h.264 fast,1/2,1/4像素运动估计与插值处理
Hadamard Transform
在1/2，1/4像素运动估计这一阶段中，对于像素残差，可以选择采用哈达玛变换来代替离散余弦变换进行高低频的分离。
- 优点：哈达玛矩阵全是+1，-1，因此只需要进行加减法就可以得到变换结果，比离散余弦变换更高效
- 缺点：高低频分离效果没有离散余弦变换好，原始数据越是均匀分布，经转换后的数据越集中于边角，反之集中力越差
在采用了哈达玛变换的情况，为了达到更精确的估计效果，计算像素残差的SAD需要更变为SATD，即对像素残差进行哈达玛矩阵变换后得到的矩阵元素的绝对值之和
1/4像素参考图像像素采集，然后与当前块进行SAD方式如下：
/*!
************************************************************************
* \brief
* Functions for fast fractional分数、小数 pel motion estimation.
* 1. int AddUpSADQuarter() returns SADT of a fractiona pel MV
* 2. int FastSubPelBlockMotionSearch () proceed the fast fractional pel ME
* \authors: Zhibo Chen
* Dept.of EE, Tsinghua Univ.
* \date : 2003.4 计算分块代价
************************************************************************
*/
int AddUpSADQuarter(int pic_pix_x,int pic_pix_y,int blocksize_x,int blocksize_y,
int cand_mv_x,int cand_mv_y, StorablePicture *ref_picture, pel_t** orig_pic,
int Mvmcost, int min_mcost,int useABT)//adaptive block transform
{
int abort_search, y0, x0, rx0, ry0, ry;
pel_t *orig_line;
int diff[16], *d;
int mcost = Mvmcost;
int yy,kk,xx;
int curr_diff[MB_BLOCK_SIZE][MB_BLOCK_SIZE]; // for ABT SATD calculation
//2004.3.3
pel_t **ref_pic = ref_picture->imgY_ups;//经过插值的图像 参考图像
int img_width = ref_picture->size_x;
int img_height = ref_picture->size_y;
for (y0=0, abort_search=0; y0<blocksize_y && !abort_search; y0+=4)//4*4块
{
ry0 = ((pic_pix_y+y0)<<2) + cand_mv_y;
for (x0=0; x0<blocksize_x; x0+=4)
{
rx0 = ((pic_pix_x+x0)<<2) + cand_mv_x;
d = diff;
orig_line = orig_pic [y0 ]; ry=ry0;
*d++ = orig_line[x0 ] - PelY_14 (ref_pic, ry, rx0 , img_height, img_width);
*d++ = orig_line[x0+1] - PelY_14 (ref_pic, ry, rx0+ 4, img_height, img_width);
*d++ = orig_line[x0+2] - PelY_14 (ref_pic, ry, rx0+ 8, img_height, img_width);
*d++ = orig_line[x0+3] - PelY_14 (ref_pic, ry, rx0+ 12, img_height, img_width);
orig_line = orig_pic [y0+1]; ry=ry0+4;
*d++ = orig_line[x0 ] - PelY_14 (ref_pic, ry, rx0 , img_height, img_width);
*d++ = orig_line[x0+1] - PelY_14 (ref_pic, ry, rx0+ 4, img_height, img_width);
*d++ = orig_line[x0+2] - PelY_14 (ref_pic, ry, rx0+ 8, img_height, img_width);
*d++ = orig_line[x0+3] - PelY_14 (ref_pic, ry, rx0+ 12, img_height, img_width);
orig_line = orig_pic [y0+2]; ry=ry0+8;
*d++ = orig_line[x0 ] - PelY_14 (ref_pic, ry, rx0 , img_height, img_width);
*d++ = orig_line[x0+1] - PelY_14 (ref_pic, ry, rx0+ 4, img_height, img_width);
*d++ = orig_line[x0+2] - PelY_14 (ref_pic, ry, rx0+ 8, img_height, img_width);
*d++ = orig_line[x0+3] - PelY_14 (ref_pic, ry, rx0+ 12, img_height, img_width);
orig_line = orig_pic [y0+3]; ry=ry0+12;
*d++ = orig_line[x0 ] - PelY_14 (ref_pic, ry, rx0 , img_height, img_width);
*d++ = orig_line[x0+1] - PelY_14 (ref_pic, ry, rx0+ 4, img_height, img_width);
*d++ = orig_line[x0+2] - PelY_14 (ref_pic, ry, rx0+ 8, img_height, img_width);
*d = orig_line[x0+3] - PelY_14 (ref_pic, ry, rx0+ 12, img_height, img_width);
if (!useABT)
{
if ((mcost += SATD (diff, input->hadamard)) > min_mcost)
{
abort_search = 1;//终止搜索
break;
}
}
