| 1 |
/************************************************************************** |
/************************************************************************** |
| 2 |
* |
* |
| 3 |
* Modifications: |
* XVID MPEG-4 VIDEO CODEC |
| 4 |
|
* motion estimation |
| 5 |
* |
* |
| 6 |
* 08.02.2002 split up PMVfast into three routines: PMVFast, PMVFast_MainLoop |
* This program is an implementation of a part of one or more MPEG-4 |
| 7 |
* PMVFast_Refine to support multiple searches with different start points |
* Video tools as specified in ISO/IEC 14496-2 standard. Those intending |
| 8 |
* 07.01.2002 uv-block-based interpolation |
* to use this software module in hardware or software products are |
| 9 |
* 06.01.2002 INTER/INTRA-decision is now done before any SEARCH8 (speedup) |
* advised that its use may infringe existing patents or copyrights, and |
| 10 |
* changed INTER_BIAS to 150 (as suggested by suxen_drol) |
* any such use would be at such party's own risk. The original |
| 11 |
* removed halfpel refinement step in PMVfastSearch8 + quality=5 |
* developer of this software module and his/her company, and subsequent |
| 12 |
* added new quality mode = 6 which performs halfpel refinement |
* editors and their companies, will have no liability for use of this |
| 13 |
* filesize difference between quality 5 and 6 is smaller than 1% |
* software or modifications or derivatives thereof. |
|
* (Isibaar) |
|
|
* 31.12.2001 PMVfastSearch16 and PMVfastSearch8 (gruel) |
|
|
* 30.12.2001 get_range/MotionSearchX simplified; blue/green bug fix |
|
|
* 22.12.2001 commented best_point==99 check |
|
|
* 19.12.2001 modified get_range (purple bug fix) |
|
|
* 15.12.2001 moved pmv displacement from mbprediction |
|
|
* 02.12.2001 motion estimation/compensation split (Isibaar) |
|
|
* 16.11.2001 rewrote/tweaked search algorithms; pross@cs.rmit.edu.au |
|
|
* 10.11.2001 support for sad16/sad8 functions |
|
|
* 28.08.2001 reactivated MODE_INTER4V for EXT_MODE |
|
|
* 24.08.2001 removed MODE_INTER4V_Q, disabled MODE_INTER4V for EXT_MODE |
|
|
* 22.08.2001 added MODE_INTER4V_Q |
|
|
* 20.08.2001 added pragma to get rid of internal compiler error with VC6 |
|
|
* idea by Cyril. Thanks. |
|
| 14 |
* |
* |
| 15 |
* Michael Militzer <isibaar@videocoding.de> |
* This program is free software; you can redistribute it and/or modify |
| 16 |
|
* it under the terms of the GNU General Public License as published by |
| 17 |
|
* the Free Software Foundation; either version 2 of the License, or |
| 18 |
|
* (at your option) any later version. |
| 19 |
* |
* |
| 20 |
**************************************************************************/ |
* This program is distributed in the hope that it will be useful, |
| 21 |
|
* but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 22 |
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 23 |
|
* GNU General Public License for more details. |
| 24 |
|
* |
| 25 |
|
* You should have received a copy of the GNU General Public License |
| 26 |
|
* along with this program; if not, write to the Free Software |
| 27 |
|
* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
| 28 |
|
* |
| 29 |
|
*************************************************************************/ |
| 30 |
|
|
| 31 |
#include <assert.h> |
#include <assert.h> |
| 32 |
#include <stdio.h> |
#include <stdio.h> |
| 33 |
|
#include <stdlib.h> |
| 34 |
|
#include <string.h> // memcpy |
| 35 |
|
|
| 36 |
#include "../encoder.h" |
#include "../encoder.h" |
| 37 |
#include "../utils/mbfunctions.h" |
#include "../utils/mbfunctions.h" |
| 38 |
#include "../prediction/mbprediction.h" |
#include "../prediction/mbprediction.h" |
| 39 |
#include "../global.h" |
#include "../global.h" |
| 40 |
#include "../utils/timer.h" |
#include "../utils/timer.h" |
| 41 |
|
#include "../image/interpolate8x8.h" |
| 42 |
|
#include "motion_est.h" |
| 43 |
|
#include "motion.h" |
| 44 |
#include "sad.h" |
#include "sad.h" |
| 45 |
|
#include "../utils/emms.h" |
| 46 |
|
|
| 47 |
// very large value |
#define INITIAL_SKIP_THRESH (10) |
| 48 |
#define MV_MAX_ERROR (4096 * 256) |
#define FINAL_SKIP_THRESH (50) |
| 49 |
|
#define MAX_SAD00_FOR_SKIP (20) |
| 50 |
// stop search if sdelta < THRESHOLD |
#define MAX_CHROMA_SAD_FOR_SKIP (22) |
| 51 |
#define MV16_THRESHOLD 192 |
|
| 52 |
#define MV8_THRESHOLD 56 |
#define CHECK_CANDIDATE(X,Y,D) { \ |
| 53 |
|
(*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } |
| 54 |
/* sad16(0,0) bias; mpeg4 spec suggests nb/2+1 */ |
|
| 55 |
/* nb = vop pixels * 2^(bpp-8) */ |
static __inline int |
| 56 |
#define MV16_00_BIAS (128+1) |
d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv) |
| 57 |
|
{ |
| 58 |
/* INTER bias for INTER/INTRA decision; mpeg4 spec suggests 2*nb */ |
int xb, yb; |
| 59 |
#define INTER_BIAS 512 |
if (qpel) { x *= 2; y *= 2;} |
| 60 |
|
else if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); } |
| 61 |
/* Parameters which control inter/inter4v decision */ |
x = pred.x - x; |
| 62 |
#define IMV16X16 5 |
y = pred.y - y; |
| 63 |
|
|
| 64 |
|
if (x == 0) xb = 1; |
| 65 |
|
else { |
| 66 |
|
if (x < 0) x = -x; |
| 67 |
|
x += (1 << (iFcode - 1)) - 1; |
| 68 |
|
x >>= (iFcode - 1); |
| 69 |
|
if (x > 32) x = 32; |
| 70 |
|
xb = mvtab[x] + iFcode; |
| 71 |
|
} |
| 72 |
|
|
| 73 |
|
if (y == 0) yb = 1; |
| 74 |
|
else { |
| 75 |
|
if (y < 0) y = -y; |
| 76 |
|
y += (1 << (iFcode - 1)) - 1; |
| 77 |
|
y >>= (iFcode - 1); |
| 78 |
|
if (y > 32) y = 32; |
| 79 |
|
yb = mvtab[y] + iFcode; |
| 80 |
|
} |
| 81 |
|
return xb + yb; |
| 82 |
|
} |
| 83 |
|
|
| 84 |
|
static int32_t |
| 85 |
|
ChromaSAD(int dx, int dy, const SearchData * const data) |
| 86 |
|
{ |
| 87 |
|
int sad; |
| 88 |
|
dx = (dx >> 1) + roundtab_79[dx & 0x3]; |
| 89 |
|
dy = (dy >> 1) + roundtab_79[dy & 0x3]; |
| 90 |
|
|
| 91 |
|
if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently |
| 92 |
|
|
| 93 |
|
switch (((dx & 1) << 1) | (dy & 1)) { |
| 94 |
|
case 0: |
| 95 |
|
sad = sad8(data->CurU, data->RefCU + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); |
| 96 |
|
sad += sad8(data->CurV, data->RefCV + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); |
| 97 |
|
break; |
| 98 |
|
case 1: |
| 99 |
|
dx = dx / 2; dy = (dy - 1) / 2; |
| 100 |
|
sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); |
| 101 |
|
sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); |
| 102 |
|
break; |
| 103 |
|
case 2: |
| 104 |
|
dx = (dx - 1) / 2; dy = dy / 2; |
| 105 |
|
sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); |
| 106 |
|
sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); |
| 107 |
|
break; |
| 108 |
|
default: |
| 109 |
|
dx = (dx - 1) / 2; dy = (dy - 1) / 2; |
| 110 |
|
interpolate8x8_halfpel_hv(data->RefQ, |
| 111 |
|
data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, |
| 112 |
|
data->rounding); |
| 113 |
|
sad = sad8(data->CurU, data->RefQ, data->iEdgedWidth/2); |
| 114 |
|
interpolate8x8_halfpel_hv(data->RefQ, |
| 115 |
|
data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, |
| 116 |
|
data->rounding); |
| 117 |
|
sad += sad8(data->CurV, data->RefQ, data->iEdgedWidth/2); |
| 118 |
|
break; |
| 119 |
|
} |
| 120 |
|
data->temp[5] = dx; data->temp[6] = dy; data->temp[7] = sad; //backup |
| 121 |
|
return sad; |
| 122 |
|
} |
| 123 |
|
|
| 124 |
|
static __inline const uint8_t * |
| 125 |
|
GetReference(const int x, const int y, const int dir, const SearchData * const data) |
| 126 |
|
{ |
| 127 |
|
// dir : 0 = forward, 1 = backward |
| 128 |
|
switch ( (dir << 2) | ((x&1)<<1) | (y&1) ) { |
| 129 |
|
case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth); |
| 130 |
|
case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); |
| 131 |
|
case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); |
| 132 |
|
case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); |
| 133 |
|
case 4 : return data->bRef + x/2 + (y/2)*(data->iEdgedWidth); |
| 134 |
|
case 5 : return data->bRefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); |
| 135 |
|
case 6 : return data->bRefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); |
| 136 |
|
default : return data->bRefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); |
| 137 |
|
} |
| 138 |
|
} |
| 139 |
|
|
| 140 |
|
static uint8_t * |
| 141 |
|
Interpolate8x8qpel(const int x, const int y, const int block, const int dir, const SearchData * const data) |
| 142 |
|
{ |
| 143 |
|
// create or find a qpel-precision reference picture; return pointer to it |
| 144 |
|
uint8_t * Reference = (uint8_t *)data->RefQ + 16*dir; |
| 145 |
|
const int32_t iEdgedWidth = data->iEdgedWidth; |
| 146 |
|
const uint32_t rounding = data->rounding; |
| 147 |
|
const int halfpel_x = x/2; |
| 148 |
|
const int halfpel_y = y/2; |
| 149 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
| 150 |
|
|
| 151 |
|
ref1 = GetReference(halfpel_x, halfpel_y, dir, data); |
| 152 |
|
ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
| 153 |
|
switch( ((x&1)<<1) + (y&1) ) { |
| 154 |
|
case 0: // pure halfpel position |
| 155 |
|
return (uint8_t *) ref1; |
| 156 |
|
break; |
| 157 |
|
|
| 158 |
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
| 159 |
|
ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
| 160 |
|
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
| 161 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
| 162 |
|
break; |
| 163 |
|
|
| 164 |
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
| 165 |
|
ref2 = GetReference(x - halfpel_x, halfpel_y, dir, data); |
| 166 |
|
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
| 167 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
| 168 |
|
break; |
| 169 |
|
|
| 170 |
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
| 171 |
|
// bottom left/right) during qpel refinement |
| 172 |
|
ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
| 173 |
|
ref3 = GetReference(x - halfpel_x, halfpel_y, dir, data); |
| 174 |
|
ref4 = GetReference(x - halfpel_x, y - halfpel_y, dir, data); |
| 175 |
|
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
| 176 |
|
ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
| 177 |
|
ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
| 178 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
| 179 |
|
break; |
| 180 |
|
} |
| 181 |
|
return Reference; |
| 182 |
|
} |
| 183 |
|
|
| 184 |
|
static uint8_t * |
| 185 |
|
Interpolate16x16qpel(const int x, const int y, const int dir, const SearchData * const data) |
| 186 |
|
{ |
| 187 |
|
// create or find a qpel-precision reference picture; return pointer to it |
| 188 |
|
uint8_t * Reference = (uint8_t *)data->RefQ + 16*dir; |
| 189 |
|
const int32_t iEdgedWidth = data->iEdgedWidth; |
| 190 |
|
const uint32_t rounding = data->rounding; |
| 191 |
|
const int halfpel_x = x/2; |
| 192 |
|
const int halfpel_y = y/2; |
| 193 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
| 194 |
|
|
| 195 |
|
ref1 = GetReference(halfpel_x, halfpel_y, dir, data); |
| 196 |
|
switch( ((x&1)<<1) + (y&1) ) { |
| 197 |
|
case 0: // pure halfpel position |
| 198 |
|
return (uint8_t *) ref1; |
| 199 |
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
| 200 |
|
ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
| 201 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
| 202 |
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
| 203 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); |
| 204 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
| 205 |
|
break; |
| 206 |
|
|
| 207 |
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
| 208 |
|
ref2 = GetReference(x - halfpel_x, halfpel_y, dir, data); |
| 209 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
| 210 |
|
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); |
| 212 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
| 213 |
|
break; |
| 214 |
|
|
| 215 |
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
| 216 |
|
// bottom left/right) during qpel refinement |
| 217 |
|
ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); |
| 218 |
|
ref3 = GetReference(x - halfpel_x, halfpel_y, dir, data); |
| 219 |
|
ref4 = GetReference(x - halfpel_x, y - halfpel_y, dir, data); |
| 220 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
| 221 |
|
interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); |
| 222 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); |
| 223 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); |
| 224 |
|
break; |
| 225 |
|
} |
| 226 |
|
return Reference; |
| 227 |
|
} |
| 228 |
|
|
| 229 |
|
/* CHECK_CANDIATE FUNCTIONS START */ |
| 230 |
|
|
| 231 |
|
static void |
| 232 |
|
CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
| 233 |
|
{ |
| 234 |
|
int t, xc, yc; |
| 235 |
|
const uint8_t * Reference; |
| 236 |
|
VECTOR * current; |
| 237 |
|
|
| 238 |
|
if (( x > data->max_dx) || ( x < data->min_dx) |
| 239 |
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
| 240 |
|
|
| 241 |
|
if (data->qpel_precision) { // x and y are in 1/4 precision |
| 242 |
|
Reference = Interpolate16x16qpel(x, y, 0, data); |
| 243 |
|
xc = x/2; yc = y/2; //for chroma sad |
| 244 |
|
current = data->currentQMV; |
| 245 |
|
} else { |
| 246 |
|
Reference = GetReference(x, y, 0, data); |
| 247 |
|
current = data->currentMV; |
| 248 |
|
xc = x; yc = y; |
| 249 |
|
} |
| 250 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0); |
| 251 |
|
|
| 252 |
|
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
| 253 |
|
|
| 254 |
|
data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; |
| 255 |
|
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; |
| 256 |
|
|
| 257 |
|
if (data->chroma) data->temp[0] += ChromaSAD(xc, yc, data); |
| 258 |
|
|
| 259 |
|
if (data->temp[0] < data->iMinSAD[0]) { |
| 260 |
|
data->iMinSAD[0] = data->temp[0]; |
| 261 |
|
current[0].x = x; current[0].y = y; |
| 262 |
|
*dir = Direction; } |
| 263 |
|
|
| 264 |
|
if (data->temp[1] < data->iMinSAD[1]) { |
| 265 |
|
data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y= y; } |
| 266 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
| 267 |
|
data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; } |
| 268 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
| 269 |
|
data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; } |
| 270 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
| 271 |
|
data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; } |
| 272 |
|
|
| 273 |
|
} |
| 274 |
|
|
| 275 |
|
static void |
| 276 |
|
CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
| 277 |
|
{ |
| 278 |
|
int t; |
| 279 |
|
const uint8_t * Reference; |
| 280 |
|
|
| 281 |
|
if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero integer value |
| 282 |
|
( x > data->max_dx) || ( x < data->min_dx) |
| 283 |
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
| 284 |
|
|
| 285 |
|
Reference = GetReference(x, y, 0, data); |
| 286 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1); |
| 287 |
|
|
| 288 |
|
