31 |
#include <assert.h> |
#include <assert.h> |
32 |
#include <stdio.h> |
#include <stdio.h> |
33 |
#include <stdlib.h> |
#include <stdlib.h> |
34 |
|
#include <string.h> // memcpy |
35 |
|
#include <math.h> // lrint |
36 |
|
|
37 |
#include "../encoder.h" |
#include "../encoder.h" |
38 |
#include "../utils/mbfunctions.h" |
#include "../utils/mbfunctions.h" |
44 |
#include "motion.h" |
#include "motion.h" |
45 |
#include "sad.h" |
#include "sad.h" |
46 |
#include "../utils/emms.h" |
#include "../utils/emms.h" |
47 |
|
#include "../dct/fdct.h" |
48 |
|
|
49 |
#define INITIAL_SKIP_THRESH (10) |
#define INITIAL_SKIP_THRESH (10) |
50 |
#define FINAL_SKIP_THRESH (50) |
#define FINAL_SKIP_THRESH (50) |
51 |
#define MAX_SAD00_FOR_SKIP (20) |
#define MAX_SAD00_FOR_SKIP (20) |
52 |
#define MAX_CHROMA_SAD_FOR_SKIP (22) |
#define MAX_CHROMA_SAD_FOR_SKIP (22) |
|
#define SKIP_THRESH_B (25) |
|
53 |
|
|
54 |
#define CHECK_CANDIDATE(X,Y,D) { \ |
#define CHECK_CANDIDATE(X,Y,D) { \ |
55 |
(*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } |
CheckCandidate((X),(Y), (D), &iDirection, data ); } |
56 |
|
|
57 |
#define GET_REFERENCE(X, Y, REF) { \ |
static __inline uint32_t |
58 |
switch ( (((X)&1)<<1) + ((Y)&1) ) \ |
d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv) |
|
{ \ |
|
|
case 0 : REF = (uint8_t *)data->Ref + (X)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ |
|
|
case 1 : REF = (uint8_t *)data->RefV + (X)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ |
|
|
case 2 : REF = (uint8_t *)data->RefH + ((X)-1)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ |
|
|
default : REF = (uint8_t *)data->RefHV + ((X)-1)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ |
|
|
} \ |
|
|
} |
|
|
// I hate those macros :/ |
|
|
#define GET_REFERENCE2(X, Y, REF) { \ |
|
|
switch ( (((X)&1)<<1) + ((Y)&1) ) \ |
|
|
{ \ |
|
|
case 0 : REF = (uint8_t *)data->bRef + (X)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ |
|
|
case 1 : REF = (uint8_t *)data->bRefV + (X)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ |
|
|
case 2 : REF = (uint8_t *)data->bRefH + ((X)-1)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ |
|
|
default : REF = (uint8_t *)data->bRefHV + ((X)-1)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ |
|
|
} \ |
|
|
} |
|
|
|
|
|
|
|
|
#define iDiamondSize 2 |
|
|
|
|
|
static __inline int |
|
|
d_mv_bits(int x, int y, const uint32_t iFcode) |
|
59 |
{ |
{ |
60 |
int xb, yb; |
int xb, yb; |
61 |
|
x = qpel ? x<<1 : x; |
62 |
|
y = qpel ? y<<1 : y; |
63 |
|
if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); } |
64 |
|
|
65 |
if (x == 0) xb = 1; |
x -= pred.x; |
66 |
else { |
y -= pred.y; |
67 |
if (x < 0) x = -x; |
|
68 |
|
if (x) { |
69 |
|
x = ABS(x); |
70 |
x += (1 << (iFcode - 1)) - 1; |
x += (1 << (iFcode - 1)) - 1; |
71 |
x >>= (iFcode - 1); |
x >>= (iFcode - 1); |
72 |
if (x > 32) x = 32; |
if (x > 32) x = 32; |
73 |
xb = mvtab[x] + iFcode; |
xb = mvtab[x] + iFcode; |
74 |
} |
} else xb = 1; |
75 |
|
|
76 |
if (y == 0) yb = 1; |
if (y) { |
77 |
else { |
y = ABS(y); |
|
if (y < 0) y = -y; |
|
78 |
y += (1 << (iFcode - 1)) - 1; |
y += (1 << (iFcode - 1)) - 1; |
79 |
y >>= (iFcode - 1); |
y >>= (iFcode - 1); |
80 |
if (y > 32) y = 32; |
if (y > 32) y = 32; |
81 |
yb = mvtab[y] + iFcode; |
yb = mvtab[y] + iFcode; |
82 |
} |
} else yb = 1; |
83 |
return xb + yb; |
return xb + yb; |
84 |
} |
} |
85 |
|
|
86 |
|
static int32_t ChromaSAD2(int fx, int fy, int bx, int by, const SearchData * const data) |
87 |
|
{ |
88 |
|
int sad; |
89 |
|
const uint32_t stride = data->iEdgedWidth/2; |
90 |
|
uint8_t * f_refu = data->RefQ, |
91 |
|
* f_refv = data->RefQ + 8, |
92 |
|
* b_refu = data->RefQ + 16, |
93 |
|
* b_refv = data->RefQ + 24; |
94 |
|
|
95 |
|
switch (((fx & 1) << 1) | (fy & 1)) { |
96 |
|
case 0: |
97 |
|
fx = fx / 2; fy = fy / 2; |
98 |
|
f_refu = (uint8_t*)data->RefCU + fy * stride + fx, stride; |
99 |
|
f_refv = (uint8_t*)data->RefCV + fy * stride + fx, stride; |
100 |
|
break; |
101 |
|
case 1: |
102 |
|
fx = fx / 2; fy = (fy - 1) / 2; |
103 |
|
interpolate8x8_halfpel_v(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding); |
104 |
|
interpolate8x8_halfpel_v(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding); |
105 |
|
break; |
106 |
|
case 2: |
107 |
|
fx = (fx - 1) / 2; fy = fy / 2; |
108 |
|
interpolate8x8_halfpel_h(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding); |
109 |
|
interpolate8x8_halfpel_h(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding); |
110 |
|
break; |
111 |
|
default: |
112 |
|
fx = (fx - 1) / 2; fy = (fy - 1) / 2; |
113 |
|
interpolate8x8_halfpel_hv(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding); |
114 |
|
interpolate8x8_halfpel_hv(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding); |
115 |
|
break; |
116 |
|
} |
117 |
|
|
118 |
|
switch (((bx & 1) << 1) | (by & 1)) { |
119 |
|
case 0: |
120 |
|
bx = bx / 2; by = by / 2; |
121 |
|
b_refu = (uint8_t*)data->b_RefCU + by * stride + bx, stride; |
122 |
|
b_refv = (uint8_t*)data->b_RefCV + by * stride + bx, stride; |
123 |
|
break; |
124 |
|
case 1: |
125 |
|
bx = bx / 2; by = (by - 1) / 2; |
126 |
|
interpolate8x8_halfpel_v(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding); |
127 |
|
interpolate8x8_halfpel_v(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding); |
128 |
|
break; |
129 |
|
case 2: |
130 |
|
bx = (bx - 1) / 2; by = by / 2; |
131 |
|
interpolate8x8_halfpel_h(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding); |
132 |
|
interpolate8x8_halfpel_h(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding); |
133 |
|
break; |
134 |
|
default: |
135 |
|
bx = (bx - 1) / 2; by = (by - 1) / 2; |
136 |
|
interpolate8x8_halfpel_hv(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding); |
137 |
|
interpolate8x8_halfpel_hv(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding); |
138 |
|
break; |
139 |
|
} |
140 |
|
|
141 |
|
sad = sad8bi(data->CurU, b_refu, f_refu, stride); |
142 |
|
sad += sad8bi(data->CurV, b_refv, f_refv, stride); |
143 |
|
|
144 |
|
return sad; |
145 |
|
} |
146 |
|
|
147 |
|
|
148 |
static int32_t |
static int32_t |
149 |
ChromaSAD(int dx, int dy, const SearchData * const data) |
ChromaSAD(int dx, int dy, const SearchData * const data) |
150 |
{ |
{ |
151 |
int sad; |
int sad; |
152 |
dx = (dx >> 1) + roundtab_79[dx & 0x3]; |
const uint32_t stride = data->iEdgedWidth/2; |
|
dy = (dy >> 1) + roundtab_79[dy & 0x3]; |
|
153 |
|
|
154 |
switch (((dx & 1) << 1) + (dy & 1)) { // ((dx%2)?2:0)+((dy%2)?1:0) |
if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently |
155 |
|
data->temp[5] = dx; data->temp[6] = dy; // backup |
156 |
|
|
157 |
|
switch (((dx & 1) << 1) | (dy & 1)) { |
158 |
case 0: |
case 0: |
159 |
sad = sad8(data->CurU, data->RefCU + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); |
dx = dx / 2; dy = dy / 2; |
160 |
sad += sad8(data->CurV, data->RefCV + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); |
sad = sad8(data->CurU, data->RefCU + dy * stride + dx, stride); |
161 |
|
sad += sad8(data->CurV, data->RefCV + dy * stride + dx, stride); |
162 |
break; |
break; |
163 |
case 1: |
case 1: |
164 |
dx = dx / 2; dy = (dy - 1) / 2; |
dx = dx / 2; dy = (dy - 1) / 2; |
165 |
sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); |
sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + (dy+1) * stride + dx, stride); |
166 |
sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); |
sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + (dy+1) * stride + dx, stride); |
167 |
break; |
break; |
168 |
case 2: |
case 2: |
169 |
dx = (dx - 1) / 2; dy = dy / 2; |
dx = (dx - 1) / 2; dy = dy / 2; |
170 |
sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); |
sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + dy * stride + dx+1, stride); |
171 |
sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); |
sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + dy * stride + dx+1, stride); |
172 |
break; |
break; |
173 |
default: |
default: |
174 |
dx = (dx - 1) / 2; dy = (dy - 1) / 2; |
dx = (dx - 1) / 2; dy = (dy - 1) / 2; |
175 |
interpolate8x8_halfpel_hv(data->RefQ, |
interpolate8x8_halfpel_hv(data->RefQ, data->RefCU + dy * stride + dx, stride, data->rounding); |
176 |
data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, |
sad = sad8(data->CurU, data->RefQ, stride); |
177 |
data->rounding); |
|
178 |
sad = sad8(data->CurU, data->RefQ, data->iEdgedWidth/2); |
interpolate8x8_halfpel_hv(data->RefQ, data->RefCV + dy * stride + dx, stride, data->rounding); |
179 |
interpolate8x8_halfpel_hv(data->RefQ, |
sad += sad8(data->CurV, data->RefQ, stride); |
|
data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, |
|
|
data->rounding); |
|
|
sad += sad8(data->CurV, data->RefQ, data->iEdgedWidth/2); |
|
180 |
break; |
break; |
181 |
} |
} |
182 |
|
data->temp[7] = sad; //backup, part 2 |
183 |
return sad; |
return sad; |
184 |
} |
} |
185 |
|
|
186 |
|
static __inline const uint8_t * |
187 |
|
GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data) |
188 |
|
{ |
189 |
|
// dir : 0 = forward, 1 = backward |
190 |
|
switch ( (dir << 2) | ((x&1)<<1) | (y&1) ) { |
191 |
|
case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth); |
192 |
|
case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); |
193 |
|
case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); |
194 |
|
case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); |
195 |
|
case 4 : return data->bRef + x/2 + (y/2)*(data->iEdgedWidth); |
196 |
|
case 5 : return data->bRefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); |
197 |
|
case 6 : return data->bRefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); |
198 |
|
default : return data->bRefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); |
199 |
|
} |
200 |
|
} |
201 |
|
|
202 |
/* CHECK_CANDIATE FUNCTIONS START */ |
// this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate |
203 |
|
static __inline const uint8_t * |
204 |
|
GetReference(const int x, const int y, const SearchData * const data) |
205 |
|
{ |
206 |
|
switch ( ((x&1)<<1) | (y&1) ) { |
207 |
|
case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth); |
208 |
|
case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); |
209 |
|
case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); |
210 |
|
default : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); //case 2 |
211 |
|
} |
212 |
|
} |
213 |
|
|
214 |
|
static uint8_t * |
215 |
|
Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data) |
216 |
|
{ |
217 |
|
// create or find a qpel-precision reference picture; return pointer to it |
218 |
|
uint8_t * Reference = data->RefQ + 16*dir; |
219 |
|
const uint32_t iEdgedWidth = data->iEdgedWidth; |
220 |
|
const uint32_t rounding = data->rounding; |
221 |
|
const int halfpel_x = x/2; |
222 |
|
const int halfpel_y = y/2; |
223 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
224 |
|
|
225 |
|
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
226 |
|
ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
227 |
|
switch( ((x&1)<<1) + (y&1) ) { |
228 |
|
case 0: // pure halfpel position |
229 |
|
return (uint8_t *) ref1; |
230 |
|
break; |
231 |
|
|
232 |
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
233 |
|
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
234 |
|
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
235 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
236 |
|
break; |
237 |
|
|
238 |
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
239 |
|
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
240 |
|
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
241 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
242 |
|
break; |
243 |
|
|
244 |
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
245 |
|
// bottom left/right) during qpel refinement |
246 |
|
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
247 |
|
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
248 |
|
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
249 |
|
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
250 |
|
ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
251 |
|
ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
252 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
253 |
|
break; |
254 |
|
} |
255 |
|
return Reference; |
256 |
|
} |
257 |
|
|
258 |
|
static uint8_t * |
259 |
|
Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data) |
260 |
|
{ |
261 |
|
// create or find a qpel-precision reference picture; return pointer to it |
262 |
|
uint8_t * Reference = data->RefQ + 16*dir; |
263 |
|
const uint32_t iEdgedWidth = data->iEdgedWidth; |
264 |
|
const uint32_t rounding = data->rounding; |
265 |
|
const int halfpel_x = x/2; |
266 |
|
const int halfpel_y = y/2; |
267 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
268 |
|
|
269 |
|
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
270 |
|
switch( ((x&1)<<1) + (y&1) ) { |
271 |
|
case 3: // x and y in qpel resolution - the "corners" (top left/right and |
272 |
|
// bottom left/right) during qpel refinement |
273 |
|
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
274 |
|
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
275 |
|
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
276 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
277 |
|
interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); |
278 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); |
279 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); |
280 |
|
break; |
281 |
|
|
282 |
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
283 |
|
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
284 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
285 |
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
286 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); |
287 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
288 |
|
break; |
289 |
|
|
290 |
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
291 |
|
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
292 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
293 |
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
294 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); |
295 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
296 |
|
break; |
297 |
|
|
298 |
|
case 0: // pure halfpel position |
299 |
|
return (uint8_t *) ref1; |
300 |
|
} |
301 |
|
return Reference; |
302 |
|
} |
303 |
|
|
304 |
|
/* CHECK_CANDIATE FUNCTIONS START */ |
305 |
|
|
306 |
static void |
static void |
307 |
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) |
308 |
{ |
{ |
309 |
int t; |
int xc, yc; |
310 |
const uint8_t * Reference; |
const uint8_t * Reference; |
311 |
|
VECTOR * current; |
312 |
|
int32_t sad; uint32_t t; |
313 |
|
|
314 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if (( x > data->max_dx) || ( x < data->min_dx) |
315 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
316 |
|
|
317 |
switch ( ((x&1)<<1) + (y&1) ) { |
if (!data->qpel_precision) { |
318 |
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
Reference = GetReference(x, y, data); |
319 |
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
current = data->currentMV; |
320 |
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
xc = x; yc = y; |
321 |
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
} else { // x and y are in 1/4 precision |
322 |
|
Reference = Interpolate16x16qpel(x, y, 0, data); |
323 |
|
xc = x/2; yc = y/2; //for chroma sad |
324 |
|
current = data->currentQMV; |
325 |
} |
} |
326 |
|
|
327 |
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
328 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
329 |
|
|
330 |
if (data->qpel) t = d_mv_bits(2*x - data->predQMV.x, 2*y - data->predQMV.y, data->iFcode); |
sad += (data->lambda16 * t * sad)>>10; |
331 |
else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; |
332 |
|
|
333 |
data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; |
if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
334 |
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; |
(yc >> 1) + roundtab_79[yc & 0x3], data); |
335 |
|
|
336 |
if (data->chroma) data->temp[0] += ChromaSAD(x, y, data); |
if (sad < data->iMinSAD[0]) { |
337 |
|
data->iMinSAD[0] = sad; |
338 |
if (data->temp[0] < data->iMinSAD[0]) { |
current[0].x = x; current[0].y = y; |
339 |
data->iMinSAD[0] = data->temp[0]; |
*dir = Direction; |
340 |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
} |
|
*dir = Direction; } |
|
341 |
|
|
342 |
if (data->temp[1] < data->iMinSAD[1]) { |
if (data->temp[1] < data->iMinSAD[1]) { |
343 |
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y = y; } |
344 |
if (data->temp[2] < data->iMinSAD[2]) { |
if (data->temp[2] < data->iMinSAD[2]) { |
345 |
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; } |
346 |
if (data->temp[3] < data->iMinSAD[3]) { |
if (data->temp[3] < data->iMinSAD[3]) { |
347 |
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; } |
348 |
if (data->temp[4] < data->iMinSAD[4]) { |
if (data->temp[4] < data->iMinSAD[4]) { |
349 |
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; } |
350 |
|
|
351 |
} |
} |
352 |
|
|
353 |
static void |
static void |
354 |
CheckCandidate16no4v(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) |
355 |
{ |
{ |
356 |
int32_t sad; |
int32_t sad; uint32_t t; |
357 |
const uint8_t * Reference; |
const uint8_t * Reference; |
358 |
|
|
359 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if (( x > data->max_dx) || ( x < data->min_dx) |
360 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
361 |
|
|
362 |
switch ( ((x&1)<<1) + (y&1) ) |
if (!data->qpel_precision) Reference = GetReference(x, y, data); |
363 |
{ |
else Reference = Interpolate8x8qpel(x, y, 0, 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; |
|
|
} |
|
364 |
|
|
365 |
sad = sad16(data->Cur, Reference, data->iEdgedWidth, MV_MAX_ERROR); |
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
366 |
if (data->qpel) //only to be used in b-frames' ME |
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
367 |
sad += (data->lambda16 * d_mv_bits(2*x - data->predMV.x, 2*y - data->predMV.y, data->iFcode) * sad)/1000; |
|
368 |
else |
sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10; |
|
sad += (data->lambda16 * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode) * sad)/1000; |
|
369 |
|
|
370 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
371 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
372 |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
data->currentMV->x = x; data->currentMV->y = y; |
373 |
*dir = Direction; } |
*dir = Direction; |
374 |
|
} |
375 |
} |
} |
376 |
|
|
|
static void |
|
|
CheckCandidate16_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
377 |
|
|
378 |
// CheckCandidate16 variant which expects x and y in quarter pixel resolution |
static void |
379 |
// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!) |
CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
// around currentMV! |
|
380 |
{ |
{ |
381 |
int t; |
uint32_t t; |
382 |
uint8_t * Reference = (uint8_t *)data->RefQ; |
const uint8_t * Reference; |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
|
|
VECTOR halfpelMV = *(data->currentMV); |
|
|
|
|
|
int32_t iEdgedWidth = data->iEdgedWidth; |
|
|
uint32_t rounding = data->rounding; |
|
383 |
|
|
384 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero integer value |
385 |
|
(x > data->max_dx) || (x < data->min_dx) |
386 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
387 |
|
|
388 |
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this refenrence is used in all cases |
Reference = GetReference(x, y, data); |
389 |
switch( ((x&1)<<1) + (y&1) ) |
t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1); |
|
{ |
|
|
case 0: // pure halfpel position - shouldn't happen during a refinement step |
|
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, Reference); |
|
|
break; |
|
|
|
|
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
|
|
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
|
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
|
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); |
|
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); |