else // copy diff to curr_diff for ABT SATD calculation
{
for (yy=y0,kk=0; yy<y0+4; yy++)
for (xx=x0; xx<x0+4; xx++, kk++)
curr_diff[yy][xx] = diff[kk];
}
}
}
return mcost;
}
View Code
快速子像素运动估计
采用1/4像素为基准进行运动向量搜索，搜索起始点周围7x7的1/4像素范围（即当前像素与周围像素点之间的1/4像素点）。该模式只在子宏块（blocktype>3）时使用
在进行搜索之前，需要确定搜索中心点，在前面的整像素搜索之后，得到的最佳搜索位置为整像素点（mvInt），但是预测运动向量（mvp）有可能不是处于整像素的位置，因此在确定搜索中心时，需要在mvInt与mvp间取最低cost的一个作为搜索中心
搜索方式为the third step with a small search pattern（小菱形模板反复搜索）
小范围的子像素搜索是线性的（JVT-F017r1 3.1.2），即基本上都是朝着一个方向进行的，因此循环次数取边界长度7，而不会去搜索该范围内所有的位置
int // ==> minimum motion cost after search
FastSubPelBlockMotionSearch (pel_t** orig_pic, // <-- original pixel values for the AxB block
int ref, // <-- reference frame (0... or -1 (backward))
int list,
int pic_pix_x, // <-- absolute x-coordinate of regarded AxB block
int pic_pix_y, // <-- absolute y-coordinate of regarded AxB block
int blocktype, // <-- block type (1-16x16 ... 7-4x4)
int pred_mv_x, // <-- motion vector predictor (x) in sub-pel units
int pred_mv_y, // <-- motion vector predictor (y) in sub-pel units
int* mv_x, // <--> in: search center (x) / out: motion vector (x) - in pel units
int* mv_y, // <--> in: search center (y) / out: motion vector (y) - in pel units
int search_pos2, // <-- search positions for half-pel search (default: 9)
int search_pos4, // <-- search positions for quarter-pel search (default: 9)
int min_mcost, // <-- minimum motion cost (cost for center or huge value)
double lambda,
int useABT) // <-- lagrangian parameter for determining motion cost
{
static int Diamond_x[4] = {-1, 0, 1, 0};//菱形搜索
static int Diamond_y[4] = {0, 1, 0, -1};
int mcost;
int cand_mv_x, cand_mv_y;
int incr = list==1 ? ((!img->fld_type)&&(enc_picture!=enc_frame_picture)&&(img->type==B_SLICE)) : (enc_picture==enc_frame_picture)&&(img->type==B_SLICE) ;//未见使用
int list_offset = ((img->MbaffFrameFlag)&&(img->mb_data[img->current_mb_nr].mb_field))? img->current_mb_nr%2 ? 4 : 2 : 0;
StorablePicture *ref_picture = listX[list+list_offset][ref];
pel_t **ref_pic = ref_picture->imgY_ups;
int lambda_factor = LAMBDA_FACTOR (lambda);
int mv_shift = 0;
int check_position0 = (blocktype==1 && *mv_x==0 && *mv_y==0 && input->hadamard && !input->rdopt && img->type!=B_SLICE && ref==0);
int blocksize_x = input->blc_size[blocktype][0];
int blocksize_y = input->blc_size[blocktype][1];
int pic4_pix_x = (pic_pix_x << 2);
int pic4_pix_y = (pic_pix_y << 2);
int max_pos_x4 = ((ref_picture->size_x/*img->width*/-blocksize_x+1)<<2);
int max_pos_y4 = ((ref_picture->size_y/*img->height*/-blocksize_y+1)<<2);
int min_pos2 = (input->hadamard ? 0 : 1);
int max_pos2 = (input->hadamard ? max(1,search_pos2) : search_pos2);
int search_range_dynamic,iXMinNow,iYMinNow,i;
int iSADLayer,m,currmv_x,currmv_y,iCurrSearchRange;
int search_range = input->search_range;
int pred_frac_mv_x,pred_frac_mv_y,abort_search;
int mv_cost;
int pred_frac_up_mv_x, pred_frac_up_mv_y;
*mv_x <<= 2;
*mv_y <<= 2;
if ((pic4_pix_x + *mv_x > 1) && (pic4_pix_x + *mv_x < max_pos_x4 - 2) &&