data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
| 289 |
|
|
| 290 |
|
data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; |
| 291 |
|
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; |
| 292 |
|
|
| 293 |
|
if (data->temp[0] < data->iMinSAD[0]) { |
| 294 |
|
data->iMinSAD[0] = data->temp[0]; |
| 295 |
|
data->currentMV[0].x = x; data->currentMV[0].y = y; |
| 296 |
|
*dir = Direction; } |
| 297 |
|
|
| 298 |
|
if (data->temp[1] < data->iMinSAD[1]) { |
| 299 |
|
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
| 300 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
| 301 |
|
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
| 302 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
| 303 |
|
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
| 304 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
| 305 |
|
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
| 306 |
|
} |
| 307 |
|
|
| 308 |
|
static void |
| 309 |
|
CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
| 310 |
|
{ |
| 311 |
|
int32_t sad; |
| 312 |
|
const uint8_t * Reference; |
| 313 |
|
int t; |
| 314 |
|
VECTOR * current; |
| 315 |
|
|
| 316 |
|
if (( x > data->max_dx) || ( x < data->min_dx) |
| 317 |
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
| 318 |
|
|
| 319 |
|
if (data->rrv) |
| 320 |
|
if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) ) return; //non-zero integer value |
| 321 |
|
|
| 322 |
|
if (data->qpel_precision) { // x and y are in 1/4 precision |
| 323 |
|
Reference = Interpolate16x16qpel(x, y, 0, data); |
| 324 |
|
current = data->currentQMV; |
| 325 |
|
} else { |
| 326 |
|
Reference = GetReference(x, y, 0, data); |
| 327 |
|
current = data->currentMV; |
| 328 |
|
} |
| 329 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, |
| 330 |
|
data->qpel && !data->qpel_precision && !data->rrv, data->rrv); |
| 331 |
|
|
| 332 |
|
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
| 333 |
|
sad += (data->lambda16 * t * sad)/1000; |
| 334 |
|
|
| 335 |
|
if (sad < *(data->iMinSAD)) { |
| 336 |
|
*(data->iMinSAD) = sad; |
| 337 |
|
current->x = x; current->y = y; |
| 338 |
|
*dir = Direction; } |
| 339 |
|
} |
| 340 |
|
|
| 341 |
|
static void |
| 342 |
|
CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
| 343 |
|
{ |
| 344 |
|
// maximum speed - for P/B/I decision |
| 345 |
|
|
| 346 |
|
if (( x > data->max_dx) || ( x < data->min_dx) |
| 347 |
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
| 348 |
|
|
| 349 |
|
data->temp[0] = sad32v_c(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), |
| 350 |
|
data->iEdgedWidth, data->temp+1); |
| 351 |
|
if (data->temp[0] < *(data->iMinSAD)) { |
| 352 |
|
*(data->iMinSAD) = data->temp[0]; |
| 353 |
|
data->currentMV[0].x = x; data->currentMV[0].y = y; |
| 354 |
|
*dir = Direction; } |
| 355 |
|
if (data->temp[1] < data->iMinSAD[1]) { |
| 356 |
|
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
| 357 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
| 358 |
|
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
| 359 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
| 360 |
|
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
| 361 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
| 362 |
|
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
| 363 |
|
|
| 364 |
|
} |
| 365 |
|
|
| 366 |
|
|
| 367 |
|
static void |
| 368 |
|
CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) |
| 369 |
|
{ |
| 370 |
|
int32_t sad; |
| 371 |
|
int xb, yb, t; |
| 372 |
|
const uint8_t *ReferenceF, *ReferenceB; |
| 373 |
|
VECTOR *current; |
| 374 |
|
|
| 375 |
|
if (( xf > data->max_dx) || ( xf < data->min_dx) |
| 376 |
|
|| ( yf > data->max_dy) || (yf < data->min_dy)) return; |
| 377 |
|
|
| 378 |
|
if (data->qpel_precision) { |
| 379 |
|
ReferenceF = Interpolate16x16qpel(xf, yf, 0, data); |
| 380 |
|
xb = data->currentQMV[1].x; yb = data->currentQMV[1].y; |
| 381 |
|
current = data->currentQMV; |
| 382 |
|
ReferenceB = Interpolate16x16qpel(xb, yb, 1, data); |
| 383 |
|
} else { |
| 384 |
|
ReferenceF = GetReference(xf, yf, 0, data); |
| 385 |
|
xb = data->currentMV[1].x; yb = data->currentMV[1].y; |
| 386 |
|
ReferenceB = GetReference(xb, yb, 1, data); |
| 387 |
|
current = data->currentMV; |
| 388 |
|
} |
| 389 |
|
|
| 390 |
|
t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0) |
| 391 |
|
+ d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel && !data->qpel_precision, 0); |
| 392 |
|
|
| 393 |
|
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
| 394 |
|
sad += (data->lambda16 * t * sad)/1000; |
| 395 |
|
|
| 396 |
|
if (sad < *(data->iMinSAD)) { |
| 397 |
|
*(data->iMinSAD) = sad; |
| 398 |
|
current->x = xf; current->y = yf; |
| 399 |
|
*dir = Direction; } |
| 400 |
|
} |
| 401 |
|
|
| 402 |
|
static void |
| 403 |
|
CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
| 404 |
|
{ |
| 405 |
|
int32_t sad = 0; |
| 406 |
|
int k; |
| 407 |
|
const uint8_t *ReferenceF; |
| 408 |
|
const uint8_t *ReferenceB; |
| 409 |
|
VECTOR mvs, b_mvs; |
| 410 |
|
const VECTOR zeroMV={0,0}; |
| 411 |
|
|
| 412 |
|
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
| 413 |
|
|
| 414 |
|
for (k = 0; k < 4; k++) { |
| 415 |
|
mvs.x = data->directmvF[k].x + x; |
| 416 |
|
b_mvs.x = ((x == 0) ? |
| 417 |
|
data->directmvB[k].x |
| 418 |
|
: mvs.x - data->referencemv[k].x); |
| 419 |
|
|
| 420 |
|
mvs.y = data->directmvF[k].y + y; |
| 421 |
|
b_mvs.y = ((y == 0) ? |
| 422 |
|
data->directmvB[k].y |
| 423 |
|
: mvs.y - data->referencemv[k].y); |
| 424 |
|
|
| 425 |
|
if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
| 426 |
|
|| ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) |
| 427 |
|
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
| 428 |
|
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
| 429 |
|
|
| 430 |
|
if (!data->qpel) { |
| 431 |
|
mvs.x *= 2; mvs.y *= 2; |
| 432 |
|
b_mvs.x *= 2; b_mvs.y *= 2; //we move to qpel precision anyway |
| 433 |
|
} |
| 434 |
|
ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); |
| 435 |
|
ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); |
| 436 |
|
|
| 437 |
|
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
| 438 |
|
ReferenceF, ReferenceB, |
| 439 |
|
data->iEdgedWidth); |
| 440 |
|
if (sad > *(data->iMinSAD)) return; |
| 441 |
|
} |
| 442 |
|
|
| 443 |
|
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)/1000; |
| 444 |
|
|
| 445 |
|
if (sad < *(data->iMinSAD)) { |
| 446 |
|
*(data->iMinSAD) = sad; |
| 447 |
|
data->currentMV->x = x; data->currentMV->y = y; |
| 448 |
|
*dir = Direction; } |
| 449 |
|
} |
| 450 |
|
|
| 451 |
|
static void |
| 452 |
|
CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
| 453 |
|
{ |
| 454 |
|
int32_t sad; |
| 455 |
|
const uint8_t *ReferenceF; |
| 456 |
|
const uint8_t *ReferenceB; |
| 457 |
|
VECTOR mvs, b_mvs; |
| 458 |
|
const VECTOR zeroMV = {0,0}; |
| 459 |
|
|
| 460 |
|
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
| 461 |
|
|
| 462 |
|
mvs.x = data->directmvF[0].x + x; |
| 463 |
|
b_mvs.x = ((x == 0) ? |
| 464 |
|
data->directmvB[0].x |
| 465 |
|
: mvs.x - data->referencemv[0].x); |
| 466 |
|
|
| 467 |
|
mvs.y = data->directmvF[0].y + y; |
| 468 |
|
b_mvs.y = ((y == 0) ? |
| 469 |
|
data->directmvB[0].y |
| 470 |
|
: mvs.y - data->referencemv[0].y); |
| 471 |
|
|
| 472 |
|
if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
| 473 |
|
|| ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) |
| 474 |
|
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
| 475 |
|
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
| 476 |
|
|
| 477 |
|
if (!data->qpel) { |
| 478 |
|
mvs.x *= 2; mvs.y *= 2; |
| 479 |
|
b_mvs.x *= 2; b_mvs.y *= 2; //we move to qpel precision anyway |
| 480 |
|
} |
| 481 |
|
ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data); |
| 482 |
|
ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); |
| 483 |
|
|
| 484 |
|
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
| 485 |
|
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)/1000; |
| 486 |
|
|
| 487 |
|
if (sad < *(data->iMinSAD)) { |
| 488 |
|
*(data->iMinSAD) = sad; |
| 489 |
|
data->currentMV->x = x; data->currentMV->y = y; |
| 490 |
|
*dir = Direction; } |
| 491 |
|
} |
| 492 |
|
|
| 493 |
|
static void |
| 494 |
|
CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
| 495 |
|
{ |
| 496 |
|
int32_t sad; int t; |
| 497 |
|
const uint8_t * Reference; |
| 498 |
|
|
| 499 |
|
if (( x > data->max_dx) || ( x < data->min_dx) |
| 500 |
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
| 501 |
|
|
| 502 |
|
if (data->qpel) Reference = Interpolate16x16qpel(x, y, 0, data); |
| 503 |
|
else Reference = GetReference(x, y, 0, data); |
| 504 |
|
|
| 505 |
|
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
| 506 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0); |
| 507 |
|
|
| 508 |
|
sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))/100; |
| 509 |
|
|
| 510 |
|
if (sad < *(data->iMinSAD)) { |
| 511 |
|
*(data->iMinSAD) = sad; |
| 512 |
|
data->currentMV->x = x; data->currentMV->y = y; |
| 513 |
|
*dir = Direction; } |
| 514 |
|
} |
| 515 |
|
|
| 516 |
|
/* CHECK_CANDIATE FUNCTIONS END */ |
| 517 |
|
|
| 518 |
|
/* MAINSEARCH FUNCTIONS START */ |
| 519 |
|
|
| 520 |
|
static void |
| 521 |
|
AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection) |
| 522 |
|
{ |
| 523 |
|
|
| 524 |
|
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
| 525 |
|
|
| 526 |
|
int iDirection; |
| 527 |
|
|
| 528 |
|
for(;;) { //forever |
| 529 |
|
iDirection = 0; |
| 530 |
|
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
| 531 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
| 532 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
| 533 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
| 534 |
|
|
| 535 |
|
/* now we're doing diagonal checks near our candidate */ |
| 536 |
|
|
| 537 |
|
if (iDirection) { //checking if anything found |
| 538 |
|
bDirection = iDirection; |
| 539 |
|
iDirection = 0; |
| 540 |
|
x = data->currentMV->x; y = data->currentMV->y; |
| 541 |
|
if (bDirection & 3) { //our candidate is left or right |
| 542 |
|
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
| 543 |
|
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
| 544 |
|
} else { // what remains here is up or down |
| 545 |
|
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
| 546 |
|
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
| 547 |
|
} |
| 548 |
|
|
| 549 |
|
if (iDirection) { |
| 550 |
|
bDirection += iDirection; |
| 551 |
|
x = data->currentMV->x; y = data->currentMV->y; |
| 552 |
|
} |
| 553 |
|
} else { //about to quit, eh? not so fast.... |
| 554 |
|
switch (bDirection) { |
| 555 |
|
case 2: |
| 556 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
| 557 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
| 558 |
|
break; |
| 559 |
|
case 1: |
| 560 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
| 561 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
| 562 |
|
break; |
| 563 |
|
case 2 + 4: |
| 564 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
| 565 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
| 566 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
| 567 |
|
break; |
| 568 |
|
case 4: |
| 569 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
| 570 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
| 571 |
|
break; |
| 572 |
|
case 8: |
| 573 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
| 574 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
| 575 |
|
break; |
| 576 |
|
case 1 + 4: |
| 577 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
| 578 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
| 579 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
| 580 |
|
break; |
| 581 |
|
case 2 + 8: |
| 582 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
| 583 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
| 584 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
| 585 |
|
break; |
| 586 |
|
case 1 + 8: |
| 587 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
| 588 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
| 589 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
| 590 |
|
break; |
| 591 |
|
default: //1+2+4+8 == we didn't find anything at all |
| 592 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
| 593 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
| 594 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
| 595 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
| 596 |
|
break; |
| 597 |
|
} |
| 598 |
|
if (!iDirection) break; //ok, the end. really |
| 599 |
|
bDirection = iDirection; |
| 600 |
|
x = data->currentMV->x; y = data->currentMV->y; |
| 601 |
|
} |
| 602 |
|
} |
| 603 |
|
} |
| 604 |
|
|
| 605 |
|
static void |
| 606 |
|
SquareSearch(int x, int y, const SearchData * const data, int bDirection) |
| 607 |
|
{ |
| 608 |
|
int iDirection; |
| 609 |
|
|
| 610 |
|
do { |
| 611 |
|
iDirection = 0; |
| 612 |
|
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64); |
| 613 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128); |
| 614 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32); |
| 615 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128); |
| 616 |
|
if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64); |
| 617 |
|
if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128); |
| 618 |
|
if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128); |
| 619 |
|
if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128); |
| 620 |
|
|
| 621 |
|
bDirection = iDirection; |
| 622 |
|
x = data->currentMV->x; y = data->currentMV->y; |
| 623 |
|
} while (iDirection); |
| 624 |
|
} |
| 625 |
|
|
| 626 |
|
static void |
| 627 |
|
DiamondSearch(int x, int y, const SearchData * const data, int bDirection) |
| 628 |
|
{ |
| 629 |
|
|
| 630 |
|
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
| 631 |
|
|
| 632 |
|
int iDirection; |
| 633 |
|
|
| 634 |
|
do { |
| 635 |
|
iDirection = 0; |
| 636 |
|
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
| 637 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
| 638 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
| 639 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
| 640 |
|
|
| 641 |
|
/* now we're doing diagonal checks near our candidate */ |
| 642 |
|
|
| 643 |
|
if (iDirection) { //checking if anything found |
| 644 |
|
bDirection = iDirection; |
| 645 |
|
iDirection = 0; |
| 646 |
|
x = data->currentMV->x; y = data->currentMV->y; |
| 647 |
|
if (bDirection & 3) { //our candidate is left or right |
| 648 |
|
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
| 649 |
|
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
| 650 |
|
} else { // what remains here is up or down |
| 651 |
|
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
| 652 |
|
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
| 653 |
|
} |
| 654 |
|
bDirection += iDirection; |
| 655 |
|
