-
Yunqing Wang authored
Changed motion search in vp8_find_best_half_pixel_step() to be the same as in vp8_find_best_sub_pixel_step(), which checks 5 points instead of 8 points. This only affects real-time mode with cpu-used >=9. Tests showed it gives 2% encoding speedup with a quality loss(psnr) of up to 0.5%. Change-Id: I16049cad1535002346d46cfdfad345bfc3dc5146
b9f19f89
mcomp.c 60.44 KiB
/* * Copyright (c) 2010 The WebM project authors. All Rights Reserved. * * Use of this source code is governed by a BSD-style license * that can be found in the LICENSE file in the root of the source * tree. An additional intellectual property rights grant can be found * in the file PATENTS. All contributing project authors may * be found in the AUTHORS file in the root of the source tree. */#include "mcomp.h"#include "vpx_mem/vpx_mem.h"#include "vpx_ports/config.h"#include <stdio.h>#include <limits.h>#include <math.h>#include "vp8/common/findnearmv.h"#ifdef ENTROPY_STATSstatic int mv_ref_ct [31] [4] [2];static int mv_mode_cts [4] [2];#endifint vp8_mv_bit_cost(int_mv *mv, int_mv *ref, int *mvcost[2], int Weight){ // MV costing is based on the distribution of vectors in the previous frame and as such will tend to // over state the cost of vectors. In addition coding a new vector can have a knock on effect on the // cost of subsequent vectors and the quality of prediction from NEAR and NEAREST for subsequent blocks. // The "Weight" parameter allows, to a limited extent, for some account to be taken of these factors. return ((mvcost[0][(mv->as_mv.row - ref->as_mv.row) >> 1] + mvcost[1][(mv->as_mv.col - ref->as_mv.col) >> 1]) * Weight) >> 7;}static int mv_err_cost(int_mv *mv, int_mv *ref, int *mvcost[2], int error_per_bit){ return ((mvcost[0][(mv->as_mv.row - ref->as_mv.row) >> 1] + mvcost[1][(mv->as_mv.col - ref->as_mv.col) >> 1]) * error_per_bit + 128) >> 8;}static int mvsad_err_cost(int_mv *mv, int_mv *ref, int *mvsadcost[2], int error_per_bit){ /* Calculate sad error cost on full pixel basis. */ return ((mvsadcost[0][(mv->as_mv.row - ref->as_mv.row)] + mvsadcost[1][(mv->as_mv.col - ref->as_mv.col)]) * error_per_bit + 128) >> 8;}void vp8_init_dsmotion_compensation(MACROBLOCK *x, int stride){ int Len; int search_site_count = 0; // Generate offsets for 4 search sites per step. Len = MAX_FIRST_STEP; x->ss[search_site_count].mv.col = 0; x->ss[search_site_count].mv.row = 0; x->ss[search_site_count].offset = 0; search_site_count++; while (Len > 0) { // Compute offsets for search sites. x->ss[search_site_count].mv.col = 0; x->ss[search_site_count].mv.row = -Len; x->ss[search_site_count].offset = -Len * stride; search_site_count++;7172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140
// Compute offsets for search sites.
x->ss[search_site_count].mv.col = 0;
x->ss[search_site_count].mv.row = Len;
x->ss[search_site_count].offset = Len * stride;
search_site_count++;
// Compute offsets for search sites.
x->ss[search_site_count].mv.col = -Len;
x->ss[search_site_count].mv.row = 0;
x->ss[search_site_count].offset = -Len;
search_site_count++;
// Compute offsets for search sites.
x->ss[search_site_count].mv.col = Len;
x->ss[search_site_count].mv.row = 0;
x->ss[search_site_count].offset = Len;
search_site_count++;
// Contract.
Len /= 2;
}
x->ss_count = search_site_count;
x->searches_per_step = 4;
}
void vp8_init3smotion_compensation(MACROBLOCK *x, int stride)
{
int Len;
int search_site_count = 0;
// Generate offsets for 8 search sites per step.
Len = MAX_FIRST_STEP;
x->ss[search_site_count].mv.col = 0;
x->ss[search_site_count].mv.row = 0;
x->ss[search_site_count].offset = 0;
search_site_count++;
while (Len > 0)
{
// Compute offsets for search sites.
x->ss[search_site_count].mv.col = 0;
x->ss[search_site_count].mv.row = -Len;
x->ss[search_site_count].offset = -Len * stride;
search_site_count++;
// Compute offsets for search sites.
x->ss[search_site_count].mv.col = 0;
x->ss[search_site_count].mv.row = Len;
x->ss[search_site_count].offset = Len * stride;
search_site_count++;
// Compute offsets for search sites.
x->ss[search_site_count].mv.col = -Len;
x->ss[search_site_count].mv.row = 0;
x->ss[search_site_count].offset = -Len;
search_site_count++;
// Compute offsets for search sites.