(pic4_pix_y + *mv_y > 1) && (pic4_pix_y + *mv_y < max_pos_y4 - 2) )//判断是否出界
{
PelY_14 = FastPelY_14;
}
else
{
PelY_14 = UMVPelY_14;
}
search_range_dynamic = 3;//dynamic动态
pred_frac_mv_x = (pred_mv_x - *mv_x)%4;
pred_frac_mv_y = (pred_mv_y - *mv_y)%4;
pred_frac_up_mv_x = (pred_MV_uplayer[0] - *mv_x)%4;
pred_frac_up_mv_y = (pred_MV_uplayer[1] - *mv_y)%4;
//SearchState 小数 pel search 插值像素搜索标记 ,搜索过的位置会被标记，被标记过的位置不会重新搜索
memset(SearchState[0],0,(2*search_range_dynamic+1)*(2*search_range_dynamic+1));
if(input->hadamard)
{
cand_mv_x = *mv_x;
cand_mv_y = *mv_y;
mv_cost = MV_COST (lambda_factor, mv_shift, cand_mv_x, cand_mv_y, pred_mv_x, pred_mv_y);
mcost = AddUpSADQuarter(pic_pix_x,pic_pix_y,blocksize_x,blocksize_y,cand_mv_x,cand_mv_y,ref_picture/*ref_pic*//*Wenfang Fu 2004.3.12*/,orig_pic,mv_cost,min_mcost,useABT);
//计算分块和MV总代价
SearchState[search_range_dynamic][search_range_dynamic] = 1;
if (mcost < min_mcost)
{
min_mcost = mcost;
currmv_x = cand_mv_x;
currmv_y = cand_mv_y;
}
}
else
{
SearchState[search_range_dynamic][search_range_dynamic] = 1;
currmv_x = *mv_x;
currmv_y = *mv_y;
}
if(pred_frac_mv_x!=0 || pred_frac_mv_y!=0)//mvp
{
cand_mv_x = *mv_x + pred_frac_mv_x;
cand_mv_y = *mv_y + pred_frac_mv_y;
mv_cost = MV_COST (lambda_factor, mv_shift, cand_mv_x, cand_mv_y, pred_mv_x, pred_mv_y);
mcost = AddUpSADQuarter(pic_pix_x,pic_pix_y,blocksize_x,blocksize_y,cand_mv_x,cand_mv_y,ref_picture/*ref_pic*//*Wenfang Fu 2004.3.12*/,orig_pic,mv_cost,min_mcost,useABT);
SearchState[cand_mv_y -*mv_y + search_range_dynamic][cand_mv_x - *mv_x + search_range_dynamic] = 1;
if (mcost < min_mcost)
{
min_mcost = mcost;
currmv_x = cand_mv_x;
currmv_y = cand_mv_y;
}
}
iXMinNow = currmv_x;//iXMinNow当前搜索中心 currmv_x最佳MV
iYMinNow = currmv_y;
iCurrSearchRange = 2*search_range_dynamic+1;
//采用iCurrSearchRange,总面积是iCurrSearchRange*iCurrSearchRange
//而采用一条边作为搜索范围，表明是单一方向的搜索，
//为什么呢，因为小范围内的搜索一般都是低频，即搜索方向不会有太大变化
//JVT-F017r1 3.1.2
for(i=0;i<iCurrSearchRange;i++)
{
abort_search=1;//终止搜索标记 1终止 0继续
iSADLayer = 65536;
for (m = 0; m < 4; m++)
{
cand_mv_x = iXMinNow + Diamond_x[m];
cand_mv_y = iYMinNow + Diamond_y[m];
if(abs(cand_mv_x - *mv_x) <=search_range_dynamic && abs(cand_mv_y - *mv_y)<= search_range_dynamic)
{
if(!SearchState[cand_mv_y -*mv_y+ search_range_dynamic][cand_mv_x -*mv_x+ search_range_dynamic])//未被搜索
{
mv_cost = MV_COST (lambda_factor, mv_shift, cand_mv_x, cand_mv_y, pred_mv_x, pred_mv_y);
mcost = AddUpSADQuarter(pic_pix_x,pic_pix_y,blocksize_x,blocksize_y,cand_mv_x,cand_mv_y,ref_picture/*ref_pic*//*Wenfang Fu 2004.3.12*/,orig_pic,mv_cost,min_mcost,useABT);
SearchState[cand_mv_y - *mv_y + search_range_dynamic][cand_mv_x - *mv_x + search_range_dynamic] = 1;
if (mcost < min_mcost)
{
min_mcost = mcost;
currmv_x = cand_mv_x;
currmv_y = cand_mv_y;
abort_search = 0;
}
}
}
}
iXMinNow = currmv_x;
iYMinNow = currmv_y;
if(abort_search)
break;
}
*mv_x = currmv_x;
*mv_y = currmv_y;
//===== return minimum motion cost =====
return min_mcost;
}
View Code
子像素运动估计
搜索分为两个步骤：
- 搜索中心周围（包括中心）1/2像素的9个像素点，从中选择最佳位置
- 以最佳位置为中心，搜索周围1/4像素的9个像素点，从中选择最佳位置
*!