|
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); |
|
|
break; |
|
|
|
|
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
|
|
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); |
|
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
|
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); |
|
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); |
|
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); |
|
|
break; |
|
|
|
|
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
|
|
// bottom left/right) during qpel refinement |
|
|
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
|
|
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); |
|
|
GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); |
|
|
|
|
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
|
|
interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); |
|
|
interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); |
|
|
interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); |
|
|
break; |
|
|
} |
|
390 |
|
|
391 |
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp+1); |
data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
392 |
|
|
393 |
t = d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode); |
data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10; |
394 |
data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; |
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; |
|
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; |
|
|
|
|
|
if (data->chroma) |
|
|
data->temp[0] += ChromaSAD(x/2, y/2, data); |
|
395 |
|
|
396 |
if (data->temp[0] < data->iMinSAD[0]) { |
if (data->temp[0] < data->iMinSAD[0]) { |
397 |
data->iMinSAD[0] = data->temp[0]; |
data->iMinSAD[0] = data->temp[0]; |
398 |
data->currentQMV[0].x = x; data->currentQMV[0].y = y; |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
399 |
/* *dir = Direction;*/ } |
*dir = Direction; } |
400 |
|
|
401 |
if (data->temp[1] < data->iMinSAD[1]) { |
if (data->temp[1] < data->iMinSAD[1]) { |
402 |
data->iMinSAD[1] = data->temp[1]; data->currentQMV[1].x = x; data->currentQMV[1].y = y; } |
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
403 |
if (data->temp[2] < data->iMinSAD[2]) { |
if (data->temp[2] < data->iMinSAD[2]) { |
404 |
data->iMinSAD[2] = data->temp[2]; data->currentQMV[2].x = x; data->currentQMV[2].y = y; } |
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
405 |
if (data->temp[3] < data->iMinSAD[3]) { |
if (data->temp[3] < data->iMinSAD[3]) { |
406 |
data->iMinSAD[3] = data->temp[3]; data->currentQMV[3].x = x; data->currentQMV[3].y = y; } |
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
407 |
if (data->temp[4] < data->iMinSAD[4]) { |
if (data->temp[4] < data->iMinSAD[4]) { |
408 |
data->iMinSAD[4] = data->temp[4]; data->currentQMV[4].x = x; data->currentQMV[4].y = y; } |
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
409 |
} |
} |
410 |
|
|
411 |
static void |
static void |
412 |
CheckCandidate16no4v_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
|
|
|
// CheckCandidate16no4v variant which expects x and y in quarter pixel resolution |
|
|
// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!) |
|
|
// around currentMV! |
|
|
// this function is for B-frames' search only |
|
|
{ |
|
|
uint8_t * Reference = (uint8_t *)data->RefQ; |
|
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
|
|
VECTOR halfpelMV = *(data->currentMV); |
|
|
|
|
|
int32_t iEdgedWidth = data->iEdgedWidth; |
|
|
int32_t sad; |
|
|
|
|
|
if (( x > data->max_dx) || ( x < data->min_dx) |
|
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|
|
|
|
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this refenrence is used in all cases |
|
|
switch( ((x&1)<<1) + (y&1) ) |
|
413 |
{ |
{ |
414 |
case 0: // pure halfpel position - shouldn't happen during a refinement step |
int32_t sad, xc, yc; |
415 |
GET_REFERENCE(halfpelMV.x, halfpelMV.y, Reference); |
const uint8_t * Reference; |
416 |
break; |
uint32_t t; |
417 |
|
VECTOR * current; |
418 |
|
|
419 |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
if ( (x > data->max_dx) | ( x < data->min_dx) |
420 |
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
| (y > data->max_dy) | (y < data->min_dy) ) return; |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); |
|
|
break; |
|
421 |
|
|
422 |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; //non-zero even value |
|
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); |
|
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); |
|
|
break; |
|
423 |
|
|
424 |
default: // x and y in qpel resolution - the "corners" (top left/right and |
if (data->qpel_precision) { // x and y are in 1/4 precision |
425 |
// bottom left/right) during qpel refinement |
Reference = Interpolate16x16qpel(x, y, 0, data); |
426 |
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
current = data->currentQMV; |
427 |
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); |
xc = x/2; yc = y/2; |
428 |
GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); |
} else { |
429 |
|
Reference = GetReference(x, y, data); |
430 |
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, 0); |
current = data->currentMV; |
431 |
interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, 0); |
xc = x; yc = y; |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, 0); |
|
|
interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, 0); |
|
|
break; |
|
432 |
} |
} |
433 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, |
434 |
|
data->qpel^data->qpel_precision, data->rrv); |
435 |
|
|
436 |
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
437 |
sad += (data->lambda16 * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode) * sad)/1000; |
sad += (data->lambda16 * t * sad)>>10; |
438 |
|
|
439 |
if (sad < data->iMinSAD[0]) { |
if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
440 |
data->iMinSAD[0] = sad; |
(yc >> 1) + roundtab_79[yc & 0x3], data); |
441 |
data->currentQMV[0].x = x; data->currentQMV[0].y = y; |
|
442 |
/* *dir = Direction;*/ } |
if (sad < *(data->iMinSAD)) { |
443 |
|
*(data->iMinSAD) = sad; |
444 |
|
current->x = x; current->y = y; |
445 |
|
*dir = Direction; |
446 |
|
} |
447 |
} |
} |
448 |
|
|
449 |
static void |
static void |
450 |
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) |
451 |
{ |
{ |
452 |
// maximum speed - for P/B/I decision |
// maximum speed - for P/B/I decision |
453 |
int32_t sad; |
int32_t sad; |
455 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if (( x > data->max_dx) || ( x < data->min_dx) |
456 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
457 |
|
|
458 |
sad = sad16(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), |
sad = sad32v_c(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), |
459 |
data->iEdgedWidth, 256*4096); |
data->iEdgedWidth, data->temp+1); |
460 |
|
|
461 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
462 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
463 |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
464 |
*dir = Direction; } |
*dir = Direction; |
465 |
} |
} |
466 |
|
if (data->temp[1] < data->iMinSAD[1]) { |
467 |
|
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
468 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
469 |
|
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
470 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
471 |
|
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
472 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
473 |
|
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
474 |
|
|
475 |
|
} |
476 |
|
|
477 |
static void |
static void |
478 |
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) |
479 |
{ |
{ |
480 |
int32_t sad; |
int32_t sad, xb, yb, xcf, ycf, xcb, ycb; |
481 |
const int xb = data->currentMV[1].x; |
uint32_t t; |
|
const int yb = data->currentMV[1].y; |
|
482 |
const uint8_t *ReferenceF, *ReferenceB; |
const uint8_t *ReferenceF, *ReferenceB; |
483 |
|
VECTOR *current; |
484 |
|
|
485 |
if (( xf > data->max_dx) || ( xf < data->min_dx) |
if ( (xf > data->max_dx) | (xf < data->min_dx) |
486 |
|| ( yf > data->max_dy) || (yf < data->min_dy)) return; |
| (yf > data->max_dy) | (yf < data->min_dy) ) return; |
487 |
|
|
488 |
switch ( ((xf&1)<<1) + (yf&1) ) { |
if (!data->qpel_precision) { |
489 |
case 0 : ReferenceF = data->Ref + xf/2 + (yf/2)*(data->iEdgedWidth); break; |
ReferenceF = GetReference(xf, yf, data); |
490 |
case 1 : ReferenceF = data->RefV + xf/2 + ((yf-1)/2)*(data->iEdgedWidth); break; |
xb = data->currentMV[1].x; yb = data->currentMV[1].y; |
491 |
case 2 : ReferenceF = data->RefH + (xf-1)/2 + (yf/2)*(data->iEdgedWidth); break; |
ReferenceB = GetReferenceB(xb, yb, 1, data); |
492 |
default : ReferenceF = data->RefHV + (xf-1)/2 + ((yf-1)/2)*(data->iEdgedWidth); break; |
current = data->currentMV; |
493 |
|
xcf = xf; ycf = yf; |
494 |
|
xcb = xb; ycb = yb; |
495 |
|
} else { |
496 |
|
ReferenceF = Interpolate16x16qpel(xf, yf, 0, data); |
497 |
|
xb = data->currentQMV[1].x; yb = data->currentQMV[1].y; |
498 |
|
current = data->currentQMV; |
499 |
|
ReferenceB = Interpolate16x16qpel(xb, yb, 1, data); |
500 |
|
xcf = xf/2; ycf = yf/2; |
501 |
|
xcb = xb/2; ycb = yb/2; |
502 |
} |
} |
503 |
|
|
504 |
switch ( ((xb&1)<<1) + (yb&1) ) { |
t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0) |
505 |
case 0 : ReferenceB = data->bRef + xb/2 + (yb/2)*(data->iEdgedWidth); break; |
+ d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
|
case 1 : ReferenceB = data->bRefV + xb/2 + ((yb-1)/2)*(data->iEdgedWidth); break; |
|
|
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; |
|
|
} |
|
506 |
|
|
507 |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
508 |
|
sad += (data->lambda16 * t * sad)>>10; |
509 |
|
|
510 |
if (data->qpel) |
if (data->chroma) sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3], |
511 |
sad += (data->lambda16 * |
(ycf >> 1) + roundtab_79[ycf & 0x3], |
512 |
( d_mv_bits(2*xf - data->predMV.x, 2*yf - data->predMV.y, data->iFcode) + |
(xcb >> 1) + roundtab_79[xcb & 0x3], |
513 |
d_mv_bits(2*xb - data->bpredMV.x, 2*yb - data->bpredMV.y, data->iFcode)) * sad)/1000; |
(ycb >> 1) + roundtab_79[ycb & 0x3], data); |
|
else |
|
|
sad += (data->lambda16 * |
|
|
( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) + |
|
|
d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode)) * sad)/1000; |
|
514 |
|
|
515 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
516 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
517 |
data->currentMV->x = xf; data->currentMV->y = yf; |
current->x = xf; current->y = yf; |
518 |
*dir = Direction; } |
*dir = Direction; |
|
} |
|
|
|
|
|
|
|
|
static void |
|
|
CheckCandidateInt_qpel(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) |
|
|
{ |
|
|
// CheckCandidateInt variant which expects x and y in quarter pixel resolution |
|
|
|
|
|
int32_t sad; |
|
|
const int xb = data->currentQMV[1].x; |
|
|
const int yb = data->currentQMV[1].y; |
|
|
uint8_t * ReferenceF = (uint8_t *)data->RefQ; |
|
|
uint8_t * ReferenceB = (uint8_t *)data->RefQ + 16; |
|
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
|
|
VECTOR halfpelMV; |
|
|
const int32_t iEdgedWidth = data->iEdgedWidth; |
|
|
|
|
|
if (( xf > data->max_dx) || ( xf < data->min_dx) |
|
|
|| ( yf > data->max_dy) || (yf < data->min_dy)) return; |
|
|
|
|
|
halfpelMV.x = xf/2; //forward first |
|
|
halfpelMV.y = yf/2; |
|
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this reference is used in all cases |
|
|
switch( ((xf&1)<<1) + (yf&1) ) |
|
|
{ |
|
|
case 0: // pure halfpel position - shouldn't happen during a refinement step |
|
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ReferenceF); |
|
|
break; |
|
|
|
|
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
|
|
GET_REFERENCE(halfpelMV.x, yf - halfpelMV.y, ref2); |
|
|
interpolate8x8_avg2(ReferenceF, ref1, ref2, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceF+8, ref1+8, ref2+8, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceF+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceF+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); |
|
|
break; |
|
|
|
|
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
|
|
GET_REFERENCE(xf - halfpelMV.x, halfpelMV.y, ref2); |
|
|
interpolate8x8_avg2(ReferenceF, ref1, ref2, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceF+8, ref1+8, ref2+8, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceF+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceF+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); |
|
|
break; |
|
|
|
|
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
|
|
// bottom left/right) during qpel refinement |
|
|
GET_REFERENCE(halfpelMV.x, yf - halfpelMV.y, ref2); |
|
|
GET_REFERENCE(xf - halfpelMV.x, halfpelMV.y, ref3); |
|
|
GET_REFERENCE(xf - halfpelMV.x, yf - halfpelMV.y, ref4); |
|
|
|
|
|
interpolate8x8_avg4(ReferenceF, ref1, ref2, ref3, ref4, iEdgedWidth, 0); |
|
|
interpolate8x8_avg4(ReferenceF+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, 0); |
|
|
interpolate8x8_avg4(ReferenceF+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, 0); |
|
|
interpolate8x8_avg4(ReferenceF+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, 0); |
|
|
break; |
|
|
} |
|
|
|
|
|
halfpelMV.x = xb/2; //backward |
|
|
halfpelMV.y = yb/2; |
|
|
GET_REFERENCE2(halfpelMV.x, halfpelMV.y, ref1); // this reference is used in all cases |
|
|
switch( ((xb&1)<<1) + (yb&1) ) |
|
|
{ |
|
|
case 0: // pure halfpel position - shouldn't happen during a refinement step |
|
|
GET_REFERENCE2(halfpelMV.x, halfpelMV.y, ReferenceB); |
|
|
break; |
|
|
|
|
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
|
|
GET_REFERENCE2(halfpelMV.x, yb - halfpelMV.y, ref2); |
|
|
interpolate8x8_avg2(ReferenceB, ref1, ref2, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceB+8, ref1+8, ref2+8, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceB+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceB+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); |
|
|
break; |
|
|
|
|
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
|
|
GET_REFERENCE2(xb - halfpelMV.x, halfpelMV.y, ref2); |
|
|
interpolate8x8_avg2(ReferenceB, ref1, ref2, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceB+8, ref1+8, ref2+8, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceB+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceB+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); |
|
|
break; |
|
|
|
|
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
|
|
// bottom left/right) during qpel refinement |
|
|
GET_REFERENCE2(halfpelMV.x, yb - halfpelMV.y, ref2); |
|
|
GET_REFERENCE2(xb - halfpelMV.x, halfpelMV.y, ref3); |
|
|
GET_REFERENCE2(xb - halfpelMV.x, yb - halfpelMV.y, ref4); |
|
|
|
|
|
interpolate8x8_avg4(ReferenceB, ref1, ref2, ref3, ref4, iEdgedWidth, 0); |
|
|
interpolate8x8_avg4(ReferenceB+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, 0); |
|
|
interpolate8x8_avg4(ReferenceB+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, 0); |
|
|
interpolate8x8_avg4(ReferenceB+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, 0); |
|
|
break; |
|
519 |
} |
} |
|
|
|
|
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
|
|
|
|
|
sad += (data->lambda16 * |
|
|
( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) + |
|
|
d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode)) * sad)/1000; |
|
|
|
|
|
if (sad < *(data->iMinSAD)) { |
|
|
*(data->iMinSAD) = sad; |
|
|
data->currentQMV->x = xf; data->currentQMV->y = yf; |
|
|
*dir = Direction; } |
|
520 |
} |
} |
521 |
|
|
522 |
static void |
static void |
523 |
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) |
524 |
{ |
{ |
525 |
int32_t sad = 0; |
int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0; |
526 |
int k; |
uint32_t k; |
527 |
const uint8_t *ReferenceF; |
const uint8_t *ReferenceF; |
528 |
const uint8_t *ReferenceB; |
const uint8_t *ReferenceB; |
529 |
VECTOR mvs, b_mvs; |
VECTOR mvs, b_mvs; |
530 |
|
|
531 |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
if (( x > 31) | ( x < -32) | ( y > 31) | (y < -32)) return; |
532 |
|
|
533 |
for (k = 0; k < 4; k++) { |
for (k = 0; k < 4; k++) { |
534 |
mvs.x = data->directmvF[k].x + x; |
mvs.x = data->directmvF[k].x + x; |
541 |
data->directmvB[k].y |
data->directmvB[k].y |
542 |
: mvs.y - data->referencemv[k].y); |
: mvs.y - data->referencemv[k].y); |
543 |
|
|
544 |
if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
if ( (mvs.x > data->max_dx) | (mvs.x < data->min_dx) |
545 |
|| ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) |
| (mvs.y > data->max_dy) | (mvs.y < data->min_dy) |
546 |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
| (b_mvs.x > data->max_dx) | (b_mvs.x < data->min_dx) |
547 |
|| ( 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; |
548 |
|
|
549 |
switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
if (data->qpel) { |
550 |
case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
xcf += mvs.x/2; ycf += mvs.y/2; |
551 |
case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
xcb += b_mvs.x/2; ycb += b_mvs.y/2; |
552 |
case 2 : ReferenceF = data->RefH + (mvs.x-1)/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
} else { |
553 |
default : ReferenceF = data->RefHV + (mvs.x-1)/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
xcf += mvs.x; ycf += mvs.y; |
554 |
} |
xcb += b_mvs.x; ycb += b_mvs.y; |
555 |
|
mvs.x *= 2; mvs.y *= 2; //we move to qpel precision anyway |
556 |
switch ( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { |
b_mvs.x *= 2; b_mvs.y *= 2; |
|
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; |
|
557 |
} |
} |
558 |
|
|
559 |
|
ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); |
560 |
|
ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); |
561 |
|
|
562 |
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
563 |
ReferenceF + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
ReferenceF, ReferenceB, data->iEdgedWidth); |
|
ReferenceB + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
|
|
data->iEdgedWidth); |
|
564 |
if (sad > *(data->iMinSAD)) return; |
if (sad > *(data->iMinSAD)) return; |
565 |
} |
} |
566 |
|
|
567 |
sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000; |
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
568 |
|
|
569 |
|
if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
570 |
|
(ycf >> 3) + roundtab_76[ycf & 0xf], |
571 |
|
(xcb >> 3) + roundtab_76[xcb & 0xf], |
572 |
|
(ycb >> 3) + roundtab_76[ycb & 0xf], data); |
573 |
|
|
574 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
575 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
576 |
data->currentMV->x = x; data->currentMV->y = y; |
data->currentMV->x = x; data->currentMV->y = y; |
577 |
*dir = Direction; } |
*dir = Direction; |
578 |
|
} |
579 |
} |
} |
|
|
|
580 |
|
|
581 |
static void |
static void |
582 |
CheckCandidateDirect_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
583 |
{ |
{ |
584 |
int32_t sad = 0; |
int32_t sad, xcf, ycf, xcb, ycb; |
585 |
int k; |
const uint8_t *ReferenceF; |
586 |
VECTOR mvs, b_mvs, halfpelMV; |
const uint8_t *ReferenceB; |
587 |
const uint8_t *ref1, *ref2, *ref3, *ref4; |
VECTOR mvs, b_mvs; |
|
uint8_t *ReferenceF, *ReferenceB; |
|
|
const uint32_t iEdgedWidth = data->iEdgedWidth; |
|
|
|
|
|
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
|
588 |
|
|
589 |
for (k = 0; k < 4; k++) { |
if (( x > 31) | ( x < -32) | ( y > 31) | (y < -32)) return; |
|
ReferenceF = (uint8_t *)data->RefQ; |
|
|
ReferenceB = (uint8_t *)data->RefQ + 64; |
|
590 |
|
|
591 |
mvs.x = data->directmvF[k].x + x; |
mvs.x = data->directmvF[0].x + x; |
592 |
b_mvs.x = ((x == 0) ? |
b_mvs.x = ((x == 0) ? |
593 |
data->directmvB[k].x |
data->directmvB[0].x |
594 |
: mvs.x - data->referencemv[k].x); |
: mvs.x - data->referencemv[0].x); |
595 |
|
|
596 |
mvs.y = data->directmvF[k].y + y; |
mvs.y = data->directmvF[0].y + y; |
|
b_mvs.y = ((y == 0) ? |
|
|
data->directmvB[k].y |
|
|
: mvs.y - data->referencemv[k].y); |
|
|
|
|
|
if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
|
|
|| ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) |
|
|
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
|
|
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
