31 |
#include <assert.h> |
#include <assert.h> |
32 |
#include <stdio.h> |
#include <stdio.h> |
33 |
#include <stdlib.h> |
#include <stdlib.h> |
34 |
|
#include <string.h> // memcpy |
35 |
|
|
36 |
#include "../encoder.h" |
#include "../encoder.h" |
37 |
#include "../utils/mbfunctions.h" |
#include "../utils/mbfunctions.h" |
48 |
#define FINAL_SKIP_THRESH (50) |
#define FINAL_SKIP_THRESH (50) |
49 |
#define MAX_SAD00_FOR_SKIP (20) |
#define MAX_SAD00_FOR_SKIP (20) |
50 |
#define MAX_CHROMA_SAD_FOR_SKIP (22) |
#define MAX_CHROMA_SAD_FOR_SKIP (22) |
|
#define SKIP_THRESH_B (25) |
|
51 |
|
|
52 |
#define CHECK_CANDIDATE(X,Y,D) { \ |
#define CHECK_CANDIDATE(X,Y,D) { \ |
53 |
(*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } |
(*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } |
54 |
|
|
55 |
#define iDiamondSize 2 |
static __inline uint32_t |
56 |
|
d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv) |
|
static VECTOR |
|
|
GlobalMotionEst(const MACROBLOCK * const pMBs, |
|
|
const MBParam * const pParam, const uint32_t iFcode); |
|
|
|
|
|
|
|
|
static __inline int |
|
|
d_mv_bits(int x, int y, const uint32_t iFcode) |
|
57 |
{ |
{ |
58 |
int xb, yb; |
int xb, yb; |
59 |
|
x += x * qpel; y += y * qpel; |
60 |
|
if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); } |
61 |
|
x -= pred.x; |
62 |
|
y -= pred.y; |
63 |
|
|
64 |
if (x == 0) xb = 1; |
if (x) { |
65 |
else { |
x = ABS(x); |
|
if (x < 0) x = -x; |
|
66 |
x += (1 << (iFcode - 1)) - 1; |
x += (1 << (iFcode - 1)) - 1; |
67 |
x >>= (iFcode - 1); |
x >>= (iFcode - 1); |
68 |
if (x > 32) x = 32; |
if (x > 32) x = 32; |
69 |
xb = mvtab[x] + iFcode; |
xb = mvtab[x] + iFcode; |
70 |
} |
} else xb = 1; |
71 |
|
|
72 |
if (y == 0) yb = 1; |
if (y) { |
73 |
else { |
y = ABS(y); |
|
if (y < 0) y = -y; |
|
74 |
y += (1 << (iFcode - 1)) - 1; |
y += (1 << (iFcode - 1)) - 1; |
75 |
y >>= (iFcode - 1); |
y >>= (iFcode - 1); |
76 |
if (y > 32) y = 32; |
if (y > 32) y = 32; |
77 |
yb = mvtab[y] + iFcode; |
yb = mvtab[y] + iFcode; |
78 |
} |
} else yb = 1; |
79 |
return xb + yb; |
return xb + yb; |
80 |
} |
} |
81 |
|
|
83 |
ChromaSAD(int dx, int dy, const SearchData * const data) |
ChromaSAD(int dx, int dy, const SearchData * const data) |
84 |
{ |
{ |
85 |
int sad; |
int sad; |
86 |
dx = (dx >> 1) + roundtab_79[dx & 0x3]; |
const uint32_t stride = data->iEdgedWidth/2; |
|
dy = (dy >> 1) + roundtab_79[dy & 0x3]; |
|
87 |
|
|
88 |
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 |
89 |
|
data->temp[5] = dx; data->temp[6] = dy; // backup |
90 |
|
|
91 |
|
switch (((dx & 1) << 1) | (dy & 1)) { |
92 |
case 0: |
case 0: |
93 |
sad = sad8(data->CurU, data->RefCU + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); |
dx = dx / 2; dy = dy / 2; |
94 |
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); |
95 |
|
sad += sad8(data->CurV, data->RefCV + dy * stride + dx, stride); |
96 |
break; |
break; |
97 |
case 1: |
case 1: |
98 |
dx = dx / 2; dy = (dy - 1) / 2; |
dx = dx / 2; dy = (dy - 1) / 2; |
99 |
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); |
100 |
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); |
101 |
break; |
break; |
102 |
case 2: |
case 2: |
103 |
dx = (dx - 1) / 2; dy = dy / 2; |
dx = (dx - 1) / 2; dy = dy / 2; |
104 |
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); |
105 |
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); |
106 |
break; |
break; |
107 |
default: |
default: |
108 |
dx = (dx - 1) / 2; dy = (dy - 1) / 2; |
dx = (dx - 1) / 2; dy = (dy - 1) / 2; |
109 |
interpolate8x8_halfpel_hv(data->RefQ, |
interpolate8x8_halfpel_hv(data->RefQ, data->RefCU + dy * stride + dx, stride, data->rounding); |
110 |
data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, |
sad = sad8(data->CurU, data->RefQ, stride); |
111 |
data->rounding); |
|
112 |
sad = sad8(data->CurU, data->RefQ, data->iEdgedWidth/2); |
interpolate8x8_halfpel_hv(data->RefQ, data->RefCV + dy * stride + dx, stride, data->rounding); |
113 |
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); |
|
114 |
break; |
break; |
115 |
} |
} |
116 |
|
data->temp[7] = sad; //backup, part 2 |
117 |
return sad; |
return sad; |
118 |
} |
} |
119 |
|
|
120 |
static __inline const uint8_t * |
static __inline const uint8_t * |
121 |
GetReference(const int x, const int y, const int dir, const SearchData * const data) |
GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data) |
122 |
{ |
{ |
123 |
// dir : 0 = forward, 1 = backward |
// dir : 0 = forward, 1 = backward |
124 |
switch ( (dir << 2) | ((x&1)<<1) | (y&1) ) { |
switch ( (dir << 2) | ((x&1)<<1) | (y&1) ) { |
130 |
case 5 : return data->bRefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); |
case 5 : return data->bRefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); |
131 |
case 6 : return data->bRefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); |
case 6 : return data->bRefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); |
132 |
default : return data->bRefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); |
default : return data->bRefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); |
133 |
|
} |
134 |
|
} |
135 |
|
|
136 |
|
// this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate |
137 |
|
static __inline const uint8_t * |
138 |
|
GetReference(const int x, const int y, const SearchData * const data) |
139 |
|
{ |
140 |
|
switch ( ((x&1)<<1) | (y&1) ) { |
141 |
|
case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth); |
142 |
|
case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); |
143 |
|
case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); |
144 |
|
default : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); |
145 |
} |
} |
146 |
} |
} |
147 |
|
|
148 |
static uint8_t * |
static uint8_t * |
149 |
Interpolate8x8qpel(const int x, const int y, const int block, const int dir, const SearchData * const data) |
Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data) |
150 |
{ |
{ |
151 |
// create or find a qpel-precision reference picture; return pointer to it |
// create or find a qpel-precision reference picture; return pointer to it |
152 |
uint8_t * Reference = (uint8_t *)data->RefQ + 16*dir; |
uint8_t * Reference = data->RefQ + 16*dir; |
153 |
const int32_t iEdgedWidth = data->iEdgedWidth; |
const uint32_t iEdgedWidth = data->iEdgedWidth; |
154 |
const uint32_t rounding = data->rounding; |
const uint32_t rounding = data->rounding; |
155 |
const int halfpel_x = x/2; |
const int halfpel_x = x/2; |
156 |
const int halfpel_y = y/2; |
const int halfpel_y = y/2; |
157 |
const uint8_t *ref1, *ref2, *ref3, *ref4; |
const uint8_t *ref1, *ref2, *ref3, *ref4; |
158 |
|
|
159 |
ref1 = GetReference(halfpel_x, halfpel_y, dir, data); // this reference is used in all cases |
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
160 |
ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
161 |
switch( ((x&1)<<1) + (y&1) ) { |
switch( ((x&1)<<1) + (y&1) ) { |
162 |
case 0: // pure halfpel position |
case 0: // pure halfpel position |
163 |
Reference = (uint8_t *) GetReference(halfpel_x, halfpel_y, dir, data); |
return (uint8_t *) ref1; |
|
Reference += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
|
164 |
break; |
break; |
165 |
|
|
166 |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
167 |
ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
168 |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
169 |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
170 |
break; |
break; |
171 |
|
|
172 |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
173 |
ref2 = GetReference(x - halfpel_x, halfpel_y, dir, data); |
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
174 |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
175 |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
176 |
break; |
break; |
177 |
|
|
178 |
default: // x and y in qpel resolution - the "corners" (top left/right and |
default: // x and y in qpel resolution - the "corners" (top left/right and |
179 |
// bottom left/right) during qpel refinement |
// bottom left/right) during qpel refinement |
180 |
ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
181 |
ref3 = GetReference(x - halfpel_x, halfpel_y, dir, data); |
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
182 |
ref4 = GetReference(x - halfpel_x, y - halfpel_y, dir, data); |
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
183 |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
184 |
ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
185 |
ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
190 |
} |
} |
191 |
|
|
192 |
static uint8_t * |
static uint8_t * |
193 |
Interpolate16x16qpel(const int x, const int y, const int dir, const SearchData * const data) |
Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data) |
194 |
{ |
{ |
195 |
// create or find a qpel-precision reference picture; return pointer to it |
// create or find a qpel-precision reference picture; return pointer to it |
196 |
uint8_t * Reference = (uint8_t *)data->RefQ + 16*dir; |
uint8_t * Reference = data->RefQ + 16*dir; |
197 |
const int32_t iEdgedWidth = data->iEdgedWidth; |
const uint32_t iEdgedWidth = data->iEdgedWidth; |
198 |
const uint32_t rounding = data->rounding; |
const uint32_t rounding = data->rounding; |
199 |
const int halfpel_x = x/2; |
const int halfpel_x = x/2; |
200 |
const int halfpel_y = y/2; |
const int halfpel_y = y/2; |
201 |
const uint8_t *ref1, *ref2, *ref3, *ref4; |
const uint8_t *ref1, *ref2, *ref3, *ref4; |
202 |
|
|
203 |
ref1 = GetReference(halfpel_x, halfpel_y, dir, data); // this reference is used in all cases |
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
204 |
switch( ((x&1)<<1) + (y&1) ) { |
switch( ((x&1)<<1) + (y&1) ) { |
205 |
case 0: // pure halfpel position |
case 0: // pure halfpel position |
206 |
return (uint8_t *) GetReference(halfpel_x, halfpel_y, dir, data); |
return (uint8_t *) ref1; |
207 |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
208 |
ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
209 |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
210 |
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
211 |
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); |
213 |
break; |
break; |
214 |
|
|
215 |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
216 |
ref2 = GetReference(x - halfpel_x, halfpel_y, dir, data); |
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
217 |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
218 |
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
219 |
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); |
222 |
|
|
223 |
default: // x and y in qpel resolution - the "corners" (top left/right and |
default: // x and y in qpel resolution - the "corners" (top left/right and |
224 |
// bottom left/right) during qpel refinement |
// bottom left/right) during qpel refinement |
225 |
ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
226 |
ref3 = GetReference(x - halfpel_x, halfpel_y, dir, data); |