***********************************************************************
* \brief以确定的最佳像素点为中心，进行半像素估计，在以最佳半像素点为中心，进行
1/4像素估计
* Sub pixel block motion search
***********************************************************************
*/
int // ==> minimum motion cost after search
SubPelBlockMotionSearch (pel_t** orig_pic, // <-- original pixel values for the AxB block
int ref, // <-- reference frame (0... or -1 (backward))
int list, // <-- reference picture list
int pic_pix_x, // <-- absolute x-coordinate of regarded AxB block
int pic_pix_y, // <-- absolute y-coordinate of regarded AxB block
int blocktype, // <-- block type (1-16x16 ... 7-4x4)
int pred_mv_x, // <-- motion vector predictor (x) in sub-pel units
int pred_mv_y, // <-- motion vector predictor (y) in sub-pel units
int* mv_x, // <--> in: search center (x) / out: motion vector (x) - in pel units
int* mv_y, // <--> in: search center (y) / out: motion vector (y) - in pel units
int search_pos2, // <-- 半像素搜索点的位置 (default: 9)
int search_pos4, // <-- search positions for quarter-pel search (default: 9)
int min_mcost, // <-- minimum motion cost (cost for center or huge value)
double lambda // <-- lagrangian parameter for determining motion cost
)
{
int diff[16], *d;
int pos, best_pos, mcost, abort_search;
int y0, x0, ry0, rx0, ry;
int cand_mv_x, cand_mv_y;
int max_pos_x4, max_pos_y4;
pel_t *orig_line;
pel_t **ref_pic;
StorablePicture *ref_picture;
int lambda_factor = LAMBDA_FACTOR (lambda);
int mv_shift = 0;
int check_position0 = (blocktype==1 && *mv_x==0 && *mv_y==0 && input->hadamard && !input->rdopt && img->type!=B_SLICE && ref==0);
int blocksize_x = input->blc_size[blocktype][0];
int blocksize_y = input->blc_size[blocktype][1];
int pic4_pix_x = (pic_pix_x << 2);//换算成以1/4像素为单位的坐标
int pic4_pix_y = (pic_pix_y << 2);
int min_pos2 = (input->hadamard ? 0 : 1);
int max_pos2 = (input->hadamard ? max(1,search_pos2) : search_pos2);
int list_offset = ((img->MbaffFrameFlag)&&(img->mb_data[img->current_mb_nr].mb_field))? img->current_mb_nr%2 ? 4 : 2 : 0;
int apply_weights = ( (active_pps->weighted_pred_flag && (img->type == P_SLICE || img->type == SP_SLICE)) ||
(active_pps->weighted_bipred_idc && (img->type == B_SLICE))); //是否使用权重计算代价
int img_width, img_height;
ref_picture = listX[list+list_offset][ref];
if (apply_weights)
{
ref_pic = listX[list+list_offset][ref]->imgY_ups_w;
}
else
ref_pic = listX[list+list_offset][ref]->imgY_ups; //做过1/4像素插值的图像放在imgY_ups
img_width = ref_picture->size_x;
img_height = ref_picture->size_y;//未做插值的图像的宽高
max_pos_x4 = ((ref_picture->size_x - blocksize_x+1)<<2);//最大1/4像素搜索位置
max_pos_y4 = ((ref_picture->size_y - blocksize_y+1)<<2);
/*********************************
***** *****
***** HALF-PEL REFINEMENT *****
***** *****
*********************************/
//===== convert search center to quarter-pel units ===== 先变成1/4像素的运动向量
*mv_x <<= 2;
*mv_y <<= 2;
//===== set function for getting pixel values =====
if ((pic4_pix_x + *mv_x > 1) && (pic4_pix_x + *mv_x < max_pos_x4 - 2) &&
(pic4_pix_y + *mv_y > 1) && (pic4_pix_y + *mv_y < max_pos_y4 - 2) )//判断半像素是否出界
{
PelY_14 = FastPelY_14;//没出界
}
else
{
PelY_14 = UMVPelY_14;
}
//===== loop over search positions =====
for (best_pos = 0, pos = min_pos2; pos < max_pos2; pos++) //半像素搜索，只搜索最近的9个点(max_pos2==9)
{
cand_mv_x = *mv_x + (spiral_search_x[pos] << 1); // quarter-pel units
cand_mv_y = *mv_y + (spiral_search_y[pos] << 1); // quarter-pel units