|
|
|
|
|
halfpelMV.x = mvs.x/2; //forward first |
|
|
halfpelMV.y = mvs.y/2; |
|
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this reference is used in all cases |
|
|
switch( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
|
|
case 0: // pure halfpel position |
|
|
GET_REFERENCE(halfpelMV.x + 16*(k&1), halfpelMV.y + 16*(k>>1), ReferenceF); |
|
|
break; |
|
|
|
|
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
|
|
GET_REFERENCE(halfpelMV.x, mvs.y - halfpelMV.y, ref2); |
|
|
interpolate8x8_avg2(ReferenceF, ref1+8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
|
|
ref2+ 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0); |
|
|
break; |
|
|
|
|
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
|
|
GET_REFERENCE(mvs.x - halfpelMV.x, halfpelMV.y, ref2); |
|
|
interpolate8x8_avg2(ReferenceF, ref1 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
|
|
ref2 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0); |
|
|
break; |
|
|
|
|
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
|
|
// bottom left/right) during qpel refinement |
|
|
GET_REFERENCE(halfpelMV.x, mvs.y - halfpelMV.y, ref2); |
|
|
GET_REFERENCE(mvs.x - halfpelMV.x, halfpelMV.y, ref3); |
|
|
GET_REFERENCE(mvs.x - halfpelMV.x, mvs.y - halfpelMV.y, ref4); |
|
|
interpolate8x8_avg4(ReferenceF, ref1 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
|
|
ref2 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
|
|
ref3 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
|
|
ref4 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0); |
|
|
break; |
|
|
} |
|
|
|
|
|
halfpelMV.x = b_mvs.x/2; |
|
|
halfpelMV.y = b_mvs.y/2; |
|
|
GET_REFERENCE2(halfpelMV.x, halfpelMV.y, ref1); // this reference is used in most cases |
|
|
switch( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { |
|
|
case 0: // pure halfpel position |
|
|
GET_REFERENCE2(halfpelMV.x + 16*(k&1), halfpelMV.y + 16*(k>>1), ReferenceB); |
|
|
break; |
|
|
|
|
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
|
|
GET_REFERENCE2(halfpelMV.x, b_mvs.y - halfpelMV.y, ref2); |
|
|
interpolate8x8_avg2(ReferenceB, ref1+8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
|
|
ref2+ 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0); |
|
|
break; |
|
|
|
|
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
|
|
GET_REFERENCE2(b_mvs.x - halfpelMV.x, halfpelMV.y, ref2); |
|
|
interpolate8x8_avg2(ReferenceB, ref1 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
|
|
ref2 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0); |
|
|
break; |
|
|
|
|
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
|
|
// bottom left/right) during qpel refinement |
|
|
GET_REFERENCE2(halfpelMV.x, b_mvs.y - halfpelMV.y, ref2); |
|
|
GET_REFERENCE2(b_mvs.x - halfpelMV.x, halfpelMV.y, ref3); |
|
|
GET_REFERENCE2(b_mvs.x - halfpelMV.x, b_mvs.y - halfpelMV.y, ref4); |
|
|
interpolate8x8_avg4(ReferenceB, ref1 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
|
|
ref2 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
|
|
ref3 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
|
|
ref4 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0); |
|
|
break; |
|
|
} |
|
|
|
|
|
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
|
|
ReferenceF, |
|
|
ReferenceB, |
|
|
data->iEdgedWidth); |
|
|
if (sad > *(data->iMinSAD)) return; |
|
|
} |
|
|
|
|
|
sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000; |
|
|
|
|
|
if (sad < *(data->iMinSAD)) { |
|
|
*(data->iMinSAD) = sad; |
|
|
data->currentMV->x = x; data->currentMV->y = y; |
|
|
*dir = Direction; } |
|
|
} |
|
|
|
|
|
static void |
|
|
CheckCandidateDirectno4v_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
|
{ |
|
|
int32_t sad = 0; |
|
|
VECTOR mvs, b_mvs, halfpelMV; |
|
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
|
|
const uint32_t iEdgedWidth = data->iEdgedWidth; |
|
|
uint8_t * ReferenceF = (uint8_t *)data->RefQ; |
|
|
uint8_t * ReferenceB = (uint8_t *)data->RefQ + 64; |
|
|
|
|
|
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
|
|
|
|
|
mvs.x = data->directmvF[0].x + x; |
|
|
b_mvs.x = ((x == 0) ? |
|
|
data->directmvB[0].x |
|
|
: mvs.x - data->referencemv[0].x); |
|
|
|
|
|
mvs.y = data->directmvF[0].y + y; |
|
597 |
b_mvs.y = ((y == 0) ? |
b_mvs.y = ((y == 0) ? |
598 |
data->directmvB[0].y |
data->directmvB[0].y |
599 |
: mvs.y - data->referencemv[0].y); |
: mvs.y - data->referencemv[0].y); |
600 |
|
|
601 |
if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
if ( (mvs.x > data->max_dx) | (mvs.x < data->min_dx) |
602 |
|| ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) |
| (mvs.y > data->max_dy) | (mvs.y < data->min_dy) |
603 |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
| (b_mvs.x > data->max_dx) | (b_mvs.x < data->min_dx) |
604 |
|| ( 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; |
605 |
|
|
606 |
halfpelMV.x = mvs.x/2; //forward first |
if (data->qpel) { |
607 |
halfpelMV.y = mvs.y/2; |
xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2); |
608 |
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this reference is used in all cases |
xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2); |
609 |
switch( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data); |
610 |
case 0: // pure halfpel position |
ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); |
611 |
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ReferenceF); |
} else { |
612 |
break; |
xcf = 4*mvs.x; ycf = 4*mvs.y; |
613 |
|
xcb = 4*b_mvs.x; ycb = 4*b_mvs.y; |
614 |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
ReferenceF = GetReference(mvs.x, mvs.y, data); |
615 |
GET_REFERENCE(halfpelMV.x, mvs.y - halfpelMV.y, ref2); |
ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data); |
|
interpolate8x8_avg2(ReferenceF, ref1, ref2, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceF+8, ref1+8, ref2+8, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceF+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceF+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); |
|
|
break; |
|
|
|
|
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
|
|
GET_REFERENCE(mvs.x - halfpelMV.x, halfpelMV.y, ref2); |
|
|
interpolate8x8_avg2(ReferenceF, ref1, ref2, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceF+8, ref1+8, ref2+8, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceF+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceF+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); |
|
|
break; |
|
|
|
|
|
default: // x and y in qpel resolution |
|
|
GET_REFERENCE(halfpelMV.x, mvs.y - halfpelMV.y, ref2); |
|
|
GET_REFERENCE(mvs.x - halfpelMV.x, halfpelMV.y, ref3); |
|
|
GET_REFERENCE(mvs.x - halfpelMV.x, mvs.y - halfpelMV.y, ref4); |
|
|
|
|
|
interpolate8x8_avg4(ReferenceF, ref1, ref2, ref3, ref4, iEdgedWidth, 0); |
|
|
interpolate8x8_avg4(ReferenceF+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, 0); |
|
|
interpolate8x8_avg4(ReferenceF+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, 0); |
|
|
interpolate8x8_avg4(ReferenceF+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, 0); |
|
|
break; |
|
|
} |
|
|
|
|
|
halfpelMV.x = b_mvs.x/2; //backward |
|
|
halfpelMV.y = b_mvs.y/2; |
|
|
GET_REFERENCE2(halfpelMV.x, halfpelMV.y, ref1); |
|
|
switch( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) |
|
|
{ |
|
|
case 0: // pure halfpel position |
|
|
GET_REFERENCE2(halfpelMV.x, halfpelMV.y, ReferenceB); |
|
|
break; |
|
|
|
|
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
|
|
GET_REFERENCE2(halfpelMV.x, b_mvs.y - halfpelMV.y, ref2); |
|
|
interpolate8x8_avg2(ReferenceB, ref1, ref2, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceB+8, ref1+8, ref2+8, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceB+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceB+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); |
|
|
break; |
|
|
|
|
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
|
|
GET_REFERENCE2(b_mvs.x - halfpelMV.x, halfpelMV.y, ref2); |
|
|
interpolate8x8_avg2(ReferenceB, ref1, ref2, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceB+8, ref1+8, ref2+8, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceB+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); |
|
|
interpolate8x8_avg2(ReferenceB+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); |
|
|
break; |
|
|
|
|
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
|
|
// bottom left/right) during qpel refinement |
|
|
GET_REFERENCE2(halfpelMV.x, b_mvs.y - halfpelMV.y, ref2); |
|
|
GET_REFERENCE2(b_mvs.x - halfpelMV.x, halfpelMV.y, ref3); |
|
|
GET_REFERENCE2(b_mvs.x - halfpelMV.x, b_mvs.y - halfpelMV.y, ref4); |
|
|
|
|
|
interpolate8x8_avg4(ReferenceB, ref1, ref2, ref3, ref4, iEdgedWidth, 0); |
|
|
interpolate8x8_avg4(ReferenceB+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, 0); |
|
|
interpolate8x8_avg4(ReferenceB+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, 0); |
|
|
interpolate8x8_avg4(ReferenceB+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, 0); |
|
|
break; |
|
616 |
} |
} |
617 |
|
|
618 |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
619 |
sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000; |
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
|
|
|
|
if (sad < *(data->iMinSAD)) { |
|
|
*(data->iMinSAD) = sad; |
|
|
data->currentMV->x = x; data->currentMV->y = y; |
|
|
*dir = Direction; } |
|
|
} |
|
|
|
|
|
|
|
|
static void |
|
|
CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
|
{ |
|
|
int32_t sad; |
|
|
const uint8_t *ReferenceF; |
|
|
const uint8_t *ReferenceB; |
|
|
VECTOR mvs, b_mvs; |
|
|
|
|
|
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
|
|
|
|
|
mvs.x = data->directmvF[0].x + x; |
|
|
b_mvs.x = ((x == 0) ? |
|
|
data->directmvB[0].x |
|
|
: mvs.x - data->referencemv[0].x); |
|
|
|
|
|
mvs.y = data->directmvF[0].y + y; |
|
|
b_mvs.y = ((y == 0) ? |
|
|
data->directmvB[0].y |
|
|
: mvs.y - data->referencemv[0].y); |
|
620 |
|
|
621 |
if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
622 |
|| ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) |
(ycf >> 3) + roundtab_76[ycf & 0xf], |
623 |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
(xcb >> 3) + roundtab_76[xcb & 0xf], |
624 |
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
(ycb >> 3) + roundtab_76[ycb & 0xf], data); |
|
|
|
|
switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
|
|
case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
|
|
case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
|
|
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; |
|
|
} |
|
|
|
|
|
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
|
|
sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000; |
|
625 |
|
|
626 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
627 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
628 |
data->currentMV->x = x; data->currentMV->y = y; |
data->currentMV->x = x; data->currentMV->y = y; |
629 |
*dir = Direction; } |
*dir = Direction; |
630 |
} |
} |
|
|
|
|
static void |
|
|
CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
|
{ |
|
|
int32_t sad; int t; |
|
|
const uint8_t * Reference; |
|
|
|
|
|
if (( x > data->max_dx) || ( x < data->min_dx) |
|
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|
|
|
|
|
switch ( ((x&1)<<1) + (y&1) ) |
|
|
{ |
|
|
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
|
|
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
|
|
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
|
|
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
|
631 |
} |
} |
632 |
|
|
|
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
|
|
if (data->qpel) t = d_mv_bits(2 * x - data->predQMV.x, 2 * y - data->predQMV.y, data->iFcode); |
|
|
else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
|
|
|
|
|
sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))/100; |
|
|
|
|
|
if (sad < *(data->iMinSAD)) { |
|
|
*(data->iMinSAD) = sad; |
|
|
data->currentMV->x = x; data->currentMV->y = y; |
|
|
*dir = Direction; } |
|
|
} |
|
633 |
|
|
634 |
static void |
static void |
635 |
CheckCandidate8_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidateBits16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
// CheckCandidate8 variant which expects x and y in quarter pixel resolution |
|
|
// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!) |
|
|
// around currentMV! |
|
|
|
|
636 |
{ |
{ |
|
int32_t sad; |
|
|
uint8_t *Reference = (uint8_t *) data->RefQ; |
|
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
|
|
VECTOR halfpelMV = *(data->currentMV); |
|
637 |
|
|
638 |
int32_t iEdgedWidth = data->iEdgedWidth; |
static int16_t in[64], coeff[64]; |
639 |
uint32_t rounding = data->rounding; |
int32_t bits = 0, sum; |
640 |
|
VECTOR * current; |
641 |
|
const uint8_t * ptr; |
642 |
|
int i, cbp = 0, t, xc, yc; |
643 |
|
|
644 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if (( x > data->max_dx) || ( x < data->min_dx) |
645 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
646 |
|
|
647 |
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); |
if (!data->qpel_precision) { |
648 |
switch( ((x&1)<<1) + (y&1) ) |
ptr = GetReference(x, y, data); |
649 |
{ |
current = data->currentMV; |
650 |
case 0: // pure halfpel position - shouldn't happen during a refinement step |
xc = x; yc = y; |
651 |
GET_REFERENCE(halfpelMV.x, halfpelMV.y, Reference); |
} else { // x and y are in 1/4 precision |
652 |
break; |
ptr = Interpolate16x16qpel(x, y, 0, data); |
653 |
|
current = data->currentQMV; |
654 |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
xc = x/2; yc = y/2; |
655 |
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
} |
656 |
|
|
657 |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
for(i = 0; i < 4; i++) { |
658 |
break; |
int s = 8*((i&1) + (i>>1)*data->iEdgedWidth); |
659 |
|
transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth); |
660 |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
fdct(in); |
661 |
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); |
if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); |
662 |
|
else sum = quant4_inter(coeff, in, data->lambda16); |
663 |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
if (sum > 0) { |
664 |
break; |
cbp |= 1 << (5 - i); |
665 |
|
bits += data->temp[i] = CodeCoeffInter_CalcBits(coeff, scan_tables[0]); |
666 |
|
} else data->temp[i] = 0; |
667 |
|
} |
668 |
|
|
669 |
|
bits += t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
670 |
|
|
671 |
|
if (bits < data->iMinSAD[0]) { // there is still a chance, adding chroma |
672 |
|
xc = (xc >> 1) + roundtab_79[xc & 0x3]; |
673 |
|
yc = (yc >> 1) + roundtab_79[yc & 0x3]; |
674 |
|
|
675 |
|
//chroma U |
676 |
|
ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefCU, 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); |
677 |
|
transfer_8to16subro(in, ptr, data->CurU, data->iEdgedWidth/2); |
678 |
|
fdct(in); |
679 |
|
if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); |
680 |
|
else sum = quant4_inter(coeff, in, data->lambda16); |
681 |
|
if (sum > 0) { |
682 |
|
cbp |= 1 << (5 - 4); |
683 |
|
bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); |
684 |
|
} |
685 |
|
|
686 |
|
if (bits < data->iMinSAD[0]) { |
687 |
|
//chroma V |
688 |
|
ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefCV, 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); |
689 |
|
transfer_8to16subro(in, ptr, data->CurV, data->iEdgedWidth/2); |
690 |
|
fdct(in); |
691 |
|
if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); |
692 |
|
else sum = quant4_inter(coeff, in, data->lambda16); |
693 |
|
if (sum > 0) { |
694 |
|
cbp |= 1 << (5 - 5); |
695 |
|
bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); |
696 |
|
} |
697 |
|
} |
698 |
|
} |
699 |
|
|
700 |
|
bits += cbpy_tab[15-(cbp>>2)].len; |
701 |
|
bits += mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; |
702 |
|
|
703 |
|
if (bits < data->iMinSAD[0]) { |
704 |
|
data->iMinSAD[0] = bits; |
705 |
|
current[0].x = x; current[0].y = y; |
706 |
|
*dir = Direction; |
707 |
|
} |
708 |
|
|
709 |
|
if (data->temp[0] + t < data->iMinSAD[1]) { |
710 |
|
data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; } |
711 |
|
if (data->temp[1] < data->iMinSAD[2]) { |
712 |
|
data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; } |
713 |
|
if (data->temp[2] < data->iMinSAD[3]) { |
714 |
|
data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; } |
715 |
|
if (data->temp[3] < data->iMinSAD[4]) { |
716 |
|
data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; } |
717 |
|
|
718 |
|
} |
719 |
|
static void |
720 |
|
CheckCandidateBits8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
721 |
|
{ |
722 |
|
|
723 |
|
static int16_t in[64], coeff[64]; |
724 |
|
int32_t sum, bits; |
725 |
|
VECTOR * current; |
726 |
|
const uint8_t * ptr; |
727 |
|
int cbp; |
728 |
|
|
729 |
default: // x and y in qpel resolution - the "corners" (top left/right and |
if ( (x > data->max_dx) || (x < data->min_dx) |
730 |
// bottom left/right) during qpel refinement |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
|
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
|
|
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); |
|
|
GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); |
|
731 |
|
|
732 |
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
if (!data->qpel_precision) { |
733 |
break; |
ptr = GetReference(x, y, data); |
734 |
|
current = data->currentMV; |
735 |
|
} else { // x and y are in 1/4 precision |
736 |
|
ptr = Interpolate8x8qpel(x, y, 0, 0, data); |
737 |
|
current = data->currentQMV; |
738 |
|
} |
739 |
|
|
740 |
|
transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth); |
741 |
|
fdct(in); |
742 |
|
if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); |
743 |
|
else sum = quant4_inter(coeff, in, data->lambda16); |
744 |
|
if (sum > 0) { |
745 |
|
bits = CodeCoeffInter_CalcBits(coeff, scan_tables[0]); |
746 |
|
cbp = 1; |
747 |
|
} else cbp = bits = 0; |
748 |
|
|
749 |
|
bits += sum = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
750 |
|
|
751 |
|
if (bits < data->iMinSAD[0]) { |
752 |
|
data->temp[0] = cbp; |
753 |
|
data->iMinSAD[0] = bits; |
754 |
|
current[0].x = x; current[0].y = y; |
755 |
|
*dir = Direction; |
756 |
} |
} |
|
|
|
|
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
|
|
sad += (data->lambda8 * d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode) * (sad+NEIGH_8X8_BIAS))/100; |
|
|
|
|
|
if (sad < *(data->iMinSAD)) { |
|
|
*(data->iMinSAD) = sad; |
|
|
data->currentQMV->x = x; data->currentQMV->y = y; |
|
|
*dir = Direction; } |
|
757 |
} |
} |
758 |
|
|
759 |
/* CHECK_CANDIATE FUNCTIONS END */ |
/* CHECK_CANDIATE FUNCTIONS END */ |
768 |
|
|
769 |
int iDirection; |
int iDirection; |
770 |
|
|
771 |
do { |
for(;;) { //forever |
772 |
iDirection = 0; |
iDirection = 0; |
773 |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
774 |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
777 |
|
|
778 |
/* now we're doing diagonal checks near our candidate */ |
/* now we're doing diagonal checks near our candidate */ |
779 |
|
|
780 |
if (iDirection) { //checking if anything found |
if (iDirection) { //if anything found |
781 |
bDirection = iDirection; |
bDirection = iDirection; |
782 |
iDirection = 0; |
iDirection = 0; |
783 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
786 |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
787 |
} else { // what remains here is up or down |
} else { // what remains here is up or down |
788 |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
789 |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); } |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
790 |
|
} |
791 |
|
|
792 |
if (iDirection) { |
if (iDirection) { |
793 |
bDirection += iDirection; |
bDirection += iDirection; |
794 |
x = data->currentMV->x; y = data->currentMV->y; } |
x = data->currentMV->x; y = data->currentMV->y; |
795 |
|
} |
796 |
} else { //about to quit, eh? not so fast.... |
} else { //about to quit, eh? not so fast.... |
797 |
switch (bDirection) { |
switch (bDirection) { |
798 |
case 2: |
case 2: |
843 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
844 |
} |
} |
845 |
} |
} |
|
while (1); //forever |
|
846 |