x->ss[search_site_count].mv.col = Len;
x->ss[search_site_count].mv.row = 0;
x->ss[search_site_count].offset = Len;
search_site_count++;
// Compute offsets for search sites.
x->ss[search_site_count].mv.col = -Len;
x->ss[search_site_count].mv.row = -Len;
x->ss[search_site_count].offset = -Len * stride - Len;
search_site_count++;
141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210
// Compute offsets for search sites.
x->ss[search_site_count].mv.col = Len;
x->ss[search_site_count].mv.row = -Len;
x->ss[search_site_count].offset = -Len * stride + Len;
search_site_count++;
// Compute offsets for search sites.
x->ss[search_site_count].mv.col = -Len;
x->ss[search_site_count].mv.row = Len;
x->ss[search_site_count].offset = Len * stride - Len;
search_site_count++;
// Compute offsets for search sites.
x->ss[search_site_count].mv.col = Len;
x->ss[search_site_count].mv.row = Len;
x->ss[search_site_count].offset = Len * stride + Len;
search_site_count++;
// Contract.
Len /= 2;
}
x->ss_count = search_site_count;
x->searches_per_step = 8;
}
/*
* To avoid the penalty for crossing cache-line read, preload the reference
* area in a small buffer, which is aligned to make sure there won't be crossing
* cache-line read while reading from this buffer. This reduced the cpu
* cycles spent on reading ref data in sub-pixel filter functions.
* TODO: Currently, since sub-pixel search range here is -3 ~ 3, copy 22 rows x
* 32 cols area that is enough for 16x16 macroblock. Later, for SPLITMV, we
* could reduce the area.
*/
#define MVC(r,c) (((mvcost[0][(r)-rr] + mvcost[1][(c) - rc]) * error_per_bit + 128 )>>8 ) // estimated cost of a motion vector (r,c)
#define PRE(r,c) (y + (((r)>>2) * y_stride + ((c)>>2) -(offset))) // pointer to predictor base of a motionvector
#define SP(x) (((x)&3)<<1) // convert motion vector component to offset for svf calc
#define DIST(r,c) vfp->svf( PRE(r,c), y_stride, SP(c),SP(r), z,b->src_stride,&sse) // returns subpixel variance error function.
#define IFMVCV(r,c,s,e) if ( c >= minc && c <= maxc && r >= minr && r <= maxr) s else e;
#define ERR(r,c) (MVC(r,c)+DIST(r,c)) // returns distortion + motion vector cost
#define CHECK_BETTER(v,r,c) IFMVCV(r,c,{thismse = DIST(r,c); if((v = (MVC(r,c)+thismse)) < besterr) { besterr = v; br=r; bc=c; *distortion = thismse; *sse1 = sse; }}, v=INT_MAX;)// checks if (r,c) has better score than previous best
#define MIN(x,y) (((x)<(y))?(x):(y))
#define MAX(x,y) (((x)>(y))?(x):(y))
int vp8_find_best_sub_pixel_step_iteratively(MACROBLOCK *x, BLOCK *b, BLOCKD *d,
int_mv *bestmv, int_mv *ref_mv,
int error_per_bit,
const vp8_variance_fn_ptr_t *vfp,
int *mvcost[2], int *distortion,
unsigned int *sse1)
{
unsigned char *z = (*(b->base_src) + b->src);
int rr = ref_mv->as_mv.row >> 1, rc = ref_mv->as_mv.col >> 1;
int br = bestmv->as_mv.row << 2, bc = bestmv->as_mv.col << 2;
int tr = br, tc = bc;
unsigned int besterr = INT_MAX;
unsigned int left, right, up, down, diag;
unsigned int sse;
unsigned int whichdir;
unsigned int halfiters = 4;
unsigned int quarteriters = 4;
int thismse;
int minc = MAX(x->mv_col_min << 2, (ref_mv->as_mv.col >> 1) - ((1 << mvlong_width) - 1));
int maxc = MIN(x->mv_col_max << 2, (ref_mv->as_mv.col >> 1) + ((1 << mvlong_width) - 1));
int minr = MAX(x->mv_row_min << 2, (ref_mv->as_mv.row >> 1) - ((1 << mvlong_width) - 1));
