iMinSAD = 4096*256; |
1636 |
fData->iEdgedWidth = bData.iEdgedWidth = iEdgedWidth; |
bData.currentMV++; bData.currentQMV++; |
|
bData.currentMV = fData->currentMV + 1; |
|
|
bData.iQuant = fData->iQuant; |
|
1637 |
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
1638 |
|
|
1639 |
bData.bRef = fData->Ref = f_Ref + (x + y * iEdgedWidth) * 16; |
i = (x + y * fData->iEdgedWidth) * 16; |
1640 |
bData.bRefH = fData->RefH = f_RefH + (x + y * iEdgedWidth) * 16; |
bData.bRef = fData->Ref = f_Ref + i; |
1641 |
bData.bRefV = fData->RefV = f_RefV + (x + y * iEdgedWidth) * 16; |
bData.bRefH = fData->RefH = f_RefH + i; |
1642 |
bData.bRefHV = fData->RefHV = f_RefHV + (x + y * iEdgedWidth) * 16; |
bData.bRefV = fData->RefV = f_RefV + i; |
1643 |
bData.Ref = fData->bRef = b_Ref + (x + y * iEdgedWidth) * 16; |
bData.bRefHV = fData->RefHV = f_RefHV + i; |
1644 |
bData.RefH = fData->bRefH = b_RefH + (x + y * iEdgedWidth) * 16; |
bData.Ref = fData->bRef = b_Ref + i; |
1645 |
bData.RefV = fData->bRefV = b_RefV + (x + y * iEdgedWidth) * 16; |
bData.RefH = fData->bRefH = b_RefH + i; |
1646 |
bData.RefHV = fData->bRefHV = b_RefHV + (x + y * iEdgedWidth) * 16; |
bData.RefV = fData->bRefV = b_RefV + i; |
1647 |
|
bData.RefHV = fData->bRefHV = b_RefHV + i; |
1648 |
|
|
1649 |
bData.bpredMV = fData->predMV = *f_predMV; |
bData.bpredMV = fData->predMV = *f_predMV; |
1650 |
fData->bpredMV = bData.predMV = *b_predMV; |
fData->bpredMV = bData.predMV = *b_predMV; |
1651 |
|
fData->currentMV[0] = fData->currentMV[2]; |
1652 |
|
|
1653 |
fData->currentMV[0] = pMB->mvs[0]; |
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode - fData->qpel, 0, 0); |
1654 |
fData->currentMV[1] = pMB->b_mvs[0]; |
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode - fData->qpel, 0, 0); |
|
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode); |
|
|
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode); |
|
1655 |
|
|
1656 |
if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; |
if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; |
1657 |
if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dy; |
if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx; |
1658 |
if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dx; |
if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy; |
1659 |
if (fData->currentMV[0].y > fData->min_dy) fData->currentMV[0].y = fData->min_dy; |
if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy; |
1660 |
|
|
1661 |
if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx; |
if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx; |
1662 |
if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dy; |
if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx; |
1663 |
if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dx; |
if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy; |
1664 |
if (fData->currentMV[1].y > bData.min_dy) fData->currentMV[1].y = bData.min_dy; |
if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy; |
1665 |
|
|
1666 |
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
1667 |
|
|
1680 |
// backward MV moves |
// backward MV moves |
1681 |
i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
1682 |
fData->currentMV[2] = fData->currentMV[0]; |
fData->currentMV[2] = fData->currentMV[0]; |
|
|
|
1683 |
CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); |
CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); |
1684 |
CheckCandidateInt(i, j + 1, 0, &iDirection, &bData); |
CheckCandidateInt(i, j + 1, 0, &iDirection, &bData); |
1685 |
CheckCandidateInt(i - 1, j, 0, &iDirection, &bData); |
CheckCandidateInt(i - 1, j, 0, &iDirection, &bData); |
1687 |
|
|
1688 |
} while (!(iDirection)); |