} |
} |
847 |
|
|
848 |
static void |
static void |
892 |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
893 |
} else { // what remains here is up or down |
} else { // what remains here is up or down |
894 |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
895 |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); } |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
896 |
|
} |
897 |
bDirection += iDirection; |
bDirection += iDirection; |
898 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
899 |
} |
} |
903 |
|
|
904 |
/* MAINSEARCH FUNCTIONS END */ |
/* MAINSEARCH FUNCTIONS END */ |
905 |
|
|
|
/* HALFPELREFINE COULD BE A MAINSEARCH FUNCTION, BUT THERE IS NO NEED FOR IT */ |
|
|
|
|
|
static void |
|
|
HalfpelRefine(const SearchData * const data) |
|
|
{ |
|
|
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
|
|
|
|
|
VECTOR backupMV = *(data->currentMV); |
|
|
int iDirection; //not needed |
|
|
|
|
|
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); |
|
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); |
|
|
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); |
|
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0); |
|
|
|
|
|
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); |
|
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); |
|
|
|
|
|
CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); |
|
|
CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); |
|
|
} |
|
|
|
|
|
|
|
906 |
static void |
static void |
907 |
QuarterpelRefine(const SearchData * const data) |
SubpelRefine(const SearchData * const data) |
908 |
{ |
{ |
909 |
/* Perform quarter pixel refinement*/ |
/* Do a half-pel or q-pel refinement */ |
910 |
|
const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV; |
911 |
VECTOR backupMV = *(data->currentQMV); |
int iDirection; //only needed because macro expects it |
912 |
int iDirection; //not needed |
|
913 |
|
CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0); |
914 |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); |
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0); |
915 |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); |
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0); |
916 |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); |
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0); |
917 |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0); |
CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0); |
918 |
|
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0); |
919 |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); |
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0); |
920 |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); |
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0); |
|
|
|
|
CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); |
|
|
CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); |
|
|
|
|
921 |
} |
} |
922 |
|
|
923 |
static __inline int |
static __inline int |
924 |
SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
925 |
const int x, const int y, |
const int x, const int y, |
926 |
const uint32_t iEdgedWidth, const uint32_t iQuant) |
const uint32_t stride, const uint32_t iQuant, int rrv) |
927 |
|
|
928 |
{ |
{ |
929 |
/* keep repeating checks for all b-frames before this P frame, |
if(!rrv) { |
930 |
to make sure that SKIP is possible (todo) |
uint32_t sadC = sad8(current->u + x*8 + y*stride*8, |
931 |
how: if skip is not possible set sad00 to a very high value */ |
reference->u + x*8 + y*stride*8, stride); |
|
|
|
|
uint32_t sadC = sad8(current->u + x*8 + y*(iEdgedWidth/2)*8, |
|
|
reference->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); |
|
932 |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
933 |
sadC += sad8(current->v + (x + y*(iEdgedWidth/2))*8, |
sadC += sad8(current->v + (x + y*stride)*8, |
934 |
reference->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); |
reference->v + (x + y*stride)*8, stride); |
935 |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
936 |
|
return 1; |
937 |
|
|
938 |
|
} else { |
939 |
|
uint32_t sadC = sad16(current->u + x*16 + y*stride*16, |
940 |
|
reference->u + x*16 + y*stride*16, stride, 256*4096); |
941 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
942 |
|
sadC += sad16(current->v + (x + y*stride)*16, |
943 |
|
reference->v + (x + y*stride)*16, stride, 256*4096); |
944 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
945 |
return 1; |
return 1; |
946 |
} |
} |
947 |
|
} |
948 |
|
|
949 |
static __inline void |
static __inline void |
950 |
SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
951 |
{ |
{ |
952 |
pMB->mode = MODE_NOT_CODED; |
pMB->mode = MODE_NOT_CODED; |
953 |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV; |
954 |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; |
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV; |
|
|
|
|
pMB->qmvs[0].x = pMB->qmvs[1].x = pMB->qmvs[2].x = pMB->qmvs[3].x = 0; |
|
|
pMB->qmvs[0].y = pMB->qmvs[1].y = pMB->qmvs[2].y = pMB->qmvs[3].y = 0; |
|
|
|
|
955 |
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; |
956 |
} |
} |
957 |
|
|
968 |
const IMAGE *const pCurrent = ¤t->image; |
const IMAGE *const pCurrent = ¤t->image; |
969 |
const IMAGE *const pRef = &reference->image; |
const IMAGE *const pRef = &reference->image; |
970 |
|
|
971 |
const VECTOR zeroMV = { 0, 0 }; |
uint32_t mb_width = pParam->mb_width; |
972 |
|
uint32_t mb_height = pParam->mb_height; |
973 |
|
const uint32_t iEdgedWidth = pParam->edged_width; |
974 |
|
const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->global_flags); |
975 |
|
|
976 |
uint32_t x, y; |
uint32_t x, y; |
977 |
uint32_t iIntra = 0; |
uint32_t iIntra = 0; |
978 |
int32_t InterBias, quant = current->quant, sad00; |
int32_t quant = current->quant, sad00; |
|
uint8_t *qimage; |
|
979 |
|
|
980 |
// some pre-initialized thingies for SearchP |
// some pre-initialized thingies for SearchP |
981 |
int32_t temp[5]; |
int32_t temp[8]; |
982 |
VECTOR currentMV[5]; |
VECTOR currentMV[5]; |
983 |
VECTOR currentQMV[5]; |
VECTOR currentQMV[5]; |
984 |
int32_t iMinSAD[5]; |
int32_t iMinSAD[5]; |
985 |
SearchData Data; |
SearchData Data; |
986 |
Data.iEdgedWidth = pParam->edged_width; |
memset(&Data, 0, sizeof(SearchData)); |
987 |
|
Data.iEdgedWidth = iEdgedWidth; |
988 |
Data.currentMV = currentMV; |
Data.currentMV = currentMV; |
989 |
Data.currentQMV = currentQMV; |
Data.currentQMV = currentQMV; |
990 |
Data.iMinSAD = iMinSAD; |
Data.iMinSAD = iMinSAD; |
992 |
Data.iFcode = current->fcode; |
Data.iFcode = current->fcode; |
993 |
Data.rounding = pParam->m_rounding_type; |
Data.rounding = pParam->m_rounding_type; |
994 |
Data.qpel = pParam->m_quarterpel; |
Data.qpel = pParam->m_quarterpel; |
995 |
Data.chroma = current->global_flags & XVID_ME_COLOUR; |
Data.chroma = MotionFlags & PMV_CHROMA16; |
996 |
|
Data.rrv = current->global_flags & XVID_REDUCED; |
997 |
|
|
998 |
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
if ((current->global_flags & XVID_REDUCED)) { |
999 |
return 1; // allocate some mem for qpel interpolated blocks |
mb_width = (pParam->width + 31) / 32; |
1000 |
// somehow this is dirty since I think we shouldn't use malloc outside |
mb_height = (pParam->height + 31) / 32; |
1001 |
// encoder_create() - so please fix me! |
Data.qpel = 0; |
1002 |
Data.RefQ = qimage; |
} |
1003 |
|
|
1004 |
|
Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose) |
1005 |
if (sadInit) (*sadInit) (); |
if (sadInit) (*sadInit) (); |
1006 |
|
|
1007 |
for (y = 0; y < pParam->mb_height; y++) { |
for (y = 0; y < mb_height; y++) { |
1008 |
for (x = 0; x < pParam->mb_width; x++) { |
for (x = 0; x < mb_width; x++) { |
1009 |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
1010 |
|
|
1011 |
pMB->sad16 |
if (!Data.rrv) pMB->sad16 = |
1012 |
= sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, |
sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16, |
1013 |
pRef->y + (x + y * pParam->edged_width) * 16, |
pRef->y + (x + y * iEdgedWidth) * 16, |
1014 |
pParam->edged_width, pMB->sad8 ); |
pParam->edged_width, pMB->sad8 ); |
1015 |
|
|
1016 |
if (Data.chroma) { |
else pMB->sad16 = |
1017 |
pMB->sad16 += sad8(pCurrent->u + x*8 + y*(pParam->edged_width/2)*8, |
sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32, |
1018 |
pRef->u + x*8 + y*(pParam->edged_width/2)*8, pParam->edged_width/2); |
pRef->y + (x + y * iEdgedWidth) * 32, |
1019 |
|
pParam->edged_width, pMB->sad8 ); |
1020 |
|
|
1021 |
pMB->sad16 += sad8(pCurrent->v + (x + y*(pParam->edged_width/2))*8, |
if (Data.chroma) { |
1022 |
pRef->v + (x + y*(pParam->edged_width/2))*8, pParam->edged_width/2); |
Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8, |
1023 |
|
pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2) |
1024 |
|
+ sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8, |
1025 |
|
pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); |
1026 |
|
pMB->sad16 += Data.temp[7]; |
1027 |
} |
} |
1028 |
|
|
1029 |
sad00 = pMB->sad16; //if no gmc; else sad00 = (..) |
sad00 = pMB->sad16; |
1030 |
|
|
1031 |
if (!(current->global_flags & XVID_LUMIMASKING)) { |
if (!(current->global_flags & XVID_LUMIMASKING)) { |
1032 |
pMB->dquant = NO_CHANGE; |
pMB->dquant = NO_CHANGE; |
|
pMB->quant = current->quant; |
|
1033 |
} else { |
} else { |
1034 |
if (pMB->dquant != NO_CHANGE) { |
if (pMB->dquant != NO_CHANGE) { |
1035 |
quant += DQtab[pMB->dquant]; |
quant += DQtab[pMB->dquant]; |
1036 |
if (quant > 31) quant = 31; |
if (quant > 31) quant = 31; |
1037 |
else if (quant < 1) quant = 1; |
else if (quant < 1) quant = 1; |
1038 |
} |
} |
|
pMB->quant = quant; |
|
1039 |
} |
} |
1040 |
|
pMB->quant = current->quant; |
1041 |
|
|
1042 |
//initial skip decision |
//initial skip decision |
1043 |
/* no early skip for GMC (global vector = skip vector is unknown!) */ |
/* no early skip for GMC (global vector = skip vector is unknown!) */ |
1044 |
if (current->coding_type == P_VOP) { /* no fast SKIP for S(GMC)-VOPs */ |
if (!(current->global_flags & XVID_GMC)) { /* no fast SKIP for S(GMC)-VOPs */ |
1045 |
if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH) |
if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH * (Data.rrv ? 4:1) ) |
1046 |
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) { |
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) { |
1047 |
SkipMacroblockP(pMB, sad00); |
SkipMacroblockP(pMB, sad00); |
1048 |
continue; |
continue; |
1049 |
} |
} |
1050 |
} |
} |
1051 |
|
|
1052 |
SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
1053 |
y, current->motion_flags, pMB->quant, |
y, MotionFlags, current->global_flags, pMB->quant, |
1054 |
&Data, pParam, pMBs, reference->mbs, |
&Data, pParam, pMBs, reference->mbs, |
1055 |
current->global_flags & XVID_INTER4V, pMB); |
current->global_flags & XVID_INTER4V, pMB); |
1056 |
|
|
1057 |
/* 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?" */ |
1058 |
if (current->coding_type == P_VOP) { |
if (!(current->global_flags & XVID_GMC)) { |
1059 |
if ( (pMB->dquant == NO_CHANGE) && (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
if ( pMB->dquant == NO_CHANGE && sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) { |
1060 |
&& ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) ) |
if (!(current->global_flags & XVID_MODEDECISION_BITS)) { |
1061 |
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) { |
if ( (100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1) ) |
1062 |
|
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) |
1063 |
SkipMacroblockP(pMB, sad00); |
SkipMacroblockP(pMB, sad00); |
1064 |
continue; |
} else { // BITS mode decision |
1065 |
} |
if (pMB->sad16 > 10) |
1066 |
} |
SkipMacroblockP(pMB, sad00); // more than 10 bits would be used for this MB - skip |
|
|
|
|
/* finally, intra decision */ |
|
|
|
|
|
InterBias = MV16_INTER_BIAS; |
|
|
if (pMB->quant > 8) InterBias += 100 * (pMB->quant - 8); // to make high quants work |
|
|
if (y != 0) |
|
|
if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; |
|
|
if (x != 0) |
|
|
if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; |
|
1067 |
|
|
|
if (Data.chroma) InterBias += 50; // to compensate bigger SAD |
|
|
|
|
|
if (InterBias < pMB->sad16) { |
|
|
const int32_t deviation = |
|
|
dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
|
|
pParam->edged_width); |
|
|
|
|
|
if (deviation < (pMB->sad16 - InterBias)) { |
|
|
if (++iIntra >= iLimit) { free(qimage); return 1; } |
|
|
pMB->mode = MODE_INTRA; |
|
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = |
|
|
pMB->mvs[3] = zeroMV; |
|
|
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = |
|
|
pMB->qmvs[3] = zeroMV; |
|
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = |
|
|
pMB->sad8[3] = 0; |
|
1068 |
} |
} |
1069 |
} |
} |
1070 |
} |
} |
1071 |
|
if (pMB->mode == MODE_INTRA) |
1072 |
|
if (++iIntra > iLimit) return 1; |
1073 |
|
} |
1074 |
} |
} |
|
free(qimage); |
|
|
|
|
|
if (current->coding_type == S_VOP) /* first GMC step only for S(GMC)-VOPs */ |
|
|
current->GMC_MV = GlobalMotionEst( pMBs, pParam, current->fcode ); |
|
|
else |
|
|
current->GMC_MV = zeroMV; |
|
1075 |
|
|
1076 |
|
if (current->global_flags & XVID_GMC ) /* GMC only for S(GMC)-VOPs */ |
1077 |
|
{ |
1078 |
|
current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV); |
1079 |
|
} |
1080 |
return 0; |
return 0; |
1081 |
} |
} |
1082 |
|
|
1083 |
|
|
|
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
|
|
|
|
1084 |
static __inline int |
static __inline int |
1085 |
make_mask(const VECTOR * const pmv, const int i) |
make_mask(const VECTOR * const pmv, const int i) |
1086 |
{ |
{ |
1088 |
for (j = 0; j < i; j++) { |
for (j = 0; j < i; j++) { |
1089 |
if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already |
if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already |
1090 |
if (pmv[i].x == pmv[j].x) { |
if (pmv[i].x == pmv[j].x) { |
1091 |
if (pmv[i].y == pmv[j].y + iDiamondSize) { mask &= ~4; continue; } |
if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4; |
1092 |
if (pmv[i].y == pmv[j].y - iDiamondSize) { mask &= ~8; continue; } |
else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8; |
1093 |
} else |
} else |
1094 |
if (pmv[i].y == pmv[j].y) { |
if (pmv[i].y == pmv[j].y) { |
1095 |
if (pmv[i].x == pmv[j].x + iDiamondSize) { mask &= ~1; continue; } |
if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1; |
1096 |
if (pmv[i].x == pmv[j].x - iDiamondSize) { mask &= ~2; continue; } |
else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2; |
1097 |
} |
} |
1098 |
} |
} |
1099 |
return mask; |
return mask; |
1100 |
} |
} |
1101 |
|
|
1102 |
static __inline void |
static __inline void |
1103 |
PreparePredictionsP(VECTOR * const pmv, int x, int y, const int iWcount, |
PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, |
1104 |
const int iHcount, const MACROBLOCK * const prevMB) |
int iHcount, const MACROBLOCK * const prevMB, int rrv) |
1105 |
{ |
{ |
1106 |
|
|
1107 |
//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 |
1108 |
|
if (rrv) { iWcount /= 2; iHcount /= 2; } |
1109 |
|
|
1110 |
if ( (y != 0) && (x != (iWcount-1)) ) { // [5] top-right neighbour |
if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour |
1111 |
pmv[5].x = EVEN(pmv[3].x); |
pmv[5].x = EVEN(pmv[3].x); |
1112 |
pmv[5].y = EVEN(pmv[3].y); |
pmv[5].y = EVEN(pmv[3].y); |
1113 |
} else pmv[5].x = pmv[5].y = 0; |
} else pmv[5].x = pmv[5].y = 0; |
1126 |
pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
1127 |
pmv[2].y = EVEN(prevMB->mvs[0].y); |
pmv[2].y = EVEN(prevMB->mvs[0].y); |
1128 |
|
|
1129 |
if ((x != iWcount-1) && (y != iHcount-1)) { |
if ((x < iWcount-1) && (y < iHcount-1)) { |
1130 |
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 |
1131 |
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
1132 |
} else pmv[6].x = pmv[6].y = 0; |
} else pmv[6].x = pmv[6].y = 0; |
1133 |
|
|
1134 |
|
if (rrv) { |
1135 |
|
int i; |
1136 |
|
for (i = 0; i < 7; i++) { |
1137 |
|
pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); |
1138 |
|
pmv[i].y = RRV_MV_SCALEUP(pmv[i].y); |
1139 |
|
} |
1140 |
|
} |
1141 |
|
} |
1142 |
|
|
1143 |
|
static int |
1144 |
|
ModeDecision(const uint32_t iQuant, SearchData * const Data, |
1145 |
|
int inter4v, |
1146 |
|
MACROBLOCK * const pMB, |
1147 |
|
const MACROBLOCK * const pMBs, |
1148 |
|
const int x, const int y, |
1149 |
|
const MBParam * const pParam, |
1150 |
|
const uint32_t MotionFlags, |
1151 |
|
const uint32_t GlobalFlags) |
1152 |
|
{ |
1153 |
|
|
1154 |
|
int mode = MODE_INTER; |
1155 |
|
|
1156 |
|
if (!(GlobalFlags & XVID_MODEDECISION_BITS)) { //normal, fast, SAD-based mode decision |
1157 |
|
int intra = 0; |
1158 |
|
int sad; |
1159 |
|
int InterBias = MV16_INTER_BIAS; |
1160 |
|
if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
1161 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) { |
1162 |
|
mode = 0; //inter |
1163 |
|
sad = Data->iMinSAD[0]; |
1164 |
|
} else { |
1165 |
|
mode = MODE_INTER4V; |
1166 |
|
sad = Data->iMinSAD[1] + Data->iMinSAD[2] + |
1167 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant; |
1168 |
|
Data->iMinSAD[0] = sad; |
1169 |
|
} |
1170 |
|
|
1171 |
|
/* intra decision */ |
1172 |
|
|
1173 |
|
if (iQuant > 8) InterBias += 100 * (iQuant - 8); // to make high quants work |
1174 |
|
if (y != 0) |
1175 |
|
if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; |
1176 |
|
if (x != 0) |
1177 |
|
if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; |
1178 |
|
|
1179 |
|
if (Data->chroma) InterBias += 50; // to compensate bigger SAD |
1180 |
|
if (Data->rrv) InterBias *= 4; |
1181 |
|
|
1182 |
|
if (InterBias < pMB->sad16) { |
1183 |
|
int32_t deviation; |
1184 |
|
if (!Data->rrv) deviation = dev16(Data->Cur, Data->iEdgedWidth); |
1185 |
|
else deviation = dev16(Data->Cur, Data->iEdgedWidth) + |
1186 |
|
dev16(Data->Cur+8, Data->iEdgedWidth) + |
1187 |
|
dev16(Data->Cur + 8*Data->iEdgedWidth, Data->iEdgedWidth) + |
1188 |
|
dev16(Data->Cur+8+8*Data->iEdgedWidth, Data->iEdgedWidth); |
1189 |
|
|
1190 |
|
if (deviation < (sad - InterBias)) return MODE_INTRA;// intra |
1191 |
|
} |
1192 |
|
return mode; |
1193 |
|
|
1194 |
|
} else { |
1195 |
|
|
1196 |
|
int bits, intra, i; |
1197 |
|
VECTOR backup[5], *v; |
1198 |
|
Data->lambda16 = iQuant; |
1199 |
|
Data->lambda8 = pParam->m_quant_type; |
1200 |
|
|
1201 |
|
v = Data->qpel ? Data->currentQMV : Data->currentMV; |
1202 |
|
for (i = 0; i < 5; i++) { |
1203 |
|
Data->iMinSAD[i] = 256*4096; |
1204 |
|
backup[i] = v[i]; |
1205 |
|
} |
1206 |
|
|
1207 |
|
bits = CountMBBitsInter(Data, pMBs, x, y, pParam, MotionFlags); |
1208 |
|
if (bits == 0) return MODE_INTER; // quick stop |
1209 |
|
|
1210 |
|
if (inter4v) { |
1211 |
|
int inter4v = CountMBBitsInter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup); |
1212 |
|
if (inter4v < bits) { Data->iMinSAD[0] = bits = inter4v; mode = MODE_INTER4V; } |
1213 |
|
} |
1214 |
|
|
1215 |
|
|
1216 |
|
intra = CountMBBitsIntra(Data); |
1217 |
|
|
1218 |
|
if (intra < bits) { *Data->iMinSAD = bits = intra; return MODE_INTRA; } |
1219 |
|
|
1220 |
|
return mode; |
1221 |
|
} |
1222 |
} |
} |
1223 |
|
|
1224 |
static void |
static void |
1230 |
const int x, |
const int x, |
1231 |
const int y, |
const int y, |
1232 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
1233 |
|
const uint32_t GlobalFlags, |
1234 |
const uint32_t iQuant, |
const uint32_t iQuant, |
1235 |
SearchData * const Data, |
SearchData * const Data, |
1236 |
const MBParam * const pParam, |
const MBParam * const pParam, |
1243 |
int i, iDirection = 255, mask, threshA; |
int i, iDirection = 255, mask, threshA; |
1244 |
VECTOR pmv[7]; |
VECTOR pmv[7]; |
1245 |
|
|
|
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); //has to be changed to get_pmv(2)() |
|
1246 |
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, |
1247 |
pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
1248 |
|
|
1249 |
Data->predMV = pmv[0]; |
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); |
1250 |
|
|
1251 |
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16; |
Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache |
1252 |
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
i = Data->rrv ? 2 : 1; |
1253 |
Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i; |