211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280
int maxr = MIN(x->mv_row_max << 2, (ref_mv->as_mv.row >> 1) + ((1 << mvlong_width) - 1));
int y_stride;
int offset;
#if ARCH_X86 || ARCH_X86_64
MACROBLOCKD *xd = &x->e_mbd;
unsigned char *y0 = *(d->base_pre) + d->pre + (bestmv->as_mv.row) * d->pre_stride + bestmv->as_mv.col;
unsigned char *y;
int buf_r1, buf_r2, buf_c1, buf_c2;
// Clamping to avoid out-of-range data access
buf_r1 = ((bestmv->as_mv.row - 3) < x->mv_row_min)?(bestmv->as_mv.row - x->mv_row_min):3;
buf_r2 = ((bestmv->as_mv.row + 3) > x->mv_row_max)?(x->mv_row_max - bestmv->as_mv.row):3;
buf_c1 = ((bestmv->as_mv.col - 3) < x->mv_col_min)?(bestmv->as_mv.col - x->mv_col_min):3;
buf_c2 = ((bestmv->as_mv.col + 3) > x->mv_col_max)?(x->mv_col_max - bestmv->as_mv.col):3;
y_stride = 32;
/* Copy to intermediate buffer before searching. */
vfp->copymem(y0 - buf_c1 - d->pre_stride*buf_r1, d->pre_stride, xd->y_buf, y_stride, 16+buf_r1+buf_r2);
y = xd->y_buf + y_stride*buf_r1 +buf_c1;
#else
unsigned char *y = *(d->base_pre) + d->pre + (bestmv->as_mv.row) * d->pre_stride + bestmv->as_mv.col;
y_stride = d->pre_stride;
#endif
offset = (bestmv->as_mv.row) * y_stride + bestmv->as_mv.col;
// central mv
bestmv->as_mv.row <<= 3;
bestmv->as_mv.col <<= 3;
// calculate central point error
besterr = vfp->vf(y, y_stride, z, b->src_stride, sse1);
*distortion = besterr;
besterr += mv_err_cost(bestmv, ref_mv, mvcost, error_per_bit);
// TODO: Each subsequent iteration checks at least one point in common with the last iteration could be 2 ( if diag selected)
while (--halfiters)
{
// 1/2 pel
CHECK_BETTER(left, tr, tc - 2);
CHECK_BETTER(right, tr, tc + 2);
CHECK_BETTER(up, tr - 2, tc);
CHECK_BETTER(down, tr + 2, tc);
whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
switch (whichdir)
{
case 0:
CHECK_BETTER(diag, tr - 2, tc - 2);
break;
case 1:
CHECK_BETTER(diag, tr - 2, tc + 2);
break;
case 2:
CHECK_BETTER(diag, tr + 2, tc - 2);
break;
case 3:
CHECK_BETTER(diag, tr + 2, tc + 2);
break;
}
// no reason to check the same one again.
if (tr == br && tc == bc)
break;
tr = br;
tc = bc;
281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350
}
// TODO: Each subsequent iteration checks at least one point in common with the last iteration could be 2 ( if diag selected)
// 1/4 pel
while (--quarteriters)
{
CHECK_BETTER(left, tr, tc - 1);
CHECK_BETTER(right, tr, tc + 1);
CHECK_BETTER(up, tr - 1, tc);
CHECK_BETTER(down, tr + 1, tc);
whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
switch (whichdir)
{
case 0:
CHECK_BETTER(diag, tr - 1, tc - 1);
break;
case 1:
CHECK_BETTER(diag, tr - 1, tc + 1);
break;
case 2:
CHECK_BETTER(diag, tr + 1, tc - 1);
break;
case 3:
CHECK_BETTER(diag, tr + 1, tc + 1);
break;
}
// no reason to check the same one again.
if (tr == br && tc == bc)
break;
tr = br;
tc = bc;
}
bestmv->as_mv.row = br << 1;
bestmv->as_mv.col = bc << 1;
if ((abs(bestmv->as_mv.col - ref_mv->as_mv.col) > (MAX_FULL_PEL_VAL<<3)) ||
(abs(bestmv->as_mv.row - ref_mv->as_mv.row) > (MAX_FULL_PEL_VAL<<3)))
return INT_MAX;
return besterr;
}
#undef MVC
#undef PRE
#undef SP
#undef DIST
#undef IFMVCV
#undef ERR
#undef CHECK_BETTER
#undef MIN
#undef MAX
int vp8_find_best_sub_pixel_step(MACROBLOCK *x, BLOCK *b, BLOCKD *d,
int_mv *bestmv, int_mv *ref_mv,
int error_per_bit,
const vp8_variance_fn_ptr_t *vfp,