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
227 |
ref4 = GetReference(x - halfpel_x, y - halfpel_y, dir, data); |
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
228 |
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
229 |
interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); |
interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); |
230 |
interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); |
interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); |
243 |
const uint8_t * Reference; |
const uint8_t * Reference; |
244 |
VECTOR * current; |
VECTOR * current; |
245 |
|
|
246 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if ( (x > data->max_dx) | (x < data->min_dx) |
247 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
| (y > data->max_dy) | (y < data->min_dy) ) return; |
248 |
|
|
249 |
if (data->qpel_precision) { // x and y are in 1/4 precision |
if (data->qpel_precision) { // x and y are in 1/4 precision |
250 |
Reference = Interpolate16x16qpel(x, y, 0, data); |
Reference = Interpolate16x16qpel(x, y, 0, data); |
|
t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
|
251 |
xc = x/2; yc = y/2; //for chroma sad |
xc = x/2; yc = y/2; //for chroma sad |
252 |
current = data->currentQMV; |
current = data->currentQMV; |
253 |
} else { |
} else { |
254 |
switch ( ((x&1)<<1) + (y&1) ) { |
Reference = GetReference(x, y, data); |
|
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
|
|
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
|
|
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
|
|
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
|
|
} |
|
|
if (data->qpel) t = d_mv_bits(2*x - data->predMV.x, 2*y - data->predMV.y, data->iFcode); |
|
|
else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
|
255 |
current = data->currentMV; |
current = data->currentMV; |
256 |
xc = x; yc = y; |
xc = x; yc = y; |
257 |
} |
} |
258 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
259 |
|
|
260 |
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
261 |
|
|
262 |
data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; |
data->temp[0] += (data->lambda16 * t * data->temp[0])>>10; |
263 |
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; |
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; |
264 |
|
|
265 |
if (data->chroma) data->temp[0] += ChromaSAD(xc, yc, data); |
if (data->chroma) data->temp[0] += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
266 |
|
(yc >> 1) + roundtab_79[yc & 0x3], data); |
267 |
|
|
268 |
if (data->temp[0] < data->iMinSAD[0]) { |
if (data->temp[0] < data->iMinSAD[0]) { |
269 |
data->iMinSAD[0] = data->temp[0]; |
data->iMinSAD[0] = data->temp[0]; |
282 |
} |
} |
283 |
|
|
284 |
static void |
static void |
285 |
|
CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
286 |
|
{ |
287 |
|
uint32_t t; |
288 |
|
const uint8_t * Reference; |
289 |
|
|
290 |
|
if ( (!(x&1) && x !=0) | (!(y&1) && y !=0) || //non-zero integer value |
291 |
|
(x > data->max_dx) | (x < data->min_dx) |
292 |
|
| (y > data->max_dy) | (y < data->min_dy) ) return; |
293 |
|
|
294 |
|
Reference = GetReference(x, y, data); |
295 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1); |
296 |
|
|
297 |
|
data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
298 |
|
|
299 |
|
data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10; |
300 |
|
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; |
301 |
|
|
302 |
|
if (data->temp[0] < data->iMinSAD[0]) { |
303 |
|
data->iMinSAD[0] = data->temp[0]; |
304 |
|
data->currentMV[0].x = x; data->currentMV[0].y = y; |
305 |
|
*dir = Direction; } |
306 |
|
|
307 |
|
if (data->temp[1] < data->iMinSAD[1]) { |
308 |
|
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
309 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
310 |
|
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
311 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
312 |
|
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
313 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
314 |
|
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
315 |
|
} |
316 |
|
|
317 |
|
static void |
318 |
CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
319 |
{ |
{ |
320 |
int32_t sad; |
int32_t sad; |
321 |
const uint8_t * Reference; |
const uint8_t * Reference; |
322 |
int t; |
uint32_t t; |
323 |
VECTOR * current; |
VECTOR * current; |
324 |
|
|
325 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if ( (x > data->max_dx) | ( x < data->min_dx) |
326 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
| (y > data->max_dy) | (y < data->min_dy) ) return; |
327 |
|
|
328 |
|
if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; //non-zero even value |
329 |
|
|
330 |
if (data->qpel_precision) { // x and y are in 1/4 precision |
if (data->qpel_precision) { // x and y are in 1/4 precision |
331 |
Reference = Interpolate16x16qpel(x, y, 0, data); |
Reference = Interpolate16x16qpel(x, y, 0, data); |
|
t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
|
332 |
current = data->currentQMV; |
current = data->currentQMV; |
333 |
} else { |
} else { |
334 |
switch ( ((x&1)<<1) + (y&1) ) { |
Reference = GetReference(x, y, data); |
|
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
|
|
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
|
|
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
|
|
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
|
|
} |
|
|
if (data->qpel) t = d_mv_bits(2*x - data->predMV.x, 2*y - data->predMV.y, data->iFcode); |
|
|
else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
|
335 |
current = data->currentMV; |
current = data->currentMV; |
336 |
} |
} |
337 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, |
338 |
|
data->qpel^data->qpel_precision, data->rrv); |
339 |
|
|
340 |
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
341 |
sad += (data->lambda16 * t * sad)/1000; |
sad += (data->lambda16 * t * sad)>>10; |
342 |
|
|
343 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
344 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
347 |
} |
} |
348 |
|
|
349 |
static void |
static void |
350 |
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) |
351 |
{ |
{ |
352 |
// maximum speed - for P/B/I decision |
// maximum speed - for P/B/I decision |
|
int32_t sad; |
|
353 |
|
|
354 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if ( (x > data->max_dx) | (x < data->min_dx) |
355 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
| (y > data->max_dy) | (y < data->min_dy) ) return; |
356 |
|
|
357 |
sad = sad16(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), |
data->temp[0] = sad32v_c(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), |
358 |
data->iEdgedWidth, 256*4096); |
data->iEdgedWidth, data->temp+1); |
359 |
|
|
360 |
if (sad < *(data->iMinSAD)) { |
if (data->temp[0] < *(data->iMinSAD)) { |
361 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = data->temp[0]; |
362 |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
363 |
*dir = Direction; } |
*dir = Direction; } |
364 |
} |
if (data->temp[1] < data->iMinSAD[1]) { |
365 |
|
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
366 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
367 |
|
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
368 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
369 |
|
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
370 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
371 |
|
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
372 |
|
|
373 |
|
} |
374 |
|
|
375 |
static void |
static void |
376 |
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) |
377 |
{ |
{ |
378 |
int32_t sad; |
int32_t sad, xb, yb; |
379 |
int xb, yb, t; |
uint32_t t; |
380 |
const uint8_t *ReferenceF, *ReferenceB; |
const uint8_t *ReferenceF, *ReferenceB; |
381 |
VECTOR *current; |
VECTOR *current; |
382 |
|
|
383 |
if (( xf > data->max_dx) || ( xf < data->min_dx) |
if ( (xf > data->max_dx) | (xf < data->min_dx) |
384 |
|| ( yf > data->max_dy) || (yf < data->min_dy)) return; |
| (yf > data->max_dy) | (yf < data->min_dy) ) return; |
385 |
|
|
386 |
if (data->qpel_precision) { |
if (!data->qpel_precision) { |
387 |
|
ReferenceF = GetReference(xf, yf, data); |
388 |
|
xb = data->currentMV[1].x; yb = data->currentMV[1].y; |
389 |
|
ReferenceB = GetReferenceB(xb, yb, 1, data); |
390 |
|
current = data->currentMV; |
391 |
|
} else { |
392 |
ReferenceF = Interpolate16x16qpel(xf, yf, 0, data); |
ReferenceF = Interpolate16x16qpel(xf, yf, 0, data); |
393 |
xb = data->currentQMV[1].x; yb = data->currentQMV[1].y; |
xb = data->currentQMV[1].x; yb = data->currentQMV[1].y; |
394 |
current = data->currentQMV; |
current = data->currentQMV; |
395 |
ReferenceB = Interpolate16x16qpel(xb, yb, 1, data); |
ReferenceB = Interpolate16x16qpel(xb, yb, 1, data); |
|
t = d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) |
|
|
+ d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode); |
|
|
} else { |
|
|
ReferenceF = Interpolate16x16qpel(2*xf, 2*yf, 0, data); |
|
|
xb = data->currentMV[1].x; yb = data->currentMV[1].y; |
|
|
ReferenceB = Interpolate16x16qpel(2*xb, 2*yb, 1, data); |
|
|
current = data->currentMV; |
|
|
if (data->qpel) |
|
|
t = d_mv_bits(2*xf - data->predMV.x, 2*yf - data->predMV.y, data->iFcode) |
|
|
+ d_mv_bits(2*xb - data->bpredMV.x, 2*yb - data->bpredMV.y, data->iFcode); |
|
|
else |
|
|
t = d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) |
|
|
+ d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode); |
|
396 |
} |
} |
397 |
|
|
398 |
|
t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0) |
399 |
|
+ d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
400 |
|
|
401 |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
402 |
sad += (data->lambda16 * t * sad)/1000; |
sad += (data->lambda16 * t * sad)>>10; |
403 |
|
|
404 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
405 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
411 |
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) |
412 |
{ |
{ |
413 |
int32_t sad = 0; |
int32_t sad = 0; |
414 |
int k; |
uint32_t k; |
415 |
const uint8_t *ReferenceF; |
const uint8_t *ReferenceF; |
416 |
const uint8_t *ReferenceB; |
const uint8_t *ReferenceB; |
417 |
VECTOR mvs, b_mvs; |
VECTOR mvs, b_mvs; |
418 |
|
|
419 |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
if (( x > 31) | ( x < -32) | ( y > 31) | (y < -32)) return; |
420 |
|
|
421 |
for (k = 0; k < 4; k++) { |
for (k = 0; k < 4; k++) { |
422 |
mvs.x = data->directmvF[k].x + x; |
mvs.x = data->directmvF[k].x + x; |
434 |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
435 |
|| ( 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; |
436 |
|
|
437 |
if (!data->qpel) { |
|
438 |
mvs.x *= 2; mvs.y *= 2; |
mvs.x *= 2 - data->qpel; mvs.y *= 2 - data->qpel; |
439 |
b_mvs.x *= 2; b_mvs.y *= 2; //we move to qpel precision anyway |
b_mvs.x *= 2 - data->qpel; b_mvs.y *= 2 - data->qpel; //we move to qpel precision anyway |
440 |
} |
|
441 |
ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); |
ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); |
442 |
ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); |
ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); |
443 |
|
|
444 |
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
445 |
ReferenceF, ReferenceB, |
ReferenceF, ReferenceB, data->iEdgedWidth); |
|
data->iEdgedWidth); |
|
446 |
if (sad > *(data->iMinSAD)) return; |
if (sad > *(data->iMinSAD)) return; |
447 |
} |
} |
448 |
|
|
449 |
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; |
450 |
|
|
451 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
452 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
462 |
const uint8_t *ReferenceB; |
const uint8_t *ReferenceB; |
463 |
VECTOR mvs, b_mvs; |
VECTOR mvs, b_mvs; |
464 |
|
|
465 |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
if (( x > 31) | ( x < -32) | ( y > 31) | (y < -32)) return; |
466 |
|
|
467 |
mvs.x = data->directmvF[0].x + x; |
mvs.x = data->directmvF[0].x + x; |
468 |
b_mvs.x = ((x == 0) ? |
b_mvs.x = ((x == 0) ? |
479 |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
480 |
|| ( 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; |
481 |
|
|
482 |
if (!data->qpel) { |
mvs.x *= 2 - data->qpel; mvs.y *= 2 - data->qpel; |
483 |
mvs.x *= 2; mvs.y *= 2; |
b_mvs.x *= 2 - data->qpel; b_mvs.y *= 2 - data->qpel; //we move to qpel precision anyway |
484 |
b_mvs.x *= 2; b_mvs.y *= 2; //we move to qpel precision anyway |
|
|
} |
|
485 |
ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data); |
ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data); |
486 |
ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); |
ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); |
487 |
|
|
488 |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
489 |
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; |
490 |
|
|
491 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
492 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
497 |
static void |
static void |
498 |
CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
499 |
{ |
{ |
500 |
int32_t sad; int t; |
int32_t sad; uint32_t t; |
501 |
const uint8_t * Reference; |
const uint8_t * Reference; |
502 |
|
|
503 |
if (( x > data->max_dx) || ( x < data->min_dx) |
if ( (x > data->max_dx) | (x < data->min_dx) |
504 |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
| (y > data->max_dy) | (y < data->min_dy) ) return; |
505 |
|
|
506 |
if (data->qpel) Reference = Interpolate16x16qpel(x, y, 0, data); |
if (data->qpel) Reference = Interpolate16x16qpel(x, y, 0, data); |
507 |
else Reference = Interpolate16x16qpel(2*x, 2*y, 0, data); |
else Reference = GetReference(x, y, data); |
508 |
|
|
509 |
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
510 |
if (data->qpel) t = d_mv_bits(2 * x - data->predMV.x, 2 * y - data->predMV.y, data->iFcode); |
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0); |
|
else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
|
511 |
|
|
512 |
sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))/100; |
sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10; |
513 |
|
|
514 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
515 |
*(data->iMinSAD) = sad; |
*(data->iMinSAD) = sad; |
529 |
|
|
530 |
int iDirection; |
int iDirection; |
531 |
|
|
532 |
do { |
for(;;) { //forever |
533 |
iDirection = 0; |
iDirection = 0; |
534 |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
535 |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
538 |
|
|
539 |
/* now we're doing diagonal checks near our candidate */ |
/* now we're doing diagonal checks near our candidate */ |
540 |
|
|
541 |
if (iDirection) { //checking if anything found |
if (iDirection) { //if anything found |
542 |
bDirection = iDirection; |
bDirection = iDirection; |
543 |
iDirection = 0; |
iDirection = 0; |
544 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
547 |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
548 |
} else { // what remains here is up or down |
} else { // what remains here is up or down |
549 |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
550 |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); } |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
551 |
|
} |
552 |
|
|
553 |
if (iDirection) { |
if (iDirection) { |
554 |
bDirection += iDirection; |
bDirection += iDirection; |
555 |
x = data->currentMV->x; y = data->currentMV->y; } |
x = data->currentMV->x; y = data->currentMV->y; |
556 |
|
} |
557 |
} else { //about to quit, eh? not so fast.... |
} else { //about to quit, eh? not so fast.... |
558 |
switch (bDirection) { |
switch (bDirection) { |
559 |
case 2: |
case 2: |
604 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
605 |
} |
} |
606 |
} |
} |
|
while (1); //forever |
|
607 |
} |
} |
608 |
|
|
609 |
static void |
static void |
653 |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
654 |
} else { // what remains here is up or down |
} else { // what remains here is up or down |
655 |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
656 |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); } |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
657 |
|
} |
658 |
bDirection += iDirection; |
bDirection += iDirection; |
659 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
660 |
} |
} |
664 |
|
|
665 |
/* MAINSEARCH FUNCTIONS END */ |
/* MAINSEARCH FUNCTIONS END */ |
666 |
|
|
|
/* HALFPELREFINE COULD BE A MAINSEARCH FUNCTION, BUT THERE IS NO NEED FOR IT */ |
|
|
|
|
667 |
static void |
static void |
668 |
SubpelRefine(const SearchData * const data) |
SubpelRefine(const SearchData * const data) |
669 |
{ |
{ |
670 |
/* Do a half-pel or q-pel refinement */ |
/* Do a half-pel or q-pel refinement */ |
671 |
VECTOR backupMV; |
const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV; |
672 |
int iDirection; //not needed |
int iDirection; //only needed because macro expects it |
|
|
|
|
if (data->qpel_precision) |
|
|
backupMV = *(data->currentQMV); |
|
|
else backupMV = *(data->currentMV); |
|
|
|
|
|
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); |
|
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); |
|
|
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); |
|
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0); |
|
|
|
|
|
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); |
|
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); |
|
673 |
|
|
674 |
CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); |
CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0); |
675 |
CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); |
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0); |
676 |
|
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0); |
677 |
|
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0); |
678 |
|
CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0); |
679 |
|
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0); |
680 |
|
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0); |
681 |
|
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0); |
682 |
} |
} |
683 |
|
|
684 |
static __inline int |
static __inline int |
685 |
SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
686 |
const int x, const int y, |
const int x, const int y, |
687 |
const uint32_t iEdgedWidth, const uint32_t iQuant) |
const uint32_t stride, const uint32_t iQuant, int rrv) |
688 |
|
|
689 |
{ |
{ |
690 |
/* keep repeating checks for all b-frames before this P frame, |
if(!rrv) { |
691 |
to make sure that SKIP is possible (todo) |
uint32_t sadC = sad8(current->u + x*8 + y*stride*8, |
692 |
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); |
|
693 |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
694 |
sadC += sad8(current->v + (x + y*(iEdgedWidth/2))*8, |
sadC += sad8(current->v + (x + y*stride)*8, |
695 |
reference->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); |
reference->v + (x + y*stride)*8, stride); |
696 |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
697 |
|
return 1; |
698 |
|
|
699 |
|
} else { |
700 |
|
uint32_t sadC = sad16(current->u + x*16 + y*stride*16, |
701 |
|
reference->u + x*16 + y*stride*16, stride, 256*4096); |
702 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
703 |
|
sadC += sad16(current->v + (x + y*stride)*16, |
704 |
|
reference->v + (x + y*stride)*16, stride, 256*4096); |
705 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
706 |
return 1; |
return 1; |
707 |
} |
} |
708 |
|
} |
709 |
|
|
710 |
static __inline void |
static __inline void |
711 |
SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
712 |
{ |
{ |
713 |
pMB->mode = MODE_NOT_CODED; |
pMB->mode = MODE_NOT_CODED; |
714 |
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; |
715 |
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; |
|
|
|
|
716 |
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; |
717 |
} |
} |
718 |
|
|
729 |
const IMAGE *const pCurrent = ¤t->image; |
const IMAGE *const pCurrent = ¤t->image; |
730 |
const IMAGE *const pRef = &reference->image; |
const IMAGE *const pRef = &reference->image; |
731 |
|
|
732 |
const VECTOR zeroMV = { 0, 0 }; |
uint32_t mb_width = pParam->mb_width; |
733 |
|
uint32_t mb_height = pParam->mb_height; |
734 |
int mb_width = pParam->mb_width; |
const uint32_t iEdgedWidth = pParam->edged_width; |
|
int mb_height = pParam->mb_height; |
|
735 |
|
|
736 |
uint32_t x, y; |
uint32_t x, y; |
737 |
uint32_t iIntra = 0; |
uint32_t iIntra = 0; |
738 |
int32_t InterBias, quant = current->quant, sad00; |
int32_t InterBias, quant = current->quant, sad00; |
|
uint8_t *qimage; |
|
739 |
|
|
740 |
// some pre-initialized thingies for SearchP |
// some pre-initialized thingies for SearchP |
741 |
int32_t temp[5]; |
int32_t temp[8]; |
742 |
VECTOR currentMV[5]; |
VECTOR currentMV[5]; |
743 |
VECTOR currentQMV[5]; |
VECTOR currentQMV[5]; |
744 |
int32_t iMinSAD[5]; |
int32_t iMinSAD[5]; |
745 |
SearchData Data; |
SearchData Data; |
746 |
Data.iEdgedWidth = pParam->edged_width; |
memset(&Data, 0, sizeof(SearchData)); |
747 |
|
Data.iEdgedWidth = iEdgedWidth; |
748 |
Data.currentMV = currentMV; |
Data.currentMV = currentMV; |
749 |
Data.currentQMV = currentQMV; |
Data.currentQMV = currentQMV; |
750 |
Data.iMinSAD = iMinSAD; |
Data.iMinSAD = iMinSAD; |
753 |
Data.rounding = pParam->m_rounding_type; |
Data.rounding = pParam->m_rounding_type; |
754 |
Data.qpel = pParam->m_quarterpel; |
Data.qpel = pParam->m_quarterpel; |
755 |
Data.chroma = current->global_flags & XVID_ME_COLOUR; |
Data.chroma = current->global_flags & XVID_ME_COLOUR; |
756 |
|
Data.rrv = current->global_flags & XVID_REDUCED; |
757 |
|
|
758 |
if ((current->global_flags & XVID_REDUCED)) |
if ((current->global_flags & XVID_REDUCED)) { |
|
{ |
|
759 |
mb_width = (pParam->width + 31) / 32; |
mb_width = (pParam->width + 31) / 32; |
760 |
mb_height = (pParam->height + 31) / 32; |
mb_height = (pParam->height + 31) / 32; |
761 |
|
Data.qpel = Data.chroma = 0; |
762 |
} |
} |
763 |
|
|
764 |