1254 |
|
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1255 |
Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16; |
Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1256 |
Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16; |
|
1257 |
Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16; |
Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16*i; |
1258 |
Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16; |
Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16*i; |
1259 |
Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16*i; |
1260 |
Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16*i; |
1261 |
|
Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1262 |
|
Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1263 |
|
|
1264 |
Data->lambda16 = lambda_vec16[iQuant]; |
Data->lambda16 = lambda_vec16[iQuant]; |
1265 |
Data->lambda8 = lambda_vec8[iQuant]; |
Data->lambda8 = lambda_vec8[iQuant]; |
1266 |
|
Data->qpel_precision = 0; |
|
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); } |
|
1267 |
|
|
1268 |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
1269 |
|
|
1270 |
for(i = 0; i < 5; i++) |
for(i = 0; i < 5; i++) |
1271 |
Data->currentMV[i].x = Data->currentMV[i].y = 0; |
Data->currentMV[i].x = Data->currentMV[i].y = 0; |
1272 |
|
|
1273 |
if (pParam->m_quarterpel) { |
if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
1274 |
Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
else Data->predMV = pmv[0]; |
|
i = d_mv_bits(Data->predQMV.x, Data->predQMV.y, Data->iFcode); |
|
|
} else i = d_mv_bits(Data->predMV.x, Data->predMV.y, Data->iFcode); |
|
1275 |
|
|
1276 |
Data->iMinSAD[0] = pMB->sad16 + (Data->lambda16 * i * pMB->sad16)/1000; |
i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0); |
1277 |
Data->iMinSAD[1] = pMB->sad8[0] + (Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS))/100; |
Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10); |
1278 |
|
Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10); |
1279 |
Data->iMinSAD[2] = pMB->sad8[1]; |
Data->iMinSAD[2] = pMB->sad8[1]; |
1280 |
Data->iMinSAD[3] = pMB->sad8[2]; |
Data->iMinSAD[3] = pMB->sad8[2]; |
1281 |
Data->iMinSAD[4] = pMB->sad8[3]; |
Data->iMinSAD[4] = pMB->sad8[3]; |
1282 |
|
|
1283 |
if ((x == 0) && (y == 0)) threshA = 512; |
if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) || (x | y)) { |
1284 |
else { |
threshA = Data->temp[0]; // that's where we keep this SAD atm |
|
threshA = Data->temp[0]; // that's when we keep this SAD atm |
|
1285 |
if (threshA < 512) threshA = 512; |
if (threshA < 512) threshA = 512; |
1286 |
if (threshA > 1024) threshA = 1024; } |
else if (threshA > 1024) threshA = 1024; |
1287 |
|
} else |
1288 |
|
threshA = 512; |
1289 |
|
|
1290 |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
1291 |
prevMBs + x + y * pParam->mb_width); |
prevMBs + x + y * pParam->mb_width, Data->rrv); |
1292 |
|
|
1293 |
if (inter4v || pParam->m_quarterpel || Data->chroma) CheckCandidate = CheckCandidate16; |
if (!Data->rrv) { |
1294 |
else CheckCandidate = CheckCandidate16no4v; |
if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16; |
1295 |
|
else CheckCandidate = CheckCandidate16no4v; //for extra speed |
1296 |
|
} else CheckCandidate = CheckCandidate32; |
1297 |
|
|
1298 |
/* main loop. checking all predictions */ |
/* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/ |
1299 |
|
|
1300 |
for (i = 1; i < 7; i++) { |
for (i = 1; i < 7; i++) { |
1301 |
if (!(mask = make_mask(pmv, i)) ) continue; |
if (!(mask = make_mask(pmv, i)) ) continue; |
1302 |
(*CheckCandidate)(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1303 |
if (Data->iMinSAD[0] <= threshA) break; |
if (Data->iMinSAD[0] <= threshA) break; |
1304 |
} |
} |
1305 |
|
|
1306 |
if ((Data->iMinSAD[0] <= threshA) || |
if ((Data->iMinSAD[0] <= threshA) || |
1307 |
(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
1308 |
(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { |
(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { |
1309 |
inter4v = 0; |
if (!(GlobalFlags & XVID_MODEDECISION_BITS)) inter4v = 0; } |
1310 |
} else { |
else { |
1311 |
|
|
1312 |
MainSearchFunc * MainSearchPtr; |
MainSearchFunc * MainSearchPtr; |
1313 |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1314 |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1315 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1316 |
|
|
1317 |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
1318 |
|
|
1319 |
/* extended search, diamond starting in 0,0 and in prediction. |
/* extended search, diamond starting in 0,0 and in prediction. |
1320 |
note that this search is/might be done in halfpel positions, |
note that this search is/might be done in halfpel positions, |
1323 |
if (MotionFlags & PMV_EXTSEARCH16) { |
if (MotionFlags & PMV_EXTSEARCH16) { |
1324 |
int32_t bSAD; |
int32_t bSAD; |
1325 |
VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; |
VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; |
1326 |
if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? |
if (Data->rrv) { |
1327 |
startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); |
startMV.x = RRV_MV_SCALEUP(startMV.x); |
1328 |
|
startMV.y = RRV_MV_SCALEUP(startMV.y); |
1329 |
|
} |
1330 |
if (!(MVequal(startMV, backupMV))) { |
if (!(MVequal(startMV, backupMV))) { |
1331 |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1332 |
|
|
1333 |
(*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); |
CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data); |
1334 |
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
MainSearchPtr(startMV.x, startMV.y, Data, 255); |
1335 |
if (bSAD < Data->iMinSAD[0]) { |
if (bSAD < Data->iMinSAD[0]) { |
1336 |
Data->currentMV[0] = backupMV; |
Data->currentMV[0] = backupMV; |
1337 |
Data->iMinSAD[0] = bSAD; } |
Data->iMinSAD[0] = bSAD; } |
1338 |
} |
} |
1339 |
|
|
1340 |
backupMV = Data->currentMV[0]; |
backupMV = Data->currentMV[0]; |
1341 |
if (MotionFlags & PMV_HALFPELREFINE16) startMV.x = startMV.y = 1; |
startMV.x = startMV.y = 1; |
|
else startMV.x = startMV.y = 0; |
|
1342 |
if (!(MVequal(startMV, backupMV))) { |
if (!(MVequal(startMV, backupMV))) { |
1343 |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1344 |
|
|
1345 |
(*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); |
CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data); |
1346 |
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
MainSearchPtr(startMV.x, startMV.y, Data, 255); |
1347 |
if (bSAD < Data->iMinSAD[0]) { |
if (bSAD < Data->iMinSAD[0]) { |
1348 |
Data->currentMV[0] = backupMV; |
Data->currentMV[0] = backupMV; |
1349 |
Data->iMinSAD[0] = bSAD; } |
Data->iMinSAD[0] = bSAD; } |
1351 |
} |
} |
1352 |
} |
} |
1353 |
|
|
1354 |
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data); |
if (MotionFlags & PMV_HALFPELREFINE16) |
1355 |
|
if ((!(MotionFlags & HALFPELREFINE16_BITS)) || Data->iMinSAD[0] < 200*(int)iQuant) |
1356 |
|
SubpelRefine(Data); |
1357 |
|
|
1358 |
for(i = 0; i < 5; i++) { |
for(i = 0; i < 5; i++) { |
1359 |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors |
1360 |
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
1361 |
} |
} |
1362 |
|
|
1363 |
if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { |
if (MotionFlags & PMV_QUARTERPELREFINE16) |
1364 |
|
if ((!(MotionFlags & QUARTERPELREFINE16_BITS)) || (Data->iMinSAD[0] < 200*(int)iQuant)) { |
1365 |
CheckCandidate = CheckCandidate16_qpel; |
Data->qpel_precision = 1; |
1366 |
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, |
1367 |
pParam->width, pParam->height, Data->iFcode, 0); |
pParam->width, pParam->height, Data->iFcode, 1, 0); |
1368 |
|
|
1369 |
QuarterpelRefine(Data); |
SubpelRefine(Data); |
1370 |
} |
} |
1371 |
|
|
1372 |
if (Data->iMinSAD[0] < (int32_t)iQuant * 30 ) inter4v = 0; |
if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) && (Data->iMinSAD[0] < (int32_t)iQuant * 30)) inter4v = 0; |
1373 |
if (inter4v) { |
|
1374 |
|
if (inter4v && (!(GlobalFlags & XVID_MODEDECISION_BITS) || |
1375 |
|
(!(MotionFlags & QUARTERPELREFINE8_BITS)) || (!(MotionFlags & HALFPELREFINE8_BITS)) || |
1376 |
|
((!(MotionFlags & EXTSEARCH_BITS)) && (!(MotionFlags&PMV_EXTSEARCH8)) ))) { |
1377 |
|
// if decision is BITS-based and all refinement steps will be done in BITS domain, there is no reason to call this loop |
1378 |
|
|
1379 |
SearchData Data8; |
SearchData Data8; |
1380 |
Data8.iFcode = Data->iFcode; |
memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data |
1381 |
Data8.lambda8 = Data->lambda8; |
|
|
Data8.iEdgedWidth = Data->iEdgedWidth; |
|
|
Data8.RefQ = Data->RefQ; |
|
|
Data8.qpel = Data->qpel; |
|
1382 |
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
1383 |
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); |
1384 |
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); |
1385 |
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); |
1386 |
|
|
1387 |
if (Data->chroma) { |
if ((Data->chroma) && (!(GlobalFlags & XVID_MODEDECISION_BITS))) { |
1388 |
int sum, dx, dy; |
// chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, there is no reason to compute it |
1389 |
|
int sumx = 0, sumy = 0; |
1390 |
if(pParam->m_quarterpel) { |
const int div = 1 + Data->qpel; |
1391 |
sum = pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2; |
const VECTOR * const mv = Data->qpel ? pMB->qmvs : pMB->mvs; |
|
} else sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
|
|
dy = (sum >> 3) + roundtab_76[sum & 0xf]; |
|
1392 |
|
|
1393 |
if(pParam->m_quarterpel) { |
for (i = 0; i < 4; i++) { |
1394 |
sum = pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2; |
sumx += mv[i].x / div; |
1395 |
} else sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
sumy += mv[i].y / div; |
1396 |
dx = (sum >> 3) + roundtab_76[sum & 0xf]; |
} |
1397 |
|
|
1398 |
Data->iMinSAD[1] += ChromaSAD(dx, dy, Data); |
Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf], |
1399 |
|
(sumy >> 3) + roundtab_76[sumy & 0xf], Data); |
1400 |
|
} |
1401 |
} |
} |
1402 |
|
|
1403 |
|
inter4v = ModeDecision(iQuant, Data, inter4v, pMB, pMBs, x, y, pParam, MotionFlags, GlobalFlags); |
1404 |
|
|
1405 |
|
if (Data->rrv) { |
1406 |
|
Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); |
1407 |
|
Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); |
1408 |
} |
} |
1409 |
|
|
1410 |
if (!(inter4v) || |
if (inter4v == MODE_INTER) { |
|
(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
|
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
|
|
// INTER MODE |
|
1411 |
pMB->mode = MODE_INTER; |
pMB->mode = MODE_INTER; |
1412 |
pMB->mvs[0] = pMB->mvs[1] |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
1413 |
= pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
1414 |
|
|
1415 |
|
if(Data->qpel) { |
1416 |
pMB->qmvs[0] = pMB->qmvs[1] |
pMB->qmvs[0] = pMB->qmvs[1] |
1417 |
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
1418 |
|
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; |
1419 |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; |
|
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
|
|
|
|
|
if(pParam->m_quarterpel) { |
|
|
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predQMV.x; |
|
|
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predQMV.y; |
|
1420 |
} else { |
} else { |
1421 |
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
1422 |
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
1423 |
} |
} |
1424 |
} else { |
|
1425 |
// INTER4V MODE; all other things are already set in Search8 |
} else if (inter4v == MODE_INTER4V) { |
1426 |
pMB->mode = MODE_INTER4V; |
pMB->mode = MODE_INTER4V; |
1427 |
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + |
pMB->sad16 = Data->iMinSAD[0]; |
1428 |
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant; |
} else { // INTRA mode |
1429 |
|
SkipMacroblockP(pMB, 0); // not skip, but similar enough |
1430 |
|
pMB->mode = MODE_INTRA; |
1431 |
} |
} |
1432 |
|
|
1433 |
} |
} |
1434 |
|
|
1435 |
static void |
static void |
1442 |
const int block, |
const int block, |
1443 |
SearchData * const Data) |
SearchData * const Data) |
1444 |
{ |
{ |
1445 |
|
int i = 0; |
1446 |
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
1447 |
Data->currentMV = OldData->currentMV + 1 + block; |
Data->currentMV = OldData->currentMV + 1 + block; |
1448 |
Data->currentQMV = OldData->currentQMV + 1 + block; |
Data->currentQMV = OldData->currentQMV + 1 + block; |
1449 |
|
|
1450 |
if(pParam->m_quarterpel) { |
if(Data->qpel) { |
1451 |
Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
1452 |
if (block != 0) *(Data->iMinSAD) += (Data->lambda8 * |
if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y, |
1453 |
d_mv_bits( Data->currentQMV->x - Data->predQMV.x, |
Data->predMV, Data->iFcode, 0, 0); |
|
Data->currentQMV->y - Data->predQMV.y, |
|
|
Data->iFcode) * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100; |
|
1454 |
} else { |
} else { |
1455 |
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
1456 |
if (block != 0) *(Data->iMinSAD) += (Data->lambda8 * |
if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y, |
1457 |
d_mv_bits( Data->currentMV->x - Data->predMV.x, |
Data->predMV, Data->iFcode, 0, Data->rrv); |
|
Data->currentMV->y - Data->predMV.y, |
|
|
Data->iFcode) * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100; |
|
1458 |
} |
} |
1459 |
|
|
1460 |
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { |
*(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10; |
1461 |
|
|
1462 |
|
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8|PMV_QUARTERPELREFINE8)) { |
1463 |
|
if (Data->rrv) i = 2; else i = 1; |
1464 |
|
|
1465 |
Data->Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->Ref = OldData->Ref + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1466 |
Data->RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->RefH = OldData->RefH + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1467 |
Data->RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->RefV = OldData->RefV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1468 |
Data->RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->RefHV = OldData->RefHV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1469 |
|
|
1470 |
Data->Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->Cur = OldData->Cur + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1471 |
|
Data->qpel_precision = 0; |
1472 |
|
|
1473 |
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, |
1474 |
pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel); |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
|
CheckCandidate = CheckCandidate8; |
|
1475 |
|
|
1476 |
if (MotionFlags & PMV_EXTSEARCH8) { |
if (!Data->rrv) CheckCandidate = CheckCandidate8; |
1477 |
|
else CheckCandidate = CheckCandidate16no4v; |
1478 |
|
|
1479 |
|
if (MotionFlags & PMV_EXTSEARCH8 && (!(MotionFlags & EXTSEARCH_BITS))) { |
1480 |
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
1481 |
|
|
1482 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1484 |
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
1485 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1486 |
|
|
1487 |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1488 |
|
|
1489 |
if(*(Data->iMinSAD) < temp_sad) { |
if(*(Data->iMinSAD) < temp_sad) { |
1490 |
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
1495 |
if (MotionFlags & PMV_HALFPELREFINE8) { |
if (MotionFlags & PMV_HALFPELREFINE8) { |
1496 |
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
1497 |
|
|
1498 |
HalfpelRefine(Data); // perform halfpel refine of current best vector |
SubpelRefine(Data); // perform halfpel refine of current best vector |
1499 |
|
|
1500 |
if(*(Data->iMinSAD) < temp_sad) { // we have found a better match |
if(*(Data->iMinSAD) < temp_sad) { // we have found a better match |
1501 |
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
1503 |
} |
} |
1504 |
} |
} |
1505 |
|
|
1506 |
if(pParam->m_quarterpel) { |
if (Data->qpel && MotionFlags & PMV_QUARTERPELREFINE8) { |
1507 |
if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && |
Data->qpel_precision = 1; |
|
(MotionFlags & PMV_QUARTERPELREFINE8)) { |
|
|
CheckCandidate = CheckCandidate8_qpel; |
|
1508 |
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, |
1509 |
pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel); |
pParam->width, pParam->height, Data->iFcode, 1, 0); |
1510 |
QuarterpelRefine(Data); |
SubpelRefine(Data); |
|
} |
|
1511 |
} |
} |
1512 |
} |
} |
1513 |
|
|
1514 |
if(pParam->m_quarterpel) { |
if (Data->rrv) { |
1515 |
pMB->pmvs[block].x = Data->currentQMV->x - Data->predQMV.x; |
Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x); |
1516 |
pMB->pmvs[block].y = Data->currentQMV->y - Data->predQMV.y; |
Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y); |
1517 |
} |
} |
1518 |
else { |
|
1519 |
|
if(Data->qpel) { |
1520 |
|
pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x; |
1521 |
|
pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y; |
1522 |
|
pMB->qmvs[block] = *Data->currentQMV; |
1523 |
|
} else { |
1524 |
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
1525 |
pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; |
pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; |
1526 |
} |
} |
1527 |
|
|
1528 |
pMB->mvs[block] = *(Data->currentMV); |
pMB->mvs[block] = *Data->currentMV; |
1529 |
pMB->qmvs[block] = *(Data->currentQMV); |
pMB->sad8[block] = 4 * *Data->iMinSAD; |
|
|
|
|
pMB->sad8[block] = 4 * (*Data->iMinSAD); |
|
1530 |
} |
} |
1531 |
|
|
1532 |
/* B-frames code starts here */ |
/* motion estimation for B-frames */ |
1533 |
|
|
1534 |
static __inline VECTOR |
static __inline VECTOR |
1535 |
ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) |
ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) |
1536 |
{ |
{ |
1537 |
/* the stupidiest function ever */ |
/* the stupidiest function ever */ |
1538 |
if (mode == MODE_FORWARD) return pMB->mvs[0]; |
return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]); |
|
else return pMB->b_mvs[0]; |
|
1539 |
} |
} |
1540 |
|
|
1541 |
static void __inline |
static void __inline |
1568 |
pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
1569 |
} else pmv[5].x = pmv[5].y = 0; |
} else pmv[5].x = pmv[5].y = 0; |
1570 |
|
|
1571 |
if ((x != 0)&&(y != 0)) { |
if (x != 0 && y != 0) { |
1572 |
pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); |
pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); |
1573 |
pmv[6].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y); |
1574 |
} else pmv[6].x = pmv[6].y = 0; |
} else pmv[6].x = pmv[6].y = 0; |
|
|
|
|
// more? |
|
1575 |
} |
} |
1576 |
|
|
1577 |
|
|
1578 |
/* search backward or forward, for b-frames */ |
/* search backward or forward */ |
1579 |
static void |
static void |
1580 |
SearchBF( const uint8_t * const pRef, |
SearchBF( const IMAGE * const pRef, |
1581 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1582 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
1583 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
1593 |
SearchData * const Data) |
SearchData * const Data) |
1594 |
{ |
{ |
1595 |
|
|
1596 |
const int32_t iEdgedWidth = pParam->edged_width; |
int i, iDirection = 255, mask; |
|
|
|
|
int i, iDirection, mask; |
|
1597 |
VECTOR pmv[7]; |
VECTOR pmv[7]; |
1598 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1599 |
*Data->iMinSAD = MV_MAX_ERROR; |
*Data->iMinSAD = MV_MAX_ERROR; |
1600 |
Data->iFcode = iFcode; |
Data->iFcode = iFcode; |
1601 |
|
Data->qpel_precision = 0; |
1602 |
|
Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; // reset chroma-sad cache |
1603 |
|
|
1604 |
Data->Ref = pRef + (x + y * iEdgedWidth) * 16; |
Data->Ref = pRef->y + (x + y * Data->iEdgedWidth) * 16; |
1605 |
Data->RefH = pRefH + (x + y * iEdgedWidth) * 16; |
Data->RefH = pRefH + (x + y * Data->iEdgedWidth) * 16; |
1606 |
Data->RefV = pRefV + (x + y * iEdgedWidth) * 16; |
Data->RefV = pRefV + (x + y * Data->iEdgedWidth) * 16; |
1607 |
Data->RefHV = pRefHV + (x + y * iEdgedWidth) * 16; |
Data->RefHV = pRefHV + (x + y * Data->iEdgedWidth) * 16; |
1608 |
|
Data->RefCU = pRef->u + (x + y * Data->iEdgedWidth/2) * 8; |
1609 |
|
Data->RefCV = pRef->v + (x + y * Data->iEdgedWidth/2) * 8; |