int *mvcost[2], int *distortion,
unsigned int *sse1)
{
int bestmse = INT_MAX;
int_mv startmv;
int_mv this_mv;
unsigned char *z = (*(b->base_src) + b->src);
int left, right, up, down, diag;
unsigned int sse;
int whichdir ;
int thismse;
351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420
int y_stride;
#if ARCH_X86 || ARCH_X86_64
MACROBLOCKD *xd = &x->e_mbd;
unsigned char *y0 = *(d->base_pre) + d->pre + (bestmv->as_mv.row) * d->pre_stride + bestmv->as_mv.col;
unsigned char *y;
y_stride = 32;
/* Copy 18 rows x 32 cols area to intermediate buffer before searching. */
vfp->copymem(y0 - 1 - d->pre_stride, d->pre_stride, xd->y_buf, y_stride, 18);
y = xd->y_buf + y_stride + 1;
#else
unsigned char *y = *(d->base_pre) + d->pre + (bestmv->as_mv.row) * d->pre_stride + bestmv->as_mv.col;
y_stride = d->pre_stride;
#endif
// central mv
bestmv->as_mv.row <<= 3;
bestmv->as_mv.col <<= 3;
startmv = *bestmv;
// calculate central point error
bestmse = vfp->vf(y, y_stride, z, b->src_stride, sse1);
*distortion = bestmse;
bestmse += mv_err_cost(bestmv, ref_mv, mvcost, error_per_bit);
// go left then right and check error
this_mv.as_mv.row = startmv.as_mv.row;
this_mv.as_mv.col = ((startmv.as_mv.col - 8) | 4);
thismse = vfp->svf_halfpix_h(y - 1, y_stride, z, b->src_stride, &sse);
left = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
if (left < bestmse)
{
*bestmv = this_mv;
bestmse = left;
*distortion = thismse;
*sse1 = sse;
}
this_mv.as_mv.col += 8;
thismse = vfp->svf_halfpix_h(y, y_stride, z, b->src_stride, &sse);
right = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
if (right < bestmse)
{
*bestmv = this_mv;
bestmse = right;
*distortion = thismse;
*sse1 = sse;
}
// go up then down and check error
this_mv.as_mv.col = startmv.as_mv.col;
this_mv.as_mv.row = ((startmv.as_mv.row - 8) | 4);
thismse = vfp->svf_halfpix_v(y - y_stride, y_stride, z, b->src_stride, &sse);
up = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
if (up < bestmse)
{
*bestmv = this_mv;
bestmse = up;
*distortion = thismse;
*sse1 = sse;
}
this_mv.as_mv.row += 8;
thismse = vfp->svf_halfpix_v(y, y_stride, z, b->src_stride, &sse);
down = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490
if (down < bestmse)
{
*bestmv = this_mv;
bestmse = down;
*distortion = thismse;
*sse1 = sse;
}
// now check 1 more diagonal
whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
//for(whichdir =0;whichdir<4;whichdir++)
//{
this_mv = startmv;
switch (whichdir)
{
case 0:
this_mv.as_mv.col = (this_mv.as_mv.col - 8) | 4;
this_mv.as_mv.row = (this_mv.as_mv.row - 8) | 4;
thismse = vfp->svf_halfpix_hv(y - 1 - y_stride, y_stride, z, b->src_stride, &sse);
break;
case 1:
this_mv.as_mv.col += 4;
this_mv.as_mv.row = (this_mv.as_mv.row - 8) | 4;
thismse = vfp->svf_halfpix_hv(y - y_stride, y_stride, z, b->src_stride, &sse);
break;
case 2:
this_mv.as_mv.col = (this_mv.as_mv.col - 8) | 4;
this_mv.as_mv.row += 4;
thismse = vfp->svf_halfpix_hv(y - 1, y_stride, z, b->src_stride, &sse);
break;
case 3:
default:
this_mv.as_mv.col += 4;
this_mv.as_mv.row += 4;
thismse = vfp->svf_halfpix_hv(y, y_stride, z, b->src_stride, &sse);
break;
}
diag = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
if (diag < bestmse)
{
*bestmv = this_mv;
bestmse = diag;
*distortion = thismse;
*sse1 = sse;
}
// }
// time to check quarter pels.