x = data->currentMV->x; y = data->currentMV->y; |
| 656 |
|
} |
| 657 |
|
} |
| 658 |
|
while (iDirection); |
| 659 |
|
} |
| 660 |
|
|
| 661 |
/* vector map (vlc delta size) smoother parameters */ |
/* MAINSEARCH FUNCTIONS END */ |
|
#define NEIGH_TEND_16X16 2 |
|
|
#define NEIGH_TEND_8X8 2 |
|
| 662 |
|
|
| 663 |
|
/* HALFPELREFINE COULD BE A MAINSEARCH FUNCTION, BUT THERE IS NO NEED FOR IT */ |
| 664 |
|
|
| 665 |
// fast ((A)/2)*2 |
static void |
| 666 |
#define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) |
SubpelRefine(const SearchData * const data) |
| 667 |
|
{ |
| 668 |
|
/* Do a half-pel or q-pel refinement */ |
| 669 |
|
VECTOR backupMV; |
| 670 |
|
int iDirection; //not needed |
| 671 |
|
|
| 672 |
|
if (data->qpel_precision) |
| 673 |
|
backupMV = *(data->currentQMV); |
| 674 |
|
else backupMV = *(data->currentMV); |
| 675 |
|
|
| 676 |
#define MIN(X, Y) ((X)<(Y)?(X):(Y)) |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); |
| 677 |
#define MAX(X, Y) ((X)>(Y)?(X):(Y)) |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); |
| 678 |
#define ABS(X) (((X)>0)?(X):-(X)) |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); |
| 679 |
#define SIGN(X) (((X)>0)?1:-1) |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0); |
| 680 |
|
|
| 681 |
|
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); |
| 682 |
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); |
| 683 |
|
|
| 684 |
int32_t PMVfastSearch8( |
CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); |
| 685 |
const uint8_t * const pRef, |
CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); |
| 686 |
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 int start_x, int start_y, |
|
|
const uint32_t iQuality, |
|
|
MBParam * const pParam, |
|
|
MACROBLOCK * const pMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV); |
|
| 687 |
|
|
| 688 |
int32_t PMVfastSearch16( |
static __inline int |
| 689 |
const uint8_t * const pRef, |
SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const IMAGE * const pCur, |
|
| 690 |
const int x, const int y, |
const int x, const int y, |
| 691 |
const uint32_t iQuality, |
const uint32_t iEdgedWidth, const uint32_t iQuant, int rrv) |
|
MBParam * const pParam, |
|
|
MACROBLOCK * const pMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV); |
|
|
|
|
|
|
|
|
|
|
|
/* diamond search stuff |
|
|
keep the the sequence in circular order (so optimization works) |
|
|
*/ |
|
| 692 |
|
|
|
typedef struct |
|
| 693 |
{ |
{ |
| 694 |
int32_t dx; |
/* keep repeating checks for all b-frames before this P frame, |
| 695 |
int32_t dy; |
to make sure that SKIP is possible (todo) |
| 696 |
|
how: if skip is not possible set sad00 to a very high value */ |
| 697 |
|
if(rrv) { |
| 698 |
|
uint32_t sadC = sad16(current->u + x*16 + y*(iEdgedWidth/2)*16, |
| 699 |
|
reference->u + x*16 + y*(iEdgedWidth/2)*16, iEdgedWidth/2, 256*4096); |
| 700 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
| 701 |
|
sadC += sad16(current->v + (x + y*(iEdgedWidth/2))*16, |
| 702 |
|
reference->v + (x + y*(iEdgedWidth/2))*16, iEdgedWidth/2, 256*4096); |
| 703 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
| 704 |
|
return 1; |
| 705 |
|
} else { |
| 706 |
|
uint32_t sadC = sad8(current->u + x*8 + y*(iEdgedWidth/2)*8, |
| 707 |
|
reference->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); |
| 708 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
| 709 |
|
sadC += sad8(current->v + (x + y*(iEdgedWidth/2))*8, |
| 710 |
|
reference->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); |
| 711 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
| 712 |
|
return 1; |
| 713 |
|
} |
| 714 |
} |
} |
|
DPOINT; |
|
|
|
|
|
|
|
|
static const DPOINT diamond_small[4] = |
|
|
{ |
|
|
{0, 1}, {1, 0}, {0, -1}, {-1, 0} |
|
|
}; |
|
|
|
|
| 715 |
|
|
| 716 |
static const DPOINT diamond_large[8] = |
static __inline void |
| 717 |
|
SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
| 718 |
{ |
{ |
| 719 |
{0, 2}, {1, 1}, {2, 0}, {1, -1}, {0, -2}, {-1, -1}, {-2, 0}, {-1, 1} |
pMB->mode = MODE_NOT_CODED; |
| 720 |
}; |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
| 721 |
|
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; |
|
|
|
|
// mv.length table |
|
|
static const uint32_t mvtab[33] = { |
|
|
1, 2, 3, 4, 6, 7, 7, 7, |
|
|
9, 9, 9, 10, 10, 10, 10, 10, |
|
|
10, 10, 10, 10, 10, 10, 10, 10, |
|
|
10, 11, 11, 11, 11, 11, 11, 12, 12 |
|
|
}; |
|
|
|
|
| 722 |
|
|
| 723 |
static __inline uint32_t mv_bits(int32_t component, const uint32_t iFcode) |
pMB->qmvs[0].x = pMB->qmvs[1].x = pMB->qmvs[2].x = pMB->qmvs[3].x = 0; |
| 724 |
{ |
pMB->qmvs[0].y = pMB->qmvs[1].y = pMB->qmvs[2].y = pMB->qmvs[3].y = 0; |
|
if (component == 0) |
|
|
return 1; |
|
| 725 |
|
|
| 726 |
if (component < 0) |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
| 727 |
component = -component; |
} |
| 728 |
|
|
| 729 |
if (iFcode == 1) |
bool |
| 730 |
|
MotionEstimation(MBParam * const pParam, |
| 731 |
|
FRAMEINFO * const current, |
| 732 |
|
FRAMEINFO * const reference, |
| 733 |
|
const IMAGE * const pRefH, |
| 734 |
|
const IMAGE * const pRefV, |
| 735 |
|
const IMAGE * const pRefHV, |
| 736 |
|
const uint32_t iLimit) |
| 737 |
{ |
{ |
| 738 |
if (component > 32) |
MACROBLOCK *const pMBs = current->mbs; |
| 739 |
component = 32; |
const IMAGE *const pCurrent = ¤t->image; |
| 740 |
|
const IMAGE *const pRef = &reference->image; |
| 741 |
return mvtab[component] + 1; |
|
| 742 |
|
const VECTOR zeroMV = { 0, 0 }; |
| 743 |
|
|
| 744 |
|
uint32_t mb_width = pParam->mb_width; |
| 745 |
|
uint32_t mb_height = pParam->mb_height; |
| 746 |
|
|
| 747 |
|
uint32_t x, y; |
| 748 |
|
uint32_t iIntra = 0; |
| 749 |
|
int32_t InterBias, quant = current->quant, sad00; |
| 750 |
|
uint8_t *qimage; |
| 751 |
|
|
| 752 |
|
// some pre-initialized thingies for SearchP |
| 753 |
|
int32_t temp[8]; |
| 754 |
|
VECTOR currentMV[5]; |
| 755 |
|
VECTOR currentQMV[5]; |
| 756 |
|
int32_t iMinSAD[5]; |
| 757 |
|
SearchData Data; |
| 758 |
|
memset(&Data, 0, sizeof(SearchData)); |
| 759 |
|
Data.iEdgedWidth = pParam->edged_width; |
| 760 |
|
Data.currentMV = currentMV; |
| 761 |
|
Data.currentQMV = currentQMV; |
| 762 |
|
Data.iMinSAD = iMinSAD; |
| 763 |
|
Data.temp = temp; |
| 764 |
|
Data.iFcode = current->fcode; |
| 765 |
|
Data.rounding = pParam->m_rounding_type; |
| 766 |
|
Data.qpel = pParam->m_quarterpel; |
| 767 |
|
Data.chroma = current->global_flags & XVID_ME_COLOUR; |
| 768 |
|
Data.rrv = current->global_flags & XVID_REDUCED; |
| 769 |
|
|
| 770 |
|
if ((current->global_flags & XVID_REDUCED)) { |
| 771 |
|
mb_width = (pParam->width + 31) / 32; |
| 772 |
|
mb_height = (pParam->height + 31) / 32; |
| 773 |
|
Data.qpel = Data.chroma = 0; |
| 774 |
|
} |
| 775 |
|
|
| 776 |
|
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
| 777 |
|
return 1; // allocate some mem for qpel interpolated blocks |
| 778 |
|
// somehow this is dirty since I think we shouldn't use malloc outside |
| 779 |
|
// encoder_create() - so please fix me! |
| 780 |
|
Data.RefQ = qimage; |
| 781 |
|
if (sadInit) (*sadInit) (); |
| 782 |
|
|
| 783 |
|
for (y = 0; y < mb_height; y++) { |
| 784 |
|
for (x = 0; x < mb_width; x++) { |
| 785 |
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
| 786 |
|
|
| 787 |
|
if (Data.rrv) pMB->sad16 = |
| 788 |
|
sad32v_c(pCurrent->y + (x + y * pParam->edged_width) * 32, |
| 789 |
|
pRef->y + (x + y * pParam->edged_width) * 32, |
| 790 |
|
pParam->edged_width, pMB->sad8 ); |
| 791 |
|
|
| 792 |
|
else pMB->sad16 = |
| 793 |
|
sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, |
| 794 |
|
pRef->y + (x + y * pParam->edged_width) * 16, |
| 795 |
|
pParam->edged_width, pMB->sad8 ); |
| 796 |
|
|
| 797 |
|
if (Data.chroma) { |
| 798 |
|
pMB->sad16 += sad8(pCurrent->u + x*8 + y*(pParam->edged_width/2)*8, |
| 799 |
|
pRef->u + x*8 + y*(pParam->edged_width/2)*8, pParam->edged_width/2); |
| 800 |
|
|
| 801 |
|
pMB->sad16 += sad8(pCurrent->v + (x + y*(pParam->edged_width/2))*8, |
| 802 |
|
pRef->v + (x + y*(pParam->edged_width/2))*8, pParam->edged_width/2); |
| 803 |
|
} |
| 804 |
|
|
| 805 |
|
sad00 = pMB->sad16; //if no gmc; else sad00 = (..) |
| 806 |
|
|
| 807 |
|
if (!(current->global_flags & XVID_LUMIMASKING)) { |
| 808 |
|
pMB->dquant = NO_CHANGE; |
| 809 |
|
pMB->quant = current->quant; |
| 810 |
|
} else { |
| 811 |
|
if (pMB->dquant != NO_CHANGE) { |
| 812 |
|
quant += DQtab[pMB->dquant]; |
| 813 |
|
if (quant > 31) quant = 31; |
| 814 |
|
else if (quant < 1) quant = 1; |
| 815 |
|
} |
| 816 |
|
pMB->quant = quant; |
| 817 |
|
} |
| 818 |
|
|
| 819 |
|
//initial skip decision |
| 820 |
|
/* no early skip for GMC (global vector = skip vector is unknown!) */ |
| 821 |
|
if (current->coding_type == P_VOP) { /* no fast SKIP for S(GMC)-VOPs */ |
| 822 |
|
if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH * (Data.rrv ? 4:1) ) |
| 823 |
|
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant, Data.rrv)) { |
| 824 |
|
SkipMacroblockP(pMB, sad00); |
| 825 |
|
continue; |
| 826 |
} |
} |
|
|
|
|
component += (1 << (iFcode - 1)) - 1; |
|
|
component >>= (iFcode - 1); |
|
|
|
|
|
if (component > 32) |
|
|
component = 32; |
|
|
|
|
|
return mvtab[component] + 1 + iFcode - 1; |
|
| 827 |
} |
} |
| 828 |
|
|
| 829 |
|
SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
| 830 |
static __inline uint32_t calc_delta_16(const int32_t dx, const int32_t dy, const uint32_t iFcode) |
y, current->motion_flags, pMB->quant, |
| 831 |
{ |
&Data, pParam, pMBs, reference->mbs, |
| 832 |
return NEIGH_TEND_16X16 * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); |
current->global_flags & XVID_INTER4V, pMB); |
| 833 |
|
|
| 834 |
|
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
| 835 |
|
if (current->coding_type == P_VOP) { |
| 836 |
|
if ( (pMB->dquant == NO_CHANGE) && (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
| 837 |
|
&& ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1)) ) |
| 838 |
|
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant, Data.rrv)) { |
| 839 |
|
SkipMacroblockP(pMB, sad00); |
| 840 |
|
continue; |
| 841 |
} |
} |
|
|
|
|
static __inline uint32_t calc_delta_8(const int32_t dx, const int32_t dy, const uint32_t iFcode) |
|
|
|
|
|
{ |
|
|
return NEIGH_TEND_8X8 * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); |
|
| 842 |
} |
} |
| 843 |
|
|
| 844 |
|
/* finally, intra decision */ |
| 845 |
|
|
| 846 |
|
InterBias = MV16_INTER_BIAS; |
| 847 |
|
if (pMB->quant > 8) InterBias += 100 * (pMB->quant - 8); // to make high quants work |
| 848 |
|
if (y != 0) |
| 849 |
|
if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; |
| 850 |
|
if (x != 0) |
| 851 |
|
if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; |
| 852 |
|
|
| 853 |
|
if (Data.chroma) InterBias += 50; // to compensate bigger SAD |
| 854 |
|
if (Data.rrv) InterBias *= 4; //?? |
| 855 |
|
|
| 856 |
/* calculate the min/max range (in halfpixels) |
if (InterBias < pMB->sad16) { |
| 857 |
relative to the _MACROBLOCK_ position |
int32_t deviation; |
| 858 |
*/ |
if (Data.rrv) { |
| 859 |
|
deviation = dev16(pCurrent->y + (x + y * pParam->edged_width) * 32, |
| 860 |
|
pParam->edged_width) |
| 861 |
|
+ dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16, |
| 862 |
|
pParam->edged_width) |
| 863 |
|
+ dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16 * pParam->edged_width, |
| 864 |
|
pParam->edged_width) |
| 865 |
|
+ dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16 * (pParam->edged_width+1), |
| 866 |
|
pParam->edged_width); |
| 867 |
|
} else |
| 868 |
|
deviation = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
| 869 |
|
pParam->edged_width); |
| 870 |
|
|
| 871 |
static void __inline get_range( |
if (deviation < (pMB->sad16 - InterBias)) { |
| 872 |
int32_t * const min_dx, int32_t * const max_dx, |
if (++iIntra >= iLimit) { free(qimage); return 1; } |
| 873 |
int32_t * const min_dy, int32_t * const max_dy, |
pMB->mode = MODE_INTRA; |
| 874 |
const uint32_t x, const uint32_t y, |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = |
| 875 |
const uint32_t block_sz, // block dimension, 8 or 16 |
pMB->mvs[3] = zeroMV; |
| 876 |
const uint32_t width, const uint32_t height, |
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = |
| 877 |
const uint32_t fcode) |
pMB->qmvs[3] = zeroMV; |
| 878 |
{ |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = |
| 879 |
const int search_range = 32 << (fcode - 1); |
pMB->sad8[3] = 0; |
| 880 |
const int high = search_range - 1; |
} |
|
const int low = -search_range; |
|
|
|
|
|
// convert full-pixel measurements to half pixel |
|
|
const int hp_width = 2 * width; |
|
|
const int hp_height = 2 * height; |
|
|
const int hp_edge = 2 * block_sz; |
|
|
const int hp_x = 2 * (x) * block_sz; // we need _right end_ of block, not x-coordinate |
|
|
const int hp_y = 2 * (y) * block_sz; // same for _bottom end_ |
|
|
|
|
|
*max_dx = MIN(high, hp_width - hp_x); |
|
|
*max_dy = MIN(high, hp_height - hp_y); |
|
|
*min_dx = MAX(low, -(hp_edge + hp_x)); |
|
|
*min_dy = MAX(low, -(hp_edge + hp_y)); |
|
| 881 |
} |
} |
| 882 |
|
} |
| 883 |
|
} |
| 884 |
|
free(qimage); |
| 885 |
|
|
| 886 |
|
if (current->coding_type == S_VOP) /* first GMC step only for S(GMC)-VOPs */ |
| 887 |
|
current->GMC_MV = GlobalMotionEst( pMBs, pParam, current->fcode ); |
| 888 |
|
else |
| 889 |
|
current->GMC_MV = zeroMV; |
| 890 |
|
|
| 891 |
/* getref: calculate reference image pointer |
return 0; |
|
the decision to use interpolation h/v/hv or the normal image is |
|
|
based on dx & dy. |
|
|
*/ |
|
|
|
|
|
static __inline const uint8_t * get_ref( |
|
|
const uint8_t * const refn, |
|
|
const uint8_t * const refh, |
|
|
const uint8_t * const refv, |
|
|
const uint8_t * const refhv, |
|
|
const uint32_t x, const uint32_t y, |
|
|
const uint32_t block, // block dimension, 8 or 16 |
|
|
const int32_t dx, const int32_t dy, |
|
|
const uint32_t stride) |
|
|
{ |
|
|
switch ( ((dx&1)<<1) + (dy&1) ) // ((dx%2)?2:0)+((dy%2)?1:0) |
|
|
{ |
|
|
case 0 : return refn + (x*block+dx/2) + (y*block+dy/2)*stride; |
|
|
case 1 : return refv + (x*block+dx/2) + (y*block+(dy-1)/2)*stride; |
|
|
case 2 : return refh + (x*block+(dx-1)/2) + (y*block+dy/2)*stride; |
|
|
default : |
|
|
case 3 : return refhv + (x*block+(dx-1)/2) + (y*block+(dy-1)/2)*stride; |
|
|
} |
|
| 892 |
} |
} |
| 893 |
|
|
| 894 |
|
|
| 895 |
/* This is somehow a copy of get_ref, but with MV instead of X,Y */ |
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
| 896 |
|
|
| 897 |
static __inline const uint8_t * get_ref_mv( |
static __inline int |
| 898 |
const uint8_t * const refn, |
make_mask(const VECTOR * const pmv, const int i) |
| 899 |
const uint8_t * const refh, |
{ |
| 900 |
const uint8_t * const refv, |
int mask = 255, j; |
| 901 |
const uint8_t * const refhv, |
for (j = 0; j < i; j++) { |
| 902 |
const uint32_t x, const uint32_t y, |
if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already |
| 903 |
const uint32_t block, // block dimension, 8 or 16 |