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose) |
|
return 1; // 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; |
|
765 |
if (sadInit) (*sadInit) (); |
if (sadInit) (*sadInit) (); |
766 |
|
|
767 |
for (y = 0; y < mb_height; y++) { |
for (y = 0; y < mb_height; y++) { |
768 |
for (x = 0; x < mb_width; x++) { |
for (x = 0; x < mb_width; x++) { |
769 |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
770 |
|
|
771 |
pMB->sad16 |
if (!Data.rrv) pMB->sad16 = |
772 |
= sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, |
sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16, |
773 |
pRef->y + (x + y * pParam->edged_width) * 16, |
pRef->y + (x + y * iEdgedWidth) * 16, |
774 |
pParam->edged_width, pMB->sad8 ); |
pParam->edged_width, pMB->sad8 ); |
775 |
|
|
776 |
if (Data.chroma) { |
else pMB->sad16 = |
777 |
pMB->sad16 += sad8(pCurrent->u + x*8 + y*(pParam->edged_width/2)*8, |
sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32, |
778 |
pRef->u + x*8 + y*(pParam->edged_width/2)*8, pParam->edged_width/2); |
pRef->y + (x + y * iEdgedWidth) * 32, |
779 |
|
pParam->edged_width, pMB->sad8 ); |
780 |
|
|
781 |
pMB->sad16 += sad8(pCurrent->v + (x + y*(pParam->edged_width/2))*8, |
if (Data.chroma) { |
782 |
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, |
783 |
|
pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2) |
784 |
|
+ sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8, |
785 |
|
pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); |
786 |
|
pMB->sad16 += Data.temp[7]; |
787 |
} |
} |
788 |
|
|
789 |
sad00 = pMB->sad16; //if no gmc; else sad00 = (..) |
sad00 = pMB->sad16; |
790 |
|
|
791 |
if (!(current->global_flags & XVID_LUMIMASKING)) { |
if (!(current->global_flags & XVID_LUMIMASKING)) { |
792 |
pMB->dquant = NO_CHANGE; |
pMB->dquant = NO_CHANGE; |
|
pMB->quant = current->quant; |
|
793 |
} else { |
} else { |
794 |
if (pMB->dquant != NO_CHANGE) { |
if (pMB->dquant != NO_CHANGE) { |
795 |
quant += DQtab[pMB->dquant]; |
quant += DQtab[pMB->dquant]; |
796 |
if (quant > 31) quant = 31; |
if (quant > 31) quant = 31; |
797 |
else if (quant < 1) quant = 1; |
else if (quant < 1) quant = 1; |
798 |
} |
} |
|
pMB->quant = quant; |
|
799 |
} |
} |
800 |
|
pMB->quant = current->quant; |
801 |
|
|
802 |
//initial skip decision |
//initial skip decision |
803 |
/* no early skip for GMC (global vector = skip vector is unknown!) */ |
/* no early skip for GMC (global vector = skip vector is unknown!) */ |
804 |
if (current->coding_type == P_VOP) { /* no fast SKIP for S(GMC)-VOPs */ |
if (current->coding_type == P_VOP) { /* no fast SKIP for S(GMC)-VOPs */ |
805 |
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) ) |
806 |
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)) { |
807 |
SkipMacroblockP(pMB, sad00); |
SkipMacroblockP(pMB, sad00); |
808 |
continue; |
continue; |
809 |
} |
} |
817 |
/* 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?" */ |
818 |
if (current->coding_type == P_VOP) { |
if (current->coding_type == P_VOP) { |
819 |
if ( (pMB->dquant == NO_CHANGE) && (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
if ( (pMB->dquant == NO_CHANGE) && (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
820 |
&& ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) ) |
&& ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1)) ) |
821 |
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)) { |
822 |
SkipMacroblockP(pMB, sad00); |
SkipMacroblockP(pMB, sad00); |
823 |
continue; |
continue; |
824 |
} |
} |
834 |
if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; |
if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; |
835 |
|
|
836 |
if (Data.chroma) InterBias += 50; // to compensate bigger SAD |
if (Data.chroma) InterBias += 50; // to compensate bigger SAD |
837 |
|
if (Data.rrv) InterBias *= 4; |
838 |
|
|
839 |
if (InterBias < pMB->sad16) { |
if (InterBias < pMB->sad16) { |
840 |
const int32_t deviation = |
int32_t deviation; |
841 |
dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
if (!Data.rrv) |
842 |
pParam->edged_width); |
deviation = dev16(pCurrent->y + (x + y * iEdgedWidth) * 16, iEdgedWidth); |
843 |
|
else { |
844 |
|
deviation = dev16(pCurrent->y + (x + y * iEdgedWidth) * 32, iEdgedWidth) |
845 |
|
+ dev16(pCurrent->y + (x + y * iEdgedWidth) * 32 + 16, iEdgedWidth) |
846 |
|
+ dev16(pCurrent->y + (x + y * iEdgedWidth) * 32 + 16 * iEdgedWidth, iEdgedWidth) |
847 |
|
+ dev16(pCurrent->y + (x + y * iEdgedWidth) * 32 + 16 * (iEdgedWidth+1), iEdgedWidth); |
848 |
|
} |
849 |
if (deviation < (pMB->sad16 - InterBias)) { |
if (deviation < (pMB->sad16 - InterBias)) { |
850 |
if (++iIntra >= iLimit) { free(qimage); return 1; } |
if (++iIntra >= iLimit) return 1; |
851 |
|
SkipMacroblockP(pMB, 0); //same thing |
852 |
pMB->mode = MODE_INTRA; |
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; |
|
853 |
} |
} |
854 |
} |
} |
855 |
} |
} |
856 |
} |
} |
|
free(qimage); |
|
857 |
|
|
858 |
if (current->coding_type == S_VOP) /* first GMC step only for S(GMC)-VOPs */ |
if (current->coding_type == S_VOP) /* first GMC step only for S(GMC)-VOPs */ |
859 |
current->GMC_MV = GlobalMotionEst( pMBs, pParam, current->fcode ); |
current->GMC_MV = GlobalMotionEst( pMBs, pParam, current->fcode ); |
864 |
} |
} |
865 |
|
|
866 |
|
|
|
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
|
|
|
|
867 |
static __inline int |
static __inline int |
868 |
make_mask(const VECTOR * const pmv, const int i) |
make_mask(const VECTOR * const pmv, const int i) |
869 |
{ |
{ |
871 |
for (j = 0; j < i; j++) { |
for (j = 0; j < i; j++) { |
872 |
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 |
873 |
if (pmv[i].x == pmv[j].x) { |
if (pmv[i].x == pmv[j].x) { |
874 |
if (pmv[i].y == pmv[j].y + iDiamondSize) { mask &= ~4; continue; } |
if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4; |
875 |
if (pmv[i].y == pmv[j].y - iDiamondSize) { mask &= ~8; continue; } |
else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8; |
876 |
} else |
} else |
877 |
if (pmv[i].y == pmv[j].y) { |
if (pmv[i].y == pmv[j].y) { |
878 |
if (pmv[i].x == pmv[j].x + iDiamondSize) { mask &= ~1; continue; } |
if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1; |
879 |
if (pmv[i].x == pmv[j].x - iDiamondSize) { mask &= ~2; continue; } |
else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2; |
880 |
} |
} |
881 |
} |
} |
882 |
return mask; |
return mask; |
883 |
} |
} |
884 |
|
|
885 |
static __inline void |
static __inline void |
886 |
PreparePredictionsP(VECTOR * const pmv, int x, int y, const int iWcount, |
PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, |
887 |
const int iHcount, const MACROBLOCK * const prevMB) |
int iHcount, const MACROBLOCK * const prevMB, int rrv) |
888 |
{ |
{ |
889 |
|
|
890 |
//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 |
891 |
|
if (rrv) { iWcount /= 2; iHcount /= 2; } |
892 |
|
|
893 |
if ( (y != 0) && (x != (iWcount-1)) ) { // [5] top-right neighbour |
if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour |
894 |
pmv[5].x = EVEN(pmv[3].x); |
pmv[5].x = EVEN(pmv[3].x); |
895 |
pmv[5].y = EVEN(pmv[3].y); |
pmv[5].y = EVEN(pmv[3].y); |
896 |
} else pmv[5].x = pmv[5].y = 0; |
} else pmv[5].x = pmv[5].y = 0; |
909 |
pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
910 |
pmv[2].y = EVEN(prevMB->mvs[0].y); |
pmv[2].y = EVEN(prevMB->mvs[0].y); |
911 |
|
|
912 |
if ((x != iWcount-1) && (y != iHcount-1)) { |
if ((x < iWcount-1) && (y < iHcount-1)) { |
913 |
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 |
914 |
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
915 |
} else pmv[6].x = pmv[6].y = 0; |
} else pmv[6].x = pmv[6].y = 0; |
916 |
|
|
917 |
|
if (rrv) { |
918 |
|
int i; |
919 |
|
for (i = 0; i < 7; i++) { |
920 |
|
pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); |
921 |
|
pmv[i].y = RRV_MV_SCALEUP(pmv[i].y); |
922 |
|
} |
923 |
|
} |
924 |
} |
} |
925 |
|
|
926 |
static void |
static void |
944 |
int i, iDirection = 255, mask, threshA; |
int i, iDirection = 255, mask, threshA; |
945 |
VECTOR pmv[7]; |
VECTOR pmv[7]; |
946 |
|
|
|
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); //has to be changed to get_pmv(2)() |
|
947 |
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, |
948 |
pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel); |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
949 |
|
|
950 |
|
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); |
951 |
|
|
952 |
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16; |
Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache |
953 |
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
i = Data->rrv ? 2 : 1; |
954 |
Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i; |
955 |
|
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
956 |
Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16; |
Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
957 |
Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16; |
|
958 |
Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16; |
Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16*i; |
959 |
Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16; |
Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16*i; |
960 |
Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16*i; |
961 |
Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16*i; |
962 |
|
Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
963 |
|
Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
964 |
|
|
965 |
Data->lambda16 = lambda_vec16[iQuant]; |
Data->lambda16 = lambda_vec16[iQuant]; |
966 |
Data->lambda8 = lambda_vec8[iQuant]; |
Data->lambda8 = lambda_vec8[iQuant]; |
967 |
Data->qpel_precision = 0; |
Data->qpel_precision = 0; |
968 |
|
|
|
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); } |
|
|
|
|
969 |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
970 |
|
|
971 |
for(i = 0; i < 5; i++) |
for(i = 0; i < 5; i++) |
972 |
Data->currentMV[i].x = Data->currentMV[i].y = 0; |
Data->currentMV[i].x = Data->currentMV[i].y = 0; |
973 |
|
|
974 |
if (pParam->m_quarterpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
975 |
else Data->predMV = pmv[0]; |
else Data->predMV = pmv[0]; |
976 |
|
|
977 |
i = d_mv_bits(Data->predMV.x, Data->predMV.y, Data->iFcode); |
i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0); |
978 |
Data->iMinSAD[0] = pMB->sad16 + (Data->lambda16 * i * pMB->sad16)/1000; |
Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10); |
979 |
Data->iMinSAD[1] = pMB->sad8[0] + (Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS))/100; |
Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10); |
980 |
Data->iMinSAD[2] = pMB->sad8[1]; |
Data->iMinSAD[2] = pMB->sad8[1]; |
981 |
Data->iMinSAD[3] = pMB->sad8[2]; |
Data->iMinSAD[3] = pMB->sad8[2]; |
982 |
Data->iMinSAD[4] = pMB->sad8[3]; |
Data->iMinSAD[4] = pMB->sad8[3]; |
983 |
|
|
984 |
if ((x == 0) && (y == 0)) threshA = 512; |
if (x | y) { |
|
else { |
|
985 |
threshA = Data->temp[0]; // that's when we keep this SAD atm |