//----- set motion vector cost -----
mcost = MV_COST (lambda_factor, mv_shift, cand_mv_x, cand_mv_y, pred_mv_x, pred_mv_y);
if (check_position0 && pos==0) //中央像素点位置为当前宏块位置
{
mcost -= WEIGHTED_COST (lambda_factor, 16);
}
if (mcost >= min_mcost) continue;
//----- add up SATD -----
for (y0=0, abort_search=0; y0<blocksize_y && !abort_search; y0+=4)
{
ry0 = ((pic_pix_y+y0)<<2) + cand_mv_y;//搜索中心
for (x0=0; x0<blocksize_x; x0+=4)//4*4块计算SATD
{
rx0 = ((pic_pix_x+x0)<<2) + cand_mv_x;
d = diff;//共享内存空间
//找到1/4像素的搜索中心，根据插值后的像素值与整数像素值同位置出相减
orig_line = orig_pic [y0 ]; ry=ry0;
*d++ = orig_line[x0 ] - PelY_14 (ref_pic, ry, rx0 , img_height, img_width);
*d++ = orig_line[x0+1] - PelY_14 (ref_pic, ry, rx0+ 4, img_height, img_width);
*d++ = orig_line[x0+2] - PelY_14 (ref_pic, ry, rx0+ 8, img_height, img_width);
*d++ = orig_line[x0+3] - PelY_14 (ref_pic, ry, rx0+12, img_height, img_width);
orig_line = orig_pic [y0+1]; ry=ry0+4;
*d++ = orig_line[x0 ] - PelY_14 (ref_pic, ry, rx0 , img_height, img_width);
*d++ = orig_line[x0+1] - PelY_14 (ref_pic, ry, rx0+ 4, img_height, img_width);
*d++ = orig_line[x0+2] - PelY_14 (ref_pic, ry, rx0+ 8, img_height, img_width);
*d++ = orig_line[x0+3] - PelY_14 (ref_pic, ry, rx0+12, img_height, img_width);
orig_line = orig_pic [y0+2]; ry=ry0+8;
*d++ = orig_line[x0 ] - PelY_14 (ref_pic, ry, rx0 , img_height, img_width);
*d++ = orig_line[x0+1] - PelY_14 (ref_pic, ry, rx0+ 4, img_height, img_width);
*d++ = orig_line[x0+2] - PelY_14 (ref_pic, ry, rx0+ 8, img_height, img_width);
*d++ = orig_line[x0+3] - PelY_14 (ref_pic, ry, rx0+12, img_height, img_width);
orig_line = orig_pic [y0+3]; ry=ry0+12;
*d++ = orig_line[x0 ] - PelY_14 (ref_pic, ry, rx0 , img_height, img_width);
*d++ = orig_line[x0+1] - PelY_14 (ref_pic, ry, rx0+ 4, img_height, img_width);
*d++ = orig_line[x0+2] - PelY_14 (ref_pic, ry, rx0+ 8, img_height, img_width);
*d = orig_line[x0+3] - PelY_14 (ref_pic, ry, rx0+12, img_height, img_width);
if ((mcost += SATD (diff, input->hadamard)) > min_mcost)
{
abort_search = 1;
break;
}//比最小代价大，终止搜索
}
}
if (mcost < min_mcost)
{
min_mcost = mcost;
best_pos = pos;
}
}
if (best_pos)
{
*mv_x += (spiral_search_x [best_pos] << 1);
*mv_y += (spiral_search_y [best_pos] << 1);
}
/************************************
***** *****
***** QUARTER-PEL REFINEMENT *****
***** *****
************************************/
//===== set function for getting pixel values =====
if ((pic4_pix_x + *mv_x > 1) && (pic4_pix_x + *mv_x < max_pos_x4 - 1) &&
(pic4_pix_y + *mv_y > 1) && (pic4_pix_y + *mv_y < max_pos_y4 - 1) )
{
PelY_14 = FastPelY_14;
}
else
{
PelY_14 = UMVPelY_14;
}
//===== loop over search positions =====
for (best_pos = 0, pos = 1; pos < search_pos4; pos++)//1/4像素搜索，只搜索最近的9个点(max_pos2==9)
{
cand_mv_x = *mv_x + spiral_search_x[pos]; // quarter-pel units
cand_mv_y = *mv_y + spiral_search_y[pos]; // quarter-pel units
//----- set motion vector cost -----
mcost = MV_COST (lambda_factor, mv_shift, cand_mv_x, cand_mv_y, pred_mv_x, pred_mv_y);
if (mcost >= min_mcost) continue;
//----- add up SATD -----
for (y0=0, abort_search=0; y0<blocksize_y && !abort_search; y0+=4)
{
ry0 = ((pic_pix_y+y0)<<2) + cand_mv_y;
for (x0=0; x0<blocksize_x; x0+=4)
{
rx0 = ((pic_pix_x+x0)<<2) + cand_mv_x;
d = diff;
orig_line = orig_pic [y0 ]; ry=ry0;
*d++ = orig_line[x0 ] - PelY_14 (ref_pic, ry, rx0 , img_height, img_width);