if (bestmv->as_mv.row < startmv.as_mv.row)
y -= y_stride;
if (bestmv->as_mv.col < startmv.as_mv.col)
y--;
startmv = *bestmv;
// go left then right and check error
this_mv.as_mv.row = startmv.as_mv.row;
if (startmv.as_mv.col & 7)
{
this_mv.as_mv.col = startmv.as_mv.col - 2;
491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560
thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
}
else
{
this_mv.as_mv.col = (startmv.as_mv.col - 8) | 6;
thismse = vfp->svf(y - 1, y_stride, 6, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
}
left = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
if (left < bestmse)
{
*bestmv = this_mv;
bestmse = left;
*distortion = thismse;
*sse1 = sse;
}
this_mv.as_mv.col += 4;
thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
right = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
if (right < bestmse)
{
*bestmv = this_mv;
bestmse = right;
*distortion = thismse;
*sse1 = sse;
}
// go up then down and check error
this_mv.as_mv.col = startmv.as_mv.col;
if (startmv.as_mv.row & 7)
{
this_mv.as_mv.row = startmv.as_mv.row - 2;
thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
}
else
{
this_mv.as_mv.row = (startmv.as_mv.row - 8) | 6;
thismse = vfp->svf(y - y_stride, y_stride, this_mv.as_mv.col & 7, 6, z, b->src_stride, &sse);
}
up = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
if (up < bestmse)
{
*bestmv = this_mv;
bestmse = up;
*distortion = thismse;
*sse1 = sse;
}
this_mv.as_mv.row += 4;
thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
down = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
if (down < bestmse)
{
*bestmv = this_mv;
bestmse = down;
*distortion = thismse;
*sse1 = sse;
}
// now check 1 more diagonal
whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630
// for(whichdir=0;whichdir<4;whichdir++)
// {
this_mv = startmv;
switch (whichdir)
{
case 0:
if (startmv.as_mv.row & 7)
{
this_mv.as_mv.row -= 2;
if (startmv.as_mv.col & 7)
{
this_mv.as_mv.col -= 2;
thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
}
else
{
this_mv.as_mv.col = (startmv.as_mv.col - 8) | 6;
thismse = vfp->svf(y - 1, y_stride, 6, this_mv.as_mv.row & 7, z, b->src_stride, &sse);;
}
}
else
{
this_mv.as_mv.row = (startmv.as_mv.row - 8) | 6;
if (startmv.as_mv.col & 7)
{
this_mv.as_mv.col -= 2;
thismse = vfp->svf(y - y_stride, y_stride, this_mv.as_mv.col & 7, 6, z, b->src_stride, &sse);
}
else
{
this_mv.as_mv.col = (startmv.as_mv.col - 8) | 6;
thismse = vfp->svf(y - y_stride - 1, y_stride, 6, 6, z, b->src_stride, &sse);
}
}
break;
case 1:
this_mv.as_mv.col += 2;
if (startmv.as_mv.row & 7)
{
this_mv.as_mv.row -= 2;
thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
}
else
{
this_mv.as_mv.row = (startmv.as_mv.row - 8) | 6;
thismse = vfp->svf(y - y_stride, y_stride, this_mv.as_mv.col & 7, 6, z, b->src_stride, &sse);
}
break;
case 2:
this_mv.as_mv.row += 2;
if (startmv.as_mv.col & 7)
{
this_mv.as_mv.col -= 2;
thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
}
else
{
this_mv.as_mv.col = (startmv.as_mv.col - 8) | 6;
thismse = vfp->svf(y - 1, y_stride, 6, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
}
break;
631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700
case 3:
this_mv.as_mv.col += 2;
this_mv.as_mv.row += 2;
thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
break;
}
diag = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
if (diag < bestmse)
{
*bestmv = this_mv;
bestmse = diag;
*distortion = thismse;
*sse1 = sse;
}
return bestmse;
}
int vp8_find_best_half_pixel_step(MACROBLOCK *x, BLOCK *b, BLOCKD *d,
int_mv *bestmv, int_mv *ref_mv,
int error_per_bit,
const vp8_variance_fn_ptr_t *vfp,
int *mvcost[2], int *distortion,
unsigned int *sse1)
{
int bestmse = INT_MAX;
int_mv startmv;
int_mv this_mv;
unsigned char *z = (*(b->base_src) + b->src);
int left, right, up, down, diag;
unsigned int sse;
int whichdir ;
int thismse;
int y_stride;
#if ARCH_X86 || ARCH_X86_64
MACROBLOCKD *xd = &x->e_mbd;
unsigned char *y0 = *(d->base_pre) + d->pre + (bestmv->as_mv.row) * d->pre_stride + bestmv->as_mv.col;
unsigned char *y;
y_stride = 32;
/* Copy 18 rows x 32 cols area to intermediate buffer before searching. */
vfp->copymem(y0 - 1 - d->pre_stride, d->pre_stride, xd->y_buf, y_stride, 18);
y = xd->y_buf + y_stride + 1;
#else
unsigned char *y = *(d->base_pre) + d->pre + (bestmv->as_mv.row) * d->pre_stride + bestmv->as_mv.col;
y_stride = d->pre_stride;
#endif
// central mv
bestmv->as_mv.row <<= 3;
bestmv->as_mv.col <<= 3;