} while (!(iDirection)); |
1689 |
|
|
1690 |
// two bits are needed to code interpolate mode. we treat the bits just like they were vector's |
if (fData->qpel) { |
1691 |
*fData->iMinSAD += 2 * lambda_vec16[fData->iQuant]; |
CheckCandidate = CheckCandidateInt; |
1692 |
|
fData->qpel_precision = bData.qpel_precision = 1; |
1693 |
|
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0); |
1694 |
|
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0); |
1695 |
|
fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x; |
1696 |
|
fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y; |
1697 |
|
fData->currentQMV[1].x = 2 * fData->currentMV[1].x; |
1698 |
|
fData->currentQMV[1].y = 2 * fData->currentMV[1].y; |
1699 |
|
SubpelRefine(fData); |
1700 |
|
fData->currentQMV[2] = fData->currentQMV[0]; |
1701 |
|
SubpelRefine(&bData); |
1702 |
|
} |
1703 |
|
|
1704 |
|
*fData->iMinSAD += (2+2) * fData->lambda16; // two bits are needed to code interpolate mode. |
1705 |
|
|
1706 |
if (*fData->iMinSAD < *best_sad) { |
if (*fData->iMinSAD < *best_sad) { |
1707 |
*best_sad = *fData->iMinSAD; |
*best_sad = *fData->iMinSAD; |
1708 |
pMB->mvs[0] = fData->currentMV[0]; |
pMB->mvs[0] = fData->currentMV[0]; |
1709 |
pMB->b_mvs[0] = fData->currentMV[1]; |
pMB->b_mvs[0] = fData->currentMV[1]; |
1710 |
pMB->mode = MODE_INTERPOLATE; |
pMB->mode = MODE_INTERPOLATE; |
1711 |
|
if (fData->qpel) { |
1712 |
|
pMB->qmvs[0] = fData->currentQMV[0]; |
1713 |
|
pMB->b_qmvs[0] = fData->currentQMV[1]; |
1714 |
|
pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x; |
1715 |
|
pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y; |
1716 |
|
pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x; |
1717 |
|
pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y; |
1718 |
|
} else { |
1719 |
pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; |
pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; |
1720 |
pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; |
pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; |
1721 |
pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; |
pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; |
1722 |
pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; |
pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; |
1723 |
} |
} |
1724 |
} |
} |
1725 |
|
} |
1726 |
|
|
1727 |
void |
void |
1728 |
MotionEstimationBVOP(MBParam * const pParam, |
MotionEstimationBVOP(MBParam * const pParam, |
1736 |
const IMAGE * const f_refV, |
const IMAGE * const f_refV, |
1737 |
const IMAGE * const f_refHV, |
const IMAGE * const f_refHV, |
1738 |
// backward (future) reference |
// backward (future) reference |
1739 |
const MACROBLOCK * const b_mbs, |
const FRAMEINFO * const b_reference, |
1740 |
const IMAGE * const b_ref, |
const IMAGE * const b_ref, |
1741 |
const IMAGE * const b_refH, |
const IMAGE * const b_refH, |
1742 |
const IMAGE * const b_refV, |
const IMAGE * const b_refV, |
1743 |
const IMAGE * const b_refHV) |
const IMAGE * const b_refHV) |
1744 |
{ |
{ |
1745 |
uint32_t i, j; |
uint32_t i, j; |
1746 |
int32_t best_sad, skip_sad; |
int32_t best_sad; |
1747 |
|
uint32_t skip_sad; |
1748 |
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; |
1749 |
static const VECTOR zeroMV={0,0}; |
static const VECTOR zeroMV={0,0}; |
1750 |
|
const MACROBLOCK * const b_mbs = b_reference->mbs; |
1751 |
|
|
1752 |
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
1753 |
|
|
1754 |
const int32_t TRB = time_pp - time_bp; |
const int32_t TRB = time_pp - time_bp; |
1755 |
const int32_t TRD = time_pp; |
const int32_t TRD = time_pp; |
1756 |
|
uint8_t * qimage; |
1757 |
|
|
1758 |
// some pre-inintialized data for the rest of the search |