threshA = Data->temp[0]; // that's when we keep this SAD atm |
986 |
if (threshA < 512) threshA = 512; |
if (threshA < 512) threshA = 512; |
987 |
if (threshA > 1024) threshA = 1024; } |
else if (threshA > 1024) threshA = 1024; |
988 |
|
} else threshA = 512; |
989 |
|
|
990 |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
991 |
prevMBs + x + y * pParam->mb_width); |
prevMBs + x + y * pParam->mb_width, Data->rrv); |
992 |
|
|
993 |
if (inter4v || Data->chroma) CheckCandidate = CheckCandidate16; |
if (!Data->rrv) { |
994 |
|
if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16; |
995 |
else CheckCandidate = CheckCandidate16no4v; //for extra speed |
else CheckCandidate = CheckCandidate16no4v; //for extra speed |
996 |
|
} else CheckCandidate = CheckCandidate32; |
997 |
|
|
998 |
/* main loop. checking all predictions */ |
/* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/ |
999 |
|
|
1000 |
for (i = 1; i < 7; i++) { |
for (i = 1; i < 7; i++) { |
1001 |
if (!(mask = make_mask(pmv, i)) ) continue; |
if (!(mask = make_mask(pmv, i)) ) continue; |
1005 |
|
|
1006 |
if ((Data->iMinSAD[0] <= threshA) || |
if ((Data->iMinSAD[0] <= threshA) || |
1007 |
(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
1008 |
(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { |
(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) |
1009 |
inter4v = 0; |
inter4v = 0; |
1010 |
} else { |
else { |
1011 |
|
|
1012 |
MainSearchFunc * MainSearchPtr; |
MainSearchFunc * MainSearchPtr; |
1013 |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1023 |
if (MotionFlags & PMV_EXTSEARCH16) { |
if (MotionFlags & PMV_EXTSEARCH16) { |
1024 |
int32_t bSAD; |
int32_t bSAD; |
1025 |
VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; |
VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; |
1026 |
|
if (Data->rrv) { |
1027 |
|
startMV.x = RRV_MV_SCALEUP(startMV.x); |
1028 |
|
startMV.y = RRV_MV_SCALEUP(startMV.y); |
1029 |
|
} else |
1030 |
if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? |
if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? |
1031 |
startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); |
startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); |
1032 |
if (!(MVequal(startMV, backupMV))) { |
if (!(MVequal(startMV, backupMV))) { |
1040 |
} |
} |
1041 |
|
|
1042 |
backupMV = Data->currentMV[0]; |
backupMV = Data->currentMV[0]; |
1043 |
if (MotionFlags & PMV_HALFPELREFINE16) startMV.x = startMV.y = 1; |
if (MotionFlags & PMV_HALFPELREFINE16 && !Data->rrv) startMV.x = startMV.y = 1; |
1044 |
else startMV.x = startMV.y = 0; |
else startMV.x = startMV.y = 0; |
1045 |
if (!(MVequal(startMV, backupMV))) { |
if (!(MVequal(startMV, backupMV))) { |
1046 |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1061 |
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
1062 |
} |
} |
1063 |
|
|
1064 |
if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { |
if (Data->qpel && MotionFlags & PMV_QUARTERPELREFINE16) { |
|
|
|
1065 |
Data->qpel_precision = 1; |
Data->qpel_precision = 1; |
1066 |
get_range_qpel(&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, |
1067 |
pParam->width, pParam->height, Data->iFcode); |
pParam->width, pParam->height, Data->iFcode, 1, 0); |
1068 |
|
|
1069 |
SubpelRefine(Data); |
SubpelRefine(Data); |
1070 |
} |
} |
1072 |
if (Data->iMinSAD[0] < (int32_t)iQuant * 30 ) inter4v = 0; |
if (Data->iMinSAD[0] < (int32_t)iQuant * 30 ) inter4v = 0; |
1073 |
if (inter4v) { |
if (inter4v) { |
1074 |
SearchData Data8; |
SearchData Data8; |
1075 |
Data8.iFcode = Data->iFcode; |
memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data |
1076 |
Data8.lambda8 = Data->lambda8; |
|
|
Data8.iEdgedWidth = Data->iEdgedWidth; |
|
|
Data8.RefQ = Data->RefQ; |
|
|
Data8.qpel = Data->qpel; |
|
1077 |
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
1078 |
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); |
1079 |
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); |
1080 |
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); |
1081 |
|
|
1082 |
if (Data->chroma) { |
if (Data->chroma) { |
1083 |
int sumx, sumy, dx, dy; |
int sumx = 0, sumy = 0; |
1084 |
|
const int div = 1 + Data->qpel; |
1085 |
|
const VECTOR * const mv = Data->qpel ? pMB->qmvs : pMB->mvs; |
1086 |
|
|
1087 |
if(pParam->m_quarterpel) { |
for (i = 0; i < 4; i++) { |
1088 |
sumx= 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; |
1089 |
sumy = pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2; |
sumy += mv[i].y / div; |
|
} else { |
|
|
sumx = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
|
|
sumy = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
|
1090 |
} |
} |
|
dx = (sumx >> 3) + roundtab_76[sumx & 0xf]; |
|
|
dy = (sumy >> 3) + roundtab_76[sumy & 0xf]; |
|
1091 |
|
|
1092 |
Data->iMinSAD[1] += ChromaSAD(dx, dy, Data); |
Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf], |
1093 |
|
(sumy >> 3) + roundtab_76[sumy & 0xf], Data); |
1094 |
|
} |
1095 |
} |
} |
1096 |
|
|
1097 |
|
if (Data->rrv) { |
1098 |
|
Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); |
1099 |
|
Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); |
1100 |
} |
} |
1101 |
|
|
1102 |
if (!(inter4v) || |
if (!(inter4v) || |
1104 |
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
1105 |
// INTER MODE |
// INTER MODE |
1106 |
pMB->mode = MODE_INTER; |
pMB->mode = MODE_INTER; |
1107 |
pMB->mvs[0] = pMB->mvs[1] |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
1108 |
= 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]; |
|
|
|
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
|
|
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
|
1109 |
|
|
1110 |
if(pParam->m_quarterpel) { |
if(Data->qpel) { |
1111 |
pMB->qmvs[0] = pMB->qmvs[1] |
pMB->qmvs[0] = pMB->qmvs[1] |
1112 |
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
1113 |
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; |
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; |
1134 |
const int block, |
const int block, |
1135 |
SearchData * const Data) |
SearchData * const Data) |
1136 |
{ |
{ |
1137 |
|
int i = 0; |
1138 |
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
1139 |
Data->currentMV = OldData->currentMV + 1 + block; |
Data->currentMV = OldData->currentMV + 1 + block; |
1140 |
Data->currentQMV = OldData->currentQMV + 1 + block; |
Data->currentQMV = OldData->currentQMV + 1 + block; |
1141 |
|
|
1142 |
if(pParam->m_quarterpel) { |
if(Data->qpel) { |
1143 |
Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
1144 |
if (block != 0) *(Data->iMinSAD) += (Data->lambda8 * |
if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y, |
1145 |
d_mv_bits( Data->currentQMV->x - Data->predMV.x, |
Data->predMV, Data->iFcode, 0, 0); |
|
Data->currentQMV->y - Data->predMV.y, |
|
|
Data->iFcode) * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100; |
|
1146 |
} else { |
} else { |
1147 |
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); |
1148 |
if (block != 0) *(Data->iMinSAD) += (Data->lambda8 * |
if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y, |
1149 |
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; |
|
1150 |
} |
} |
1151 |
|
|
1152 |
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { |
*(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10; |
1153 |
|
|
1154 |
|
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8|PMV_QUARTERPELREFINE8)) { |
1155 |
|
if (Data->rrv) i = 2; else i = 1; |
1156 |
|
|
1157 |
Data->Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->Ref = OldData->Ref + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1158 |
Data->RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->RefH = OldData->RefH + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1159 |
Data->RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->RefV = OldData->RefV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1160 |
Data->RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->RefHV = OldData->RefHV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1161 |
|
|
1162 |
Data->Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
Data->Cur = OldData->Cur + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); |
1163 |
Data->qpel_precision = 0; |
Data->qpel_precision = 0; |
1164 |
|
|
1165 |
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, |
1166 |
pParam->width, pParam->height, OldData->iFcode - pParam->m_quarterpel); |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
1167 |
CheckCandidate = CheckCandidate8; |
|
1168 |
|
if (!Data->rrv) CheckCandidate = CheckCandidate8; |
1169 |
|
else CheckCandidate = CheckCandidate16no4v; |
1170 |
|
|
1171 |
if (MotionFlags & PMV_EXTSEARCH8) { |
if (MotionFlags & PMV_EXTSEARCH8) { |
1172 |
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
1195 |
} |
} |
1196 |
} |
} |
1197 |
|
|
1198 |
if(pParam->m_quarterpel) { |
if (Data->qpel && MotionFlags & PMV_QUARTERPELREFINE8) { |
|
if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && |
|
|
(MotionFlags & PMV_QUARTERPELREFINE8)) { |
|
1199 |
Data->qpel_precision = 1; |
Data->qpel_precision = 1; |
1200 |
get_range_qpel(&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, |
1201 |
pParam->width, pParam->height, OldData->iFcode); |
pParam->width, pParam->height, Data->iFcode, 1, 0); |
1202 |
SubpelRefine(Data); |
SubpelRefine(Data); |
1203 |
} |
} |
1204 |
} |
} |
1205 |
|
|
1206 |
|
if (Data->rrv) { |
1207 |
|
Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x); |
1208 |
|
Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y); |
1209 |
} |
} |
1210 |
|
|
1211 |
if(pParam->m_quarterpel) { |
if(Data->qpel) { |
1212 |
pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x; |
pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x; |
1213 |
pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y; |
pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y; |
1214 |
pMB->qmvs[block] = *(Data->currentQMV); |
pMB->qmvs[block] = *Data->currentQMV; |
1215 |
} |
} else { |
|
else { |
|
1216 |
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
1217 |
pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; |
pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; |
1218 |
} |
} |
1219 |
|
|
1220 |
pMB->mvs[block] = *(Data->currentMV); |
pMB->mvs[block] = *Data->currentMV; |
1221 |
pMB->sad8[block] = 4 * (*Data->iMinSAD); |
pMB->sad8[block] = 4 * *Data->iMinSAD; |
1222 |
} |
} |
1223 |
|
|
1224 |
/* B-frames code starts here */ |
/* motion estimation for B-frames */ |
1225 |
|
|
1226 |
static __inline VECTOR |
static __inline VECTOR |
1227 |
ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) |
ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) |
1228 |
{ |
{ |
1229 |
/* the stupidiest function ever */ |
/* the stupidiest function ever */ |
1230 |
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]; |
|
1231 |
} |
} |
1232 |
|
|
1233 |
static void __inline |
static void __inline |
1260 |
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); |
1261 |
} else pmv[5].x = pmv[5].y = 0; |