*d++ = orig_line[x0+1] - PelY_14 (ref_pic, ry, rx0+ 4, img_height, img_width);
*d++ = orig_line[x0+2] - PelY_14 (ref_pic, ry, rx0+ 8, img_height, img_width);
*d++ = orig_line[x0+3] - PelY_14 (ref_pic, ry, rx0+12, img_height, img_width);
orig_line = orig_pic [y0+1]; ry=ry0+4;
*d++ = orig_line[x0 ] - PelY_14 (ref_pic, ry, rx0 , img_height, img_width);
*d++ = orig_line[x0+1] - PelY_14 (ref_pic, ry, rx0+ 4, img_height, img_width);
*d++ = orig_line[x0+2] - PelY_14 (ref_pic, ry, rx0+ 8, img_height, img_width);
*d++ = orig_line[x0+3] - PelY_14 (ref_pic, ry, rx0+12, img_height, img_width);
orig_line = orig_pic [y0+2]; ry=ry0+8;
*d++ = orig_line[x0 ] - PelY_14 (ref_pic, ry, rx0 , img_height, img_width);
*d++ = orig_line[x0+1] - PelY_14 (ref_pic, ry, rx0+ 4, img_height, img_width);
*d++ = orig_line[x0+2] - PelY_14 (ref_pic, ry, rx0+ 8, img_height, img_width);
*d++ = orig_line[x0+3] - PelY_14 (ref_pic, ry, rx0+12, img_height, img_width);
orig_line = orig_pic [y0+3]; ry=ry0+12;
*d++ = orig_line[x0 ] - PelY_14 (ref_pic, ry, rx0 , img_height, img_width);
*d++ = orig_line[x0+1] - PelY_14 (ref_pic, ry, rx0+ 4, img_height, img_width);
*d++ = orig_line[x0+2] - PelY_14 (ref_pic, ry, rx0+ 8, img_height, img_width);
*d = orig_line[x0+3] - PelY_14 (ref_pic, ry, rx0+12, img_height, img_width);
if ((mcost += SATD (diff, input->hadamard)) > min_mcost)
{
abort_search = 1;
break;
}
}
}
if (mcost < min_mcost)
{
min_mcost = mcost;
best_pos = pos;
}
}
if (best_pos)
{
*mv_x += spiral_search_x [best_pos];
*mv_y += spiral_search_y [best_pos];
}
//===== return minimum motion cost =====
return min_mcost;
}
View Code
1/4像素插值处理
在做插值之前，需要先对原图像添加边框（Pad），即在原图上下左右四个方向分别添加边界像素点，像素点的取值为边界像素点的值，可以参考Edge Process
在JM8.6时，HPad与VPad都是4，而在JM18.6，HPad为32，VPad为20
插值过程可以分为6个步骤：
- 水平1/2像素通过6抽头滤波器插值
- 垂直1/2像素通过6抽头滤波器插值
- 水平1/4像素内插
- 垂直1/4像素内插
- 45°内插
- 反向45°内插
1/2像素插值
1/4像素水平、垂直插值 1/4像素正反45°插值
以上步骤是以JM8.6为蓝本，JM18.6有自己的方式，具体可以参考下面的code
JM8.6
*!
************************************************************************
* \brief
* Upsample 4 times, store them in out4x. Color is simply copied
*
* \par Input:
* srcy, srcu, srcv, out4y, out4u, out4v
*
* \par Side Effects_
* Uses (writes) img4Y_tmp. This should be moved to a static variable
* in this module
************************************************************************/
void UnifiedOneForthPix (StorablePicture *s)
{
int is;
int i, j, j4;
int ie2, je2, jj, maxy;
byte **out4Y;
byte *ref11;
byte **imgY = s->imgY;
int img_width =s->size_x;
int img_height=s->size_y;
// don't upsample twice
if (s->imgY_ups || s->imgY_11)
return;
s->imgY_11 = malloc ((s->size_x * s->size_y) * sizeof (byte));
if (NULL == s->imgY_11)
no_mem_exit("alloc_storable_picture: s->imgY_11");
get_mem2D (&(s->imgY_ups), (2*IMG_PAD_SIZE + s->size_y)*4, (2*IMG_PAD_SIZE + s->size_x)*4);
if (input->WeightedPrediction || input->WeightedBiprediction)
{
s->imgY_11_w = malloc ((s->size_x * s->size_y) * sizeof (byte));
get_mem2D (&(s->imgY_ups_w), (2*IMG_PAD_SIZE + s->size_y)*4, (2*IMG_PAD_SIZE + s->size_x)*4);
}
out4Y = s->imgY_ups;
ref11 = s->imgY_11;
//chj 水平半像素内插
for (j = -IMG_PAD_SIZE; j < s->size_y + IMG_PAD_SIZE; j++)
{
for (i = -IMG_PAD_SIZE; i < s->size_x + IMG_PAD_SIZE; i++)
{
jj = max (0, min (s->size_y - 1, j));
is =
(ONE_FOURTH_TAP[0][0] *
(imgY[jj][max (0, min (s->size_x - 1, i))] +
imgY[jj][max (0, min (s->size_x - 1, i + 1))]) +
ONE_FOURTH_TAP[1][0] *
(imgY[jj][max (0, min (s->size_x - 1, i - 1))] +