// some pre-inintialized data for the rest of the search |
1759 |
|
|
1760 |
SearchData Data; |
SearchData Data; |
1761 |
int32_t iMinSAD; |
int32_t iMinSAD; |
1762 |
VECTOR currentMV[3]; |
VECTOR currentMV[3]; |
1763 |
|
VECTOR currentQMV[3]; |
1764 |
|
memset(&Data, 0, sizeof(SearchData)); |
1765 |
Data.iEdgedWidth = pParam->edged_width; |
Data.iEdgedWidth = pParam->edged_width; |
1766 |
Data.currentMV = currentMV; |
Data.currentMV = currentMV; Data.currentQMV = currentQMV; |
1767 |
Data.iMinSAD = &iMinSAD; |
Data.iMinSAD = &iMinSAD; |
1768 |
Data.iQuant = frame->quant; |
Data.lambda16 = lambda_vec16[frame->quant]; |
1769 |
|
Data.chroma = frame->quant; |
1770 |
|
Data.qpel = pParam->m_quarterpel; |
1771 |
|
Data.rounding = 0; |
1772 |
|
|
1773 |
|
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
1774 |
|
return; // allocate some mem for qpel interpolated blocks |
1775 |
|
// somehow this is dirty since I think we shouldn't use malloc outside |
1776 |
|
// encoder_create() - so please fix me! |
1777 |
|
Data.RefQ = qimage; |
1778 |
|
|
1779 |
// note: i==horizontal, j==vertical |
// note: i==horizontal, j==vertical |
|
|
|
1780 |
for (j = 0; j < pParam->mb_height; j++) { |
for (j = 0; j < pParam->mb_height; j++) { |
1781 |
|
|
1782 |
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
1785 |
MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; |
MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; |
1786 |
const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; |
const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; |
1787 |
|
|
1788 |
/* 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 */ |
1789 |
|
if (b_reference->coding_type != S_VOP) |
1790 |
if (b_mb->mode == MODE_NOT_CODED) { |
if (b_mb->mode == MODE_NOT_CODED) { |
1791 |
pMB->mode = MODE_NOT_CODED; |
pMB->mode = MODE_NOT_CODED; |
1792 |
continue; |
continue; |
1793 |
} |
} |
1794 |
|
|
1795 |
Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
1796 |
|
pMB->quant = frame->quant; |
1797 |
|
|
1798 |
/* direct search comes first, because it (1) checks for SKIP-mode |
/* direct search comes first, because it (1) checks for SKIP-mode |
1799 |
and (2) sets very good predictions for forward and backward search */ |
and (2) sets very good predictions for forward and backward search */ |
|
|
|
1800 |
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, |
1801 |
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
1802 |
&frame->image, |
&frame->image, |
1810 |
|
|
1811 |
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
1812 |
|
|
|
// best_sad = 256*4096; //uncomment to disable Directsearch. |
|
|
// To disable any other mode, just comment the function call |
|
|
|
|
1813 |
// forward search |
// forward search |
1814 |
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, |
1815 |
&frame->image, i, j, |
&frame->image, i, j, |
1827 |
MODE_BACKWARD, &Data); |
MODE_BACKWARD, &Data); |
1828 |
|
|
1829 |
// interpolate search comes last, because it uses data from forward and backward as prediction |
// interpolate search comes last, because it uses data from forward and backward as prediction |
|
|
|
1830 |
SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
1831 |
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
1832 |
&frame->image, |
&frame->image, |
1838 |
pMB, &best_sad, |
pMB, &best_sad, |
1839 |
&Data); |
&Data); |
1840 |
|
|
1841 |
|
// final skip decision |
1842 |
|
if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP*2) |
1843 |
|
&& ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) |
1844 |
|
SkipDecisionB(&frame->image, f_ref, b_ref, pMB,frame->quant, i, j, &Data); |
1845 |
|
|
1846 |