startmv = *bestmv;
// calculate central point error
bestmse = vfp->vf(y, y_stride, z, b->src_stride, sse1);
*distortion = bestmse;
bestmse += mv_err_cost(bestmv, ref_mv, mvcost, error_per_bit);
// go left then right and check error
this_mv.as_mv.row = startmv.as_mv.row;
this_mv.as_mv.col = ((startmv.as_mv.col - 8) | 4);
thismse = vfp->svf_halfpix_h(y - 1, y_stride, z, b->src_stride, &sse);
left = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
if (left < bestmse)
{
*bestmv = this_mv;
701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770
bestmse = left;
*distortion = thismse;
*sse1 = sse;
}
this_mv.as_mv.col += 8;
thismse = vfp->svf_halfpix_h(y, y_stride, z, b->src_stride, &sse);
right = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
if (right < bestmse)
{
*bestmv = this_mv;
bestmse = right;
*distortion = thismse;
*sse1 = sse;
}
// go up then down and check error
this_mv.as_mv.col = startmv.as_mv.col;
this_mv.as_mv.row = ((startmv.as_mv.row - 8) | 4);
thismse = vfp->svf_halfpix_v(y - y_stride, y_stride, z, b->src_stride, &sse);
up = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
if (up < bestmse)
{
*bestmv = this_mv;
bestmse = up;
*distortion = thismse;
*sse1 = sse;
}
this_mv.as_mv.row += 8;
thismse = vfp->svf_halfpix_v(y, y_stride, z, b->src_stride, &sse);
down = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
if (down < bestmse)
{
*bestmv = this_mv;
bestmse = down;
*distortion = thismse;
*sse1 = sse;
}
// now check 1 more diagonal -
whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
this_mv = startmv;
switch (whichdir)
{
case 0:
this_mv.as_mv.col = (this_mv.as_mv.col - 8) | 4;
this_mv.as_mv.row = (this_mv.as_mv.row - 8) | 4;
thismse = vfp->svf_halfpix_hv(y - 1 - y_stride, y_stride, z, b->src_stride, &sse);
break;
case 1:
this_mv.as_mv.col += 4;
this_mv.as_mv.row = (this_mv.as_mv.row - 8) | 4;
thismse = vfp->svf_halfpix_hv(y - y_stride, y_stride, z, b->src_stride, &sse);
break;
case 2:
this_mv.as_mv.col = (this_mv.as_mv.col - 8) | 4;
this_mv.as_mv.row += 4;
thismse = vfp->svf_halfpix_hv(y - 1, y_stride, z, b->src_stride, &sse);
break;
case 3:
default:
this_mv.as_mv.col += 4;
this_mv.as_mv.row += 4;
thismse = vfp->svf_halfpix_hv(y, y_stride, z, b->src_stride, &sse);
break;
771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840
}
diag = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
if (diag < bestmse)
{
*bestmv = this_mv;
bestmse = diag;
*distortion = thismse;
*sse1 = sse;
}
return bestmse;
}
#define CHECK_BOUNDS(range) \
{\
all_in = 1;\
all_in &= ((br-range) >= x->mv_row_min);\
all_in &= ((br+range) <= x->mv_row_max);\
all_in &= ((bc-range) >= x->mv_col_min);\
all_in &= ((bc+range) <= x->mv_col_max);\
}
#define CHECK_POINT \
{\
if (this_mv.as_mv.col < x->mv_col_min) continue;\
if (this_mv.as_mv.col > x->mv_col_max) continue;\
if (this_mv.as_mv.row < x->mv_row_min) continue;\
if (this_mv.as_mv.row > x->mv_row_max) continue;\
}
#define CHECK_BETTER \
{\
if (thissad < bestsad)\
{\
thissad += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, sad_per_bit);\
if (thissad < bestsad)\
{\
bestsad = thissad;\
best_site = i;\
}\
}\
}
static const MV next_chkpts[6][3] =
{
{{ -2, 0}, { -1, -2}, {1, -2}},
{{ -1, -2}, {1, -2}, {2, 0}},
{{1, -2}, {2, 0}, {1, 2}},
{{2, 0}, {1, 2}, { -1, 2}},
{{1, 2}, { -1, 2}, { -2, 0}},
{{ -1, 2}, { -2, 0}, { -1, -2}}
};
int vp8_hex_search
(
MACROBLOCK *x,
BLOCK *b,
BLOCKD *d,
int_mv *ref_mv,
int_mv *best_mv,
int search_param,
int sad_per_bit,
const vp8_variance_fn_ptr_t *vfp,
int *mvsadcost[2],
int *mvcost[2],
int_mv *center_mv
)
{
841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910
MV hex[6] = { { -1, -2}, {1, -2}, {2, 0}, {1, 2}, { -1, 2}, { -2, 0} } ;
MV neighbors[4] = {{0, -1}, { -1, 0}, {1, 0}, {0, 1}} ;
int i, j;
unsigned char *what = (*(b->base_src) + b->src);
int what_stride = b->src_stride;
int in_what_stride = d->pre_stride;
int br, bc;
int_mv this_mv;
unsigned int bestsad = 0x7fffffff;
unsigned int thissad;
unsigned char *base_offset;
unsigned char *this_offset;
int k = -1;
int all_in;
int best_site = -1;
int_mv fcenter_mv;
fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
// adjust ref_mv to make sure it is within MV range
vp8_clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
br = ref_mv->as_mv.row;
bc = ref_mv->as_mv.col;
// Work out the start point for the search
base_offset = (unsigned char *)(*(d->base_pre) + d->pre);
this_offset = base_offset + (br * (d->pre_stride)) + bc;