1610 |
|
|
1611 |
Data->predMV = *predMV; |
Data->predMV = *predMV; |
1612 |
|
|
1613 |
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, |
1614 |
pParam->width, pParam->height, iFcode, pParam->m_quarterpel); |
pParam->width, pParam->height, iFcode - Data->qpel, 0, 0); |
1615 |
|
|
1616 |
pmv[0] = Data->predMV; |
pmv[0] = Data->predMV; |
1617 |
if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } |
if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } |
1618 |
|
|
1619 |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
1620 |
|
|
1621 |
Data->currentMV->x = Data->currentMV->y = 0; |
Data->currentMV->x = Data->currentMV->y = 0; |
|
|
|
1622 |
CheckCandidate = CheckCandidate16no4v; |
CheckCandidate = CheckCandidate16no4v; |
1623 |
|
|
1624 |
// main loop. checking all predictions |
// main loop. checking all predictions |
1625 |
for (i = 0; i < 8; i++) { |
for (i = 0; i < 7; i++) { |
1626 |
if (!(mask = make_mask(pmv, i)) ) continue; |
if (!(mask = make_mask(pmv, i)) ) continue; |
1627 |
CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1628 |
} |
} |
1629 |
|
|
1630 |
if (MotionFlags & PMV_USESQUARES16) |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1631 |
MainSearchPtr = SquareSearch; |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
|
|
MainSearchPtr = AdvDiamondSearch; |
|
1632 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1633 |
|
|
1634 |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
1635 |
|
|
1636 |
HalfpelRefine(Data); |
SubpelRefine(Data); |
1637 |
|
|
1638 |
if (Data->qpel) { |
if (Data->qpel && *Data->iMinSAD < *best_sad + 300) { |
1639 |
Data->currentQMV->x = 2*Data->currentMV->x; |
Data->currentQMV->x = 2*Data->currentMV->x; |
1640 |
Data->currentQMV->y = 2*Data->currentMV->y; |
Data->currentQMV->y = 2*Data->currentMV->y; |
1641 |
CheckCandidate = CheckCandidate16no4v_qpel; |
Data->qpel_precision = 1; |
1642 |
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, |
1643 |
pParam->width, pParam->height, iFcode, pParam->m_quarterpel); |
pParam->width, pParam->height, iFcode, 1, 0); |
1644 |
QuarterpelRefine(Data); |
SubpelRefine(Data); |
1645 |
} |
} |
1646 |
|
|
1647 |
// three bits are needed to code backward mode. four for forward |
// three bits are needed to code backward mode. four for forward |
1648 |
// we treat the bits just like they were vector's |
|
1649 |
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
1650 |
else *Data->iMinSAD += 3 * Data->lambda16; |
else *Data->iMinSAD += 3 * Data->lambda16; |
1651 |
|
|
1663 |
pMB->pmvs[0].x = Data->currentMV->x - predMV->x; |
pMB->pmvs[0].x = Data->currentMV->x - predMV->x; |
1664 |
pMB->pmvs[0].y = Data->currentMV->y - predMV->y; |
pMB->pmvs[0].y = Data->currentMV->y - predMV->y; |
1665 |
} |
} |
1666 |
if (mode_current == MODE_FORWARD) |
if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; |
1667 |
pMB->mvs[0] = *(Data->currentMV+2) = *Data->currentMV; |
else pMB->b_mvs[0] = *Data->currentMV; |
1668 |
else |
} |
|
pMB->b_mvs[0] = *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search |
|
1669 |
|
|
1670 |
|
if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; |
1671 |
|
else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search |
1672 |
} |
} |
1673 |
|
|
1674 |
|
static void |
1675 |
|
SkipDecisionB(const IMAGE * const pCur, |
1676 |
|
const IMAGE * const f_Ref, |
1677 |
|
const IMAGE * const b_Ref, |
1678 |
|
MACROBLOCK * const pMB, |
1679 |
|
const uint32_t x, const uint32_t y, |
1680 |
|
const SearchData * const Data) |
1681 |
|
{ |
1682 |
|
int dx = 0, dy = 0, b_dx = 0, b_dy = 0; |
1683 |
|
int32_t sum; |
1684 |
|
const int div = 1 + Data->qpel; |
1685 |
|
int k; |
1686 |
|
const uint32_t stride = Data->iEdgedWidth/2; |
1687 |
|
//this is not full chroma compensation, only it's fullpel approximation. should work though |
1688 |
|
|
1689 |
|
for (k = 0; k < 4; k++) { |
1690 |
|
dy += Data->directmvF[k].y / div; |
1691 |
|
dx += Data->directmvF[0].x / div; |
1692 |
|
b_dy += Data->directmvB[0].y / div; |
1693 |
|
b_dx += Data->directmvB[0].x / div; |
1694 |
} |
} |
1695 |
|
|
1696 |
static int32_t |
dy = (dy >> 3) + roundtab_76[dy & 0xf]; |
1697 |
|
dx = (dx >> 3) + roundtab_76[dx & 0xf]; |
1698 |
|
b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf]; |
1699 |
|
b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf]; |
1700 |
|
|
1701 |
|
sum = sad8bi(pCur->u + 8 * x + 8 * y * stride, |
1702 |
|
f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2, |
1703 |
|
b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, |
1704 |
|
stride); |
1705 |
|
|
1706 |
|
if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; //no skip |
1707 |
|
|
1708 |
|
sum += sad8bi(pCur->v + 8*x + 8 * y * stride, |
1709 |
|
f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2, |
1710 |
|
b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, |
1711 |
|
stride); |
1712 |
|
|
1713 |
|
if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) pMB->mode = MODE_DIRECT_NONE_MV; //skipped |
1714 |
|
} |
1715 |
|
|
1716 |
|
static __inline uint32_t |
1717 |
SearchDirect(const IMAGE * const f_Ref, |
SearchDirect(const IMAGE * const f_Ref, |
1718 |
const uint8_t * const f_RefH, |
const uint8_t * const f_RefH, |
1719 |
const uint8_t * const f_RefV, |
const uint8_t * const f_RefV, |
1734 |
|
|
1735 |
{ |
{ |
1736 |
int32_t skip_sad; |
int32_t skip_sad; |
1737 |
int k; |
int k = (x + Data->iEdgedWidth*y) * 16; |
|
|
|
1738 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1739 |
|
|
1740 |
*Data->iMinSAD = 256*4096; |
*Data->iMinSAD = 256*4096; |
1741 |
|
Data->Ref = f_Ref->y + k; |
1742 |
|
Data->RefH = f_RefH + k; |
1743 |
|
Data->RefV = f_RefV + k; |
1744 |
|
Data->RefHV = f_RefHV + k; |
1745 |
|
Data->bRef = b_Ref->y + k; |
1746 |
|
Data->bRefH = b_RefH + k; |
1747 |
|
Data->bRefV = b_RefV + k; |
1748 |
|
Data->bRefHV = b_RefHV + k; |
1749 |
|
Data->RefCU = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; |
1750 |
|
Data->RefCV = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; |
1751 |
|
Data->b_RefCU = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; |
1752 |
|
Data->b_RefCV = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; |
1753 |
|
|
1754 |
|
k = Data->qpel ? 4 : 2; |
1755 |
|
Data->max_dx = k * (pParam->width - x * 16); |
1756 |
|
Data->max_dy = k * (pParam->height - y * 16); |
1757 |
|
Data->min_dx = -k * (16 + x * 16); |
1758 |
|
Data->min_dy = -k * (16 + y * 16); |
1759 |
|
|
1760 |
Data->Ref = f_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs; |
1761 |
Data->RefH = f_RefH + (x + Data->iEdgedWidth*y) * 16; |
Data->qpel_precision = 0; |
|
Data->RefV = f_RefV + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefHV = f_RefHV + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->bRef = b_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->bRefH = b_RefH + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->bRefV = b_RefV + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->bRefHV = b_RefHV + (x + Data->iEdgedWidth*y) * 16; |
|
|
|
|
|
Data->max_dx = 2 * pParam->width - 2 * (x) * 16; |
|
|
Data->max_dy = 2 * pParam->height - 2 * (y) * 16; |
|
|
Data->min_dx = -(2 * 16 + 2 * (x) * 16); |
|
|
Data->min_dy = -(2 * 16 + 2 * (y) * 16); |
|
|
if (Data->qpel) { //we measure in qpixels |
|
|
Data->max_dx *= 2; |
|
|
Data->max_dy *= 2; |
|
|
Data->min_dx *= 2; |
|
|
Data->min_dy *= 2; |
|
|
Data->referencemv = b_mb->qmvs; |
|
|
} else Data->referencemv = b_mb->mvs; |
|
1762 |
|
|
1763 |
for (k = 0; k < 4; k++) { |
for (k = 0; k < 4; k++) { |
1764 |
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); |
1766 |
pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); |
pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); |
1767 |
pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; |
pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; |
1768 |
|
|
1769 |
if ( ( pMB->b_mvs[k].x > Data->max_dx ) || ( pMB->b_mvs[k].x < Data->min_dx ) |
if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx) |
1770 |
|| ( pMB->b_mvs[k].y > Data->max_dy ) || ( pMB->b_mvs[k].y < Data->min_dy )) { |
| (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) { |
1771 |
|
|
1772 |
*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 |
1773 |
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" |
1774 |
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
1775 |
return 0; |
return 256*4096; |
1776 |
} |
} |
1777 |
if (b_mb->mode != MODE_INTER4V) { |
if (b_mb->mode != MODE_INTER4V) { |
1778 |
pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; |
pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; |
1783 |
} |
} |
1784 |
} |
} |
1785 |
|
|
1786 |
if (Data->qpel) { |
CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v; |
|
if (b_mb->mode == MODE_INTER4V) |
|
|
CheckCandidate = CheckCandidateDirect_qpel; |
|
|
else CheckCandidate = CheckCandidateDirectno4v_qpel; |
|
|
} else { |
|
|
if (b_mb->mode == MODE_INTER4V) CheckCandidate = CheckCandidateDirect; |
|
|
else CheckCandidate = CheckCandidateDirectno4v; |
|
|
} |
|
|
|
|
|
(*CheckCandidate)(0, 0, 255, &k, Data); |
|
1787 |
|
|
1788 |
// skip decision |
CheckCandidate(0, 0, 255, &k, Data); |
|
if (*Data->iMinSAD < pMB->quant * SKIP_THRESH_B) { |
|
|
//possible skip - checking chroma. everything copied from MC |
|
|
//this is not full chroma compensation, only it's fullpel approximation. should work though |
|
|
int sum, dx, dy, b_dx, b_dy; |
|
|
|
|
|
if (Data->qpel) { |
|
|
sum = pMB->mvs[0].y/2 + pMB->mvs[1].y/2 + pMB->mvs[2].y/2 + pMB->mvs[3].y/2; |
|
|
dy = (sum >> 3) + roundtab_76[sum & 0xf]; |
|
|
sum = pMB->mvs[0].x/2 + pMB->mvs[1].x/2 + pMB->mvs[2].x/2 + pMB->mvs[3].x/2; |
|
|
dx = (sum >> 3) + roundtab_76[sum & 0xf]; |
|
|
|
|
|
sum = pMB->b_mvs[0].y/2 + pMB->b_mvs[1].y/2 + pMB->b_mvs[2].y/2 + pMB->b_mvs[3].y/2; |
|
|
b_dy = (sum >> 3) + roundtab_76[sum & 0xf]; |
|
|
sum = pMB->b_mvs[0].x/2 + pMB->b_mvs[1].x/2 + pMB->b_mvs[2].x/2 + pMB->b_mvs[3].x/2; |
|
|
b_dx = (sum >> 3) + roundtab_76[sum & 0xf]; |
|
1789 |
|
|
1790 |
} else { |
// initial (fast) skip decision |
1791 |
sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (2 + Data->chroma?1:0)) { |
1792 |
dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
//possible skip |
1793 |
sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
if (Data->chroma) { |
|
dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
|
|
|
|
sum = pMB->b_mvs[0].x + pMB->b_mvs[1].x + pMB->b_mvs[2].x + pMB->b_mvs[3].x; |
|
|
b_dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
|
sum = pMB->b_mvs[0].y + pMB->b_mvs[1].y + pMB->b_mvs[2].y + pMB->b_mvs[3].y; |
|
|
b_dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
|
} |
|
|
sum = sad8bi(pCur->u + 8*x + 8*y*(Data->iEdgedWidth/2), |
|
|
f_Ref->u + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
|
|
b_Ref->u + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
|
|
Data->iEdgedWidth/2); |
|
|
sum += sad8bi(pCur->v + 8*x + 8*y*(Data->iEdgedWidth/2), |
|
|
f_Ref->v + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
|
|
b_Ref->v + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
|
|
Data->iEdgedWidth/2); |
|
|
|
|
|
if (sum < MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) { |
|
1794 |
pMB->mode = MODE_DIRECT_NONE_MV; |
pMB->mode = MODE_DIRECT_NONE_MV; |
1795 |
return *Data->iMinSAD; |
return *Data->iMinSAD; // skip. |
1796 |
|
} else { |
1797 |
|
SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data); |
1798 |
|
if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip. |
1799 |
} |
} |
1800 |
} |
} |
1801 |
|
|
1808 |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1809 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1810 |
|
|
1811 |
(*MainSearchPtr)(0, 0, Data, 255); |
MainSearchPtr(0, 0, Data, 255); |
1812 |
|
|
1813 |
HalfpelRefine(Data); //or qpel refine, if we're in qpel mode |
SubpelRefine(Data); |
1814 |
|
|
|
*Data->iMinSAD += 1 * Data->lambda16; // one bit is needed to code direct mode |
|
1815 |
*best_sad = *Data->iMinSAD; |
*best_sad = *Data->iMinSAD; |
1816 |
|
|
1817 |
if (b_mb->mode == MODE_INTER4V) |
if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT; |
|
pMB->mode = MODE_DIRECT; |
|
1818 |
else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
1819 |
|
|
1820 |
pMB->pmvs[3] = *Data->currentMV; |
pMB->pmvs[3] = *Data->currentMV; |
1846 |
return skip_sad; |
return skip_sad; |
1847 |
} |
} |
1848 |
|
|
1849 |
|
static void |
1850 |
static __inline void |
SearchInterpolate(const IMAGE * const f_Ref, |
|
SearchInterpolate(const uint8_t * const f_Ref, |
|
1851 |
const uint8_t * const f_RefH, |
const uint8_t * const f_RefH, |
1852 |
const uint8_t * const f_RefV, |
const uint8_t * const f_RefV, |
1853 |
const uint8_t * const f_RefHV, |
const uint8_t * const f_RefHV, |
1854 |
const uint8_t * const b_Ref, |
const IMAGE * const b_Ref, |
1855 |
const uint8_t * const b_RefH, |
const uint8_t * const b_RefH, |
1856 |
const uint8_t * const b_RefV, |
const uint8_t * const b_RefV, |
1857 |
const uint8_t * const b_RefHV, |
const uint8_t * const b_RefHV, |
1869 |
|
|
1870 |
{ |
{ |
1871 |
|
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
1872 |
int iDirection, i, j; |
int iDirection, i, j; |
1873 |
SearchData bData; |
SearchData bData; |
1874 |
|
|
1875 |
*(bData.iMinSAD = fData->iMinSAD) = 4096*256; |
fData->qpel_precision = 0; |
1876 |
bData.Cur = fData->Cur; |
memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data |
1877 |
fData->iEdgedWidth = bData.iEdgedWidth = iEdgedWidth; |
*fData->iMinSAD = 4096*256; |
1878 |
bData.currentMV = fData->currentMV + 1; bData.currentQMV = fData->currentQMV + 1; |
bData.currentMV++; bData.currentQMV++; |
|
bData.lambda16 = fData->lambda16; |
|
1879 |
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
1880 |
|
|
1881 |
bData.bRef = fData->Ref = f_Ref + (x + y * iEdgedWidth) * 16; |
i = (x + y * fData->iEdgedWidth) * 16; |
1882 |
bData.bRefH = fData->RefH = f_RefH + (x + y * iEdgedWidth) * 16; |
bData.bRef = fData->Ref = f_Ref->y + i; |
1883 |
bData.bRefV = fData->RefV = f_RefV + (x + y * iEdgedWidth) * 16; |
bData.bRefH = fData->RefH = f_RefH + i; |
1884 |
bData.bRefHV = fData->RefHV = f_RefHV + (x + y * iEdgedWidth) * 16; |
bData.bRefV = fData->RefV = f_RefV + i; |
1885 |
bData.Ref = fData->bRef = b_Ref + (x + y * iEdgedWidth) * 16; |
bData.bRefHV = fData->RefHV = f_RefHV + i; |
1886 |
bData.RefH = fData->bRefH = b_RefH + (x + y * iEdgedWidth) * 16; |
bData.Ref = fData->bRef = b_Ref->y + i; |
1887 |
bData.RefV = fData->bRefV = b_RefV + (x + y * iEdgedWidth) * 16; |
bData.RefH = fData->bRefH = b_RefH + i; |
1888 |
bData.RefHV = fData->bRefHV = b_RefHV + (x + y * iEdgedWidth) * 16; |
bData.RefV = fData->bRefV = b_RefV + i; |
1889 |
bData.RefQ = fData->RefQ; |
bData.RefHV = fData->bRefHV = b_RefHV + i; |
1890 |
|
bData.b_RefCU = fData->RefCU = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; |
1891 |
|
bData.b_RefCV = fData->RefCV = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; |
1892 |
|
bData.RefCU = fData->b_RefCU = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; |
1893 |
|
bData.RefCV = fData->b_RefCV = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; |
1894 |
|
|
1895 |
|
|
1896 |
bData.bpredMV = fData->predMV = *f_predMV; |
bData.bpredMV = fData->predMV = *f_predMV; |
1897 |
fData->bpredMV = bData.predMV = *b_predMV; |
fData->bpredMV = bData.predMV = *b_predMV; |
1898 |
|
fData->currentMV[0] = fData->currentMV[2]; |
1899 |
|
|
1900 |
fData->currentMV[0] = fData->currentMV[3]; |
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); |
1901 |
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, pParam->m_quarterpel); |
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode - fData->qpel, 0, 0); |
|
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, pParam->m_quarterpel); |
|
1902 |
|
|
1903 |
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; |
1904 |
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; |
1905 |
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; |
1906 |
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; |
1907 |
|
|
1908 |
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; |
1909 |
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; |
1910 |
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; |
1911 |
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; |
1912 |
|
|
1913 |
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
1914 |
|
|
1915 |
//diamond. I wish we could use normal mainsearch functions (square, advdiamond) |
//diamond |
|
|
|
1916 |
do { |
do { |
1917 |
iDirection = 255; |
iDirection = 255; |
1918 |
// forward MV moves |
// forward MV moves |
1933 |
|
|
1934 |
} while (!(iDirection)); |
} while (!(iDirection)); |
1935 |
|
|
1936 |
*fData->iMinSAD += 2 * fData->lambda16; // two bits are needed to code interpolate mode. |
//qpel refinement |
|
|
|
1937 |
if (fData->qpel) { |
if (fData->qpel) { |
1938 |
CheckCandidate = CheckCandidateInt_qpel; |
if (*fData->iMinSAD > *best_sad + 500) return; |
1939 |
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 0); |
CheckCandidate = CheckCandidateInt; |
1940 |
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 0); |
fData->qpel_precision = bData.qpel_precision = 1; |
1941 |
|
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0); |
1942 |
|
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0); |
1943 |
fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x; |
fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x; |
1944 |
fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y; |
fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y; |
1945 |
fData->currentQMV[1].x = 2 * fData->currentMV[1].x; |
fData->currentQMV[1].x = 2 * fData->currentMV[1].x; |
1946 |
fData->currentQMV[1].y = 2 * fData->currentMV[1].y; |
fData->currentQMV[1].y = 2 * fData->currentMV[1].y; |
1947 |
QuarterpelRefine(fData); |
SubpelRefine(fData); |
1948 |
|
if (*fData->iMinSAD > *best_sad + 300) return; |
1949 |
fData->currentQMV[2] = fData->currentQMV[0]; |
fData->currentQMV[2] = fData->currentQMV[0]; |
1950 |
QuarterpelRefine(&bData); |
SubpelRefine(&bData); |
1951 |
} |
} |
1952 |
|
|
1953 |
|
*fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode. |
1954 |
|
|
1955 |
if (*fData->iMinSAD < *best_sad) { |
if (*fData->iMinSAD < *best_sad) { |
1956 |
*best_sad = *fData->iMinSAD; |
*best_sad = *fData->iMinSAD; |
1957 |
pMB->mvs[0] = fData->currentMV[0]; |
pMB->mvs[0] = fData->currentMV[0]; |
1992 |
const IMAGE * const b_refHV) |
const IMAGE * const b_refHV) |
1993 |
{ |
{ |
1994 |
uint32_t i, j; |
uint32_t i, j; |
1995 |
int32_t best_sad, skip_sad; |
int32_t best_sad; |
1996 |
|
uint32_t skip_sad; |
1997 |
int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; |
int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; |
|
static const VECTOR zeroMV={0,0}; |
|
1998 |
const MACROBLOCK * const b_mbs = b_reference->mbs; |
const MACROBLOCK * const b_mbs = b_reference->mbs; |
1999 |
|
|
2000 |
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
2001 |
|
|
2002 |
const int32_t TRB = time_pp - time_bp; |
const int32_t TRB = time_pp - time_bp; |
2003 |
const int32_t TRD = time_pp; |
const int32_t TRD = time_pp; |
|
uint8_t * qimage; |
|
2004 |
|
|
2005 |
// some pre-inintialized data for the rest of the search |
// some pre-inintialized data for the rest of the search |
2006 |
|
|
2008 |
int32_t iMinSAD; |
int32_t iMinSAD; |
2009 |
VECTOR currentMV[3]; |
VECTOR currentMV[3]; |
2010 |
VECTOR currentQMV[3]; |
VECTOR currentQMV[3]; |
2011 |
|
int32_t temp[8]; |
2012 |
|
memset(&Data, 0, sizeof(SearchData)); |
2013 |
Data.iEdgedWidth = pParam->edged_width; |
Data.iEdgedWidth = pParam->edged_width; |
2014 |
Data.currentMV = currentMV; Data.currentQMV = currentQMV; |
Data.currentMV = currentMV; Data.currentQMV = currentQMV; |
2015 |
Data.iMinSAD = &iMinSAD; |
Data.iMinSAD = &iMinSAD; |
2016 |
Data.lambda16 = lambda_vec16[frame->quant]; |
Data.lambda16 = lambda_vec16[frame->quant]; |
2017 |
Data.qpel = pParam->m_quarterpel; |
Data.qpel = pParam->m_quarterpel; |
2018 |
|
Data.rounding = 0; |
2019 |
|
Data.chroma = frame->motion_flags & PMV_CHROMA8; |
2020 |
|
Data.temp = temp; |
2021 |
|
|
2022 |
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose) |
|
return; // allocate some mem for qpel interpolated blocks |
|
|