switch (pMB->mode) { |
switch (pMB->mode) { |
1847 |
case MODE_FORWARD: |
case MODE_FORWARD: |
1848 |
f_count++; |
f_count++; |
1849 |
f_predMV = pMB->mvs[0]; |
if (Data.qpel) f_predMV = pMB->qmvs[0]; |
1850 |
|
else f_predMV = pMB->mvs[0]; |
1851 |
break; |
break; |
1852 |
case MODE_BACKWARD: |
case MODE_BACKWARD: |
1853 |
b_count++; |
b_count++; |
1854 |
b_predMV = pMB->b_mvs[0]; |
if (Data.qpel) b_predMV = pMB->b_qmvs[0]; |
1855 |
|
else b_predMV = pMB->b_mvs[0]; |
1856 |
break; |
break; |
1857 |
case MODE_INTERPOLATE: |
case MODE_INTERPOLATE: |
1858 |
i_count++; |
i_count++; |
1859 |
|
if (Data.qpel) { |
1860 |
|
f_predMV = pMB->qmvs[0]; |
1861 |
|
b_predMV = pMB->b_qmvs[0]; |
1862 |
|
} else { |
1863 |
f_predMV = pMB->mvs[0]; |
f_predMV = pMB->mvs[0]; |
1864 |
b_predMV = pMB->b_mvs[0]; |
b_predMV = pMB->b_mvs[0]; |
1865 |
|
} |
1866 |
break; |
break; |
1867 |
case MODE_DIRECT: |
case MODE_DIRECT: |
1868 |
case MODE_DIRECT_NO4V: |
case MODE_DIRECT_NO4V: |
1869 |
d_count++; |
d_count++; |
|
break; |
|
1870 |
default: |
default: |
1871 |
break; |
break; |
1872 |
} |
} |
1873 |
} |
} |
1874 |
} |
} |
1875 |
|
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); |
|
|
|
|
1876 |
} |
} |
1877 |
|
|
|
/* Hinted ME starts here */ |
|
|
|
|
1878 |
static __inline void |
static __inline void |
1879 |
Search8hinted( const SearchData * const OldData, |
MEanalyzeMB ( const uint8_t * const pRef, |
1880 |
const int x, const int y, |
const uint8_t * const pCur, |
|
const uint32_t MotionFlags, |
|
|
const MBParam * const pParam, |
|
|
MACROBLOCK * const pMB, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const int block) |
|
|
{ |
|
|
SearchData Data; |
|
|
MainSearchFunc *MainSearchPtr; |
|
|
|
|
|
Data.predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
|
|
Data.iMinSAD = OldData->iMinSAD + 1 + block; |
|
|
Data.currentMV = OldData->currentMV+1+block; |
|
|
Data.iFcode = OldData->iFcode; |
|
|
Data.iQuant = OldData->iQuant; |
|
|
|
|
|
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.iEdgedWidth = pParam->edged_width; |
|
|
Data.Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
|
|
|
|
CheckCandidate = CheckCandidate8; |
|
|
|
|
|
if (block != 0) |
|
|
*(Data.iMinSAD) += lambda_vec8[Data.iQuant] * |
|
|
d_mv_bits( Data.currentMV->x - Data.predMV.x, |
|
|
Data.currentMV->y - Data.predMV.y, |
|
|
Data.iFcode); |
|
|
|
|
|
|
|
|
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 8, |
|
|
pParam->width, pParam->height, OldData->iFcode); |
|
|
|
|
|
if (pMB->mode == MODE_INTER4V) { |
|
|
int dummy; |
|
|
CheckCandidate8(pMB->mvs[block].x, pMB->mvs[block].y, 0, &dummy, &Data); } |
|
|
|
|
|
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 (MotionFlags & PMV_HALFPELREFINE8) HalfpelRefine(&Data); |
|
|
|
|
|
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->sad8[block] = 4 * (*(Data.iMinSAD)); |
|
|
} |
|
|
|
|
|
|
|
|
static void |
|
|
SearchPhinted ( const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const IMAGE * const pCur, |
|
1881 |
const int x, |
const int x, |
1882 |
const int y, |
const int y, |
|
const uint32_t MotionFlags, |
|
|
const uint32_t iQuant, |
|
1883 |
const MBParam * const pParam, |
const MBParam * const pParam, |
1884 |
const MACROBLOCK * const pMBs, |
MACROBLOCK * const pMBs, |
|
int inter4v, |
|
|
MACROBLOCK * const pMB, |
|
1885 |
SearchData * const Data) |
SearchData * const Data) |
1886 |
{ |
{ |
1887 |
|
|
1888 |
const int32_t iEdgedWidth = pParam->edged_width; |
int i, mask; |
1889 |
|
VECTOR pmv[3]; |
1890 |
int i, t; |
MACROBLOCK * pMB = &pMBs[x + y * pParam->mb_width]; |
|
MainSearchFunc * MainSearchPtr; |
|
|
|
|
|
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