} else pmv[5].x = pmv[5].y = 0; |
1262 |
|
|
1263 |
if ((x != 0)&&(y != 0)) { |
if (x != 0 && y != 0) { |
1264 |
pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); |
pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); |
1265 |
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); |
1266 |
} else pmv[6].x = pmv[6].y = 0; |
} else pmv[6].x = pmv[6].y = 0; |
|
|
|
|
// more? |
|
1267 |
} |
} |
1268 |
|
|
1269 |
|
|
1270 |
/* search backward or forward, for b-frames */ |
/* search backward or forward */ |
1271 |
static void |
static void |
1272 |
SearchBF( const uint8_t * const pRef, |
SearchBF( const uint8_t * const pRef, |
1273 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1285 |
SearchData * const Data) |
SearchData * const Data) |
1286 |
{ |
{ |
1287 |
|
|
1288 |
const int32_t iEdgedWidth = pParam->edged_width; |
int i, iDirection = 255, mask; |
|
|
|
|
int i, iDirection, mask; |
|
1289 |
VECTOR pmv[7]; |
VECTOR pmv[7]; |
1290 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1291 |
*Data->iMinSAD = MV_MAX_ERROR; |
*Data->iMinSAD = MV_MAX_ERROR; |
1292 |
Data->iFcode = iFcode; |
Data->iFcode = iFcode; |
1293 |
Data->qpel_precision = 0; |
Data->qpel_precision = 0; |
1294 |
|
|
1295 |
Data->Ref = pRef + (x + y * iEdgedWidth) * 16; |
Data->Ref = pRef + (x + y * Data->iEdgedWidth) * 16; |
1296 |
Data->RefH = pRefH + (x + y * iEdgedWidth) * 16; |
Data->RefH = pRefH + (x + y * Data->iEdgedWidth) * 16; |
1297 |
Data->RefV = pRefV + (x + y * iEdgedWidth) * 16; |
Data->RefV = pRefV + (x + y * Data->iEdgedWidth) * 16; |
1298 |
Data->RefHV = pRefHV + (x + y * iEdgedWidth) * 16; |
Data->RefHV = pRefHV + (x + y * Data->iEdgedWidth) * 16; |
1299 |
|
|
1300 |
Data->predMV = *predMV; |
Data->predMV = *predMV; |
1301 |
|
|
1302 |
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, |
1303 |
pParam->width, pParam->height, iFcode - pParam->m_quarterpel); |
pParam->width, pParam->height, iFcode - Data->qpel, 0, 0); |
1304 |
|
|
1305 |
pmv[0] = Data->predMV; |
pmv[0] = Data->predMV; |
1306 |
if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } |
if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } |
1307 |
|
|
1308 |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
1309 |
|
|
1310 |
Data->currentMV->x = Data->currentMV->y = 0; |
Data->currentMV->x = Data->currentMV->y = 0; |
1311 |
CheckCandidate = CheckCandidate16no4v; |
CheckCandidate = CheckCandidate16no4v; |
1312 |
|
|
1313 |
// main loop. checking all predictions |
// main loop. checking all predictions |
1314 |
for (i = 0; i < 8; i++) { |
for (i = 0; i < 7; i++) { |
1315 |
if (!(mask = make_mask(pmv, i)) ) continue; |
if (!(mask = make_mask(pmv, i)) ) continue; |
1316 |
CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1317 |
} |
} |
1318 |
|
|
1319 |
if (MotionFlags & PMV_USESQUARES16) |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1320 |
MainSearchPtr = SquareSearch; |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
|
|
MainSearchPtr = AdvDiamondSearch; |
|
1321 |
else MainSearchPtr = DiamondSearch; |
else MainSearchPtr = DiamondSearch; |
1322 |
|
|
1323 |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
1324 |
|
|
1325 |
SubpelRefine(Data); |
SubpelRefine(Data); |
1326 |
|
|
1327 |
if (Data->qpel) { |
if (Data->qpel && *Data->iMinSAD < *best_sad + 300) { |
1328 |
Data->currentQMV->x = 2*Data->currentMV->x; |
Data->currentQMV->x = 2*Data->currentMV->x; |
1329 |
Data->currentQMV->y = 2*Data->currentMV->y; |
Data->currentQMV->y = 2*Data->currentMV->y; |
1330 |
Data->qpel_precision = 1; |
Data->qpel_precision = 1; |
1331 |
get_range_qpel(&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, |
1332 |
pParam->width, pParam->height, iFcode); |
pParam->width, pParam->height, iFcode, 1, 0); |
1333 |
SubpelRefine(Data); |
SubpelRefine(Data); |
1334 |
} |
} |
1335 |
|
|
1336 |
// three bits are needed to code backward mode. four for forward |
// three bits are needed to code backward mode. four for forward |
1337 |
// we treat the bits just like they were vector's |
|
1338 |
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
1339 |
else *Data->iMinSAD += 3 * Data->lambda16; |
else *Data->iMinSAD += 3 * Data->lambda16; |
1340 |
|
|
1352 |
pMB->pmvs[0].x = Data->currentMV->x - predMV->x; |
pMB->pmvs[0].x = Data->currentMV->x - predMV->x; |
1353 |
pMB->pmvs[0].y = Data->currentMV->y - predMV->y; |
pMB->pmvs[0].y = Data->currentMV->y - predMV->y; |
1354 |
} |
} |
1355 |
if (mode_current == MODE_FORWARD) |
if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; |
1356 |
pMB->mvs[0] = *(Data->currentMV+2) = *Data->currentMV; |
else pMB->b_mvs[0] = *Data->currentMV; |
1357 |
else |
} |
|
pMB->b_mvs[0] = *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search |
|
1358 |
|
|
1359 |
|
if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; |
1360 |
|
else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search |
1361 |
} |
} |
1362 |
|
|
1363 |
|
static void |
1364 |
|
SkipDecisionB(const IMAGE * const pCur, |
1365 |
|
const IMAGE * const f_Ref, |
1366 |
|
const IMAGE * const b_Ref, |
1367 |
|
MACROBLOCK * const pMB, |
1368 |
|
const uint32_t x, const uint32_t y, |
1369 |
|
const SearchData * const Data) |
1370 |
|
{ |
1371 |
|
int dx = 0, dy = 0, b_dx = 0, b_dy = 0; |
1372 |
|
int32_t sum; |
1373 |
|
const int div = 1 + Data->qpel; |
1374 |
|
int k; |
1375 |
|
const uint32_t stride = Data->iEdgedWidth/2; |
1376 |
|
//this is not full chroma compensation, only it's fullpel approximation. should work though |
1377 |
|
|
1378 |
|
for (k = 0; k < 4; k++) { |
1379 |
|
dy += Data->directmvF[k].y / div; |
1380 |
|
dx += Data->directmvF[0].x / div; |
1381 |
|
b_dy += Data->directmvB[0].y / div; |
1382 |
|
b_dx += Data->directmvB[0].x / div; |
1383 |
} |
} |
1384 |
|
|
1385 |
static int32_t |
dy = (dy >> 3) + roundtab_76[dy & 0xf]; |
1386 |
|
dx = (dx >> 3) + roundtab_76[dx & 0xf]; |
1387 |
|
b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf]; |
1388 |
|
b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf]; |
1389 |
|
|
1390 |
|
sum = sad8bi(pCur->u + 8 * x + 8 * y * stride, |
1391 |
|
f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2, |
1392 |
|
b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, |
1393 |
|
stride); |
1394 |
|
|
1395 |
|
if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; //no skip |
1396 |
|
|
1397 |
|
sum += sad8bi(pCur->v + 8*x + 8 * y * stride, |
1398 |
|
f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2, |
1399 |
|
b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, |
1400 |
|
stride); |
1401 |
|
|
1402 |
|
if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) pMB->mode = MODE_DIRECT_NONE_MV; //skipped |
1403 |
|
} |
1404 |
|
|
1405 |
|
static __inline uint32_t |
1406 |
SearchDirect(const IMAGE * const f_Ref, |
SearchDirect(const IMAGE * const f_Ref, |
1407 |
const uint8_t * const f_RefH, |
const uint8_t * const f_RefH, |
1408 |
const uint8_t * const f_RefV, |
const uint8_t * const f_RefV, |
1423 |
|
|
1424 |
{ |
{ |
1425 |
int32_t skip_sad; |
int32_t skip_sad; |
1426 |
int k; |
int k = (x + Data->iEdgedWidth*y) * 16; |
|
|
|
1427 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1428 |
|
|
1429 |
*Data->iMinSAD = 256*4096; |
*Data->iMinSAD = 256*4096; |
1430 |
|
Data->Ref = f_Ref->y + k; |
1431 |
|
Data->RefH = f_RefH + k; |
1432 |
|
Data->RefV = f_RefV + k; |
1433 |
|
Data->RefHV = f_RefHV + k; |
1434 |
|
Data->bRef = b_Ref->y + k; |
1435 |
|
Data->bRefH = b_RefH + k; |
1436 |
|
Data->bRefV = b_RefV + k; |
1437 |
|
Data->bRefHV = b_RefHV + k; |
1438 |
|
|
1439 |
|
k = Data->qpel ? 4 : 2; |
1440 |
|
Data->max_dx = k * (pParam->width - x * 16); |
1441 |
|
Data->max_dy = k * (pParam->height - y * 16); |
1442 |
|
Data->min_dx = -k * (16 + x * 16); |
1443 |
|
Data->min_dy = -k * (16 + y * 16); |
1444 |
|
|
1445 |
Data->Ref = f_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs; |
1446 |
Data->RefH = f_RefH + (x + Data->iEdgedWidth*y) * 16; |
Data->qpel_precision = 0; |
|
Data->RefV = f_RefV + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefHV = f_RefHV + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->bRef = b_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->bRefH = b_RefH + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->bRefV = b_RefV + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->bRefHV = b_RefHV + (x + Data->iEdgedWidth*y) * 16; |
|
|
|
|
|
Data->max_dx = 2 * pParam->width - 2 * (x) * 16; |
|
|
Data->max_dy = 2 * pParam->height - 2 * (y) * 16; |
|
|
Data->min_dx = -(2 * 16 + 2 * (x) * 16); |
|
|
Data->min_dy = -(2 * 16 + 2 * (y) * 16); |
|
|
if (Data->qpel) { //we measure in qpixels |
|
|
Data->max_dx *= 2; |
|
|
Data->max_dy *= 2; |
|
|
Data->min_dx *= 2; |
|
|
Data->min_dy *= 2; |
|
|
Data->referencemv = b_mb->qmvs; |
|
|
} else Data->referencemv = b_mb->mvs; |
|
|
Data->qpel_precision = 0; // it's a trick. it's 1 not 0, but we need 0 here |
|
1447 |
|
|
1448 |
for (k = 0; k < 4; k++) { |
for (k = 0; k < 4; k++) { |
1449 |
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); |
1451 |
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); |
1452 |
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; |
1453 |
|
|
1454 |
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) |
1455 |
|| ( 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) ) { |
1456 |
|
|
1457 |
*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 |
1458 |
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" |
1459 |
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
1460 |
return 0; |
return 256*4096; |
1461 |
} |
} |
1462 |
if (b_mb->mode != MODE_INTER4V) { |
if (b_mb->mode != MODE_INTER4V) { |
1463 |
pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; |
pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; |
1468 |
} |
} |
1469 |
} |
} |
1470 |
|
|
1471 |
|
CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v; |
|
if (b_mb->mode == MODE_INTER4V) CheckCandidate = CheckCandidateDirect; |
|
|
else CheckCandidate = CheckCandidateDirectno4v; |
|
1472 |
|
|
1473 |
(*CheckCandidate)(0, 0, 255, &k, Data); |
(*CheckCandidate)(0, 0, 255, &k, Data); |
1474 |
|
|
1475 |
// skip decision |
// initial (fast) skip decision |
1476 |
if (*Data->iMinSAD < pMB->quant * SKIP_THRESH_B) { |
if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * 2) { |
1477 |
//possible skip - checking chroma. everything copied from MC |
//possible skip - checking chroma |
1478 |
//this is not full chroma compensation, only it's fullpel approximation. should work though |
SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data); |
1479 |
int sum, dx, dy, b_dx, b_dy; |
if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip. |
|
|
|
|
if (Data->qpel) { |
|
|
sum = pMB->mvs[0].y/2 + pMB->mvs[1].y/2 + pMB->mvs[2].y/2 + pMB->mvs[3].y/2; |
|
|
dy = (sum >> 3) + roundtab_76[sum & 0xf]; |
|
|
sum = pMB->mvs[0].x/2 + pMB->mvs[1].x/2 + pMB->mvs[2].x/2 + pMB->mvs[3].x/2; |
|
|
dx = (sum >> 3) + roundtab_76[sum & 0xf]; |
|
|
|
|
|
sum = pMB->b_mvs[0].y/2 + pMB->b_mvs[1].y/2 + pMB->b_mvs[2].y/2 + pMB->b_mvs[3].y/2; |