if (pmv[i].x == pmv[j].x) { |
| 904 |
const VECTOR* mv, // measured in half-pel! |
if (pmv[i].y == pmv[j].y + iDiamondSize) { mask &= ~4; continue; } |
| 905 |
const uint32_t stride) |
if (pmv[i].y == pmv[j].y - iDiamondSize) { mask &= ~8; continue; } |
| 906 |
{ |
} else |
| 907 |
switch ( (((mv->x)&1)<<1) + ((mv->y)&1) ) |
if (pmv[i].y == pmv[j].y) { |
| 908 |
{ |
if (pmv[i].x == pmv[j].x + iDiamondSize) { mask &= ~1; continue; } |
| 909 |
case 0 : return refn + (x*block+(mv->x)/2) + (y*block+(mv->y)/2)*stride; |
if (pmv[i].x == pmv[j].x - iDiamondSize) { mask &= ~2; continue; } |
| 910 |
case 1 : return refv + (x*block+(mv->x)/2) + (y*block+((mv->y)-1)/2)*stride; |
} |
|
case 2 : return refh + (x*block+((mv->x)-1)/2) + (y*block+(mv->y)/2)*stride; |
|
|
default : |
|
|
case 3 : return refhv + (x*block+((mv->x)-1)/2) + (y*block+((mv->y)-1)/2)*stride; |
|
| 911 |
} |
} |
| 912 |
|
return mask; |
| 913 |
} |
} |
| 914 |
|
|
| 915 |
#ifndef SEARCH16 |
static __inline void |
| 916 |
#define SEARCH16 PMVfastSearch16 |
PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, |
| 917 |
#endif |
int iHcount, const MACROBLOCK * const prevMB, int rrv) |
| 918 |
|
{ |
| 919 |
|
|
| 920 |
#ifndef SEARCH8 |
//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself |
| 921 |
#define SEARCH8 PMVfastSearch8 |
if (rrv) { iWcount /= 2; iHcount /= 2; } |
|
#endif |
|
| 922 |
|
|
| 923 |
bool MotionEstimation( |
if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour |
| 924 |
MACROBLOCK * const pMBs, |
pmv[5].x = EVEN(pmv[3].x); |
| 925 |
MBParam * const pParam, |
pmv[5].y = EVEN(pmv[3].y); |
| 926 |
const IMAGE * const pRef, |
} else pmv[5].x = pmv[5].y = 0; |
|
const IMAGE * const pRefH, |
|
|
const IMAGE * const pRefV, |
|
|
const IMAGE * const pRefHV, |
|
|
IMAGE * const pCurrent, |
|
|
const uint32_t iLimit) |
|
| 927 |
|
|
| 928 |
{ |
if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour |
| 929 |
const uint32_t iWcount = pParam->mb_width; |
else pmv[3].x = pmv[3].y = 0; |
|
const uint32_t iHcount = pParam->mb_height; |
|
| 930 |
|
|
| 931 |
uint32_t i, j, iIntra = 0; |
if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour |
| 932 |
|
else pmv[4].x = pmv[4].y = 0; |
| 933 |
|
|
| 934 |
VECTOR mv16; |
// [1] median prediction |
| 935 |
VECTOR pmv16; |
if (rrv) { //median is in halfzero-precision |
| 936 |
|
pmv[1].x = RRV_MV_SCALEUP(pmv[0].x); |
| 937 |
|
pmv[1].y = RRV_MV_SCALEUP(pmv[0].y); |
| 938 |
|
} else { pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); } |
| 939 |
|
|
| 940 |
int32_t sad8 = 0; |
pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask |
|
int32_t sad16; |
|
|
int32_t deviation; |
|
| 941 |
|
|
| 942 |
// note: i==horizontal, j==vertical |
pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
| 943 |
for (i = 0; i < iHcount; i++) |
pmv[2].y = EVEN(prevMB->mvs[0].y); |
|
for (j = 0; j < iWcount; j++) |
|
|
{ |
|
|
MACROBLOCK *pMB = &pMBs[j + i * iWcount]; |
|
| 944 |
|
|
| 945 |
sad16 = SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
if ((x < iWcount-1) && (y < iHcount-1)) { |
| 946 |
j, i, pParam->motion_flags, |
pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame |
| 947 |
pParam, pMBs, &mv16, &pmv16); |
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
| 948 |
pMB->sad16=sad16; |
} else pmv[6].x = pmv[6].y = 0; |
| 949 |
|
|
| 950 |
|
if (rrv) { |
| 951 |
|
int i; |
| 952 |
|
for (i = 0; i < 7; i++) { |
| 953 |
|
pmv[i].x = RRV_MV_SCALEDOWN(pmv[i].x); |
| 954 |
|
pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); // a trick |
| 955 |
|
} |
| 956 |
|
} |
| 957 |
|
} |
| 958 |
|
|
| 959 |
/* decide: MODE_INTER or MODE_INTRA |
static void |
| 960 |
if (dev_intra < sad_inter - 2 * nb) use_intra |
SearchP(const IMAGE * const pRef, |
| 961 |
*/ |
const uint8_t * const pRefH, |
| 962 |
|
const uint8_t * const pRefV, |
| 963 |
|
const uint8_t * const pRefHV, |
| 964 |
|
const IMAGE * const pCur, |
| 965 |
|
const int x, |
| 966 |
|
const int y, |
| 967 |
|
const uint32_t MotionFlags, |
| 968 |
|
const uint32_t iQuant, |
| 969 |
|
SearchData * const Data, |
| 970 |
|
const MBParam * const pParam, |
| 971 |
|
const MACROBLOCK * const pMBs, |
| 972 |
|
const MACROBLOCK * const prevMBs, |
| 973 |
|
int inter4v, |
| 974 |
|
MACROBLOCK * const pMB) |
| 975 |
|
{ |
| 976 |
|
|
| 977 |
|
int i, iDirection = 255, mask, threshA; |
| 978 |
|
VECTOR pmv[7]; |
| 979 |
|
|
| 980 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
| 981 |
|
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
| 982 |
|
|
| 983 |
|
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); //has to be changed to get_pmv(2)() |
| 984 |
|
|
| 985 |
|
Data->temp[5] = Data->temp[7] = 256*4096; // to reset chroma-sad cache |
| 986 |
|
if (Data->rrv) i = 2; else i = 1; |
| 987 |
|
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i; |
| 988 |
|
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
| 989 |
|
Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
| 990 |
|
|
| 991 |
|
Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16*i; |
| 992 |
|
Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16*i; |
| 993 |
|
Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16*i; |
| 994 |
|
Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16*i; |
| 995 |
|
Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
| 996 |
|
Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
| 997 |
|
|
| 998 |
|
Data->lambda16 = lambda_vec16[iQuant]; |
| 999 |
|
Data->lambda8 = lambda_vec8[iQuant]; |
| 1000 |
|
Data->qpel_precision = 0; |
| 1001 |
|
|
| 1002 |
|
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
| 1003 |
|
|
| 1004 |
|
for(i = 0; i < 5; i++) |
| 1005 |
|
Data->currentMV[i].x = Data->currentMV[i].y = 0; |
| 1006 |
|
|
| 1007 |
|
if (pParam->m_quarterpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
| 1008 |
|
else Data->predMV = pmv[0]; |
| 1009 |
|
|
| 1010 |
|
i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0); |
| 1011 |
|
Data->iMinSAD[0] = pMB->sad16 + (Data->lambda16 * i * pMB->sad16)/1000; |
| 1012 |
|
Data->iMinSAD[1] = pMB->sad8[0] + (Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS))/100; |
| 1013 |
|
Data->iMinSAD[2] = pMB->sad8[1]; |
| 1014 |
|
Data->iMinSAD[3] = pMB->sad8[2]; |
| 1015 |
|
Data->iMinSAD[4] = pMB->sad8[3]; |
| 1016 |
|
|
| 1017 |
|
if ((x == 0) && (y == 0)) threshA = 512; |
| 1018 |
|
else { |
| 1019 |
|
threshA = Data->temp[0]; // that's when we keep this SAD atm |
| 1020 |
|
if (threshA < 512) threshA = 512; |
| 1021 |
|
if (threshA > 1024) threshA = 1024; } |
| 1022 |
|
|
| 1023 |
deviation = dev16(pCurrent->y + j*16 + i*16*pParam->edged_width, pParam->edged_width); |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
| 1024 |
|
prevMBs + x + y * pParam->mb_width, Data->rrv); |
| 1025 |
|
|
| 1026 |
if (deviation < (sad16 - INTER_BIAS)) |
if (Data->rrv) CheckCandidate = CheckCandidate32; |
| 1027 |
{ |
else if (inter4v || Data->chroma) CheckCandidate = CheckCandidate16; |
| 1028 |
pMB->mode = MODE_INTRA; |
else CheckCandidate = CheckCandidate16no4v; //for extra speed |
|
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
|
|
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; |
|
| 1029 |
|
|
| 1030 |
iIntra++; |
/* main loop. checking all predictions */ |
|
if(iIntra >= iLimit) |
|
|
return 1; |
|
| 1031 |
|
|
| 1032 |
continue; |
for (i = 1; i < 7; i++) { |
| 1033 |
|
if (!(mask = make_mask(pmv, i)) ) continue; |
| 1034 |
|
(*CheckCandidate)(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
| 1035 |
|
if (Data->iMinSAD[0] <= threshA) break; |
| 1036 |
} |
} |
| 1037 |
|
|
| 1038 |
if (pParam->global_flags & XVID_INTER4V) |
if ((Data->iMinSAD[0] <= threshA) || |
| 1039 |
{ |
(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
| 1040 |
pMB->sad8[0] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { |
| 1041 |
2 * j, 2 * i, mv16.x, mv16.y, pParam->motion_flags, |
inter4v = 0; |
| 1042 |
pParam, pMBs, &pMB->mvs[0], &pMB->pmvs[0]); |
} else { |
| 1043 |
|
|
| 1044 |
pMB->sad8[1] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
MainSearchFunc * MainSearchPtr; |
| 1045 |
2 * j + 1, 2 * i, mv16.x, mv16.y, pParam->motion_flags, |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
| 1046 |
pParam, pMBs, &pMB->mvs[1], &pMB->pmvs[1]); |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
| 1047 |
|
else MainSearchPtr = DiamondSearch; |
| 1048 |
|
|
| 1049 |
pMB->sad8[2] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
|
2 * j, 2 * i + 1, mv16.x, mv16.y, pParam->motion_flags, |
|
|
pParam, pMBs, &pMB->mvs[2], &pMB->pmvs[2]); |
|
| 1050 |
|
|
| 1051 |
pMB->sad8[3] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
/* extended search, diamond starting in 0,0 and in prediction. |
| 1052 |
2 * j + 1, 2 * i + 1, mv16.x, mv16.y, pParam->motion_flags, |
note that this search is/might be done in halfpel positions, |
| 1053 |
pParam, pMBs, &pMB->mvs[3], &pMB->pmvs[3]); |
which makes it more different than the diamond above */ |
|
|
|
|
sad8 = pMB->sad8[0] + pMB->sad8[1] + pMB->sad8[2] + pMB->sad8[3]; |
|
|
} |
|
| 1054 |
|
|
| 1055 |
|
if (MotionFlags & PMV_EXTSEARCH16) { |
| 1056 |
|
int32_t bSAD; |
| 1057 |
|
VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; |
| 1058 |
|
if (Data->rrv) { |
| 1059 |
|
startMV.x = RRV_MV_SCALEUP(startMV.x); |
| 1060 |
|
startMV.y = RRV_MV_SCALEUP(startMV.y); |
| 1061 |
|
} else |
| 1062 |
|
if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? |
| 1063 |
|
startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); |
| 1064 |
|
if (!(MVequal(startMV, backupMV))) { |
| 1065 |
|
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
| 1066 |
|
|
| 1067 |
/* decide: MODE_INTER or MODE_INTER4V |
(*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); |
| 1068 |
mpeg4: if (sad8 < sad16 - nb/2+1) use_inter4v |
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
| 1069 |
*/ |
if (bSAD < Data->iMinSAD[0]) { |
| 1070 |
|
Data->currentMV[0] = backupMV; |
| 1071 |
|
Data->iMinSAD[0] = bSAD; } |
| 1072 |
|
} |
| 1073 |
|
|
| 1074 |
if (pMB->dquant == NO_CHANGE) { |
backupMV = Data->currentMV[0]; |
| 1075 |
if (((pParam->global_flags & XVID_INTER4V)==0) || |
if (!MotionFlags & PMV_HALFPELREFINE16 || Data->rrv) startMV.x = startMV.y = 0; |
| 1076 |
(sad16 < (sad8 + (int32_t)(IMV16X16 * pParam->quant)))) { |
else startMV.x = startMV.y = 1; |
| 1077 |
|
if (!(MVequal(startMV, backupMV))) { |
| 1078 |
|
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
| 1079 |
|
|
| 1080 |
sad8 = sad16; |
(*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); |
| 1081 |
pMB->mode = MODE_INTER; |
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
| 1082 |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; |
if (bSAD < Data->iMinSAD[0]) { |
| 1083 |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; |
Data->currentMV[0] = backupMV; |
| 1084 |
pMB->pmvs[0].x = pmv16.x; |
Data->iMinSAD[0] = bSAD; } |
|
pMB->pmvs[0].y = pmv16.y; |
|
|
} |
|
|
else |
|
|
pMB->mode = MODE_INTER4V; |
|
| 1085 |
} |
} |
|
else |
|
|
{ |
|
|
sad8 = sad16; |
|
|
pMB->mode = MODE_INTER; |
|
|
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; |
|
|
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; |
|
|
pMB->pmvs[0].x = pmv16.x; |
|
|
pMB->pmvs[0].y = pmv16.y; |
|
| 1086 |
} |
} |
| 1087 |
} |
} |
| 1088 |
|
|
| 1089 |
return 0; |
if (MotionFlags & PMV_HALFPELREFINE16) SubpelRefine(Data); |
|
} |
|
|
|
|
|
#define MVzero(A) ( ((A).x)==(0) && ((A).y)==(0) ) |
|
| 1090 |
|
|
| 1091 |
#define MVequal(A,B) ( ((A).x)==((B).x) && ((A).y)==((B).y) ) |
for(i = 0; i < 5; i++) { |
| 1092 |
|
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors |
| 1093 |
|
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
| 1094 |
|
} |
| 1095 |
|
|
| 1096 |
|
if((!Data->rrv) && (pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { |
| 1097 |
|
|
| 1098 |
#define CHECK_MV16_ZERO {\ |
Data->qpel_precision = 1; |
| 1099 |
if ( (0 <= max_dx) && (0 >= min_dx) \ |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
| 1100 |
&& (0 <= max_dy) && (0 >= min_dy) ) \ |
pParam->width, pParam->height, Data->iFcode, 1, 0); |
|
{ \ |
|
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0, 0 , iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); \ |
|
|
iSAD += calc_delta_16(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD <= iQuant * 96) \ |
|
|
iSAD -= MV16_00_BIAS; \ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } } \ |
|
|
} |
|
|
|
|
|
|
|
|
#define CHECK_MV16_CANDIDATE(X,Y) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
|
|
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV16_CANDIDATE_DIR(X,Y,D) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
|
|
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV16_CANDIDATE_FOUND(X,Y,D) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
|
|
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
|
|
} |
|
|
|
|
|
|
|
|
#define CHECK_MV8_ZERO {\ |
|
|
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, 0, 0 , iEdgedWidth), iEdgedWidth); \ |
|
|
iSAD += calc_delta_8(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } \ |
|
|
} |
|
|
|
|
|
|
|
|
#define CHECK_MV8_CANDIDATE(X,Y) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
|
|
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV8_CANDIDATE_DIR(X,Y,D) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
|
|
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV8_CANDIDATE_FOUND(X,Y,D) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
|
|
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
|
|
} |
|
|
|
|
|
/* too slow and not fully functional at the moment */ |
|
|
/* |
|
|
int32_t ZeroSearch16( |
|
|
const uint8_t * 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, |
|
|
MBParam * const pParam, |
|
|
MACROBLOCK * const pMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV) |
|
|
{ |
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
|
const int32_t iQuant = pParam->quant; |
|
|
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
|
|
int32_t iSAD; |
|
|
int32_t pred_x,pred_y; |
|
| 1101 |
|
|
| 1102 |
get_pmv(pMBs, x, y, pParam->mb_width, 0, &pred_x, &pred_y); |
SubpelRefine(Data); |
| 1103 |
|
} |
| 1104 |
|
|
| 1105 |
iSAD = sad16( cur, |
if (Data->iMinSAD[0] < (int32_t)iQuant * 30 ) inter4v = 0; |
| 1106 |
get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0,0, iEdgedWidth), |
if (inter4v) { |
| 1107 |
iEdgedWidth, MV_MAX_ERROR); |
SearchData Data8; |
| 1108 |
if (iSAD <= iQuant * 96) |
Data8.iFcode = Data->iFcode; |
| 1109 |
iSAD -= MV16_00_BIAS; |
Data8.lambda8 = Data->lambda8; |
| 1110 |
|
Data8.iEdgedWidth = Data->iEdgedWidth; |
| 1111 |
|
Data8.RefQ = Data->RefQ; |
| 1112 |
|
Data8.qpel = Data->qpel; |
| 1113 |
|
Data8.rrv = Data->rrv; |
| 1114 |
|
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
| 1115 |
|
Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); |
| 1116 |
|
Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); |
| 1117 |
|
Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); |
| 1118 |
|
|
| 1119 |
currMV->x = 0; |
if (Data->chroma) { |
| 1120 |
currMV->y = 0; |
int sumx, sumy, dx, dy; |
|
currPMV->x = -pred_x; |
|
|
currPMV->y = -pred_y; |
|
| 1121 |
|
|
| 1122 |
return iSAD; |
if(pParam->m_quarterpel) { |
| 1123 |
|
sumx= pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2; |
| 1124 |
|
sumy = pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2; |
| 1125 |
|
} else { |
| 1126 |
|
sumx = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
| 1127 |
|
sumy = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
| 1128 |
|
} |
| 1129 |
|
dx = (sumx >> 3) + roundtab_76[sumx & 0xf]; |
| 1130 |
|
dy = (sumy >> 3) + roundtab_76[sumy & 0xf]; |
| 1131 |
|
|
| 1132 |
|
Data->iMinSAD[1] += ChromaSAD(dx, dy, Data); |
| 1133 |
|
} |
| 1134 |
} |
} |
|
*/ |
|
| 1135 |
|
|
| 1136 |
int32_t PMVfastSearch16_MainSearch( |
if (Data->rrv) { |
| 1137 |
const uint8_t * const pRef, |
Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); |
| 1138 |
const uint8_t * const pRefH, |
Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); |
| 1139 |
const uint8_t * const pRefV, |
} |
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, const int y, |
|
|
int32_t startx, int32_t starty, |
|
|
int32_t iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const VECTOR * const pmv, |
|
|
const int32_t min_dx, const int32_t max_dx, |
|
|
const int32_t min_dy, const int32_t max_dy, |
|
|
const int32_t iEdgedWidth, |
|
|
const int32_t iDiamondSize, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
int iFound) |
|
|
{ |
|
|
/* Do a diamond search around given starting point, return SAD of best */ |
|
| 1140 |
|
|
| 1141 |
int32_t iDirection=0; |
if (!(inter4v) || |
| 1142 |
int32_t iSAD; |
(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
| 1143 |
VECTOR backupMV; |
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
| 1144 |
backupMV.x = startx; |
// INTER MODE |
| 1145 |
backupMV.y = starty; |
pMB->mode = MODE_INTER; |
| 1146 |
|
pMB->mvs[0] = pMB->mvs[1] |
| 1147 |
|
= pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
| 1148 |
|
|
| 1149 |
/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
| 1150 |
|
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
| 1151 |
|
|
| 1152 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
if(pParam->m_quarterpel) { |
| 1153 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
pMB->qmvs[0] = pMB->qmvs[1] |
| 1154 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
| 1155 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; |
| 1156 |
|
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; |
| 1157 |
if (iDirection) |
} else { |
| 1158 |
while (!iFound) |
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
| 1159 |
{ |
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
|
iFound = 1; |
|
|
backupMV=*currMV; |
|
|
|
|
|
if ( iDirection != 2) |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
|
|
if ( iDirection != 1) |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x+iDiamondSize,backupMV.y,2); |
|
|
if ( iDirection != 4) |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x,backupMV.y-iDiamondSize,3); |
|
|
if ( iDirection != 3) |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x,backupMV.y+iDiamondSize,4); |
|
|
} |
|
|
else |
|
|
{ |
|
|
currMV->x = startx; |
|
|
currMV->y = starty; |
|
| 1160 |
} |
} |
| 1161 |
return iMinSAD; |
} else { |
| 1162 |
|
// INTER4V MODE; all other things are already set in Search8 |
| 1163 |
|
pMB->mode = MODE_INTER4V; |
| 1164 |
|
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + |
| 1165 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant; |
| 1166 |
} |
} |
|
|
|
|
int32_t PMVfastSearch16_Refine( |
|
|
const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, const int y, |
|
|
VECTOR * const currMV, |
|
|
int32_t iMinSAD, |
|
|
const VECTOR * const pmv, |
|
|
const int32_t min_dx, const int32_t max_dx, |
|
|
const int32_t min_dy, const int32_t max_dy, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
const int32_t iEdgedWidth) |
|
|
{ |
|
|
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
|
|
|
|
|
int32_t iSAD; |
|
|
VECTOR backupMV = *currMV; |
|
|
|
|
|
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y-1); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y-1); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y-1); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y+1); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y+1); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y+1); |
|
|
|
|
|
return iMinSAD; |
|
| 1167 |
} |
} |
| 1168 |
|
|
| 1169 |
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
static void |
| 1170 |
|
Search8(const SearchData * const OldData, |
|
int32_t PMVfastSearch16( |
|
|
const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const IMAGE * const pCur, |
|
| 1171 |
const int x, const int y, |
const int x, const int y, |
| 1172 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
| 1173 |
MBParam * const pParam, |
const MBParam * const pParam, |
| 1174 |
MACROBLOCK * const pMBs, |
MACROBLOCK * const pMB, |
| 1175 |
VECTOR * const currMV, |
const MACROBLOCK * const pMBs, |
| 1176 |
VECTOR * const currPMV) |
const int block, |
| 1177 |
|
SearchData * const Data) |
| 1178 |
{ |
{ |
| 1179 |
const uint32_t iWcount = pParam->mb_width; |
int i = 0; |
| 1180 |
const int32_t iFcode = pParam->fixed_code; |
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
| 1181 |
const int32_t iQuant = pParam->quant; |
Data->currentMV = OldData->currentMV + 1 + block; |
| 1182 |
const int32_t iWidth = pParam->width; |
Data->currentQMV = OldData->currentQMV + 1 + block; |
|
const int32_t iHeight = pParam->height; |
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
|
|
|
|
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
|
|
|
|
|
int32_t iDiamondSize; |
|
| 1183 |
|
|
| 1184 |
int32_t min_dx; |
if(pParam->m_quarterpel) { |
| 1185 |
int32_t max_dx; |
Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
| 1186 |
int32_t min_dy; |
if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y, |
| 1187 |
int32_t max_dy; |
Data->predMV, Data->iFcode, 0, 0); |
| 1188 |
|
|
| 1189 |
int32_t iFound; |
} else { |
| 1190 |
|
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
| 1191 |
VECTOR newMV; |
if (block != 0) { |
| 1192 |
VECTOR backupMV; /* just for PMVFAST */ |
if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y, |
| 1193 |
|
Data->predMV, Data->iFcode, 0, Data->rrv); |
| 1194 |
|
} |
| 1195 |
|
} |
| 1196 |
|
|
| 1197 |
VECTOR pmv[4]; |
*(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100; |
|
int32_t psad[4]; |
|
| 1198 |
|
|
| 1199 |
MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { |
| 1200 |
|
if (Data->rrv) i = 2; else i = 1; |
| 1201 |
|
|
| 1202 |
static int32_t threshA,threshB; |
Data->Ref = OldData->Ref + i*8 * ((block&1) + pParam->edged_width*(block>>1)); |
| 1203 |
int32_t bPredEq; |
Data->RefH = OldData->RefH + i*8 * ((block&1) + pParam->edged_width*(block>>1)); |
| 1204 |
int32_t iMinSAD,iSAD; |
Data->RefV = OldData->RefV + i*8 * ((block&1) + pParam->edged_width*(block>>1)); |
| 1205 |
|
Data->RefHV = OldData->RefHV + i*8 * ((block&1) + pParam->edged_width*(block>>1)); |
| 1206 |
|
|
| 1207 |
/* Get maximum range */ |
Data->Cur = OldData->Cur + i*8 * ((block&1) + pParam->edged_width*(block>>1)); |
| 1208 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
Data->qpel_precision = 0; |
|
x, y, 16, iWidth, iHeight, iFcode); |
|
| 1209 |
|
|
| 1210 |
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
| 1211 |
|
pParam->width, pParam->height, OldData->iFcode - Data->qpel, 0, Data->rrv); |
| 1212 |
|
|
| 1213 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
if (Data->rrv) CheckCandidate = CheckCandidate16no4v; |
| 1214 |
{ min_dx = EVEN(min_dx); |
else CheckCandidate = CheckCandidate8; |
|
max_dx = EVEN(max_dx); |
|
|
min_dy = EVEN(min_dy); |
|
|
max_dy = EVEN(max_dy); |
|
|
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
|
| 1215 |
|
|
| 1216 |
|
if (MotionFlags & PMV_EXTSEARCH8) { |
| 1217 |
|
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
| 1218 |
|
|
| 1219 |
bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
MainSearchFunc *MainSearchPtr; |
| 1220 |
|
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
| 1221 |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
| 1222 |
|
else MainSearchPtr = DiamondSearch; |
| 1223 |
|
|
| 1224 |
if ((x==0) && (y==0) ) |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
|
{ |
|
|
threshA = 512; |
|
|
threshB = 1024; |
|
| 1225 |
|
|
| 1226 |
|
if(*(Data->iMinSAD) < temp_sad) { |
| 1227 |
|
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
| 1228 |
|
Data->currentQMV->y = 2 * Data->currentMV->y; |
| 1229 |
} |
} |
|
else |
|
|
{ |
|
|
threshA = psad[0]; |
|
|
threshB = threshA+256; |
|
|
if (threshA< 512) threshA = 512; |
|
|
if (threshA>1024) threshA = 1024; |
|
|
if (threshB>1792) threshB = 1792; |
|
| 1230 |
} |
} |
| 1231 |
|
|
| 1232 |
iFound=0; |
if (MotionFlags & PMV_HALFPELREFINE8) { |
| 1233 |
|
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
|
/* Step 2: Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
|
|
vector of the median. |
|
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
|
|
*/ |
|
|
|
|
|
if ((bPredEq) && (MVequal(pmv[0],pMB->mvs[0]) ) ) |
|
|
iFound=2; |
|
|
|
|
|
/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
|
|
Otherwise select large Diamond Search. |
|
|
*/ |
|
|
|
|
|
if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536) || (bPredEq) ) |
|
|
iDiamondSize=1; // halfpel! |
|
|
else |
|
|
iDiamondSize=2; // halfpel! |
|
|
|
|
|
if (!(MotionFlags & PMV_HALFPELDIAMOND16) ) |
|
|
iDiamondSize*=2; |
|
| 1234 |
|
|
| 1235 |
/* Step 4: Calculate SAD around the Median prediction. |
SubpelRefine(Data); // perform halfpel refine of current best vector |
|
MinSAD=SAD |
|
|
If Motion Vector equal to Previous frame motion vector |
|
|
and MinSAD<PrevFrmSAD goto Step 10. |
|
|
If SAD<=256 goto Step 10. |
|
|
*/ |
|
| 1236 |
|
|
| 1237 |
|
if(*(Data->iMinSAD) < temp_sad) { // we have found a better match |
| 1238 |
// Prepare for main loop |
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
| 1239 |
|
Data->currentQMV->y = 2 * Data->currentMV->y; |
|
*currMV=pmv[0]; /* current best := prediction */ |
|
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
|
|
{ /* This should NOT be necessary! */ |
|
|
currMV->x = EVEN(currMV->x); |
|
|
currMV->y = EVEN(currMV->y); |
|
| 1240 |
} |
} |
|
|
|
|
if (currMV->x > max_dx) |
|
|
{ |
|
|
currMV->x=max_dx; |
|
| 1241 |
} |
} |
| 1242 |
if (currMV->x < min_dx) |
|
| 1243 |
{ |
if(!Data->rrv && Data->qpel) { |
| 1244 |
currMV->x=min_dx; |
if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && |
| 1245 |
|
(MotionFlags & PMV_QUARTERPELREFINE8)) { |
| 1246 |
|
Data->qpel_precision = 1; |
| 1247 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
| 1248 |
|
pParam->width, pParam->height, OldData->iFcode, 1, 0); |
| 1249 |
|
SubpelRefine(Data); |
| 1250 |
} |
} |
|
if (currMV->y > max_dy) |
|
|
{ |
|
|
currMV->y=max_dy; |
|
| 1251 |
} |
} |
|
if (currMV->y < min_dy) |
|
|
{ |
|
|
currMV->y=min_dy; |
|
| 1252 |
} |
} |
| 1253 |
|
|
| 1254 |
iMinSAD = sad16( cur, |
if (Data->rrv) { |
| 1255 |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x); |
| 1256 |
iEdgedWidth, MV_MAX_ERROR); |
Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y); |
|
iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; |
|
|
|
|
|
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,pMB->mvs[0])) && (iMinSAD < pMB->sad16) ) ) |
|
|
{ |
|
|
|
|
|
if (MotionFlags & PMV_QUICKSTOP16) |
|
|
goto step10b; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto step10; |
|
| 1257 |
} |
} |
| 1258 |
|
|
| 1259 |
/* |
if(Data->qpel) { |
| 1260 |
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x; |
| 1261 |
Also calculate (0,0) but do not subtract offset. |
pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y; |
| 1262 |
Let MinSAD be the smallest SAD up to this point. |
pMB->qmvs[block] = *(Data->currentQMV); |
| 1263 |
If MV is (0,0) subtract offset. ******** WHAT'S THIS 'OFFSET' ??? *********** |
} else { |
| 1264 |
*/ |
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
| 1265 |
|
pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; |
| 1266 |
// (0,0) is always possible |
} |
| 1267 |
|
|
| 1268 |
CHECK_MV16_ZERO; |
pMB->mvs[block] = *(Data->currentMV); |
| 1269 |
|
pMB->sad8[block] = 4 * (*Data->iMinSAD); |
| 1270 |
|
} |
| 1271 |
|
|
| 1272 |
// previous frame MV is always possible |
/* B-frames code starts here */ |
|
CHECK_MV16_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); |
|
| 1273 |
|
|
| 1274 |
// left neighbour, if allowed |
static __inline VECTOR |
| 1275 |
if (x != 0) |
ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) |
| 1276 |
{ |
{ |
| 1277 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
/* the stupidiest function ever */ |
| 1278 |
{ pmv[1].x = EVEN(pmv[1].x); |
if (mode == MODE_FORWARD) return pMB->mvs[0]; |
| 1279 |
pmv[1].y = EVEN(pmv[1].y); |
else return pMB->b_mvs[0]; |
|
} |
|
|
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
|
| 1280 |
} |
} |
| 1281 |
|
|
| 1282 |
// top neighbour, if allowed |
static void __inline |
| 1283 |
if (y != 0) |
PreparePredictionsBF(VECTOR * const pmv, const int x, const int y, |
| 1284 |
|
const uint32_t iWcount, |
| 1285 |
|
const MACROBLOCK * const pMB, |
| 1286 |
|
const uint32_t mode_curr) |
| 1287 |
{ |
{ |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
|
|
{ pmv[2].x = EVEN(pmv[2].x); |
|
|
pmv[2].y = EVEN(pmv[2].y); |
|
|
} |
|
|
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
|
| 1288 |
|
|
| 1289 |
// top right neighbour, if allowed |
// [0] is prediction |
| 1290 |
if (x != (iWcount-1)) |
pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); |
|
{ |
|
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
|
|
{ pmv[3].x = EVEN(pmv[3].x); |
|
|
pmv[3].y = EVEN(pmv[3].y); |
|
|
} |
|
|
CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); |
|
|
} |
|
|
} |
|
| 1291 |
|
|
| 1292 |
/* Step 6: If MinSAD <= thresa goto Step 10. |
pmv[1].x = pmv[1].y = 0; // [1] is zero |
|
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
|
|
*/ |
|
| 1293 |
|
|
| 1294 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,pMB->mvs[0]) && (iMinSAD < pMB->sad16) ) ) |
pmv[2] = ChoosePred(pMB, mode_curr); |
| 1295 |
{ |
pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); |
|
if (MotionFlags & PMV_QUICKSTOP16) |
|
|
goto step10b; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto step10; |
|
|
} |
|
| 1296 |
|
|
| 1297 |
|
if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour |
| 1298 |
|
pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); |
| 1299 |
|
pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); |
| 1300 |
|
} else pmv[3].x = pmv[3].y = 0; |
| 1301 |
|
|
| 1302 |
/************ (Diamond Search) **************/ |