|
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
|
|
pParam->width, pParam->height, Data->iFcode); |
|
|
|
|
|
Data->Cur = pCur->y + (x + y * iEdgedWidth) * 16; |
|
|
Data->Ref = pRef + (x + iEdgedWidth*y)*16; |
|
|
Data->RefH = pRefH + (x + iEdgedWidth*y) * 16; |
|
|
Data->RefV = pRefV + (x + iEdgedWidth*y) * 16; |
|
|
Data->RefHV = pRefHV + (x + iEdgedWidth*y) * 16; |
|
|
Data->iQuant = iQuant; |
|
|
|
|
|
if (!(MotionFlags & PMV_HALFPEL16)) { |
|
|
Data->min_dx = EVEN(Data->min_dx); |
|
|
Data->max_dx = EVEN(Data->max_dx); |
|
|
Data->min_dy = EVEN(Data->min_dy); |
|
|
Data->max_dy = EVEN(Data->max_dy); |
|
|
} |
|
1891 |
|
|
1892 |
for(i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
for(i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
1893 |
|
|
1894 |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
//median is only used as prediction. it doesn't have to be real |
1895 |
|
if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; |
1896 |
if (inter4v) |
else |
1897 |
CheckCandidate = CheckCandidate16; |
if (x == 1) //left macroblock does not have any vector now |
1898 |
else CheckCandidate = CheckCandidate16no4v; |
Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median |
1899 |
|
else if (y == 1) // top macroblock doesn't have it's vector |
1900 |
|
Data->predMV = (pMB - 1)->mvs[0]; // left instead of median |
1901 |
pMB->mvs[0].x = EVEN(pMB->mvs[0].x); |
else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median |
|
pMB->mvs[0].y = EVEN(pMB->mvs[0].y); |
|
|
if (pMB->mvs[0].x > Data->max_dx) pMB->mvs[0].x = Data->max_dx; // this is in case iFcode changed |
|
|
if (pMB->mvs[0].x < Data->min_dx) pMB->mvs[0].x = Data->min_dx; |
|
|
if (pMB->mvs[0].y > Data->max_dy) pMB->mvs[0].y = Data->max_dy; |
|
|
if (pMB->mvs[0].y < Data->min_dy) pMB->mvs[0].y = Data->min_dy; |
|
|
|
|
|
(*CheckCandidate)(pMB->mvs[0].x, pMB->mvs[0].y, 0, &t, Data); |
|
|
|
|
|
if (pMB->mode == MODE_INTER4V) |
|
|
for (i = 1; i < 4; i++) { // all four vectors will be used as four predictions for 16x16 search |
|
|
pMB->mvs[i].x = EVEN(pMB->mvs[i].x); |
|
|
pMB->mvs[i].y = EVEN(pMB->mvs[i].y); |
|
|
if (!(make_mask(pMB->mvs, i))) |
|
|
(*CheckCandidate)(pMB->mvs[i].x, pMB->mvs[i].y, 0, &t, Data); |
|
|
} |
|
1902 |
|
|
1903 |
if (MotionFlags & PMV_USESQUARES16) |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1904 |
MainSearchPtr = SquareSearch; |
pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, Data->rrv); |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
|
|
MainSearchPtr = AdvDiamondSearch; |
|
|
else MainSearchPtr = DiamondSearch; |
|
1905 |
|
|
1906 |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
1907 |
|
Data->Ref = pRef + (x + y * pParam->edged_width) * 16; |
1908 |
|
|
1909 |
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data); |
pmv[1].x = EVEN(pMB->mvs[0].x); |
1910 |
|
pmv[1].y = EVEN(pMB->mvs[0].y); |
1911 |
|
pmv[2].x = EVEN(Data->predMV.x); |
1912 |
|
pmv[2].y = EVEN(Data->predMV.y); |
1913 |
|
pmv[0].x = pmv[0].y = 0; |
1914 |
|
|
1915 |
if (inter4v) |
CheckCandidate32I(0, 0, 255, &i, Data); |
|
for(i = 0; i < 4; i++) |
|
|
Search8hinted(Data, 2*x+(i&1), 2*y+(i>>1), MotionFlags, pParam, pMB, pMBs, i); |
|
1916 |
|
|
1917 |
if (!(inter4v) || |
if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) { |
|
(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] + |
|
|
Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
|
|
// INTER MODE |
|
1918 |
|
|
1919 |
pMB->mode = MODE_INTER; |
if (!(mask = make_mask(pmv, 1))) |
1920 |
pMB->mvs[0] = pMB->mvs[1] |
CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data); |
1921 |
= pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