|
|
b_dy = (sum >> 3) + roundtab_76[sum & 0xf]; |
|
|
sum = pMB->b_mvs[0].x/2 + pMB->b_mvs[1].x/2 + pMB->b_mvs[2].x/2 + pMB->b_mvs[3].x/2; |
|
|
b_dx = (sum >> 3) + roundtab_76[sum & 0xf]; |
|
|
|
|
|
} else { |
|
|
sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
|
|
dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
|
sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
|
|
dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
|
|
|
|
sum = pMB->b_mvs[0].x + pMB->b_mvs[1].x + pMB->b_mvs[2].x + pMB->b_mvs[3].x; |
|
|
b_dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
|
sum = pMB->b_mvs[0].y + pMB->b_mvs[1].y + pMB->b_mvs[2].y + pMB->b_mvs[3].y; |
|
|
b_dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
|
} |
|
|
sum = sad8bi(pCur->u + 8*x + 8*y*(Data->iEdgedWidth/2), |
|
|
f_Ref->u + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
|
|
b_Ref->u + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
|
|
Data->iEdgedWidth/2); |
|
|
sum += sad8bi(pCur->v + 8*x + 8*y*(Data->iEdgedWidth/2), |
|
|
f_Ref->v + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
|
|
b_Ref->v + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
|
|
Data->iEdgedWidth/2); |
|
|
|
|
|
if (sum < MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) { |
|
|
pMB->mode = MODE_DIRECT_NONE_MV; |
|
|
return *Data->iMinSAD; |
|
|
} |
|
1480 |
} |
} |
1481 |
|
|
1482 |
skip_sad = *Data->iMinSAD; |
skip_sad = *Data->iMinSAD; |
1492 |
|
|
1493 |
SubpelRefine(Data); |
SubpelRefine(Data); |
1494 |
|
|
|
// *Data->iMinSAD += 1 * Data->lambda16; // one bit is needed to code direct mode |
|
1495 |
*best_sad = *Data->iMinSAD; |
*best_sad = *Data->iMinSAD; |
1496 |
|
|
1497 |
if (b_mb->mode == MODE_INTER4V) |
if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT; |
|
pMB->mode = MODE_DIRECT; |
|
1498 |
else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
1499 |
|
|
1500 |
pMB->pmvs[3] = *Data->currentMV; |
pMB->pmvs[3] = *Data->currentMV; |
1526 |
return skip_sad; |
return skip_sad; |
1527 |
} |
} |
1528 |
|
|
1529 |
|
static void |
|
static __inline void |
|
1530 |
SearchInterpolate(const uint8_t * const f_Ref, |
SearchInterpolate(const uint8_t * const f_Ref, |
1531 |
const uint8_t * const f_RefH, |
const uint8_t * const f_RefH, |
1532 |
const uint8_t * const f_RefV, |
const uint8_t * const f_RefV, |
1549 |
|
|
1550 |
{ |
{ |
1551 |
|
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
1552 |
int iDirection, i, j; |
int iDirection, i, j; |
1553 |
SearchData bData; |
SearchData bData; |
1554 |
|
|
1555 |
*(bData.iMinSAD = fData->iMinSAD) = 4096*256; |
fData->qpel_precision = 0; |
1556 |
bData.Cur = fData->Cur; |
memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data |
1557 |
fData->iEdgedWidth = bData.iEdgedWidth = iEdgedWidth; |
*fData->iMinSAD = 4096*256; |
1558 |
bData.currentMV = fData->currentMV + 1; bData.currentQMV = fData->currentQMV + 1; |
bData.currentMV++; bData.currentQMV++; |
|
bData.lambda16 = fData->lambda16; |
|
1559 |
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
1560 |
|
|
1561 |
bData.bRef = fData->Ref = f_Ref + (x + y * iEdgedWidth) * 16; |
i = (x + y * fData->iEdgedWidth) * 16; |
1562 |
bData.bRefH = fData->RefH = f_RefH + (x + y * iEdgedWidth) * 16; |
bData.bRef = fData->Ref = f_Ref + i; |
1563 |
bData.bRefV = fData->RefV = f_RefV + (x + y * iEdgedWidth) * 16; |
bData.bRefH = fData->RefH = f_RefH + i; |
1564 |
bData.bRefHV = fData->RefHV = f_RefHV + (x + y * iEdgedWidth) * 16; |
bData.bRefV = fData->RefV = f_RefV + i; |
1565 |
bData.Ref = fData->bRef = b_Ref + (x + y * iEdgedWidth) * 16; |
bData.bRefHV = fData->RefHV = f_RefHV + i; |
1566 |
bData.RefH = fData->bRefH = b_RefH + (x + y * iEdgedWidth) * 16; |
bData.Ref = fData->bRef = b_Ref + i; |
1567 |
bData.RefV = fData->bRefV = b_RefV + (x + y * iEdgedWidth) * 16; |
bData.RefH = fData->bRefH = b_RefH + i; |
1568 |
bData.RefHV = fData->bRefHV = b_RefHV + (x + y * iEdgedWidth) * 16; |
bData.RefV = fData->bRefV = b_RefV + i; |
1569 |
bData.RefQ = fData->RefQ; |
bData.RefHV = fData->bRefHV = b_RefHV + i; |
|
fData->qpel_precision = bData.qpel_precision = 0; bData.qpel = fData->qpel; |
|
|
bData.rounding = 0; |
|
1570 |
|
|
1571 |
bData.bpredMV = fData->predMV = *f_predMV; |
bData.bpredMV = fData->predMV = *f_predMV; |
1572 |
fData->bpredMV = bData.predMV = *b_predMV; |
fData->bpredMV = bData.predMV = *b_predMV; |
|
|
|
1573 |
fData->currentMV[0] = fData->currentMV[2]; |
fData->currentMV[0] = fData->currentMV[2]; |
1574 |
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); |
|
1575 |
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode - pParam->m_quarterpel); |
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode - fData->qpel, 0, 0); |
1576 |
|
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); |
1577 |
|
|
1578 |
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; |
1579 |
if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx; |
if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx; |
1587 |
|
|
1588 |
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
1589 |
|
|
1590 |
//diamond. I wish we could use normal mainsearch functions (square, advdiamond) |
//diamond |
|
|
|
1591 |
do { |
do { |
1592 |
iDirection = 255; |
iDirection = 255; |
1593 |
// forward MV moves |
// forward MV moves |
1608 |
|
|
1609 |
} while (!(iDirection)); |
} while (!(iDirection)); |
1610 |
|
|
1611 |
|
//qpel refinement |
1612 |
if (fData->qpel) { |
if (fData->qpel) { |
1613 |
|
if (*fData->iMinSAD > *best_sad + 500) return; |
1614 |
CheckCandidate = CheckCandidateInt; |
CheckCandidate = CheckCandidateInt; |
1615 |
fData->qpel_precision = bData.qpel_precision = 1; |
fData->qpel_precision = bData.qpel_precision = 1; |
1616 |
get_range_qpel(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode); |
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0); |
1617 |
get_range_qpel(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode); |
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0); |
1618 |
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; |
1619 |
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; |
1620 |
fData->currentQMV[1].x = 2 * fData->currentMV[1].x; |
fData->currentQMV[1].x = 2 * fData->currentMV[1].x; |
1621 |
fData->currentQMV[1].y = 2 * fData->currentMV[1].y; |
fData->currentQMV[1].y = 2 * fData->currentMV[1].y; |
1622 |
SubpelRefine(fData); |
SubpelRefine(fData); |
1623 |
|
if (*fData->iMinSAD > *best_sad + 300) return; |
1624 |
fData->currentQMV[2] = fData->currentQMV[0]; |
fData->currentQMV[2] = fData->currentQMV[0]; |
1625 |
SubpelRefine(&bData); |
SubpelRefine(&bData); |
1626 |
} |
} |
1627 |
|
|
1628 |
*fData->iMinSAD += 2 * fData->lambda16; // two bits are needed to code interpolate mode. |
*fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode. |
1629 |
|
|
1630 |
if (*fData->iMinSAD < *best_sad) { |
if (*fData->iMinSAD < *best_sad) { |
1631 |
*best_sad = *fData->iMinSAD; |
*best_sad = *fData->iMinSAD; |
1667 |
const IMAGE * const b_refHV) |
const IMAGE * const b_refHV) |
1668 |
{ |
{ |
1669 |
uint32_t i, j; |
uint32_t i, j; |
1670 |
int32_t best_sad, skip_sad; |
int32_t best_sad; |
1671 |
|
uint32_t skip_sad; |
1672 |
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}; |
|
1673 |
const MACROBLOCK * const b_mbs = b_reference->mbs; |
const MACROBLOCK * const b_mbs = b_reference->mbs; |
1674 |
|
|
1675 |
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
1676 |
|
|
1677 |
const int32_t TRB = time_pp - time_bp; |
const int32_t TRB = time_pp - time_bp; |
1678 |
const int32_t TRD = time_pp; |
const int32_t TRD = time_pp; |
|
uint8_t * qimage; |
|
1679 |
|
|
1680 |
// some pre-inintialized data for the rest of the search |
// some pre-inintialized data for the rest of the search |
1681 |
|
|
1683 |
int32_t iMinSAD; |
int32_t iMinSAD; |
1684 |
VECTOR currentMV[3]; |
VECTOR currentMV[3]; |
1685 |
VECTOR currentQMV[3]; |
VECTOR currentQMV[3]; |
1686 |
|
memset(&Data, 0, sizeof(SearchData)); |
1687 |
Data.iEdgedWidth = pParam->edged_width; |
Data.iEdgedWidth = pParam->edged_width; |
1688 |
Data.currentMV = currentMV; Data.currentQMV = currentQMV; |
Data.currentMV = currentMV; Data.currentQMV = currentQMV; |
1689 |
Data.iMinSAD = &iMinSAD; |
Data.iMinSAD = &iMinSAD; |
1690 |
Data.lambda16 = lambda_vec16[frame->quant] + 2; |
Data.lambda16 = lambda_vec16[frame->quant]; |
1691 |
|
Data.chroma = frame->quant; |
1692 |
Data.qpel = pParam->m_quarterpel; |
Data.qpel = pParam->m_quarterpel; |
1693 |
Data.rounding = 0; |
Data.rounding = 0; |
1694 |
|
|
1695 |
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; |
|
|
|
|
1696 |
// note: i==horizontal, j==vertical |
// note: i==horizontal, j==vertical |
1697 |
for (j = 0; j < pParam->mb_height; j++) { |
for (j = 0; j < pParam->mb_height; j++) { |
1698 |
|
|
1744 |
MODE_BACKWARD, &Data); |
MODE_BACKWARD, &Data); |
1745 |
|
|
1746 |
// 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 |
|
|
|
1747 |
SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
1748 |
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
1749 |
&frame->image, |
&frame->image, |
1755 |
pMB, &best_sad, |
pMB, &best_sad, |
1756 |
&Data); |
&Data); |
1757 |
|
|
1758 |
|
// final skip decision |
1759 |
|
if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP*2) |
1760 |
|
&& ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) |
1761 |
|
SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data); |
1762 |
|
|
1763 |
switch (pMB->mode) { |
switch (pMB->mode) { |
1764 |
case MODE_FORWARD: |
case MODE_FORWARD: |
1765 |
f_count++; |
f_count++; |
1766 |
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]; |
|
1767 |
break; |
break; |
1768 |
case MODE_BACKWARD: |
case MODE_BACKWARD: |
1769 |
b_count++; |
b_count++; |
1770 |
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]; |
|
1771 |
break; |
break; |
1772 |
case MODE_INTERPOLATE: |
case MODE_INTERPOLATE: |
1773 |
i_count++; |
i_count++; |
1774 |
if (pParam->m_quarterpel) { |
f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; |
1775 |
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]; |
|
|
} |
|
1776 |
break; |
break; |
1777 |
case MODE_DIRECT: |
case MODE_DIRECT: |
1778 |
case MODE_DIRECT_NO4V: |
case MODE_DIRECT_NO4V: |
1782 |
} |
} |
1783 |
} |
} |
1784 |
} |
} |
|
free(qimage); |
|
1785 |
} |
} |
1786 |
|
|
1787 |
/* Hinted ME starts here */ |
static __inline void |
|
|
|
|
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; |
|
|
|
|
|
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; |
|
|
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); |
|
|
} |
|
|
if (pParam->m_quarterpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
|
|
else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
|
|
|
|
|
for(i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
|
|
|
|
|
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
|
|
|
|
|
if (inter4v || 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) SubpelRefine(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_qpel(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