// somehow this is dirty since I think we shouldn't use malloc outside |
|
|
// encoder_create() - so please fix me! |
|
|
Data.RefQ = qimage; |
|
|
|
|
2023 |
// note: i==horizontal, j==vertical |
// note: i==horizontal, j==vertical |
2024 |
for (j = 0; j < pParam->mb_height; j++) { |
for (j = 0; j < pParam->mb_height; j++) { |
2025 |
|
|
2037 |
} |
} |
2038 |
|
|
2039 |
Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
2040 |
|
Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8; |
2041 |
|
Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8; |
2042 |
pMB->quant = frame->quant; |
pMB->quant = frame->quant; |
2043 |
|
|
2044 |
/* direct search comes first, because it (1) checks for SKIP-mode |
/* direct search comes first, because it (1) checks for SKIP-mode |
2057 |
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
2058 |
|
|
2059 |
// forward search |
// forward search |
2060 |
SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
2061 |
&frame->image, i, j, |
&frame->image, i, j, |
2062 |
frame->motion_flags, |
frame->motion_flags, |
2063 |
frame->fcode, pParam, |
frame->fcode, pParam, |
2065 |
MODE_FORWARD, &Data); |
MODE_FORWARD, &Data); |
2066 |
|
|
2067 |
// backward search |
// backward search |
2068 |
SearchBF(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y, |
2069 |
&frame->image, i, j, |
&frame->image, i, j, |
2070 |
frame->motion_flags, |
frame->motion_flags, |
2071 |
frame->bcode, pParam, |
frame->bcode, pParam, |
2073 |
MODE_BACKWARD, &Data); |
MODE_BACKWARD, &Data); |
2074 |
|
|
2075 |
// 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 |
2076 |
|
SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
2077 |
SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
|
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
2078 |
&frame->image, |
&frame->image, |
2079 |
i, j, |
i, j, |
2080 |
frame->fcode, frame->bcode, |
frame->fcode, frame->bcode, |
2084 |
pMB, &best_sad, |
pMB, &best_sad, |
2085 |
&Data); |
&Data); |
2086 |
|
|
2087 |
|
// final skip decision |
2088 |
|
if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2) |
2089 |
|
&& ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) |
2090 |
|
SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data); |
2091 |
|
|
2092 |
switch (pMB->mode) { |
switch (pMB->mode) { |
2093 |
case MODE_FORWARD: |
case MODE_FORWARD: |
2094 |
f_count++; |
f_count++; |
2095 |
if (pParam->m_quarterpel) f_predMV = pMB->qmvs[0]; |
f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; |
|
else f_predMV = pMB->mvs[0]; |
|
2096 |
break; |
break; |
2097 |
case MODE_BACKWARD: |
case MODE_BACKWARD: |
2098 |
b_count++; |
b_count++; |
2099 |
if (pParam->m_quarterpel) b_predMV = pMB->b_qmvs[0]; |
b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; |
|
else b_predMV = pMB->b_mvs[0]; |
|
2100 |
break; |
break; |
2101 |
case MODE_INTERPOLATE: |
case MODE_INTERPOLATE: |
2102 |
i_count++; |
i_count++; |
2103 |
if (pParam->m_quarterpel) { |
f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; |
2104 |
f_predMV = pMB->qmvs[0]; |
b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; |
|
b_predMV = pMB->b_qmvs[0]; |
|
|
} else { |
|
|
f_predMV = pMB->mvs[0]; |
|
|
b_predMV = pMB->b_mvs[0]; |
|
|
} |
|
2105 |
break; |
break; |
2106 |
case MODE_DIRECT: |
case MODE_DIRECT: |
2107 |
case MODE_DIRECT_NO4V: |
case MODE_DIRECT_NO4V: |
2108 |
d_count++; |
d_count++; |
|
break; |
|
2109 |
default: |
default: |
2110 |
break; |
break; |
2111 |
} |
} |
2112 |
} |
} |
2113 |
} |
} |
|
free(qimage); |
|
|
} |
|
|
|
|
|
/* Hinted ME starts here */ |
|
|
|
|
|
static void |
|
|
SearchPhinted ( const IMAGE * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const IMAGE * const pCur, |
|
|
const int x, |
|
|
const int y, |
|
|
const uint32_t MotionFlags, |
|
|
const uint32_t iQuant, |
|
|
const MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
int inter4v, |
|
|
MACROBLOCK * const pMB, |
|
|
SearchData * const Data) |
|
|
{ |
|
|
|
|
|
int i, t; |
|
|
MainSearchFunc * MainSearchPtr; |
|
|
|
|
|
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
|
|
Data->predQMV = get_qpmv2(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, pParam->m_quarterpel); |
|
|
|
|
|
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16; |
|
|
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
|
|
Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
|
|
|
|
|
Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
|
|
Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
|
|
|
|
|
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); |
|
|
} |
|
|
|
|
|
for(i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
|
|
|
|
|
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
|
|
|
|
|
if (inter4v || pParam->m_quarterpel || Data->chroma) CheckCandidate = CheckCandidate16; |
|
|
else CheckCandidate = CheckCandidate16no4v; |
|
|
|
|
|
pMB->mvs[0].x = EVEN(pMB->mvs[0].x); |
|
|
pMB->mvs[0].y = EVEN(pMB->mvs[0].y); |
|
|
if (pMB->mvs[0].x > Data->max_dx) pMB->mvs[0].x = Data->max_dx; // this is in case iFcode changed |
|
|
if (pMB->mvs[0].x < Data->min_dx) pMB->mvs[0].x = Data->min_dx; |
|
|
if (pMB->mvs[0].y > Data->max_dy) pMB->mvs[0].y = Data->max_dy; |
|
|
if (pMB->mvs[0].y < Data->min_dy) pMB->mvs[0].y = Data->min_dy; |
|
|
|
|
|
(*CheckCandidate)(pMB->mvs[0].x, pMB->mvs[0].y, 0, &t, Data); |
|
|
|
|
|
if (pMB->mode == MODE_INTER4V) |
|
|
for (i = 1; i < 4; i++) { // all four vectors will be used as four predictions for 16x16 search |
|
|
pMB->mvs[i].x = EVEN(pMB->mvs[i].x); |
|
|
pMB->mvs[i].y = EVEN(pMB->mvs[i].y); |
|
|
if (!(make_mask(pMB->mvs, i))) |
|
|
(*CheckCandidate)(pMB->mvs[i].x, pMB->mvs[i].y, 0, &t, Data); |
|
|
} |
|
|
|
|
|
if (MotionFlags & PMV_USESQUARES16) |
|
|
MainSearchPtr = SquareSearch; |
|
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
|
|
MainSearchPtr = AdvDiamondSearch; |
|
|
else MainSearchPtr = DiamondSearch; |
|
|
|
|
|
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
|
|
|
|
|
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data); |
|
|
|
|
|
for(i = 0; i < 5; i++) { |
|
|
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors |
|
|
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
|
|
} |
|
|
|
|
|
if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { |
|
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
|
|
pParam->width, pParam->height, Data->iFcode, 0); |
|
|
CheckCandidate = CheckCandidate16_qpel; |
|
|
QuarterpelRefine(Data); |
|
|
} |
|
|
|
|
|
if (inter4v) { |
|
|
SearchData Data8; |
|
|
Data8.iFcode = Data->iFcode; |
|
|
Data8.lambda8 = Data->lambda8; |
|
|
Data8.iEdgedWidth = Data->iEdgedWidth; |
|
|
Data8.RefQ = Data->RefQ; |
|
|
Data8.qpel = Data->qpel; |
|
|
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
|
|
Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); |
|
|
Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); |
|
|
Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); |
|
|
|
|
|
if (Data->chroma) { |
|
|
int sum, dx, dy; |
|
|
|
|
|
if(pParam->m_quarterpel) |
|
|
sum = (pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2); |
|
|
else sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
|
|
dy = (sum ? SIGN(sum) * |
|
|
(roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2) : 0); |
|
|
|
|
|
if(pParam->m_quarterpel) |
|
|
sum = (pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2); |
|
|
else sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
|
|
dx = (sum ? SIGN(sum) * |
|
|
(roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2) : 0); |
|
|
Data->iMinSAD[1] += ChromaSAD(dx, dy, Data); |
|
|
} |
|
|
} |
|
|
|
|
|
if (!(inter4v) || |
|
|
(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] + |
|
|
Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
|
|
// INTER MODE |
|
|
pMB->mode = MODE_INTER; |
|
|
pMB->mvs[0] = pMB->mvs[1] |
|
|
= pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
|
|
|
|
|
pMB->qmvs[0] = pMB->qmvs[1] |
|
|
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
|
|
|
|
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
|
|
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
|
|
|
|
|
if(pParam->m_quarterpel) { |
|
|
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predQMV.x; |
|
|
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predQMV.y; |
|
|
} else { |
|
|
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
|
|
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
|
|
} |
|
|
} else { |
|
|
// INTER4V MODE; all other things are already set in Search8 |
|
|
pMB->mode = MODE_INTER4V; |
|
|
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] |
|
|
+ Data->iMinSAD[4] + IMV16X16 * iQuant; |
|
|
} |
|
|
|
|
|
} |
|
|
|
|
|
void |
|
|
MotionEstimationHinted( MBParam * const pParam, |
|
|
FRAMEINFO * const current, |
|
|
FRAMEINFO * const reference, |
|
|
const IMAGE * const pRefH, |
|
|
const IMAGE * const pRefV, |
|
|
const IMAGE * const pRefHV) |
|
|
{ |
|
|
MACROBLOCK *const pMBs = current->mbs; |
|
|
const IMAGE *const pCurrent = ¤t->image; |
|
|
const IMAGE *const pRef = &reference->image; |
|
|
|
|
|
uint32_t x, y; |
|
|
uint8_t * qimage; |
|
|
int32_t temp[5], quant = current->quant; |
|
|
int32_t iMinSAD[5]; |
|
|
VECTOR currentMV[5], currentQMV[5]; |
|
|
SearchData Data; |
|
|
Data.iEdgedWidth = pParam->edged_width; |
|
|
Data.currentMV = currentMV; |
|
|
Data.currentQMV = currentQMV; |
|
|
Data.iMinSAD = iMinSAD; |
|
|
Data.temp = temp; |
|
|
Data.iFcode = current->fcode; |
|
|
Data.rounding = pParam->m_rounding_type; |
|
|
Data.qpel = pParam->m_quarterpel; |
|
|
Data.chroma = current->global_flags & XVID_ME_COLOUR; |
|
|
|
|
|
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
|
|
return; // allocate some mem for qpel interpolated blocks |
|
|
// somehow this is dirty since I think we shouldn't use malloc outside |
|
|
// encoder_create() - so please fix me! |
|
|
|
|
|
Data.RefQ = qimage; |
|
|
|
|
|
if (sadInit) (*sadInit) (); |
|
|
|
|
|
for (y = 0; y < pParam->mb_height; y++) { |
|
|
for (x = 0; x < pParam->mb_width; x++) { |
|
|
|
|
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
|
|
|
|
|
//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; } |
|
|
else { |
|
|
if (pMB->dquant != NO_CHANGE) { |
|
|
quant += DQtab[pMB->dquant]; |
|
|
if (quant > 31) quant = 31; |
|
|
else if (quant < 1) quant = 1; |
|
|
} |
|
|
pMB->quant = quant; |
|
|
} |
|
|
|
|
|
SearchPhinted(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
|
|
y, current->motion_flags, pMB->quant, |
|
|
pParam, pMBs, current->global_flags & XVID_INTER4V, pMB, |
|
|
&Data); |
|
|
|
|
|
} |
|
|
} |
|
|
free(qimage); |
|
2114 |
} |
} |
2115 |
|
|
2116 |
static __inline int |
static __inline void |
2117 |
MEanalyzeMB ( const uint8_t * const pRef, |
MEanalyzeMB ( const uint8_t * const pRef, |
2118 |
const uint8_t * const pCur, |
const uint8_t * const pCur, |
2119 |
const int x, |
const int x, |
2120 |
const int y, |
const int y, |
2121 |
const MBParam * const pParam, |
const MBParam * const pParam, |
2122 |
const MACROBLOCK * const pMBs, |
MACROBLOCK * const pMBs, |
|
MACROBLOCK * const pMB, |
|
2123 |
SearchData * const Data) |
SearchData * const Data) |
2124 |
{ |
{ |
2125 |
|
|
2126 |
int i = 255, mask; |
int i, mask; |
2127 |
VECTOR pmv[3]; |
VECTOR pmv[3]; |
2128 |
*(Data->iMinSAD) = MV_MAX_ERROR; |
MACROBLOCK * pMB = &pMBs[x + y * pParam->mb_width]; |
2129 |
|
|
2130 |
|
for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
2131 |
|
|
2132 |
//median is only used as prediction. it doesn't have to be real |
//median is only used as prediction. it doesn't have to be real |
2133 |
if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; |
if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; |
2134 |
else |
else |
2135 |
if (x == 1) //left macroblock does not have any vector now |
if (x == 1) //left macroblock does not have any vector now |
2136 |
Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median |
Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median |
2137 |
else if (y == 1) // top macroblock don't have it's vector |
else if (y == 1) // top macroblock doesn't have it's vector |
2138 |
Data->predMV = (pMB - 1)->mvs[0]; // left instead of median |
Data->predMV = (pMB - 1)->mvs[0]; // left instead of median |
2139 |
else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median |
else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median |
2140 |
|
|
2141 |
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, |
2142 |
pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); |
pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, Data->rrv); |
2143 |
|
|
2144 |
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
2145 |
Data->Ref = pRef + (x + y * pParam->edged_width) * 16; |
Data->Ref = pRef + (x + y * pParam->edged_width) * 16; |
2150 |
pmv[2].y = EVEN(Data->predMV.y); |
pmv[2].y = EVEN(Data->predMV.y); |
2151 |
pmv[0].x = pmv[0].y = 0; |
pmv[0].x = pmv[0].y = 0; |
2152 |
|
|
2153 |
(*CheckCandidate)(0, 0, 255, &i, Data); |
CheckCandidate32I(0, 0, 255, &i, Data); |
2154 |
|
|
2155 |
//early skip for 0,0 |
if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) { |
|
if (*Data->iMinSAD < MAX_SAD00_FOR_SKIP * 4) { |
|
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
|
|
pMB->mode = MODE_NOT_CODED; |
|
|
return 0; |
|
|
} |
|
2156 |
|
|
2157 |
if (!(mask = make_mask(pmv, 1))) |
if (!(mask = make_mask(pmv, 1))) |
2158 |
(*CheckCandidate)(pmv[1].x, pmv[1].y, mask, &i, Data); |
CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data); |
2159 |
if (!(mask = make_mask(pmv, 2))) |
if (!(mask = make_mask(pmv, 2))) |
2160 |
(*CheckCandidate)(pmv[2].x, pmv[2].y, mask, &i, Data); |
CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data); |
2161 |
|
|
2162 |
if (*Data->iMinSAD > MAX_SAD00_FOR_SKIP * 4) // diamond only if needed |
if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) // diamond only if needed |
2163 |
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
2164 |
|
|
2165 |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
for (i = 0; i < 4; i++) { |
2166 |
pMB->mode = MODE_INTER; |
MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width]; |
2167 |
return *(Data->iMinSAD); |
MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i]; |
2168 |
|
MB->mode = MODE_INTER; |
2169 |
|
MB->sad16 = Data->iMinSAD[i+1]; |
2170 |
|
} |
2171 |
|
} |
2172 |
} |
} |
2173 |
|
|
2174 |
#define INTRA_THRESH 1350 |
#define INTRA_BIAS 2500 |
2175 |
#define INTER_THRESH 900 |
#define INTRA_THRESH 1500 |
2176 |
|
#define INTER_THRESH 1400 |
2177 |
|
|
2178 |
int |
int |
2179 |
MEanalysis( const IMAGE * const pRef, |
MEanalysis( const IMAGE * const pRef, |
2181 |
MBParam * const pParam, |
MBParam * const pParam, |
2182 |
int maxIntra, //maximum number if non-I frames |
int maxIntra, //maximum number if non-I frames |
2183 |
int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame |
int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame |
2184 |
int bCount) // number if B frames in a row |
int bCount) // number of B frames in a row |
2185 |
{ |
{ |
2186 |
uint32_t x, y, intra = 0; |
uint32_t x, y, intra = 0; |
2187 |
int sSAD = 0; |
int sSAD = 0; |
2189 |
const IMAGE * const pCurrent = &Current->image; |
const IMAGE * const pCurrent = &Current->image; |
2190 |
int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH; |
int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH; |
2191 |
|
|
2192 |
VECTOR currentMV; |
int32_t iMinSAD[5], temp[5]; |
2193 |
int32_t iMinSAD; |
VECTOR currentMV[5]; |
2194 |
SearchData Data; |
SearchData Data; |
2195 |
Data.iEdgedWidth = pParam->edged_width; |
Data.iEdgedWidth = pParam->edged_width; |
2196 |
Data.currentMV = ¤tMV; |
Data.currentMV = currentMV; |
2197 |
Data.iMinSAD = &iMinSAD; |
Data.iMinSAD = iMinSAD; |
2198 |
Data.iFcode = Current->fcode; |
Data.iFcode = Current->fcode; |
2199 |
CheckCandidate = CheckCandidate16no4vI; |
Data.rrv = Current->global_flags & XVID_REDUCED; |
2200 |
|
Data.temp = temp; |
2201 |
|
CheckCandidate = CheckCandidate32I; |
2202 |
|
|
2203 |
if (intraCount < 10) // we're right after an I frame |
if (intraCount != 0 && intraCount < 10) // we're right after an I frame |
2204 |
IntraThresh += 4 * (intraCount - 10) * (intraCount - 10); |
IntraThresh += 4 * (intraCount - 10) * (intraCount - 10); |
2205 |
else |
else |
2206 |
if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec |
if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec |
2207 |
IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; |
IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; |
2208 |
|
|
2209 |
|
InterThresh += 400 * (1 - bCount); |
2210 |
InterThresh += 300 * (1 - bCount); |
if (InterThresh < 300) InterThresh = 300; |
|
if (InterThresh < 200) InterThresh = 200; |
|
2211 |
|
|
2212 |
if (sadInit) (*sadInit) (); |
if (sadInit) (*sadInit) (); |
2213 |
|
|
2214 |
for (y = 1; y < pParam->mb_height-1; y++) { |
for (y = 1; y < pParam->mb_height-1; y += 2) { |
2215 |
for (x = 1; x < pParam->mb_width-1; x++) { |
for (x = 1; x < pParam->mb_width-1; x += 2) { |
2216 |
int sad, dev; |
int i; |
2217 |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
|
2218 |
|
if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; |
2219 |
sad = MEanalyzeMB(pRef->y, pCurrent->y, x, y, |
|
2220 |
pParam, pMBs, pMB, &Data); |
MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data); |
2221 |
|
|
2222 |
if (sad > IntraThresh) { |
for (i = 0; i < 4; i++) { |
2223 |
dev = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
int dev; |
2224 |
|
MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width]; |
2225 |
|
if (pMB->sad16 > IntraThresh) { |
2226 |
|
dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16, |
2227 |
pParam->edged_width); |
pParam->edged_width); |
2228 |
if (dev + IntraThresh < sad) { |
if (dev + IntraThresh < pMB->sad16) { |
2229 |
pMB->mode = MODE_INTRA; |
pMB->mode = MODE_INTRA; |
2230 |
if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return 2; // I frame |
if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return I_VOP; |
2231 |
|
} |
2232 |
} |
} |
2233 |
|
sSAD += pMB->sad16; |
2234 |
} |
} |
|
sSAD += sad; |
|
2235 |
} |
} |
2236 |
} |
} |
2237 |
sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); |
sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); |
2238 |
if (sSAD > InterThresh ) return 1; //P frame |
// if (sSAD > IntraThresh + INTRA_BIAS) return I_VOP; |
2239 |
|
if (sSAD > InterThresh ) return P_VOP; |
2240 |
emms(); |
emms(); |
2241 |
return 0; // B frame |
return B_VOP; |
2242 |
|
|
2243 |
} |
} |
2244 |
|
|
2245 |
int |
|
2246 |
FindFcode( const MBParam * const pParam, |
static WARPPOINTS |
2247 |
const FRAMEINFO * const current) |
GlobalMotionEst(const MACROBLOCK * const pMBs, |
2248 |
|
const MBParam * const pParam, |
2249 |
|
const FRAMEINFO * const current, |
2250 |
|
const FRAMEINFO * const reference, |
2251 |
|
const IMAGE * const pRefH, |
2252 |
|
const IMAGE * const pRefV, |
2253 |
|
const IMAGE * const pRefHV ) |
2254 |
{ |
{ |
|
uint32_t x, y; |
|
|
int max = 0, min = 0, i; |
|
2255 |
|
|
2256 |
for (y = 0; y < pParam->mb_height; y++) { |
const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs |
2257 |
for (x = 0; x < pParam->mb_width; x++) { |
const int deltay=8; |
2258 |
|
const int grad=512; // lower bound for deviation in MB |
2259 |
|
|
2260 |
|
WARPPOINTS gmc; |
2261 |
|
|
2262 |
|
uint32_t mx, my; |
2263 |
|
|
2264 |
|
int MBh = pParam->mb_height; |
2265 |
|
int MBw = pParam->mb_width; |
2266 |
|
|
2267 |
MACROBLOCK *pMB = ¤t->mbs[x + y * pParam->mb_width]; |
int *MBmask= calloc(MBh*MBw,sizeof(int)); |
2268 |
for(i = 0; i < (pMB->mode == MODE_INTER4V ? 4:1); i++) { |
double DtimesF[4] = { 0.,0., 0., 0. }; |
2269 |
if (pMB->mvs[i].x > max) max = pMB->mvs[i].x; |
double sol[4] = { 0., 0., 0., 0. }; |
2270 |
if (pMB->mvs[i].y > max) max = pMB->mvs[i].y; |
double a,b,c,n,denom; |
2271 |
|
double meanx,meany; |
2272 |
|
int num,oldnum; |
2273 |
|
|
2274 |
|
if (!MBmask) { fprintf(stderr,"Mem error\n"); return gmc;} |
2275 |
|
|
2276 |
|
// filter mask of all blocks |
2277 |
|
|
2278 |
|
for (my = 1; my < MBh-1; my++) |
2279 |
|
for (mx = 1; mx < MBw-1; mx++) |
2280 |
|
{ |
2281 |
|
const int mbnum = mx + my * MBw; |
2282 |
|
const MACROBLOCK *pMB = &pMBs[mbnum]; |
2283 |
|
const VECTOR mv = pMB->mvs[0]; |
2284 |
|