if (!(mask = make_mask(pmv, 2))) |
1922 |
|
CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data); |
1923 |
|
|
1924 |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) // diamond only if needed |
1925 |
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
1926 |
|
|
1927 |
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
for (i = 0; i < 4; i++) { |
1928 |
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1) * pParam->mb_width)]; |
1929 |
} else { |
MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i]; |
1930 |
// INTER4V MODE; all other things are already set in Search8hinted |
MB->mode = MODE_INTER; |
1931 |
pMB->mode = MODE_INTER4V; |
MB->sad16 = Data->iMinSAD[i+1]; |
1932 |
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] |
} |
|
+ Data->iMinSAD[4] + IMV16X16 * iQuant; |
|
1933 |
} |
} |
|
|
|
1934 |
} |
} |
1935 |
|
|
1936 |
void |
#define INTRA_BIAS 2500 |
1937 |
MotionEstimationHinted( MBParam * const pParam, |
#define INTRA_THRESH 1500 |
1938 |
FRAMEINFO * const current, |
#define INTER_THRESH 1400 |
1939 |
FRAMEINFO * const reference, |
|
1940 |
const IMAGE * const pRefH, |
|
1941 |
const IMAGE * const pRefV, |
int |
1942 |
const IMAGE * const pRefHV) |
MEanalysis( const IMAGE * const pRef, |
1943 |
|
FRAMEINFO * const Current, |
1944 |
|
MBParam * const pParam, |
1945 |
|
int maxIntra, //maximum number if non-I frames |
1946 |
|
int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame |
1947 |
|
int bCount) // number if B frames in a row |
1948 |
{ |
{ |
1949 |
MACROBLOCK *const pMBs = current->mbs; |
uint32_t x, y, intra = 0; |
1950 |
const IMAGE *const pCurrent = ¤t->image; |
int sSAD = 0; |
1951 |
const IMAGE *const pRef = &reference->image; |
MACROBLOCK * const pMBs = Current->mbs; |
1952 |
|
const IMAGE * const pCurrent = &Current->image; |
1953 |
|
int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH; |
1954 |
|
const VECTOR zeroMV = {0,0}; |
1955 |
|
|
1956 |
uint32_t x, y; |
int32_t iMinSAD[5], temp[5]; |
|
int32_t temp[5]; |
|
|
int32_t iMinSAD[5]; |
|
1957 |
VECTOR currentMV[5]; |
VECTOR currentMV[5]; |
1958 |
SearchData Data; |
SearchData Data; |
1959 |
Data.iEdgedWidth = pParam->edged_width; |
Data.iEdgedWidth = pParam->edged_width; |
1960 |
Data.currentMV = currentMV; |
Data.currentMV = currentMV; |
1961 |
Data.iMinSAD = iMinSAD; |
Data.iMinSAD = iMinSAD; |
1962 |
|
Data.iFcode = Current->fcode; |
1963 |
|
Data.rrv = Current->global_flags & XVID_REDUCED; |
1964 |
Data.temp = temp; |
Data.temp = temp; |
1965 |
Data.iFcode = current->fcode; |
CheckCandidate = CheckCandidate32I; |
1966 |
|
|
1967 |
if (sadInit) (*sadInit) (); |
if (intraCount < 10) // we're right after an I frame |
1968 |
|
IntraThresh += 4 * (intraCount - 10) * (intraCount - 10); |
1969 |
|
else |
1970 |
|
if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec |
1971 |
|
IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; |
1972 |
|
|
1973 |
for (y = 0; y < pParam->mb_height; y++) { |
InterThresh += 400 * (1 - bCount); |
1974 |
for (x = 0; x < pParam->mb_width; x++) { |
if (InterThresh < 300) InterThresh = 300; |
1975 |
|
|
1976 |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
if (sadInit) (*sadInit) (); |
|
|
|
|
//intra mode is copied from the first pass. At least for the time being |
|
|
if ((pMB->mode == MODE_INTRA) || (pMB->mode == MODE_NOT_CODED) ) continue; |
|
|
|
|
|
if (!(current->global_flags & XVID_LUMIMASKING)) { |
|
|
pMB->dquant = NO_CHANGE; |
|
|
pMB->quant = current->quant; } |
|
|
|
|
|
SearchPhinted(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
|
|
y, current->motion_flags, pMB->quant, |
|
|
pParam, pMBs, current->global_flags & XVID_INTER4V, pMB, |
|