imgY[jj][max (0, min (s->size_x - 1, i + 2))]) +
ONE_FOURTH_TAP[2][0] *
(imgY[jj][max (0, min (s->size_x - 1, i - 2))] +
imgY[jj][max (0, min (s->size_x - 1, i + 3))]));
img4Y_tmp[j + IMG_PAD_SIZE][(i + IMG_PAD_SIZE) * 2] = imgY[jj][max (0, min (s->size_x - 1, i))] * 1024; // 1/1 pix pos
img4Y_tmp[j + IMG_PAD_SIZE][(i + IMG_PAD_SIZE) * 2 + 1] = is * 32; // 1/2 pix pos
}
}
//chj 垂直半像素内插
for (i = 0; i < (s->size_x + 2 * IMG_PAD_SIZE) * 2; i++)
{
for (j = 0; j < s->size_y + 2 * IMG_PAD_SIZE; j++)
{
j4 = j * 4;
maxy = s->size_y + 2 * IMG_PAD_SIZE - 1;
// change for TML4, use 6 TAP vertical filter
/*1/2像素用6抽头滤波器，系数为ONE_FOURTH_TAP (20 -5 1 1 -5 20) */
is =
(ONE_FOURTH_TAP[0][0] *
(img4Y_tmp[j][i] + img4Y_tmp[min (maxy, j + 1)][i]) +
ONE_FOURTH_TAP[1][0] * (img4Y_tmp[max (0, j - 1)][i] +
img4Y_tmp[min (maxy, j + 2)][i]) +
ONE_FOURTH_TAP[2][0] * (img4Y_tmp[max (0, j - 2)][i] +
img4Y_tmp[min (maxy, j + 3)][i])) / 32;
PutPel_14 (out4Y, (j - IMG_PAD_SIZE) * 4, (i - IMG_PAD_SIZE * 2) * 2, (pel_t) max (0, min (255, (int) ((img4Y_tmp[j][i] + 512) / 1024)))); // 1/2 pix
PutPel_14 (out4Y, (j - IMG_PAD_SIZE) * 4 + 2, (i - IMG_PAD_SIZE * 2) * 2, (pel_t) max (0, min (255, (int) ((is + 512) / 1024)))); // 1/2 pix
}
}
/* 1/4 pix */
/*1/4像素直接用平均值*/
/* luma */
ie2 = (s->size_x + 2 * IMG_PAD_SIZE - 1) * 4;
je2 = (s->size_y + 2 * IMG_PAD_SIZE - 1) * 4;
for (j = 0; j < je2 + 4; j += 2)
for (i = 0; i < ie2 + 3; i += 2)
{
/* '-' 水平内插 */
PutPel_14 (out4Y, j - IMG_PAD_SIZE * 4, i - IMG_PAD_SIZE * 4 + 1,
(pel_t) (max (0, min (255,(int) (FastPelY_14 (out4Y, j - IMG_PAD_SIZE * 4,
i - IMG_PAD_SIZE * 4, img_height, img_width) + FastPelY_14 (out4Y,
j - IMG_PAD_SIZE * 4, min (ie2 + 2, i + 2) - IMG_PAD_SIZE * 4, img_height, img_width)+1) / 2))));
}
for (i = 0; i < ie2 + 4; i++)
{
for (j = 0; j < je2 + 3; j += 2)
{
if (i % 2 == 0)
{
/* '|' 垂直内插*/
PutPel_14 (out4Y, j - IMG_PAD_SIZE * 4 + 1, i - IMG_PAD_SIZE * 4,
(pel_t) (max (0, min (255, (int) (FastPelY_14 (out4Y, j - IMG_PAD_SIZE * 4,
i - IMG_PAD_SIZE * 4, img_height, img_width) + FastPelY_14 (out4Y,
min (je2 + 2, j + 2) - IMG_PAD_SIZE * 4, i - IMG_PAD_SIZE * 4, img_height, img_width)+1) / 2))));
}
else if ((j % 4 == 0 && i % 4 == 1) || (j % 4 == 2 && i % 4 == 3))
{
/* '/' 45度内插 */
PutPel_14 (out4Y, j - IMG_PAD_SIZE * 4 + 1, i - IMG_PAD_SIZE * 4,
(pel_t) (max (0, min (255, (int) (FastPelY_14 (out4Y, j - IMG_PAD_SIZE * 4,
min (ie2 + 2, i + 1) - IMG_PAD_SIZE * 4, img_height, img_width) + FastPelY_14 (out4Y,
min (je2 + 2, j + 2) - IMG_PAD_SIZE * 4, i - IMG_PAD_SIZE * 4 - 1, img_height, img_width) + 1) / 2))));
}
else
{
/* '\' 反向45度内插 */
PutPel_14 (out4Y, j - IMG_PAD_SIZE * 4 + 1, i - IMG_PAD_SIZE * 4,
(pel_t) (max (0, min (255, (int) (FastPelY_14 (out4Y, j - IMG_PAD_SIZE * 4,
i - IMG_PAD_SIZE * 4 - 1, img_height, img_width) + FastPelY_14 (out4Y,
min (je2 + 2, j + 2) - IMG_PAD_SIZE * 4,
min (ie2 + 2, i + 1) - IMG_PAD_SIZE * 4, img_height, img_width) + 1) / 2))));
}
}
}
/* Chroma: 色度只是复制整数点，无内插 */
/* for (j = 0; j < img->height_cr; j++)
{
memcpy (outU[j], imgU[j], img->width_cr); // just copy 1/1 pix, interpolate "online"
memcpy (outV[j], imgV[j], img->width_cr);
}
*/
// Generate 1/1th pel representation (used for integer pel MV search)
/* ref11存放的是整像素resolution的图像
out4Y存放的是1/4像素resolution的图像
要把整像素值从out4Y复制到ref11对应的位置
*/
GenerateFullPelRepresentation (out4Y, ref11, s->size_x, s->size_y);
}
View Code
JM18.6
*!