|
2285 |
if (pMB->mvs[i].x < min) min = pMB->mvs[i].x; |
if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) |
2286 |
if (pMB->mvs[i].y < min) min = pMB->mvs[i].y; |
continue; |
2287 |
|
|
2288 |
|
if ( ( (ABS(mv.x - (pMB-1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-1)->mvs[0].y) < deltay) ) |
2289 |
|
&& ( (ABS(mv.x - (pMB+1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+1)->mvs[0].y) < deltay) ) |
2290 |
|
&& ( (ABS(mv.x - (pMB-MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-MBw)->mvs[0].y) < deltay) ) |
2291 |
|
&& ( (ABS(mv.x - (pMB+MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) ) |
2292 |
|
MBmask[mbnum]=1; |
2293 |
} |
} |
2294 |
|
|
2295 |
|
for (my = 1; my < MBh-1; my++) |
2296 |
|
for (mx = 1; mx < MBw-1; mx++) |
2297 |
|
{ |
2298 |
|
const uint8_t *const pCur = current->image.y + 16*my*pParam->edged_width + 16*mx; |
2299 |
|
|
2300 |
|
const int mbnum = mx + my * MBw; |
2301 |
|
if (!MBmask[mbnum]) |
2302 |
|
continue; |
2303 |
|
|
2304 |
|
if (sad16 ( pCur, pCur+1 , pParam->edged_width, 65536) <= grad ) |
2305 |
|
MBmask[mbnum] = 0; |
2306 |
|
if (sad16 ( pCur, pCur+pParam->edged_width, pParam->edged_width, 65536) <= grad ) |
2307 |
|
MBmask[mbnum] = 0; |
2308 |
|
|
2309 |
} |
} |
2310 |
|
|
2311 |
|
emms(); |
2312 |
|
|
2313 |
|
do { /* until convergence */ |
2314 |
|
|
2315 |
|
a = b = c = n = 0; |
2316 |
|
DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.; |
2317 |
|
for (my = 0; my < MBh; my++) |
2318 |
|
for (mx = 0; mx < MBw; mx++) |
2319 |
|
{ |
2320 |
|
const int mbnum = mx + my * MBw; |
2321 |
|
const MACROBLOCK *pMB = &pMBs[mbnum]; |
2322 |
|
const VECTOR mv = pMB->mvs[0]; |
2323 |
|
|
2324 |
|
if (!MBmask[mbnum]) |
2325 |
|
continue; |
2326 |
|
|
2327 |
|
n++; |
2328 |
|
a += 16*mx+8; |
2329 |
|
b += 16*my+8; |
2330 |
|
c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8); |
2331 |
|
|
2332 |
|
DtimesF[0] += (double)mv.x; |
2333 |
|
DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8); |
2334 |
|
DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8); |
2335 |
|
DtimesF[3] += (double)mv.y; |
2336 |
|
} |
2337 |
|
|
2338 |
|
denom = a*a+b*b-c*n; |
2339 |
|
|
2340 |
|
/* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */ |
2341 |
|
/* D'*E*F has been calculated in the same loop as matrix */ |
2342 |
|
|
2343 |
|
sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2]; |
2344 |
|
sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3]; |
2345 |
|
sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3]; |
2346 |
|
sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3]; |
2347 |
|
|
2348 |
|
sol[0] /= denom; |
2349 |
|
sol[1] /= denom; |
2350 |
|
sol[2] /= denom; |
2351 |
|
sol[3] /= denom; |
2352 |
|
|
2353 |
|
meanx = meany = 0.; |
2354 |
|
oldnum = 0; |
2355 |
|
for (my = 0; my < MBh; my++) |
2356 |
|
for (mx = 0; mx < MBw; mx++) |
2357 |
|
{ |
2358 |
|
const int mbnum = mx + my * MBw; |
2359 |
|
const MACROBLOCK *pMB = &pMBs[mbnum]; |
2360 |
|
const VECTOR mv = pMB->mvs[0]; |
2361 |
|
|
2362 |
|
if (!MBmask[mbnum]) |
2363 |
|
continue; |
2364 |
|
|
2365 |
|
oldnum++; |
2366 |
|
meanx += ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ); |
2367 |
|
meany += ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ); |
2368 |
} |
} |
2369 |
|
|
2370 |
min = -min; |
if (4*meanx > oldnum) /* better fit than 0.25 is useless */ |
2371 |
max += 1; |
meanx /= oldnum; |
2372 |
if (min > max) max = min; |
else |
2373 |
if (pParam->m_quarterpel) max *= 2; |
meanx = 0.25; |
2374 |
|
|
2375 |
|
if (4*meany > oldnum) |
2376 |
|
meany /= oldnum; |
2377 |
|
else |
2378 |
|
meany = 0.25; |
2379 |
|
|
2380 |
for (i = 1; (max > 32 << (i - 1)); i++); |
/* fprintf(stderr,"sol = (%8.5f, %8.5f, %8.5f, %8.5f)\n",sol[0],sol[1],sol[2],sol[3]); |
2381 |
return i; |
fprintf(stderr,"meanx = %8.5f meany = %8.5f %d\n",meanx,meany, oldnum); |
2382 |
|
*/ |
2383 |
|
num = 0; |
2384 |
|
for (my = 0; my < MBh; my++) |
2385 |
|
for (mx = 0; mx < MBw; mx++) |
2386 |
|
{ |
2387 |
|
const int mbnum = mx + my * MBw; |
2388 |
|
const MACROBLOCK *pMB = &pMBs[mbnum]; |
2389 |
|
const VECTOR mv = pMB->mvs[0]; |
2390 |
|
|
2391 |
|
if (!MBmask[mbnum]) |
2392 |
|
continue; |
2393 |
|
|
2394 |
|
if ( ( ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ) > meanx ) |
2395 |
|
|| ( ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ) > meany ) ) |
2396 |
|
MBmask[mbnum]=0; |
2397 |
|
else |
2398 |
|
num++; |
2399 |
} |
} |
2400 |
|
|
2401 |
static void |
} while ( (oldnum != num) && (num>=4) ); |
2402 |
CheckGMC(int x, int y, const int dir, int * iDirection, |
|
2403 |
const MACROBLOCK * const pMBs, uint32_t * bestcount, VECTOR * GMC, |
if (num < 4) |
2404 |
const MBParam * const pParam) |
{ |
2405 |
|
gmc.duv[0].x= gmc.duv[0].y= gmc.duv[1].x= gmc.duv[1].y= gmc.duv[2].x= gmc.duv[2].y=0; |
2406 |
|
} else { |
2407 |
|
|
2408 |
|
gmc.duv[0].x=(int)(sol[0]+0.5); |
2409 |
|
gmc.duv[0].y=(int)(sol[3]+0.5); |
2410 |
|
|
2411 |
|
gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5); |
2412 |
|
gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5); |
2413 |
|
|
2414 |
|
gmc.duv[2].x=0; |
2415 |
|
gmc.duv[2].y=0; |
2416 |
|
} |
2417 |
|
// fprintf(stderr,"wp1 = ( %4d, %4d) wp2 = ( %4d, %4d) \n", gmc.duv[0].x, gmc.duv[0].y, gmc.duv[1].x, gmc.duv[1].y); |
2418 |
|
|
2419 |
|
free(MBmask); |
2420 |
|
|
2421 |
|
return gmc; |
2422 |
|
} |
2423 |
|
|
2424 |
|
// functions which perform BITS-based search/bitcount |
2425 |
|
|
2426 |
|
static int |
2427 |
|
CountMBBitsInter(SearchData * const Data, |
2428 |
|
const MACROBLOCK * const pMBs, const int x, const int y, |
2429 |
|
const MBParam * const pParam, |
2430 |
|
const uint32_t MotionFlags) |
2431 |
{ |
{ |
2432 |
uint32_t mx, my, a, count = 0; |
int i, iDirection; |
2433 |
|
int32_t bsad[5]; |
2434 |
|
|
2435 |
|
CheckCandidate = CheckCandidateBits16; |
2436 |
|
|
2437 |
|
if (Data->qpel) { |
2438 |
|
for(i = 0; i < 5; i++) { |
2439 |
|
Data->currentMV[i].x = Data->currentQMV[i].x/2; |
2440 |
|
Data->currentMV[i].y = Data->currentQMV[i].y/2; |
2441 |
|
} |
2442 |
|
Data->qpel_precision = 1; |
2443 |
|
CheckCandidateBits16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data); |
2444 |
|
|
2445 |
|
//checking if this vector is perfect. if it is, we stop. |
2446 |
|
if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0) |
2447 |
|
return 0; //quick stop |
2448 |
|
|
2449 |
|
if (MotionFlags & (HALFPELREFINE16_BITS | EXTSEARCH_BITS)) { //we have to prepare for halfpixel-precision search |
2450 |
|
for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i]; |
2451 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2452 |
|
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
2453 |
|
Data->qpel_precision = 0; |
2454 |
|
if (Data->currentQMV->x & 1 || Data->currentQMV->y & 1) |
2455 |
|
CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); |
2456 |
|
} |
2457 |
|
|
2458 |
|
} else { // not qpel |
2459 |
|
|
2460 |
|
CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); |
2461 |
|
//checking if this vector is perfect. if it is, we stop. |
2462 |
|
if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0) { |
2463 |
|
return 0; //inter |
2464 |
|
} |
2465 |
|
} |
2466 |
|
|
2467 |
|
if (MotionFlags&EXTSEARCH_BITS) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
2468 |
|
|
2469 |
for (my = 1; my < pParam->mb_height-1; my++) |
if (MotionFlags&HALFPELREFINE16_BITS) SubpelRefine(Data); |
2470 |
for (mx = 1; mx < pParam->mb_width-1; mx++) { |
|
2471 |
VECTOR mv; |
if (Data->qpel) { |
2472 |
const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; |
if (MotionFlags&(EXTSEARCH_BITS | HALFPELREFINE16_BITS)) { // there was halfpel-precision search |
2473 |
if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) continue; |
for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) { |
2474 |
mv = pMB->mvs[0]; |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // we have found a better match |
2475 |
a = ABS(mv.x - x) + ABS(mv.y - y); |
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
|
if (a < 6) count += 6 - a; |
|
2476 |
} |
} |
2477 |
|
|
2478 |
if (count > *bestcount) { |
// preparing for qpel-precision search |
2479 |
*bestcount = count; |
Data->qpel_precision = 1; |
2480 |
*iDirection = dir; |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2481 |
GMC->x = x; GMC->y = y; |
pParam->width, pParam->height, Data->iFcode, 1, 0); |
2482 |
} |
} |
2483 |
|
if (MotionFlags&QUARTERPELREFINE16_BITS) SubpelRefine(Data); |
2484 |
} |
} |
2485 |
|
|
2486 |
|
if (MotionFlags&CHECKPREDICTION_BITS) { //let's check vector equal to prediction |
2487 |
|
VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV; |
2488 |
|
if (!(Data->predMV.x == v->x && Data->predMV.y == v->y)) |
2489 |
|
CheckCandidateBits16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data); |
2490 |
|
} |
2491 |
|
return Data->iMinSAD[0]; |
2492 |
|
} |
2493 |
|
|
2494 |
static VECTOR |
|
2495 |
GlobalMotionEst(const MACROBLOCK * const pMBs, const MBParam * const pParam, const uint32_t iFcode) |
static int |
2496 |
|
CountMBBitsInter4v(const SearchData * const Data, |
2497 |
|
MACROBLOCK * const pMB, const MACROBLOCK * const pMBs, |
2498 |
|
const int x, const int y, |
2499 |
|
const MBParam * const pParam, const uint32_t MotionFlags, |
2500 |
|
const VECTOR * const backup) |
2501 |
{ |
{ |
2502 |
|
|
2503 |
uint32_t count, bestcount = 0; |
int cbp = 0, bits = 0, t = 0, i, iDirection; |
2504 |
int x, y; |
SearchData Data2, *Data8 = &Data2; |
2505 |
VECTOR gmc = {0,0}; |
int sumx = 0, sumy = 0; |
2506 |
int step, min_x, max_x, min_y, max_y; |
int16_t in[64], coeff[64]; |
2507 |
uint32_t mx, my; |
|
2508 |
int iDirection, bDirection; |
memcpy(Data8, Data, sizeof(SearchData)); |
2509 |
|
CheckCandidate = CheckCandidateBits8; |
2510 |
|
|
2511 |
|
for (i = 0; i < 4; i++) { |
2512 |
|
Data8->iMinSAD = Data->iMinSAD + i + 1; |
2513 |
|
Data8->currentMV = Data->currentMV + i + 1; |
2514 |
|
Data8->currentQMV = Data->currentQMV + i + 1; |
2515 |
|
Data8->Cur = Data->Cur + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2516 |
|
Data8->Ref = Data->Ref + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2517 |
|
Data8->RefH = Data->RefH + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2518 |
|
Data8->RefV = Data->RefV + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2519 |
|
Data8->RefHV = Data->RefHV + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2520 |
|
|
2521 |
|
if(Data->qpel) { |
2522 |
|
Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i); |
2523 |
|
if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y, |
2524 |
|
Data8->predMV, Data8->iFcode, 0, 0); |
2525 |
|
} else { |
2526 |
|
Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i); |
2527 |
|
if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y, |
2528 |
|
Data8->predMV, Data8->iFcode, 0, 0); |
2529 |
|
} |
2530 |
|
|
2531 |
min_x = min_y = -32<<iFcode; |
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, |
2532 |
max_x = max_y = 32<<iFcode; |
pParam->width, pParam->height, Data8->iFcode, Data8->qpel, 0); |
2533 |
|
|
2534 |
//step1: let's find a rough camera panning |
*Data8->iMinSAD += t; |
|
for (step = 32; step >= 2; step /= 2) { |
|
|
bestcount = 0; |
|
|
for (y = min_y; y <= max_y; y += step) |
|
|
for (x = min_x ; x <= max_x; x += step) { |
|
|
count = 0; |
|
|
//for all macroblocks |
|
|
for (my = 1; my < pParam->mb_height-1; my++) |
|
|
for (mx = 1; mx < pParam->mb_width-1; mx++) { |
|
|
const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; |
|
|
VECTOR mv; |
|
2535 |
|
|
2536 |
if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) |
Data8->qpel_precision = Data8->qpel; |
2537 |
continue; |
// checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far) |
2538 |
|
if (Data8->qpel) { |
2539 |
|
if (!(Data8->currentQMV->x == backup[i+1].x && Data8->currentQMV->y == backup[i+1].y)) |
2540 |
|
CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8); |
2541 |
|
} else { |
2542 |
|
if (!(Data8->currentMV->x == backup[i+1].x && Data8->currentMV->y == backup[i+1].y)) |
2543 |
|
CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8); |
2544 |
|
} |
2545 |
|
|
2546 |
mv = pMB->mvs[0]; |
if (Data8->qpel) { |
2547 |
if ( ABS(mv.x - x) <= step && ABS(mv.y - y) <= step ) /* GMC translation is always halfpel-res */ |
if (MotionFlags&HALFPELREFINE8_BITS || (MotionFlags&PMV_EXTSEARCH8 && MotionFlags&EXTSEARCH_BITS)) { // halfpixel motion search follows |
2548 |
count++; |
int32_t s = *Data8->iMinSAD; |
2549 |
|
Data8->currentMV->x = Data8->currentQMV->x/2; |
2550 |
|
Data8->currentMV->y = Data8->currentQMV->y/2; |
2551 |
|
Data8->qpel_precision = 0; |
2552 |
|
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, |
2553 |
|
pParam->width, pParam->height, Data8->iFcode - 1, 0, 0); |
2554 |
|
|
2555 |
|
if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1) |
2556 |
|
CheckCandidateBits8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8); |
2557 |
|
|
2558 |
|
if (MotionFlags & PMV_EXTSEARCH8 && MotionFlags & EXTSEARCH_BITS) |
2559 |
|
SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255); |
2560 |
|
|
2561 |
|
if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8); |
2562 |
|
|
2563 |
|
if(s > *Data8->iMinSAD) { //we have found a better match |
2564 |
|
Data8->currentQMV->x = 2*Data8->currentMV->x; |
2565 |
|
Data8->currentQMV->y = 2*Data8->currentMV->y; |
2566 |
} |
} |
2567 |
if (count >= bestcount) { bestcount = count; gmc.x = x; gmc.y = y; } |
|
2568 |
|
Data8->qpel_precision = 1; |
2569 |
|
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, |
2570 |
|
pParam->width, pParam->height, Data8->iFcode, 1, 0); |
2571 |
|
|
2572 |
} |
} |
2573 |
min_x = gmc.x - step; |
if (MotionFlags & QUARTERPELREFINE8_BITS) SubpelRefine(Data8); |
2574 |
max_x = gmc.x + step; |
|
2575 |
min_y = gmc.y - step; |
} else // not qpel |
2576 |
max_y = gmc.y + step; |
if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8); //halfpel mode, halfpel refinement |
2577 |
|
|
2578 |
|
//checking vector equal to predicion |
2579 |
|
if (i != 0 && MotionFlags & CHECKPREDICTION_BITS) { |
2580 |
|
const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV; |
2581 |
|
if (!(Data8->predMV.x == v->x && Data8->predMV.y == v->y)) |
2582 |
|
CheckCandidateBits8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8); |
2583 |
} |
} |
2584 |
|
|
2585 |
if (bestcount < (pParam->mb_height-2)*(pParam->mb_width-2)/10) |
bits += *Data8->iMinSAD; |
2586 |
gmc.x = gmc.y = 0; //no camara pan, no GMC |
if (bits >= Data->iMinSAD[0]) break; // no chances for INTER4V |
2587 |
|
|
2588 |
// step2: let's refine camera panning using gradiend-descent approach. |
// MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else |
2589 |
// TODO: more warping points may be evaluated here (like in interpolate mode search - two vectors in one diamond) |
if(Data->qpel) { |
2590 |
bestcount = 0; |
pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x; |
2591 |
CheckGMC(gmc.x, gmc.y, 255, &iDirection, pMBs, &bestcount, &gmc, pParam); |
pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y; |
2592 |
do { |
pMB->qmvs[i] = *Data8->currentQMV; |
2593 |
x = gmc.x; y = gmc.y; |
sumx += Data8->currentQMV->x/2; |
2594 |
bDirection = iDirection; iDirection = 0; |
sumy += Data8->currentQMV->y/2; |
2595 |
if (bDirection & 1) CheckGMC(x - 1, y, 1+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); |
} else { |
2596 |
if (bDirection & 2) CheckGMC(x + 1, y, 2+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); |
pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x; |
2597 |
if (bDirection & 4) CheckGMC(x, y - 1, 1+2+4, &iDirection, pMBs, &bestcount, &gmc, pParam); |
pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y; |
2598 |
if (bDirection & 8) CheckGMC(x, y + 1, 1+2+8, &iDirection, pMBs, &bestcount, &gmc, pParam); |
sumx += Data8->currentMV->x; |
2599 |
|
sumy += Data8->currentMV->y; |
2600 |
|
} |
2601 |
|
pMB->mvs[i] = *Data8->currentMV; |
2602 |
|
pMB->sad8[i] = 4 * *Data8->iMinSAD; |
2603 |
|
if (Data8->temp[0]) cbp |= 1 << (5 - i); |
2604 |
|
} |
2605 |
|
|
2606 |
} while (iDirection); |
if (bits < *Data->iMinSAD) { // there is still a chance for inter4v mode. let's check chroma |
2607 |
|
const uint8_t * ptr; |
2608 |
|
sumx = (sumx >> 3) + roundtab_76[sumx & 0xf]; |
2609 |
|
sumy = (sumy >> 3) + roundtab_76[sumy & 0xf]; |
2610 |
|
|
2611 |
if (pParam->m_quarterpel) { |
//chroma U |
2612 |
gmc.x *= 2; |
ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefCU, 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
2613 |
gmc.y *= 2; /* we store the halfpel value as pseudo-qpel to make comparison easier */ |
transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2); |
2614 |
|
fdct(in); |
2615 |
|
if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16); |
2616 |
|
else i = quant4_inter(coeff, in, Data->lambda16); |
2617 |
|
if (i > 0) { |
2618 |
|
bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); |
2619 |
|
cbp |= 1 << (5 - 4); |
2620 |
} |
} |
2621 |
|
|
2622 |
return gmc; |
if (bits < *Data->iMinSAD) { // still possible |
2623 |
|
//chroma V |
2624 |
|
ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefCV, 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
2625 |
|
transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2); |
2626 |
|
fdct(in); |
2627 |
|
if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16); |
2628 |
|
else i = quant4_inter(coeff, in, Data->lambda16); |
2629 |
|
if (i > 0) { |
2630 |
|
bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); |
2631 |
|
cbp |= 1 << (5 - 5); |
2632 |
|
} |
2633 |
|
bits += cbpy_tab[15-(cbp>>2)].len; |
2634 |
|
bits += mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len; |
2635 |
|
} |
2636 |
|
} |
2637 |
|
|
2638 |
|
return bits; |
2639 |
|
} |
2640 |
|
|
2641 |
|
|
2642 |
|
static int |
2643 |
|
CountMBBitsIntra(const SearchData * const Data) |
2644 |
|
{ |
2645 |
|
int bits = 1; //this one is ac/dc prediction flag. always 1. |
2646 |
|
int cbp = 0, i, t, dc = 0, b_dc = 1024; |
2647 |
|
const uint32_t iQuant = Data->lambda16; |
2648 |
|
int16_t in[64], coeff[64]; |
2649 |
|
|
2650 |
|
for(i = 0; i < 4; i++) { |
2651 |
|
uint32_t iDcScaler = get_dc_scaler(iQuant, 1); |
2652 |
|
|
2653 |
|
int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2654 |
|
transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth); |
2655 |
|
fdct(in); |
2656 |
|
b_dc = dc; |
2657 |
|
dc = in[0]; |
2658 |
|
in[0] -= b_dc; |
2659 |
|
if (Data->lambda8 == 0) quant_intra_c(coeff, in, iQuant, iDcScaler); |
2660 |
|
else quant4_intra_c(coeff, in, iQuant, iDcScaler); |
2661 |
|
|
2662 |
|
b_dc = dc; |
2663 |
|
dc = coeff[0]; |
2664 |
|
if (i != 0) coeff[0] -= b_dc; |
2665 |
|
|
2666 |
|
bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcy_tab[coeff[0] + 255].len;; |
2667 |
|
Data->temp[i] = t; |
2668 |
|
if (t != 0) cbp |= 1 << (5 - i); |
2669 |
|
if (bits >= Data->iMinSAD[0]) break; |
2670 |
|
} |
2671 |
|
|
2672 |
|
if (bits < Data->iMinSAD[0]) { // INTRA still looks good, let's add chroma |
2673 |
|
uint32_t iDcScaler = get_dc_scaler(iQuant, 0); |
2674 |
|
//chroma U |
2675 |
|
transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2); |
2676 |
|
fdct(in); |
2677 |
|
in[0] -= 1024; |
2678 |
|
if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler); |
2679 |
|
else quant4_intra(coeff, in, iQuant, iDcScaler); |
2680 |
|
|
2681 |
|
bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len; |
2682 |
|
if (t != 0) cbp |= 1 << (5 - 4); |
2683 |
|
Data->temp[4] = t; |
2684 |
|
|
2685 |
|
if (bits < Data->iMinSAD[0]) { |
2686 |
|
//chroma V |
2687 |
|
transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2); |
2688 |
|
fdct(in); |
2689 |
|
in[0] -= 1024; |
2690 |
|
if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler); |
2691 |
|
else quant4_intra(coeff, in, iQuant, iDcScaler); |
2692 |
|
|
2693 |
|
bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len; |
2694 |
|
if (t != 0) cbp |= 1 << (5 - 5); |
2695 |
|
|
2696 |
|
Data->temp[5] = t; |
2697 |
|
|
2698 |
|
bits += t = cbpy_tab[cbp>>2].len; |
2699 |
|
Data->temp[6] = t; |
2700 |
|
|
2701 |
|
bits += t = mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len; |
2702 |
|
Data->temp[7] = t; |
2703 |
|
|
2704 |
|
} |
2705 |
|
} |
2706 |
|
|
2707 |
|
return bits; |
2708 |
} |
} |