this_mv.as_mv.row = br;
this_mv.as_mv.col = bc;
bestsad = vfp->sdf( what, what_stride, this_offset,
in_what_stride, 0x7fffffff)
+ mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, sad_per_bit);
// hex search
//j=0
CHECK_BOUNDS(2)
if(all_in)
{
for (i = 0; i < 6; i++)
{
this_mv.as_mv.row = br + hex[i].row;
this_mv.as_mv.col = bc + hex[i].col;
this_offset = base_offset + (this_mv.as_mv.row * in_what_stride) + this_mv.as_mv.col;
thissad=vfp->sdf( what, what_stride, this_offset, in_what_stride, bestsad);
CHECK_BETTER
}
}else
{
for (i = 0; i < 6; i++)
{
this_mv.as_mv.row = br + hex[i].row;
this_mv.as_mv.col = bc + hex[i].col;
CHECK_POINT
this_offset = base_offset + (this_mv.as_mv.row * in_what_stride) + this_mv.as_mv.col;
thissad=vfp->sdf( what, what_stride, this_offset, in_what_stride, bestsad);
CHECK_BETTER
}
}
if (best_site == -1)
goto cal_neighbors;
else
{
br += hex[best_site].row;
bc += hex[best_site].col;
k = best_site;
}
911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980
for (j = 1; j < 127; j++)
{
best_site = -1;
CHECK_BOUNDS(2)
if(all_in)
{
for (i = 0; i < 3; i++)
{
this_mv.as_mv.row = br + next_chkpts[k][i].row;
this_mv.as_mv.col = bc + next_chkpts[k][i].col;
this_offset = base_offset + (this_mv.as_mv.row * (in_what_stride)) + this_mv.as_mv.col;
thissad = vfp->sdf( what, what_stride, this_offset, in_what_stride, bestsad);
CHECK_BETTER
}
}else
{
for (i = 0; i < 3; i++)
{
this_mv.as_mv.row = br + next_chkpts[k][i].row;
this_mv.as_mv.col = bc + next_chkpts[k][i].col;
CHECK_POINT
this_offset = base_offset + (this_mv.as_mv.row * (in_what_stride)) + this_mv.as_mv.col;
thissad = vfp->sdf( what, what_stride, this_offset, in_what_stride, bestsad);
CHECK_BETTER
}
}
if (best_site == -1)
break;
else
{
br += next_chkpts[k][best_site].row;
bc += next_chkpts[k][best_site].col;
k += 5 + best_site;
if (k >= 12) k -= 12;
else if (k >= 6) k -= 6;
}
}
// check 4 1-away neighbors
cal_neighbors:
for (j = 0; j < 32; j++)
{
best_site = -1;
CHECK_BOUNDS(1)
if(all_in)
{
for (i = 0; i < 4; i++)
{
this_mv.as_mv.row = br + neighbors[i].row;
this_mv.as_mv.col = bc + neighbors[i].col;
this_offset = base_offset + (this_mv.as_mv.row * (in_what_stride)) + this_mv.as_mv.col;
thissad = vfp->sdf( what, what_stride, this_offset, in_what_stride, bestsad);
CHECK_BETTER
}
}else
{
for (i = 0; i < 4; i++)
{
this_mv.as_mv.row = br + neighbors[i].row;
this_mv.as_mv.col = bc + neighbors[i].col;
CHECK_POINT
this_offset = base_offset + (this_mv.as_mv.row * (in_what_stride)) + this_mv.as_mv.col;
thissad = vfp->sdf( what, what_stride, this_offset, in_what_stride, bestsad);
CHECK_BETTER
}
}
981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050
if (best_site == -1)
break;
else
{
br += neighbors[best_site].row;
bc += neighbors[best_site].col;
}
}
best_mv->as_mv.row = br;
best_mv->as_mv.col = bc;
return bestsad;
}
#undef CHECK_BOUNDS
#undef CHECK_POINT
#undef CHECK_BETTER
int vp8_diamond_search_sad
(
MACROBLOCK *x,
BLOCK *b,
BLOCKD *d,
int_mv *ref_mv,
int_mv *best_mv,
int search_param,
int sad_per_bit,
int *num00,
vp8_variance_fn_ptr_t *fn_ptr,
int *mvcost[2],
int_mv *center_mv
)
{
int i, j, step;
unsigned char *what = (*(b->base_src) + b->src);
int what_stride = b->src_stride;
unsigned char *in_what;
int in_what_stride = d->pre_stride;
unsigned char *best_address;
int tot_steps;
int_mv this_mv;
int bestsad = INT_MAX;
int best_site = 0;
int last_site = 0;
int ref_row;
int ref_col;
int this_row_offset;
int this_col_offset;
search_site *ss;
unsigned char *check_here;
int thissad;
int *mvsadcost[2] = {x->mvsadcost[0], x->mvsadcost[1]};
int_mv fcenter_mv;
fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
vp8_clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
ref_row = ref_mv->as_mv.row;
ref_col = ref_mv->as_mv.col;
*num00 = 0;
best_mv->as_mv.row = ref_row;
best_mv->as_mv.col = ref_col;
1051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120
// Work out the start point for the search
in_what = (unsigned char *)(*(d->base_pre) + d->pre + (ref_row * (d->pre_stride)) + ref_col);
best_address = in_what;
// Check the starting position
bestsad = fn_ptr->sdf(what, what_stride, in_what,
in_what_stride, 0x7fffffff)
+ mvsad_err_cost(best_mv, &fcenter_mv, mvsadcost, sad_per_bit);
// search_param determines the length of the initial step and hence the number of iterations
// 0 = initial step (MAX_FIRST_STEP) pel : 1 = (MAX_FIRST_STEP/2) pel, 2 = (MAX_FIRST_STEP/4) pel... etc.