************************************************************************
* \brief
* Creates the 4x4 = 16 images that contain quarter-pel samples
* sub-sampled at different spatial orientations;
* enables more efficient implementation
*
* \param p_Vid
* pointer to VideoParameters structure
* \param s
* pointer to StorablePicture structure
s************************************************************************
*/
void getSubImagesLuma( VideoParameters *p_Vid, StorablePicture *s )
{
imgpel ****cImgSub = s->p_curr_img_sub;
int otf_shift = ( p_Vid->p_Inp->OnTheFlyFractMCP == OTF_L1 ) ? (1) : (0) ;
// 0 1 2 3
// 4 5 6 7
// 8 9 10 11
// 12 13 14 15
//由于p_curr_img_sub是4维的 4x4xYxX,
//而上方的矩阵cImgSub就是头两维的4x4矩阵
//所以cImgSub[][]中的地址所指向的都是一整副图像(X x Y)
//// INTEGER PEL POSITIONS ////
// sub-image 0 [0][0]
// simply copy the integer pels
if (cImgSub[0][0][0] != s->p_curr_img[0])
{
getSubImageInteger( s, cImgSub[0][0], s->p_curr_img);
}
else
{
getSubImageInteger_s( s, cImgSub[0][0], s->p_curr_img);
}
//// HALF-PEL POSITIONS: SIX-TAP FILTER ////
// sub-image 2 [0][2]
// HOR interpolate (six-tap) sub-image [0][0]
getHorSubImageSixTap( p_Vid, s, cImgSub[0][2>>otf_shift], cImgSub[0][0] );
// sub-image 8 [2][0]
// VER interpolate (six-tap) sub-image [0][0]
getVerSubImageSixTap( p_Vid, s, cImgSub[2>>otf_shift][0], cImgSub[0][0]);
// sub-image 10 [2][2]
// VER interpolate (six-tap) sub-image [0][2]
getVerSubImageSixTapTmp( p_Vid, s, cImgSub[2>>otf_shift][2>>otf_shift]);
if( !p_Vid->p_Inp->OnTheFlyFractMCP )
{
//// QUARTER-PEL POSITIONS: BI-LINEAR INTERPOLATION ////
// sub-image 1 [0][1]
getSubImageBiLinear ( s, cImgSub[0][1], cImgSub[0][0], cImgSub[0][2]);
// sub-image 4 [1][0]
getSubImageBiLinear ( s, cImgSub[1][0], cImgSub[0][0], cImgSub[2][0]);
// sub-image 5 [1][1]
getSubImageBiLinear ( s, cImgSub[1][1], cImgSub[0][2], cImgSub[2][0]);
// sub-image 6 [1][2]
getSubImageBiLinear ( s, cImgSub[1][2], cImgSub[0][2], cImgSub[2][2]);
// sub-image 9 [2][1]
getSubImageBiLinear ( s, cImgSub[2][1], cImgSub[2][0], cImgSub[2][2]);
// sub-image 3 [0][3]
getHorSubImageBiLinear ( s, cImgSub[0][3], cImgSub[0][2], cImgSub[0][0]);
// sub-image 7 [1][3]
getHorSubImageBiLinear ( s, cImgSub[1][3], cImgSub[0][2], cImgSub[2][0]);
// sub-image 11 [2][3]
getHorSubImageBiLinear ( s, cImgSub[2][3], cImgSub[2][2], cImgSub[2][0]);
// sub-image 12 [3][0]
getVerSubImageBiLinear ( s, cImgSub[3][0], cImgSub[2][0], cImgSub[0][0]);
// sub-image 13 [3][1]
getVerSubImageBiLinear ( s, cImgSub[3][1], cImgSub[2][0], cImgSub[0][2]);
// sub-image 14 [3][2]
getVerSubImageBiLinear ( s, cImgSub[3][2], cImgSub[2][2], cImgSub[0][2]);
// sub-image 15 [3][3]
getDiagSubImageBiLinear( s, cImgSub[3][3], cImgSub[0][2], cImgSub[2][0]);
}
}
View Code