|
|
pParam->width, pParam->height, Data->iFcode); |
|
|
Data->qpel_precision = 1; |
|
|
SubpelRefine(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 sumx, sumy, dx, dy; |
|
|
|
|
|
if(pParam->m_quarterpel) { |
|
|
sumx= pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2; |
|
|
sumy = pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2; |
|
|
} else { |
|
|
sumx = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
|
|
sumy = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
|
|
} |
|
|
dx = (sumx >> 3) + roundtab_76[sumx & 0xf]; |
|
|
dy = (sumy >> 3) + roundtab_76[sumy & 0xf]; |
|
|
|
|
|
Data->iMinSAD[1] += ChromaSAD(dx, dy, Data); |
|
|
} |
|
|
} |
|
|
|
|
|
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->predMV.x; |
|
|
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; |
|
|
} else { |
|
|
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
|
|
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
|
|
} |
|
|
} else { |
|
|
// INTER4V MODE; all other things are already set in Search8 |
|
|
pMB->mode = MODE_INTER4V; |
|
|
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] |
|
|
+ Data->iMinSAD[4] + IMV16X16 * iQuant; |
|
|
} |
|
|
|
|
|
} |
|
|
|
|
|
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); |
|
|
} |
|
|
|
|
|
static __inline int |
|
1788 |
MEanalyzeMB ( const uint8_t * const pRef, |
MEanalyzeMB ( const uint8_t * const pRef, |
1789 |
const uint8_t * const pCur, |
const uint8_t * const pCur, |
1790 |
const int x, |
const int x, |
1791 |
const int y, |
const int y, |
1792 |
const MBParam * const pParam, |
const MBParam * const pParam, |
1793 |
const MACROBLOCK * const pMBs, |
MACROBLOCK * const pMBs, |
|
MACROBLOCK * const pMB, |
|
1794 |
SearchData * const Data) |
SearchData * const Data) |
1795 |
{ |
{ |
1796 |
|
|
1797 |
int i = 255, mask; |
int i, mask; |
1798 |
VECTOR pmv[3]; |
VECTOR pmv[3]; |
1799 |
*(Data->iMinSAD) = MV_MAX_ERROR; |
MACROBLOCK * pMB = &pMBs[x + y * pParam->mb_width]; |
1800 |
|
|
1801 |
|
for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
1802 |
|
|
1803 |
//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 |
1804 |
if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; |
if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; |
1805 |
else |
else |
1806 |
if (x == 1) //left macroblock does not have any vector now |
if (x == 1) //left macroblock does not have any vector now |
1807 |
Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median |
Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median |
1808 |
else if (y == 1) // top macroblock don't have it's vector |
else if (y == 1) // top macroblock doesn't have it's vector |
1809 |
Data->predMV = (pMB - 1)->mvs[0]; // left instead of median |
Data->predMV = (pMB - 1)->mvs[0]; // left instead of median |
1810 |
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 |
1811 |
|
|
1812 |
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, |
1813 |
pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel); |
pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, Data->rrv); |
1814 |
|
|
1815 |
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
1816 |
Data->Ref = pRef + (x + y * pParam->edged_width) * 16; |
Data->Ref = pRef + (x + y * pParam->edged_width) * 16; |
1821 |
pmv[2].y = EVEN(Data->predMV.y); |
pmv[2].y = EVEN(Data->predMV.y); |
1822 |
pmv[0].x = pmv[0].y = 0; |
pmv[0].x = pmv[0].y = 0; |
1823 |
|
|
1824 |
(*CheckCandidate)(0, 0, 255, &i, Data); |
CheckCandidate32I(0, 0, 255, &i, Data); |
1825 |
|
|
1826 |
//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; |
|
|
} |
|
1827 |
|
|
1828 |
if (!(mask = make_mask(pmv, 1))) |
if (!(mask = make_mask(pmv, 1))) |
1829 |
(*CheckCandidate)(pmv[1].x, pmv[1].y, mask, &i, Data); |
CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data); |
1830 |
if (!(mask = make_mask(pmv, 2))) |
if (!(mask = make_mask(pmv, 2))) |
1831 |
(*CheckCandidate)(pmv[2].x, pmv[2].y, mask, &i, Data); |
CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data); |
1832 |
|
|
1833 |
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 |
1834 |
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
1835 |
|
|
1836 |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
for (i = 0; i < 4; i++) { |
1837 |
pMB->mode = MODE_INTER; |
MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width]; |
1838 |
return *(Data->iMinSAD); |
MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i]; |
1839 |
|
MB->mode = MODE_INTER; |
1840 |
|
MB->sad16 = Data->iMinSAD[i+1]; |
1841 |
|
} |
1842 |
|
} |
1843 |
} |
} |
1844 |
|
|
1845 |
#define INTRA_THRESH 1350 |
#define INTRA_BIAS 2500 |
1846 |
#define INTER_THRESH 1200 |
#define INTRA_THRESH 1500 |
1847 |
|
#define INTER_THRESH 1400 |
1848 |
|
|
1849 |
int |
int |
1850 |
MEanalysis( const IMAGE * const pRef, |
MEanalysis( const IMAGE * const pRef, |
1852 |
MBParam * const pParam, |
MBParam * const pParam, |
1853 |
int maxIntra, //maximum number if non-I frames |
int maxIntra, //maximum number if non-I frames |
1854 |
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 |
1855 |
int bCount) // number if B frames in a row |
int bCount) // number of B frames in a row |
1856 |
{ |
{ |
|
int mb_width = pParam->mb_width; |
|
|
int mb_height = pParam->mb_height; |
|
|
|
|
1857 |
uint32_t x, y, intra = 0; |
uint32_t x, y, intra = 0; |
1858 |
int sSAD = 0; |
int sSAD = 0; |
1859 |
MACROBLOCK * const pMBs = Current->mbs; |
MACROBLOCK * const pMBs = Current->mbs; |
1860 |
const IMAGE * const pCurrent = &Current->image; |
const IMAGE * const pCurrent = &Current->image; |
1861 |
int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH; |
int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH; |
1862 |
|
|
1863 |
VECTOR currentMV; |
int32_t iMinSAD[5], temp[5]; |
1864 |
int32_t iMinSAD; |
VECTOR currentMV[5]; |
1865 |
SearchData Data; |
SearchData Data; |
1866 |
Data.iEdgedWidth = pParam->edged_width; |
Data.iEdgedWidth = pParam->edged_width; |
1867 |
Data.currentMV = ¤tMV; |
Data.currentMV = currentMV; |
1868 |
Data.iMinSAD = &iMinSAD; |
Data.iMinSAD = iMinSAD; |
1869 |
Data.iFcode = Current->fcode; |
Data.iFcode = Current->fcode; |
1870 |
CheckCandidate = CheckCandidate16no4vI; |
Data.rrv = Current->global_flags & XVID_REDUCED; |
1871 |
|
Data.temp = temp; |
1872 |
if ((Current->global_flags & XVID_REDUCED)) |
CheckCandidate = CheckCandidate32I; |
|
{ |
|
|
mb_width = (pParam->width + 31) / 32; |
|
|
mb_height = (pParam->height + 31) / 32; |
|
|
} |
|
|
|
|
1873 |
|
|
1874 |
if (intraCount < 10) // we're right after an I frame |
if (intraCount != 0 && intraCount < 10) // we're right after an I frame |
1875 |
IntraThresh += 4 * (intraCount - 10) * (intraCount - 10); |
IntraThresh += 4 * (intraCount - 10) * (intraCount - 10); |
1876 |
else |
else |
1877 |
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 |
1878 |
IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; |
IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; |
1879 |
|
|
|
|
|
1880 |
InterThresh += 400 * (1 - bCount); |
InterThresh += 400 * (1 - bCount); |
1881 |
if (InterThresh < 200) InterThresh = 200; |
if (InterThresh < 300) InterThresh = 300; |
1882 |
|
|
1883 |
if (sadInit) (*sadInit) (); |
if (sadInit) (*sadInit) (); |
1884 |
|
|
1885 |
for (y = 1; y < mb_height-1; y++) { |
for (y = 1; y < pParam->mb_height-1; y += 2) { |
1886 |
for (x = 1; x < mb_width-1; x++) { |
for (x = 1; x < pParam->mb_width-1; x += 2) { |
1887 |
int sad, dev; |
int i; |
1888 |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
|
1889 |
|
if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; |
1890 |
sad = MEanalyzeMB(pRef->y, pCurrent->y, x, y, |
|
1891 |
pParam, pMBs, pMB, &Data); |
MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data); |
1892 |
|
|
1893 |
if (sad > IntraThresh) { |
for (i = 0; i < 4; i++) { |
1894 |
dev = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
int dev; |
1895 |
|
MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width]; |
1896 |
|
if (pMB->sad16 > IntraThresh) { |
1897 |
|
dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16, |
1898 |
pParam->edged_width); |
pParam->edged_width); |
1899 |
if (dev + IntraThresh < sad) { |
if (dev + IntraThresh < pMB->sad16) { |
1900 |
pMB->mode = MODE_INTRA; |
pMB->mode = MODE_INTRA; |
1901 |
if (++intra > (mb_height-2)*(mb_width-2)/2) return 2; // I frame |
if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return I_VOP; |
|
} |
|
|
} |
|
|
sSAD += sad; |
|
1902 |
} |
} |
1903 |
} |
} |
1904 |
sSAD /= (mb_height-2)*(mb_width-2); |
sSAD += pMB->sad16; |
|
if (sSAD > InterThresh ) return 1; //P frame |
|
|
emms(); |
|
|
return 0; // B frame |
|
|
|
|
1905 |
} |
} |
|
|
|
|
int |
|
|
FindFcode( const MBParam * const pParam, |
|
|
const FRAMEINFO * const current) |
|
|
{ |
|
|
int mb_width = pParam->mb_width; |
|
|
int mb_height = pParam->mb_height; |
|
|
|
|
|
uint32_t x, y; |
|
|
int max = 0, min = 0, i; |
|
|
|
|
|
|
|
|
if ((current->global_flags & XVID_REDUCED)) |
|
|
{ |
|
|
mb_width = (pParam->width + 31) / 32; |
|
|
mb_height = (pParam->height + 31) / 32; |
|
1906 |
} |
} |
|
|
|
|
|
|
|
for (y = 0; y < mb_height; y++) { |
|
|
for (x = 0; x < mb_width; x++) { |
|
|
|
|
|
MACROBLOCK *pMB = ¤t->mbs[x + y * pParam->mb_width]; |
|
|
for(i = 0; i < (pMB->mode == MODE_INTER4V ? 4:1); i++) { |
|
|
if (pMB->mvs[i].x > max) max = pMB->mvs[i].x; |
|
|
if (pMB->mvs[i].y > max) max = pMB->mvs[i].y; |
|
|
|
|
|
if (pMB->mvs[i].x < min) min = pMB->mvs[i].x; |
|
|
if (pMB->mvs[i].y < min) min = pMB->mvs[i].y; |
|
1907 |
} |
} |
1908 |
} |
sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); |
1909 |
} |
if (sSAD > IntraThresh + INTRA_BIAS) return I_VOP; |
1910 |
|
if (sSAD > InterThresh ) return P_VOP; |
1911 |
min = -min; |
emms(); |
1912 |
max += 1; |
return B_VOP; |
|
if (min > max) max = min; |
|
|
if (pParam->m_quarterpel) max *= 2; |
|
1913 |
|
|
|
for (i = 1; (max > 32 << (i - 1)); i++); |
|
|
return i; |
|
1914 |
} |
} |
1915 |
|
|
1916 |
static void |
static void |
1977 |
max_x = gmc.x + step; |
max_x = gmc.x + step; |
1978 |
min_y = gmc.y - step; |
min_y = gmc.y - step; |
1979 |
max_y = gmc.y + step; |
max_y = gmc.y + step; |
|
|
|
1980 |
} |
} |
1981 |
|
|
1982 |
if (bestcount < (pParam->mb_height-2)*(pParam->mb_width-2)/10) |
if (bestcount < (pParam->mb_height-2)*(pParam->mb_width-2)/10) |
1983 |
gmc.x = gmc.y = 0; //no camara pan, no GMC |
gmc.x = gmc.y = 0; //no camara pan, no GMC |
1984 |
|
|
1985 |
// step2: let's refine camera panning using gradiend-descent approach. |
// step2: let's refine camera panning using gradiend-descent approach |
1986 |
// TODO: more warping points may be evaluated here (like in interpolate mode search - two vectors in one diamond) |
// TODO: more warping points may be evaluated here (like in interpolate mode search - two vectors in one diamond) |
1987 |
bestcount = 0; |
bestcount = 0; |
1988 |
CheckGMC(gmc.x, gmc.y, 255, &iDirection, pMBs, &bestcount, &gmc, pParam); |
CheckGMC(gmc.x, gmc.y, 255, &iDirection, pMBs, &bestcount, &gmc, pParam); |