if (y != 0) { |
| 1303 |
/* |
pmv[4] = ChoosePred(pMB-iWcount, mode_curr); |
| 1304 |
Step 7: Perform Diamond search, with either the small or large diamond. |
pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y); |
| 1305 |
If Found=2 only examine one Diamond pattern, and afterwards goto step 10 |
} else pmv[4].x = pmv[4].y = 0; |
|
Step 8: If small diamond, iterate small diamond search pattern until motion vector lies in the center of the diamond. |
|
|
If center then goto step 10. |
|
|
Step 9: If large diamond, iterate large diamond search pattern until motion vector lies in the center. |
|
|
Refine by using small diamond and goto step 10. |
|
|
*/ |
|
| 1306 |
|
|
| 1307 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
if (x != 0) { |
| 1308 |
|
pmv[5] = ChoosePred(pMB-1, mode_curr); |
| 1309 |
|
pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
| 1310 |
|
} else pmv[5].x = pmv[5].y = 0; |
| 1311 |
|
|
| 1312 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
if ((x != 0)&&(y != 0)) { |
| 1313 |
iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); |
| 1314 |
x, y, |
pmv[6].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
| 1315 |
currMV->x, currMV->y, iMinSAD, &newMV, |
} else pmv[6].x = pmv[6].y = 0; |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
| 1316 |
|
|
| 1317 |
if (iSAD < iMinSAD) |
// more? |
|
{ |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
| 1318 |
} |
} |
| 1319 |
|
|
|
if (MotionFlags & PMV_EXTSEARCH16) |
|
|
{ |
|
|
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
|
|
|
|
|
if (!(MVequal(pmv[0],backupMV)) ) |
|
|
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
| 1320 |
|
|
| 1321 |
if (iSAD < iMinSAD) |
/* search backward or forward, for b-frames */ |
| 1322 |
|
static void |
| 1323 |
|
SearchBF( const uint8_t * const pRef, |
| 1324 |
|
const uint8_t * const pRefH, |
| 1325 |
|
const uint8_t * const pRefV, |
| 1326 |
|
const uint8_t * const pRefHV, |
| 1327 |
|
const IMAGE * const pCur, |
| 1328 |
|
const int x, const int y, |
| 1329 |
|
const uint32_t MotionFlags, |
| 1330 |
|
const uint32_t iFcode, |
| 1331 |
|
const MBParam * const pParam, |
| 1332 |
|
MACROBLOCK * const pMB, |
| 1333 |
|
const VECTOR * const predMV, |
| 1334 |
|
int32_t * const best_sad, |
| 1335 |
|
const int32_t mode_current, |
| 1336 |
|
SearchData * const Data) |
| 1337 |
{ |
{ |
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
| 1338 |
|
|
| 1339 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
const int32_t iEdgedWidth = pParam->edged_width; |
|
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
0, 0, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
| 1340 |
|
|
| 1341 |
if (iSAD < iMinSAD) |
int i, iDirection = 255, mask; |
| 1342 |
{ |
VECTOR pmv[7]; |
| 1343 |
*currMV = newMV; |
MainSearchFunc *MainSearchPtr; |
| 1344 |
iMinSAD = iSAD; |
*Data->iMinSAD = MV_MAX_ERROR; |
| 1345 |
} |
Data->iFcode = iFcode; |
| 1346 |
|
Data->qpel_precision = 0; |
| 1347 |
|
|
| 1348 |
|
Data->Ref = pRef + (x + y * iEdgedWidth) * 16; |
| 1349 |
|
Data->RefH = pRefH + (x + y * iEdgedWidth) * 16; |
| 1350 |
|
Data->RefV = pRefV + (x + y * iEdgedWidth) * 16; |
| 1351 |
|
Data->RefHV = pRefHV + (x + y * iEdgedWidth) * 16; |
| 1352 |
|
|
| 1353 |
|
Data->predMV = *predMV; |
| 1354 |
|
|
| 1355 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
| 1356 |
|
pParam->width, pParam->height, iFcode - Data->qpel, 0, 0); |
| 1357 |
|
|
| 1358 |
|
pmv[0] = Data->predMV; |
| 1359 |
|
if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } |
| 1360 |
|
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
| 1361 |
|
|
| 1362 |
|
Data->currentMV->x = Data->currentMV->y = 0; |
| 1363 |
|
CheckCandidate = CheckCandidate16no4v; |
| 1364 |
|
|
| 1365 |
|
// main loop. checking all predictions |
| 1366 |
|
for (i = 0; i < 7; i++) { |
| 1367 |
|
if (!(mask = make_mask(pmv, i)) ) continue; |
| 1368 |
|
CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
| 1369 |
|
} |
| 1370 |
|
|
| 1371 |
|
if (MotionFlags & PMV_USESQUARES16) |
| 1372 |
|
MainSearchPtr = SquareSearch; |
| 1373 |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
| 1374 |
|
MainSearchPtr = AdvDiamondSearch; |
| 1375 |
|
else MainSearchPtr = DiamondSearch; |
| 1376 |
|
|
| 1377 |
|
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
| 1378 |
|
|
| 1379 |
|
SubpelRefine(Data); |
| 1380 |
|
|
| 1381 |
|
if (Data->qpel && *Data->iMinSAD < *best_sad + 300) { |
| 1382 |
|
Data->currentQMV->x = 2*Data->currentMV->x; |
| 1383 |
|
Data->currentQMV->y = 2*Data->currentMV->y; |
| 1384 |
|
Data->qpel_precision = 1; |
| 1385 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
| 1386 |
|
pParam->width, pParam->height, iFcode, 1, 0); |
| 1387 |
|
SubpelRefine(Data); |
| 1388 |
|
} |
| 1389 |
|
|
| 1390 |
|
// three bits are needed to code backward mode. four for forward |
| 1391 |
|
|
| 1392 |
|
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
| 1393 |
|
else *Data->iMinSAD += 3 * Data->lambda16; |
| 1394 |
|
|
| 1395 |
|
if (*Data->iMinSAD < *best_sad) { |
| 1396 |
|
*best_sad = *Data->iMinSAD; |
| 1397 |
|
pMB->mode = mode_current; |
| 1398 |
|
if (Data->qpel) { |
| 1399 |
|
pMB->pmvs[0].x = Data->currentQMV->x - predMV->x; |
| 1400 |
|
pMB->pmvs[0].y = Data->currentQMV->y - predMV->y; |
| 1401 |
|
if (mode_current == MODE_FORWARD) |
| 1402 |
|
pMB->qmvs[0] = *Data->currentQMV; |
| 1403 |
|
else |
| 1404 |
|
pMB->b_qmvs[0] = *Data->currentQMV; |
| 1405 |
|
} else { |
| 1406 |
|
pMB->pmvs[0].x = Data->currentMV->x - predMV->x; |
| 1407 |
|
pMB->pmvs[0].y = Data->currentMV->y - predMV->y; |
| 1408 |
} |
} |
| 1409 |
|
if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; |
| 1410 |
|
else pMB->b_mvs[0] = *Data->currentMV; |
| 1411 |
} |
} |
| 1412 |
|
if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; |
| 1413 |
|
else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search |
| 1414 |
|
|
|
/* |
|
|
Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
|
|
*/ |
|
|
|
|
|
step10: |
|
|
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
|
|
iMinSAD = PMVfastSearch16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
currMV, iMinSAD, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
|
|
|
|
|
step10b: |
|
|
currPMV->x = currMV->x - pmv[0].x; |
|
|
currPMV->y = currMV->y - pmv[0].y; |
|
|
return iMinSAD; |
|
| 1415 |
} |
} |
| 1416 |
|
|
| 1417 |
|
static void |
| 1418 |
|
SkipDecisionB(const IMAGE * const pCur, |
| 1419 |
|
const IMAGE * const f_Ref, |
| 1420 |
|
const IMAGE * const b_Ref, |
| 1421 |
|
MACROBLOCK * const pMB, |
| 1422 |
|
const uint32_t quant, |
| 1423 |
|
const uint32_t x, const uint32_t y, |
| 1424 |
|
const SearchData * const Data) |
| 1425 |
|
{ |
| 1426 |
|
int dx, dy, b_dx, b_dy; |
| 1427 |
|
uint32_t sum; |
| 1428 |
|
//this is not full chroma compensation, only it's fullpel approximation. should work though |
| 1429 |
|
if (Data->qpel) { |
| 1430 |
|
dy = Data->directmvF[0].y/2 + Data->directmvF[1].y/2 + |
| 1431 |
|
Data->directmvF[2].y/2 + Data->directmvF[3].y/2; |
| 1432 |
|
|
| 1433 |
|
dx = Data->directmvF[0].x/2 + Data->directmvF[1].x/2 + |
| 1434 |
|
Data->directmvF[2].x/2 + Data->directmvF[3].x/2; |
| 1435 |
|
|
| 1436 |
|
b_dy = Data->directmvB[0].y/2 + Data->directmvB[1].y/2 + |
| 1437 |
|
Data->directmvB[2].y/2 + Data->directmvB[3].y/2; |
| 1438 |
|
|
| 1439 |
|
b_dx = Data->directmvB[0].x/2 + Data->directmvB[1].x/2 + |
| 1440 |
|
Data->directmvB[2].x/2 + Data->directmvB[3].x/2; |
| 1441 |
|
|
| 1442 |
|
} else { |
| 1443 |
|
dy = Data->directmvF[0].y + Data->directmvF[1].y + |
| 1444 |
|
Data->directmvF[2].y + Data->directmvF[3].y; |
| 1445 |
|
|
| 1446 |
int32_t PMVfastSearch8_MainSearch( |
dx = Data->directmvF[0].x + Data->directmvF[1].x + |
| 1447 |
const uint8_t * const pRef, |
Data->directmvF[2].x + Data->directmvF[3].x; |
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, const int y, |
|
|
int32_t startx, int32_t starty, |
|
|
int32_t iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const VECTOR * const pmv, |
|
|
const int32_t min_dx, const int32_t max_dx, |
|
|
const int32_t min_dy, const int32_t max_dy, |
|
|
const int32_t iEdgedWidth, |
|
|
const int32_t iDiamondSize, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
int iFound) |
|
|
{ |
|
|
/* Do a diamond search around given starting point, return SAD of best */ |
|
|
|
|
|
int32_t iDirection=0; |
|
|
int32_t iSAD; |
|
|
VECTOR backupMV; |
|
|
backupMV.x = startx; |
|
|
backupMV.y = starty; |
|
| 1448 |
|
|
| 1449 |
/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ |
b_dy = Data->directmvB[0].y + Data->directmvB[1].y + |
| 1450 |
|
Data->directmvB[2].y + Data->directmvB[3].y; |
| 1451 |
|
|
| 1452 |
CHECK_MV8_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
b_dx = Data->directmvB[0].x + Data->directmvB[1].x + |
| 1453 |
CHECK_MV8_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
Data->directmvB[2].x + Data->directmvB[3].x; |
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
|
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
|
|
|
|
|
if (iDirection) |
|
|
while (!iFound) |
|
|
{ |
|
|
iFound = 1; |
|
|
backupMV=*currMV; // since iDirection!=0, this is well defined! |
|
|
|
|
|
if ( iDirection != 2) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
|
|
if ( iDirection != 1) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x+iDiamondSize,backupMV.y,2); |
|
|
if ( iDirection != 4) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x,backupMV.y-iDiamondSize,3); |
|
|
if ( iDirection != 3) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x,backupMV.y+iDiamondSize,4); |
|
|
} |
|
|
else |
|
|
{ |
|
|
currMV->x = startx; |
|
|
currMV->y = starty; |
|
|
} |
|
|
return iMinSAD; |
|
| 1454 |
} |
} |
| 1455 |
|
|
|
int32_t PMVfastSearch8_Refine( |
|
|
const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, const int y, |
|
|
VECTOR * const currMV, |
|
|
int32_t iMinSAD, |
|
|
const VECTOR * const pmv, |
|
|
const int32_t min_dx, const int32_t max_dx, |
|
|
const int32_t min_dy, const int32_t max_dy, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
const int32_t iEdgedWidth) |
|
|
{ |
|
|
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
|
| 1456 |
|
|
| 1457 |
int32_t iSAD; |
dy = (dy >> 3) + roundtab_76[dy & 0xf]; |
| 1458 |
VECTOR backupMV = *currMV; |
dx = (dx >> 3) + roundtab_76[dx & 0xf]; |
| 1459 |
|
b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf]; |
| 1460 |
|
b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf]; |
| 1461 |
|
|
| 1462 |
CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y-1); |
sum = sad8bi(pCur->u + 8*x + 8*y*(Data->iEdgedWidth/2), |
| 1463 |
CHECK_MV8_CANDIDATE(backupMV.x ,backupMV.y-1); |
f_Ref->u + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
| 1464 |
CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y-1); |
b_Ref->u + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
| 1465 |
CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y); |
Data->iEdgedWidth/2); |
| 1466 |
CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y); |
sum += sad8bi(pCur->v + 8*x + 8*y*(Data->iEdgedWidth/2), |
| 1467 |
CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y+1); |
f_Ref->v + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
| 1468 |
CHECK_MV8_CANDIDATE(backupMV.x ,backupMV.y+1); |
b_Ref->v + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
| 1469 |
CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y+1); |
Data->iEdgedWidth/2); |
| 1470 |
|
|
| 1471 |
return iMinSAD; |
if (sum < 2*MAX_CHROMA_SAD_FOR_SKIP * quant) pMB->mode = MODE_DIRECT_NONE_MV; //skipped |
| 1472 |
} |
} |
| 1473 |
|
|
| 1474 |
|
|
|
#define PMV_HALFPEL8 (PMV_HALFPELDIAMOND8|PMV_HALFPELREFINE8) |
|
| 1475 |
|
|
| 1476 |
int32_t PMVfastSearch8( |
static __inline uint32_t |
| 1477 |
const uint8_t * const pRef, |
SearchDirect(const IMAGE * const f_Ref, |
| 1478 |
const uint8_t * const pRefH, |
const uint8_t * const f_RefH, |
| 1479 |
const uint8_t * const pRefV, |
const uint8_t * const f_RefV, |
| 1480 |
const uint8_t * const pRefHV, |
const uint8_t * const f_RefHV, |
| 1481 |
|
const IMAGE * const b_Ref, |
| 1482 |
|
const uint8_t * const b_RefH, |
| 1483 |
|
const uint8_t * const b_RefV, |
| 1484 |
|
const uint8_t * const b_RefHV, |
| 1485 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
| 1486 |
const int x, const int y, |
const int x, const int y, |
|
const int start_x, int start_y, |
|
| 1487 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
| 1488 |
MBParam * const pParam, |
const int32_t TRB, const int32_t TRD, |
| 1489 |
MACROBLOCK * const pMBs, |
const MBParam * const pParam, |
| 1490 |
VECTOR * const currMV, |
MACROBLOCK * const pMB, |
| 1491 |
VECTOR * const currPMV) |
const MACROBLOCK * const b_mb, |
| 1492 |
{ |
int32_t * const best_sad, |
| 1493 |
const uint32_t iWcount = pParam->mb_width; |
SearchData * const Data) |
| 1494 |
|
|
| 1495 |
|
{ |
| 1496 |
|
int32_t skip_sad; |
| 1497 |
|
int k; |
| 1498 |
|
|
| 1499 |
|
MainSearchFunc *MainSearchPtr; |
| 1500 |
|
|
| 1501 |
|
*Data->iMinSAD = 256*4096; |
| 1502 |
|
|
| 1503 |
|
Data->Ref = f_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
| 1504 |
|
Data->RefH = f_RefH + (x + Data->iEdgedWidth*y) * 16; |
| 1505 |
|
Data->RefV = f_RefV + (x + Data->iEdgedWidth*y) * 16; |
| 1506 |
|
Data->RefHV = f_RefHV + (x + Data->iEdgedWidth*y) * 16; |
| 1507 |
|
Data->bRef = b_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
| 1508 |
|
Data->bRefH = b_RefH + (x + Data->iEdgedWidth*y) * 16; |
| 1509 |
|
Data->bRefV = b_RefV + (x + Data->iEdgedWidth*y) * 16; |
| 1510 |
|
Data->bRefHV = b_RefHV + (x + Data->iEdgedWidth*y) * 16; |
| 1511 |
|
|
| 1512 |
|
Data->max_dx = 2 * pParam->width - 2 * (x) * 16; |
| 1513 |
|
Data->max_dy = 2 * pParam->height - 2 * (y) * 16; |
| 1514 |
|
Data->min_dx = -(2 * 16 + 2 * (x) * 16); |
| 1515 |
|
Data->min_dy = -(2 * 16 + 2 * (y) * 16); |
| 1516 |
|
if (Data->qpel) { //we measure in qpixels |
| 1517 |
|
Data->max_dx *= 2; |
| 1518 |
|
Data->max_dy *= 2; |
| 1519 |
|
Data->min_dx *= 2; |
| 1520 |
|
Data->min_dy *= 2; |
| 1521 |
|
Data->referencemv = b_mb->qmvs; |
| 1522 |
|
} else Data->referencemv = b_mb->mvs; |
| 1523 |
|
Data->qpel_precision = 0; // it's a trick. it's 1 not 0, but we need 0 here |
| 1524 |
|
|
| 1525 |
|
for (k = 0; k < 4; k++) { |
| 1526 |
|
pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); |
| 1527 |
|
pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD; |
| 1528 |
|
pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); |
| 1529 |
|
pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; |
| 1530 |
|
|
| 1531 |
|
if ( ( pMB->b_mvs[k].x > Data->max_dx ) || ( pMB->b_mvs[k].x < Data->min_dx ) |
| 1532 |
|
|| ( pMB->b_mvs[k].y > Data->max_dy ) || ( pMB->b_mvs[k].y < Data->min_dy )) { |
| 1533 |
|
|
| 1534 |
|
*best_sad = 256*4096; // in that case, we won't use direct mode |
| 1535 |
|
pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV" |
| 1536 |
|
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