|
&Data); |
|
1977 |
|
|
1978 |
|
for (y = 1; y < pParam->mb_height-1; y+=2) { |
1979 |
|
for (x = 1; x < pParam->mb_width-1; x+=2) { |
1980 |
|
int i; |
1981 |
|
|
1982 |
|
if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; |
1983 |
|
|
1984 |
|
MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data); |
1985 |
|
|
1986 |
|
for (i = 0; i < 4; i++) { |
1987 |
|
int dev; |
1988 |
|
MACROBLOCK *pMB = &pMBs[x+(i&1) + y+(i>>1) * pParam->mb_width]; |
1989 |
|
if (pMB->sad16 > IntraThresh) { |
1990 |
|
dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1))* pParam->edged_width) * 16, |
1991 |
|
pParam->edged_width); |
1992 |
|
if (dev + IntraThresh < pMB->sad16) { |
1993 |
|
pMB->mode = MODE_INTRA; |
1994 |
|
if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return I_VOP; |
1995 |
} |
} |
1996 |
} |
} |
1997 |
|
sSAD += pMB->sad16; |
1998 |
} |
} |
1999 |
|
} |
2000 |
|
} |
2001 |
|
sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); |
2002 |
|
if (sSAD > IntraThresh + INTRA_BIAS ) return I_VOP; |
2003 |
|
if (sSAD > InterThresh ) return P_VOP; |
2004 |
|
emms(); |
2005 |
|
return B_VOP; |
2006 |
|
|
2007 |
static __inline int |
} |
|
MEanalyzeMB ( const uint8_t * const pRef, |
|
|
const uint8_t * const pCur, |
|
|
const int x, |
|
|
const int y, |
|
|
const MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
MACROBLOCK * const pMB, |
|
|
SearchData * const Data) |
|
|
{ |
|
|
|
|
|
int i, mask; |
|
|
VECTOR pmv[3]; |
|
|
|
|
|
*(Data->iMinSAD) = MV_MAX_ERROR; |
|
|
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
|
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
|
|
pParam->width, pParam->height, Data->iFcode); |
|
|
|
|
|
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
|
|
Data->Ref = pRef + (x + y * pParam->edged_width) * 16; |
|
|
|
|
|
CheckCandidate = CheckCandidate16no4vI; |
|
|
|
|
|
pmv[1].x = EVEN(pMB->mvs[0].x); |
|
|
pmv[1].y = EVEN(pMB->mvs[0].y); |
|
|
pmv[0].x = EVEN(Data->predMV.x); |
|
|
pmv[0].y = EVEN(Data->predMV.y); |
|
|
pmv[2].x = pmv[2].y = 0; |
|
|
|
|
|
CheckCandidate16no4vI(pmv[0].x, pmv[0].y, 255, &i, Data); |
|
|
if (!(mask = make_mask(pmv, 1))) |
|
|
CheckCandidate16no4vI(pmv[1].x, pmv[1].y, mask, &i, Data); |
|
|
if (!(mask = make_mask(pmv, 2))) |
|
|
CheckCandidate16no4vI(0, 0, mask, &i, Data); |
|
2008 |
|
|
2009 |
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
static void |
2010 |
|
CheckGMC(int x, int y, const int dir, int * iDirection, |
2011 |
|
const MACROBLOCK * const pMBs, uint32_t * bestcount, VECTOR * GMC, |
2012 |
|
const MBParam * const pParam) |
2013 |
|
{ |
2014 |
|
uint32_t mx, my, a, count = 0; |
2015 |
|
|
2016 |
pMB->mvs[0] = pMB->mvs[1] |
for (my = 1; my < pParam->mb_height-1; my++) |
2017 |
= pMB->mvs[2] = pMB->mvs[3] = *Data->currentMV; // all, for future get_pmv() |
for (mx = 1; mx < pParam->mb_width-1; mx++) { |
2018 |
|
VECTOR mv; |
2019 |
|
const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; |
2020 |
|
if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) continue; |
2021 |
|
mv = pMB->mvs[0]; |
2022 |
|
a = ABS(mv.x - x) + ABS(mv.y - y); |
2023 |
|
if (a < 6) count += 6 - a; |
2024 |
|
} |
2025 |
|
|
2026 |
return *(Data->iMinSAD); |
if (count > *bestcount) { |
2027 |
|
*bestcount = count; |
2028 |
|
*iDirection = dir; |
2029 |
|
GMC->x = x; GMC->y = y; |
2030 |
|
} |
2031 |
} |
} |
2032 |
|
|
|
#define INTRA_THRESH 1350 |
|
|
#define INTER_THRESH 900 |
|
2033 |
|
|
2034 |
int |
static VECTOR |
2035 |
MEanalysis( const IMAGE * const pRef, |
GlobalMotionEst(const MACROBLOCK * const pMBs, const MBParam * const pParam, const uint32_t iFcode) |
|
const IMAGE * const pCurrent, |
|
|
MBParam * const pParam, |
|
|