| 1537 |
|
return 256*4096; |
| 1538 |
|
} |
| 1539 |
|
if (b_mb->mode != MODE_INTER4V) { |
| 1540 |
|
pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; |
| 1541 |
|
pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0]; |
| 1542 |
|
Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0]; |
| 1543 |
|
Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0]; |
| 1544 |
|
break; |
| 1545 |
|
} |
| 1546 |
|
} |
| 1547 |
|
|
| 1548 |
|
if (b_mb->mode == MODE_INTER4V) CheckCandidate = CheckCandidateDirect; |
| 1549 |
|
else CheckCandidate = CheckCandidateDirectno4v; |
| 1550 |
|
|
| 1551 |
|
(*CheckCandidate)(0, 0, 255, &k, Data); |
| 1552 |
|
|
| 1553 |
|
// initial (fast) skip decision |
| 1554 |
|
if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH*2) { |
| 1555 |
|
SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data->chroma, Data); //possible skip - checking chroma |
| 1556 |
|
if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip. |
| 1557 |
|
} |
| 1558 |
|
|
| 1559 |
|
skip_sad = *Data->iMinSAD; |
| 1560 |
|
|
| 1561 |
|
// DIRECT MODE DELTA VECTOR SEARCH. |
| 1562 |
|
// This has to be made more effective, but at the moment I'm happy it's running at all |
| 1563 |
|
|
| 1564 |
|
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
| 1565 |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
| 1566 |
|
else MainSearchPtr = DiamondSearch; |
| 1567 |
|
|
| 1568 |
|
(*MainSearchPtr)(0, 0, Data, 255); |
| 1569 |
|
|
| 1570 |
|
SubpelRefine(Data); |
| 1571 |
|
|
| 1572 |
|
*best_sad = *Data->iMinSAD; |
| 1573 |
|
|
| 1574 |
|
if (b_mb->mode == MODE_INTER4V || Data->qpel) pMB->mode = MODE_DIRECT; |
| 1575 |
|
else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
| 1576 |
|
|
| 1577 |
|
pMB->pmvs[3] = *Data->currentMV; |
| 1578 |
|
|
| 1579 |
|
for (k = 0; k < 4; k++) { |
| 1580 |
|
pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x; |
| 1581 |
|
pMB->b_mvs[k].x = ( (Data->currentMV->x == 0) |
| 1582 |
|
? Data->directmvB[k].x |
| 1583 |
|
:pMB->mvs[k].x - Data->referencemv[k].x); |
| 1584 |
|
pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y); |
| 1585 |
|
pMB->b_mvs[k].y = ((Data->currentMV->y == 0) |
| 1586 |
|
? Data->directmvB[k].y |
| 1587 |
|
: pMB->mvs[k].y - Data->referencemv[k].y); |
| 1588 |
|
if (Data->qpel) { |
| 1589 |
|
pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2; |
| 1590 |
|
pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2; |
| 1591 |
|
pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2; |
| 1592 |
|
pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2; |
| 1593 |
|
} |
| 1594 |
|
|
| 1595 |
|
if (b_mb->mode != MODE_INTER4V) { |
| 1596 |
|
pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; |
| 1597 |
|
pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; |
| 1598 |
|
pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0]; |
| 1599 |
|
pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0]; |
| 1600 |
|
break; |
| 1601 |
|
} |
| 1602 |
|
} |
| 1603 |
|
return skip_sad; |
| 1604 |
|
} |
| 1605 |
|
|
| 1606 |
|
|
| 1607 |
|
static void |
| 1608 |
|
SearchInterpolate(const uint8_t * const f_Ref, |
| 1609 |
|
const uint8_t * const f_RefH, |
| 1610 |
|
const uint8_t * const f_RefV, |
| 1611 |
|
const uint8_t * const f_RefHV, |
| 1612 |
|
const uint8_t * const b_Ref, |
| 1613 |
|
const uint8_t * const b_RefH, |
| 1614 |
|
const uint8_t * const b_RefV, |
| 1615 |
|
const uint8_t * const b_RefHV, |
| 1616 |
|
const IMAGE * const pCur, |
| 1617 |
|
const int x, const int y, |
| 1618 |
|
const uint32_t fcode, |
| 1619 |
|
const uint32_t bcode, |
| 1620 |
|
const uint32_t MotionFlags, |
| 1621 |
|
const MBParam * const pParam, |
| 1622 |
|
const VECTOR * const f_predMV, |
| 1623 |
|
const VECTOR * const b_predMV, |
| 1624 |
|
MACROBLOCK * const pMB, |
| 1625 |
|
int32_t * const best_sad, |
| 1626 |
|
SearchData * const fData) |
| 1627 |
|
|
| 1628 |
|
{ |
| 1629 |
|
|
| 1630 |
|
int iDirection, i, j; |
| 1631 |
|
SearchData bData; |
| 1632 |
|
|
| 1633 |
|
fData->qpel_precision = 0; |
| 1634 |
|
memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data |
| 1635 |
|
*fData->iMinSAD = 4096*256; |
| 1636 |
|
bData.currentMV++; bData.currentQMV++; |
| 1637 |
|
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
| 1638 |
|
|
| 1639 |
|
i = (x + y * fData->iEdgedWidth) * 16; |
| 1640 |
|
bData.bRef = fData->Ref = f_Ref + i; |
| 1641 |
|
bData.bRefH = fData->RefH = f_RefH + i; |
| 1642 |
|
bData.bRefV = fData->RefV = f_RefV + i; |
| 1643 |
|
bData.bRefHV = fData->RefHV = f_RefHV + i; |
| 1644 |
|
bData.Ref = fData->bRef = b_Ref + i; |
| 1645 |
|
bData.RefH = fData->bRefH = b_RefH + i; |
| 1646 |
|
bData.RefV = fData->bRefV = b_RefV + i; |
| 1647 |
|
bData.RefHV = fData->bRefHV = b_RefHV + i; |
| 1648 |
|
|
| 1649 |
|
bData.bpredMV = fData->predMV = *f_predMV; |
| 1650 |
|
fData->bpredMV = bData.predMV = *b_predMV; |
| 1651 |
|
fData->currentMV[0] = fData->currentMV[2]; |
| 1652 |
|
|
| 1653 |
|
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode - fData->qpel, 0, 0); |
| 1654 |
|
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); |
| 1655 |
|
|
| 1656 |
|
if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; |
| 1657 |
|
if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx; |
| 1658 |
|
if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy; |
| 1659 |
|
if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy; |
| 1660 |
|
|
| 1661 |
|
if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx; |
| 1662 |
|
if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx; |
| 1663 |
|
if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy; |
| 1664 |
|
if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy; |
| 1665 |
|
|
| 1666 |
|
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
| 1667 |
|
|
| 1668 |
|
//diamond. I wish we could use normal mainsearch functions (square, advdiamond) |
| 1669 |
|
|
| 1670 |
|
do { |
| 1671 |
|
iDirection = 255; |
| 1672 |
|
// forward MV moves |
| 1673 |
|
i = fData->currentMV[0].x; j = fData->currentMV[0].y; |
| 1674 |
|
|
| 1675 |
|
CheckCandidateInt(i + 1, j, 0, &iDirection, fData); |
| 1676 |
|
CheckCandidateInt(i, j + 1, 0, &iDirection, fData); |
| 1677 |
|
CheckCandidateInt(i - 1, j, 0, &iDirection, fData); |
| 1678 |
|
CheckCandidateInt(i, j - 1, 0, &iDirection, fData); |
| 1679 |
|
|
| 1680 |
|
// backward MV moves |
| 1681 |
|
i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
| 1682 |
|
fData->currentMV[2] = fData->currentMV[0]; |
| 1683 |
|
CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); |
| 1684 |
|
CheckCandidateInt(i, j + 1, 0, &iDirection, &bData); |
| 1685 |
|
CheckCandidateInt(i - 1, j, 0, &iDirection, &bData); |
| 1686 |
|
CheckCandidateInt(i, j - 1, 0, &iDirection, &bData); |
| 1687 |
|
|
| 1688 |
|
} while (!(iDirection)); |
| 1689 |
|
|
| 1690 |
|
if (fData->qpel) { |
| 1691 |
|
if (*fData->iMinSAD > *best_sad + 500) return; |
| 1692 |
|
CheckCandidate = CheckCandidateInt; |
| 1693 |
|
fData->qpel_precision = bData.qpel_precision = 1; |
| 1694 |
|
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0); |
| 1695 |
|
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0); |
| 1696 |
|
fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x; |
| 1697 |
|
fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y; |
| 1698 |
|
fData->currentQMV[1].x = 2 * fData->currentMV[1].x; |
| 1699 |
|
fData->currentQMV[1].y = 2 * fData->currentMV[1].y; |
| 1700 |
|
SubpelRefine(fData); |
| 1701 |
|
if (*fData->iMinSAD > *best_sad + 300) return; |
| 1702 |
|
fData->currentQMV[2] = fData->currentQMV[0]; |
| 1703 |
|
SubpelRefine(&bData); |
| 1704 |
|
} |
| 1705 |
|
|
| 1706 |
|
*fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode. |
| 1707 |
|
|
| 1708 |
|
if (*fData->iMinSAD < *best_sad) { |
| 1709 |
|
*best_sad = *fData->iMinSAD; |
| 1710 |
|
pMB->mvs[0] = fData->currentMV[0]; |
| 1711 |
|
pMB->b_mvs[0] = fData->currentMV[1]; |
| 1712 |
|
pMB->mode = MODE_INTERPOLATE; |
| 1713 |
|
if (fData->qpel) { |
| 1714 |
|
pMB->qmvs[0] = fData->currentQMV[0]; |
| 1715 |
|
pMB->b_qmvs[0] = fData->currentQMV[1]; |
| 1716 |
|
pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x; |
| 1717 |
|
pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y; |
| 1718 |
|
pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x; |
| 1719 |
|
pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y; |
| 1720 |
|
} else { |
| 1721 |
|
pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; |
| 1722 |
|
pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; |
| 1723 |
|
pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; |
| 1724 |
|
pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; |
| 1725 |
|
} |
| 1726 |
|
} |
| 1727 |
|
} |
| 1728 |
|
|
| 1729 |
|
void |
| 1730 |
|
MotionEstimationBVOP(MBParam * const pParam, |
| 1731 |
|
FRAMEINFO * const frame, |
| 1732 |
|
const int32_t time_bp, |
| 1733 |
|
const int32_t time_pp, |
| 1734 |
|
// forward (past) reference |
| 1735 |
|
const MACROBLOCK * const f_mbs, |
| 1736 |
|
const IMAGE * const f_ref, |
| 1737 |
|
const IMAGE * const f_refH, |
| 1738 |
|
const IMAGE * const f_refV, |
| 1739 |
|
const IMAGE * const f_refHV, |
| 1740 |
|
// backward (future) reference |
| 1741 |
|
const FRAMEINFO * const b_reference, |
| 1742 |
|
const IMAGE * const b_ref, |
| 1743 |
|
const IMAGE * const b_refH, |
| 1744 |
|
const IMAGE * const b_refV, |
| 1745 |
|
const IMAGE * const b_refHV) |
| 1746 |
|
{ |
| 1747 |
|
uint32_t i, j; |
| 1748 |
|
int32_t best_sad; |
| 1749 |
|
uint32_t skip_sad; |
| 1750 |
|
int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; |
| 1751 |
|
static const VECTOR zeroMV={0,0}; |
| 1752 |
|
const MACROBLOCK * const b_mbs = b_reference->mbs; |
| 1753 |
|
|
| 1754 |
|
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
| 1755 |
|
|
| 1756 |
|
const int32_t TRB = time_pp - time_bp; |
| 1757 |
|
const int32_t TRD = time_pp; |
| 1758 |
|
uint8_t * qimage; |
| 1759 |
|
|
| 1760 |
|
// some pre-inintialized data for the rest of the search |
| 1761 |
|
|
| 1762 |
|
SearchData Data; |
| 1763 |
|
int32_t iMinSAD; |
| 1764 |
|
VECTOR currentMV[3]; |
| 1765 |
|
VECTOR currentQMV[3]; |
| 1766 |
|
memset(&Data, 0, sizeof(SearchData)); |
| 1767 |
|
Data.iEdgedWidth = pParam->edged_width; |
| 1768 |
|
Data.currentMV = currentMV; Data.currentQMV = currentQMV; |
| 1769 |
|
Data.iMinSAD = &iMinSAD; |
| 1770 |
|
Data.lambda16 = lambda_vec16[frame->quant]; |
| 1771 |
|
Data.chroma = frame->quant; |
| 1772 |
|
Data.qpel = pParam->m_quarterpel; |
| 1773 |
|
Data.rounding = 0; |
| 1774 |
|
|
| 1775 |
|
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
| 1776 |
|
return; // allocate some mem for qpel interpolated blocks |
| 1777 |
|
// somehow this is dirty since I think we shouldn't use malloc outside |
| 1778 |
|
// encoder_create() - so please fix me! |
| 1779 |
|
Data.RefQ = qimage; |
| 1780 |
|
|
| 1781 |
const int32_t iFcode = pParam->fixed_code; |
// note: i==horizontal, j==vertical |
| 1782 |
const int32_t iQuant = pParam->quant; |
for (j = 0; j < pParam->mb_height; j++) { |
| 1783 |
const int32_t iWidth = pParam->width; |
|
| 1784 |
const int32_t iHeight = pParam->height; |
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
| 1785 |
const int32_t iEdgedWidth = pParam->edged_width; |
|
| 1786 |
|
for (i = 0; i < pParam->mb_width; i++) { |
| 1787 |
|
MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; |
| 1788 |
|
const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; |
| 1789 |
|
|
| 1790 |
|
/* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */ |
| 1791 |
|
if (b_reference->coding_type != S_VOP) |
| 1792 |
|
if (b_mb->mode == MODE_NOT_CODED) { |
| 1793 |
|
pMB->mode = MODE_NOT_CODED; |
| 1794 |
|
continue; |
| 1795 |
|
} |
| 1796 |
|
|
| 1797 |
const uint8_t * cur = pCur->y + x*8 + y*8*iEdgedWidth; |
Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
| 1798 |
|
pMB->quant = frame->quant; |
| 1799 |
|
|
| 1800 |
int32_t iDiamondSize; |
/* direct search comes first, because it (1) checks for SKIP-mode |
| 1801 |
|
and (2) sets very good predictions for forward and backward search */ |
| 1802 |
|
skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
| 1803 |
|
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
| 1804 |
|
&frame->image, |
| 1805 |
|
i, j, |
| 1806 |
|
frame->motion_flags, |
| 1807 |
|
TRB, TRD, |
| 1808 |
|
pParam, |
| 1809 |
|
pMB, b_mb, |
| 1810 |
|
&best_sad, |
| 1811 |
|
&Data); |
| 1812 |
|
|
| 1813 |
|
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
| 1814 |
|
|
| 1815 |
|
// forward search |
| 1816 |
|
SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
| 1817 |
|
&frame->image, i, j, |
| 1818 |
|
frame->motion_flags, |
| 1819 |
|
frame->fcode, pParam, |
| 1820 |
|
pMB, &f_predMV, &best_sad, |
| 1821 |
|
MODE_FORWARD, &Data); |
| 1822 |
|
|
| 1823 |
|
// backward search |
| 1824 |
|
SearchBF(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
| 1825 |
|
&frame->image, i, j, |
| 1826 |
|
frame->motion_flags, |
| 1827 |
|
frame->bcode, pParam, |
| 1828 |
|
pMB, &b_predMV, &best_sad, |
| 1829 |
|
MODE_BACKWARD, &Data); |
| 1830 |
|
|
| 1831 |
|
// interpolate search comes last, because it uses data from forward and backward as prediction |
| 1832 |
|
SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
| 1833 |
|
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
| 1834 |
|
&frame->image, |
| 1835 |
|
i, j, |
| 1836 |
|
frame->fcode, frame->bcode, |
| 1837 |
|
frame->motion_flags, |
| 1838 |
|
pParam, |
| 1839 |
|
&f_predMV, &b_predMV, |
| 1840 |
|
pMB, &best_sad, |
| 1841 |
|
&Data); |
| 1842 |
|
|
| 1843 |
|
// final skip decision |
| 1844 |
|
if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP*2) |
| 1845 |
|
&& ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) |
| 1846 |
|
SkipDecisionB(&frame->image, f_ref, b_ref, pMB,frame->quant, i, j, &Data); |
| 1847 |
|
|
| 1848 |
|
switch (pMB->mode) { |
| 1849 |
|
case MODE_FORWARD: |
| 1850 |
|
f_count++; |
| 1851 |
|
if (Data.qpel) f_predMV = pMB->qmvs[0]; |
| 1852 |
|
else f_predMV = pMB->mvs[0]; |
| 1853 |
|
break; |
| 1854 |
|
case MODE_BACKWARD: |
| 1855 |
|
b_count++; |
| 1856 |
|
if (Data.qpel) b_predMV = pMB->b_qmvs[0]; |
| 1857 |
|
else b_predMV = pMB->b_mvs[0]; |
| 1858 |
|
break; |
| 1859 |
|
case MODE_INTERPOLATE: |
| 1860 |
|
i_count++; |
| 1861 |
|
if (Data.qpel) { |
| 1862 |
|
f_predMV = pMB->qmvs[0]; |
| 1863 |
|
b_predMV = pMB->b_qmvs[0]; |
| 1864 |
|
} else { |
| 1865 |
|
f_predMV = pMB->mvs[0]; |
| 1866 |
|
b_predMV = pMB->b_mvs[0]; |
| 1867 |
|
} |
| 1868 |
|
break; |
| 1869 |
|
case MODE_DIRECT: |
| 1870 |
|
case MODE_DIRECT_NO4V: |
| 1871 |
|
d_count++; |
| 1872 |
|
default: |
| 1873 |
|
break; |
| 1874 |
|
} |
| 1875 |
|
} |
| 1876 |
|
} |
| 1877 |
|
free(qimage); |
| 1878 |
|
} |
| 1879 |
|
|
| 1880 |
int32_t min_dx; |
static __inline void |
| 1881 |
int32_t max_dx; |
MEanalyzeMB ( const uint8_t * const pRef, |
| 1882 |
int32_t min_dy; |
const uint8_t * const pCur, |
| 1883 |
int32_t max_dy; |
const int x, |
| 1884 |
|
const int y, |
| 1885 |
VECTOR pmv[4]; |
const MBParam * const pParam, |
| 1886 |
int32_t psad[4]; |
MACROBLOCK * const pMBs, |
| 1887 |
VECTOR newMV; |
SearchData * const Data) |
| 1888 |
VECTOR backupMV; |
{ |
| 1889 |
|
|
| 1890 |
MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
int i, mask; |
| 1891 |
|
VECTOR pmv[3]; |
| 1892 |
|
MACROBLOCK * pMB = &pMBs[x + y * pParam->mb_width]; |
| 1893 |
|
|
| 1894 |
static int32_t threshA,threshB; |
for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
|
int32_t iFound,bPredEq; |
|
|
int32_t iMinSAD,iSAD; |
|
| 1895 |
|
|
| 1896 |
int32_t iSubBlock = ((y&1)<<1) + (x&1); |
//median is only used as prediction. it doesn't have to be real |
| 1897 |
|
if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; |
| 1898 |
|
else |
| 1899 |
|
if (x == 1) //left macroblock does not have any vector now |
| 1900 |
|
Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median |
| 1901 |
|
else if (y == 1) // top macroblock doesn't have it's vector |
| 1902 |
|
Data->predMV = (pMB - 1)->mvs[0]; // left instead of median |
| 1903 |
|
else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median |
| 1904 |
|
|
| 1905 |
/* Get maximum range */ |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
| 1906 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, Data->rrv); |
|
x, y, 8, iWidth, iHeight, iFcode); |
|
| 1907 |
|
|
| 1908 |
/* we work with abs. MVs, not relative to prediction, so range is relative to 0,0 */ |
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
| 1909 |
|
Data->Ref = pRef + (x + y * pParam->edged_width) * 16; |
| 1910 |
|
|
| 1911 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8 )) |
pmv[1].x = EVEN(pMB->mvs[0].x); |
| 1912 |
{ min_dx = EVEN(min_dx); |
pmv[1].y = EVEN(pMB->mvs[0].y); |
| 1913 |
max_dx = EVEN(max_dx); |
pmv[2].x = EVEN(Data->predMV.x); |
| 1914 |
min_dy = EVEN(min_dy); |
pmv[2].y = EVEN(Data->predMV.y); |
| 1915 |
max_dy = EVEN(max_dy); |
pmv[0].x = pmv[0].y = 0; |
|
} /* because we might use IF (dx>max_dx) THEN dx=max_dx; */ |
|
| 1916 |
|
|
| 1917 |
|
CheckCandidate32I(0, 0, 255, &i, Data); |
| 1918 |
|
|
| 1919 |
bPredEq = get_pmvdata(pMBs, (x>>1), (y>>1), iWcount, iSubBlock, pmv, psad); |
if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) { |
| 1920 |
|
|
| 1921 |
if ((x==0) && (y==0) ) |
if (!(mask = make_mask(pmv, 1))) |
| 1922 |
{ |
CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data); |
| 1923 |
threshA = 512/4; |
if (!(mask = make_mask(pmv, 2))) |
| 1924 |
threshB = 1024/4; |
CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data); |
| 1925 |
|
|
| 1926 |
|
if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) // diamond only if needed |
| 1927 |
|
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
| 1928 |
|
|
| 1929 |
|
for (i = 0; i < 4; i++) { |
| 1930 |
|
MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1) * pParam->mb_width)]; |
| 1931 |
|
MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i]; |
| 1932 |
|
MB->mode = MODE_INTER; |
| 1933 |
|
MB->sad16 = Data->iMinSAD[i+1]; |
| 1934 |
|
} |
| 1935 |
} |
} |
|
else |
|
|
{ |
|
|
threshA = psad[0]/4; /* good estimate */ |
|
|
threshB = threshA+256/4; |
|
|
if (threshA< 512/4) threshA = 512/4; |
|
|
if (threshA>1024/4) threshA = 1024/4; |
|
|
if (threshB>1792/4) threshB = 1792/4; |
|
| 1936 |
} |
} |
| 1937 |
|
|
| 1938 |
iFound=0; |
#define INTRA_BIAS 2500 |
| 1939 |
|
#define INTRA_THRESH 1500 |
| 1940 |
/* Step 2: Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
#define INTER_THRESH 1400 |
|
vector of the median. |
|
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
|
|
*/ |
|
| 1941 |
|
|
|
if ((bPredEq) && (MVequal(pmv[0],pMB->mvs[iSubBlock]) ) ) |
|
|
iFound=2; |
|
| 1942 |
|
|
| 1943 |
/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
int |
| 1944 |
Otherwise select large Diamond Search. |
MEanalysis( const IMAGE * const pRef, |
| 1945 |
*/ |
FRAMEINFO * const Current, |
| 1946 |
|
MBParam * const pParam, |
| 1947 |
|
int maxIntra, //maximum number if non-I frames |
| 1948 |
|
int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame |
| 1949 |
|
int bCount) // number if B frames in a row |
| 1950 |
|
{ |
| 1951 |
|
uint32_t x, y, intra = 0; |
| 1952 |
|
int sSAD = 0; |
| 1953 |
|
MACROBLOCK * const pMBs = Current->mbs; |
| 1954 |
|
const IMAGE * const pCurrent = &Current->image; |
| 1955 |
|
int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH; |
| 1956 |
|
const VECTOR zeroMV = {0,0}; |
| 1957 |
|
|
| 1958 |
|
int32_t iMinSAD[5], temp[5]; |
| 1959 |
|
VECTOR currentMV[5]; |
| 1960 |
|
SearchData Data; |
| 1961 |
|
Data.iEdgedWidth = pParam->edged_width; |
| 1962 |
|
Data.currentMV = currentMV; |
| 1963 |
|
Data.iMinSAD = iMinSAD; |
| 1964 |
|
Data.iFcode = Current->fcode; |
| 1965 |
|
Data.rrv = Current->global_flags & XVID_REDUCED; |
| 1966 |
|
Data.temp = temp; |
| 1967 |
|
CheckCandidate = CheckCandidate32I; |
| 1968 |
|
|
| 1969 |
if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536/4) || (bPredEq) ) |
if (intraCount < 10) // we're right after an I frame |
| 1970 |
iDiamondSize=1; // 1 halfpel! |
IntraThresh += 4 * (intraCount - 10) * (intraCount - 10); |
| 1971 |
else |
else |
| 1972 |
iDiamondSize=2; // 2 halfpel = 1 full pixel! |
if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec |
| 1973 |
|
IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; |
| 1974 |
|
|
| 1975 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8) ) |
InterThresh += 400 * (1 - bCount); |
| 1976 |
iDiamondSize*=2; |
if (InterThresh < 300) InterThresh = 300; |
| 1977 |
|
|
| 1978 |
/* Step 4: Calculate SAD around the Median prediction. |
if (sadInit) (*sadInit) (); |
|
MinSAD=SAD |
|
|
If Motion Vector equal to Previous frame motion vector |
|
|
and MinSAD<PrevFrmSAD goto Step 10. |
|
|
If SAD<=256 goto Step 10. |
|
|
*/ |
|
| 1979 |
|
|
| 1980 |
|
for (y = 1; y < pParam->mb_height-1; y+=2) { |
| 1981 |
|
for (x = 1; x < pParam->mb_width-1; x+=2) { |
| 1982 |
|
int i; |
| 1983 |
|
|
| 1984 |
// Prepare for main loop |
if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; |
| 1985 |
|
|
| 1986 |
currMV->x=start_x; /* start with mv16 */ |
MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data); |
|
currMV->y=start_y; |
|
| 1987 |
|
|
| 1988 |
iMinSAD = sad8( cur, |
for (i = 0; i < 4; i++) { |
| 1989 |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
int dev; |
| 1990 |
iEdgedWidth); |
MACROBLOCK *pMB = &pMBs[x+(i&1) + y+(i>>1) * pParam->mb_width]; |
| 1991 |
iMinSAD += calc_delta_8(currMV->x - pmv[0].x, currMV->y - pmv[0].y, (uint8_t)iFcode) * iQuant; |
if (pMB->sad16 > IntraThresh) { |
| 1992 |
|
dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1))* pParam->edged_width) * 16, |
| 1993 |
if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,pMB->mvs[iSubBlock])) && (iMinSAD < pMB->sad8[iSubBlock]) ) ) |
pParam->edged_width); |
| 1994 |
{ |
if (dev + IntraThresh < pMB->sad16) { |
| 1995 |
if (MotionFlags & PMV_QUICKSTOP8) |
pMB->mode = MODE_INTRA; |
| 1996 |
goto step10_8b; |
if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return I_VOP; |
|
if (MotionFlags & PMV_EARLYSTOP8) |
|
|
goto step10_8; |
|
| 1997 |
} |
} |
|
|
|
|
/* |
|
|
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
|
|
Also calculate (0,0) but do not subtract offset. |
|
|
Let MinSAD be the smallest SAD up to this point. |
|
|
If MV is (0,0) subtract offset. ******** WHAT'S THIS 'OFFSET' ??? *********** |
|
|
*/ |
|
|
|
|
|
// the prediction might be even better than mv16 |
|
|
CHECK_MV8_CANDIDATE(pmv[0].x,pmv[0].y); |
|
|
|
|
|
// (0,0) is always possible |
|
|
CHECK_MV8_ZERO; |
|
|
|
|
|
// previous frame MV is always possible |
|
|
CHECK_MV8_CANDIDATE(pMB->mvs[iSubBlock].x,pMB->mvs[iSubBlock].y); |
|
|
|
|
|
// left neighbour, if allowed |
|
|
if (psad[1] != MV_MAX_ERROR) |
|
|
{ |
|
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
|
|
{ pmv[1].x = EVEN(pmv[1].x); |
|
|
pmv[1].y = EVEN(pmv[1].y); |
|
| 1998 |
} |
} |
| 1999 |
CHECK_MV8_CANDIDATE(pmv[1].x,pmv[1].y); |
sSAD += pMB->sad16; |
| 2000 |
} |
} |
| 2001 |
|
} |
| 2002 |
|
} |
| 2003 |
|
sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); |
| 2004 |
|
if (sSAD > IntraThresh + INTRA_BIAS ) return I_VOP; |
| 2005 |
|
if (sSAD > InterThresh ) return P_VOP; |
| 2006 |
|
emms(); |
| 2007 |
|
return B_VOP; |
| 2008 |
|
|
|
// top neighbour, if allowed |
|
|
if (psad[2] != MV_MAX_ERROR) |
|
|
{ |
|
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
|
|
{ pmv[2].x = EVEN(pmv[2].x); |
|
|
pmv[2].y = EVEN(pmv[2].y); |
|
| 2009 |
} |
} |
|
CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); |
|
| 2010 |
|
|
| 2011 |
// top right neighbour, if allowed |
static void |
| 2012 |
if (psad[3] != MV_MAX_ERROR) |
CheckGMC(int x, int y, const int dir, int * iDirection, |
| 2013 |
|
const MACROBLOCK * const pMBs, uint32_t * bestcount, VECTOR * GMC, |
| 2014 |
|
const MBParam * const pParam) |
| 2015 |
{ |
{ |
| 2016 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
uint32_t mx, my, a, count = 0; |
| 2017 |
{ pmv[3].x = EVEN(pmv[3].x); |
|
| 2018 |
pmv[3].y = EVEN(pmv[3].y); |
for (my = 1; my < pParam->mb_height-1; my++) |
| 2019 |
|
for (mx = 1; mx < pParam->mb_width-1; mx++) { |
| 2020 |
|
VECTOR mv; |
| 2021 |
|
const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; |
| 2022 |
|
if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) continue; |
| 2023 |
|
mv = pMB->mvs[0]; |
| 2024 |
|
a = ABS(mv.x - x) + ABS(mv.y - y); |
| 2025 |
|
if (a < 6) count += 6 - a; |
| 2026 |
} |
} |
| 2027 |
CHECK_MV8_CANDIDATE(pmv[3].x,pmv[3].y); |
|
| 2028 |
|
if (count > *bestcount) { |
| 2029 |
|
*bestcount = count; |
| 2030 |
|
*iDirection = dir; |
| 2031 |
|
GMC->x = x; GMC->y = y; |
| 2032 |
} |
} |
| 2033 |
} |
} |
| 2034 |
|
|
|
/* Step 6: If MinSAD <= thresa goto Step 10. |
|
|
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
|
|
*/ |
|
| 2035 |
|
|
| 2036 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,pMB->mvs[iSubBlock]) && (iMinSAD < pMB->sad8[iSubBlock]) ) ) |
static VECTOR |
| 2037 |
|
GlobalMotionEst(const MACROBLOCK * const pMBs, const MBParam * const pParam, const uint32_t iFcode) |
| 2038 |
{ |
{ |
|
if (MotionFlags & PMV_QUICKSTOP8) |
|
|
goto step10_8b; |
|
|
if (MotionFlags & PMV_EARLYSTOP8) |
|
|
goto step10_8; |
|
|
} |
|
|
|
|
|
/************ (Diamond Search) **************/ |
|
|
/* |
|
|
Step 7: Perform Diamond search, with either the small or large diamond. |
|
|
If Found=2 only examine one Diamond pattern, and afterwards goto step 10 |
|
|
Step 8: If small diamond, iterate small diamond search pattern until motion vector lies in the center of the diamond. |
|
|
If center then goto step 10. |
|
|
Step 9: If large diamond, iterate large diamond search pattern until motion vector lies in the center. |
|
|
Refine by using small diamond and goto step 10. |
|
|
*/ |
|
| 2039 |
|
|
| 2040 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
uint32_t count, bestcount = 0; |
| 2041 |
|
int x, y; |
| 2042 |
|
VECTOR gmc = {0,0}; |
| 2043 |
|
int step, min_x, max_x, min_y, max_y; |
| 2044 |
|
uint32_t mx, my; |
| 2045 |
|
int iDirection, bDirection; |
| 2046 |
|
|
| 2047 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
min_x = min_y = -32<<iFcode; |
| 2048 |
iSAD = PMVfastSearch8_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
max_x = max_y = 32<<iFcode; |
|
x, y, |
|
|
currMV->x, currMV->y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) |
|
|
{ |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
| 2049 |
|
|
| 2050 |
if (MotionFlags & PMV_EXTSEARCH8) |
//step1: let's find a rough camera panning |
| 2051 |
{ |
for (step = 32; step >= 2; step /= 2) { |
| 2052 |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
bestcount = 0; |
| 2053 |
|
for (y = min_y; y <= max_y; y += step) |
| 2054 |
|
for (x = min_x ; x <= max_x; x += step) { |
| 2055 |
|
count = 0; |
| 2056 |
|
//for all macroblocks |
| 2057 |
|
for (my = 1; my < pParam->mb_height-1; my++) |
| 2058 |
|
for (mx = 1; mx < pParam->mb_width-1; mx++) { |
| 2059 |
|
const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; |
| 2060 |
|
VECTOR mv; |
| 2061 |
|
|
| 2062 |
if (!(MVequal(pmv[0],backupMV)) ) |
if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) |
| 2063 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
continue; |
|
x, y, |
|
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
| 2064 |
|
|
| 2065 |
if (iSAD < iMinSAD) |
mv = pMB->mvs[0]; |
| 2066 |
{ |
if ( ABS(mv.x - x) <= step && ABS(mv.y - y) <= step ) /* GMC translation is always halfpel-res */ |
| 2067 |
*currMV = newMV; |
count++; |
|
iMinSAD = iSAD; |
|
| 2068 |
} |
} |
| 2069 |
|
if (count >= bestcount) { bestcount = count; gmc.x = x; gmc.y = y; } |
| 2070 |
} |
} |
| 2071 |
|
min_x = gmc.x - step; |
| 2072 |
|
max_x = gmc.x + step; |
| 2073 |
|
min_y = gmc.y - step; |
| 2074 |
|
max_y = gmc.y + step; |
| 2075 |
|
|
|
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
|
|
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
0, 0, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) |
|
|
{ |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
| 2076 |
} |
} |
| 2077 |
|
|
| 2078 |
/* Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
if (bestcount < (pParam->mb_height-2)*(pParam->mb_width-2)/10) |
| 2079 |
By performing an optional local half-pixel search, we can refine this result even further. |
gmc.x = gmc.y = 0; //no camara pan, no GMC |
|
*/ |
|
| 2080 |
|
|
| 2081 |
step10_8: |
// step2: let's refine camera panning using gradiend-descent approach. |
| 2082 |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
// TODO: more warping points may be evaluated here (like in interpolate mode search - two vectors in one diamond) |
| 2083 |
iMinSAD = PMVfastSearch8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
bestcount = 0; |
| 2084 |
x, y, |
CheckGMC(gmc.x, gmc.y, 255, &iDirection, pMBs, &bestcount, &gmc, pParam); |
| 2085 |
currMV, iMinSAD, |
do { |
| 2086 |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
x = gmc.x; y = gmc.y; |
| 2087 |
|
bDirection = iDirection; iDirection = 0; |
| 2088 |
|
if (bDirection & 1) CheckGMC(x - 1, y, 1+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); |
| 2089 |
|
if (bDirection & 2) CheckGMC(x + 1, y, 2+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); |
| 2090 |
|
if (bDirection & 4) CheckGMC(x, y - 1, 1+2+4, &iDirection, pMBs, &bestcount, &gmc, pParam); |
| 2091 |
|
if (bDirection & 8) CheckGMC(x, y + 1, 1+2+8, &iDirection, pMBs, &bestcount, &gmc, pParam); |
| 2092 |
|
|
| 2093 |
step10_8b: |
} while (iDirection); |
| 2094 |
|
|
| 2095 |
currPMV->x = currMV->x - pmv[0].x; |
if (pParam->m_quarterpel) { |
| 2096 |
currPMV->y = currMV->y - pmv[0].y; |
gmc.x *= 2; |
| 2097 |
|
gmc.y *= 2; /* we store the halfpel value as pseudo-qpel to make comparison easier */ |
| 2098 |
|
} |
| 2099 |
|
|
| 2100 |
return iMinSAD; |
return gmc; |
| 2101 |
} |
} |
Parent Directory
| 
Revision Log
| 
Patch