MACROBLOCK * const pMBs, |
|
|
const uint32_t iFcode) |
|
2036 |
{ |
{ |
|
uint32_t x, y, intra = 0; |
|
|
int sSAD = 0; |
|
2037 |
|
|
2038 |
VECTOR currentMV; |
uint32_t count, bestcount = 0; |
2039 |
int32_t iMinSAD; |
int x, y; |
2040 |
SearchData Data; |
VECTOR gmc = {0,0}; |
2041 |
Data.iEdgedWidth = pParam->edged_width; |
int step, min_x, max_x, min_y, max_y; |
2042 |
Data.currentMV = ¤tMV; |
uint32_t mx, my; |
2043 |
Data.iMinSAD = &iMinSAD; |
int iDirection, bDirection; |
|
Data.iFcode = iFcode; |
|
|
Data.iQuant = 2; |
|
2044 |
|
|
2045 |
if (sadInit) (*sadInit) (); |
min_x = min_y = -32<<iFcode; |
2046 |
|
max_x = max_y = 32<<iFcode; |
2047 |
|
|
2048 |
for (y = 0; y < pParam->mb_height-1; y++) { |
//step1: let's find a rough camera panning |
2049 |
for (x = 0; x < pParam->mb_width; x++) { |
for (step = 32; step >= 2; step /= 2) { |
2050 |
int sad, dev; |
bestcount = 0; |
2051 |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
for (y = min_y; y <= max_y; y += step) |
2052 |
|
for (x = min_x ; x <= max_x; x += step) { |
2053 |
|
count = 0; |
2054 |
|
//for all macroblocks |
2055 |
|
for (my = 1; my < pParam->mb_height-1; my++) |
2056 |
|
for (mx = 1; mx < pParam->mb_width-1; mx++) { |
2057 |
|
const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; |
2058 |
|
VECTOR mv; |
2059 |
|
|
2060 |
sad = MEanalyzeMB(pRef->y, pCurrent->y, x, y, |
if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) |
2061 |
pParam, pMBs, pMB, &Data); |
continue; |
2062 |
|
|
2063 |
if ( x != 0 && y != 0 && x != pParam->mb_width-1 ) { //no edge macroblocks, they just don't work |
mv = pMB->mvs[0]; |
2064 |
if (sad > INTRA_THRESH) { |
if ( ABS(mv.x - x) <= step && ABS(mv.y - y) <= step ) /* GMC translation is always halfpel-res */ |
2065 |
dev = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
count++; |
|
pParam->edged_width); |
|
|
if (dev + INTRA_THRESH < sad) intra++; |
|
|
if (intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return 2; // I frame |
|
2066 |
} |
} |
2067 |
sSAD += sad; |
if (count >= bestcount) { bestcount = count; gmc.x = x; gmc.y = y; } |
2068 |
} |
} |
2069 |
|
min_x = gmc.x - step; |
2070 |
|
max_x = gmc.x + step; |
2071 |
|
min_y = gmc.y - step; |
2072 |
|
max_y = gmc.y + step; |
2073 |
|
|
2074 |
} |
} |
2075 |
|
|
2076 |
|
if (bestcount < (pParam->mb_height-2)*(pParam->mb_width-2)/10) |
2077 |
|
gmc.x = gmc.y = 0; //no camara pan, no GMC |
2078 |
|
|
2079 |
|
// step2: let's refine camera panning using gradiend-descent approach. |
2080 |
|
// TODO: more warping points may be evaluated here (like in interpolate mode search - two vectors in one diamond) |
2081 |
|
bestcount = 0; |
2082 |
|
CheckGMC(gmc.x, gmc.y, 255, &iDirection, pMBs, &bestcount, &gmc, pParam); |
2083 |
|
do { |
2084 |
|
x = gmc.x; y = gmc.y; |
2085 |
|
bDirection = iDirection; iDirection = 0; |
2086 |
|
if (bDirection & 1) CheckGMC(x - 1, y, 1+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); |
2087 |
|
if (bDirection & 2) CheckGMC(x + 1, y, 2+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); |
2088 |
|
if (bDirection & 4) CheckGMC(x, y - 1, 1+2+4, &iDirection, pMBs, &bestcount, &gmc, pParam); |
2089 |
|
if (bDirection & 8) CheckGMC(x, y + 1, 1+2+8, &iDirection, pMBs, &bestcount, &gmc, pParam); |
2090 |
|
|
2091 |
|
} while (iDirection); |
2092 |
|
|
2093 |
|
if (pParam->m_quarterpel) { |
2094 |
|
gmc.x *= 2; |
2095 |
|
gmc.y *= 2; /* we store the halfpel value as pseudo-qpel to make comparison easier */ |
2096 |
} |
} |
|
sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); |
|
|
if (sSAD > INTER_THRESH ) return 1; //P frame |
|
|
emms(); |
|
|
return 0; // B frame |
|
2097 |
|
|
2098 |
|
return gmc; |
2099 |
} |
} |