ss = &x->ss[search_param * x->searches_per_step];
tot_steps = (x->ss_count / x->searches_per_step) - search_param;
i = 1;
for (step = 0; step < tot_steps ; step++)
{
for (j = 0 ; j < x->searches_per_step ; j++)
{
// Trap illegal vectors
this_row_offset = best_mv->as_mv.row + ss[i].mv.row;
this_col_offset = best_mv->as_mv.col + ss[i].mv.col;
if ((this_col_offset > x->mv_col_min) && (this_col_offset < x->mv_col_max) &&
(this_row_offset > x->mv_row_min) && (this_row_offset < x->mv_row_max))
{
check_here = ss[i].offset + best_address;
thissad = fn_ptr->sdf(what, what_stride, check_here , in_what_stride, bestsad);
if (thissad < bestsad)
{
this_mv.as_mv.row = this_row_offset;
this_mv.as_mv.col = this_col_offset;
thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
mvsadcost, sad_per_bit);
if (thissad < bestsad)
{
bestsad = thissad;
best_site = i;
}
}
}
i++;
}
if (best_site != last_site)
{
best_mv->as_mv.row += ss[best_site].mv.row;
best_mv->as_mv.col += ss[best_site].mv.col;
best_address += ss[best_site].offset;
last_site = best_site;
}
else if (best_address == in_what)
(*num00)++;
}
this_mv.as_mv.row = best_mv->as_mv.row << 3;
this_mv.as_mv.col = best_mv->as_mv.col << 3;
if (bestsad == INT_MAX)
return INT_MAX;
return fn_ptr->vf(what, what_stride, best_address, in_what_stride, (unsigned int *)(&thissad))
+ mv_err_cost(&this_mv, center_mv, mvcost, x->errorperbit);
}
1121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190
int vp8_diamond_search_sadx4
(
MACROBLOCK *x,
BLOCK *b,
BLOCKD *d,
int_mv *ref_mv,
int_mv *best_mv,
int search_param,
int sad_per_bit,
int *num00,
vp8_variance_fn_ptr_t *fn_ptr,
int *mvcost[2],
int_mv *center_mv
)
{
int i, j, step;
unsigned char *what = (*(b->base_src) + b->src);
int what_stride = b->src_stride;
unsigned char *in_what;
int in_what_stride = d->pre_stride;
unsigned char *best_address;
int tot_steps;
int_mv this_mv;
int bestsad = INT_MAX;
int best_site = 0;
int last_site = 0;
int ref_row;
int ref_col;
int this_row_offset;
int this_col_offset;
search_site *ss;
unsigned char *check_here;
unsigned int thissad;
int *mvsadcost[2] = {x->mvsadcost[0], x->mvsadcost[1]};
int_mv fcenter_mv;
fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
vp8_clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
ref_row = ref_mv->as_mv.row;
ref_col = ref_mv->as_mv.col;
*num00 = 0;
best_mv->as_mv.row = ref_row;
best_mv->as_mv.col = ref_col;
// Work out the start point for the search
in_what = (unsigned char *)(*(d->base_pre) + d->pre + (ref_row * (d->pre_stride)) + ref_col);
best_address = in_what;
// Check the starting position
bestsad = fn_ptr->sdf(what, what_stride,
in_what, in_what_stride, 0x7fffffff)
+ mvsad_err_cost(best_mv, &fcenter_mv, mvsadcost, sad_per_bit);
// search_param determines the length of the initial step and hence the number of iterations
// 0 = initial step (MAX_FIRST_STEP) pel : 1 = (MAX_FIRST_STEP/2) pel, 2 = (MAX_FIRST_STEP/4) pel... etc.
ss = &x->ss[search_param * x->searches_per_step];
tot_steps = (x->ss_count / x->searches_per_step) - search_param;
i = 1;
for (step = 0; step < tot_steps ; step++)
{
int all_in = 1, t;
1191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260
// To know if all neighbor points are within the bounds, 4 bounds checking are enough instead of
// checking 4 bounds for each points.
all_in &= ((best_mv->as_mv.row + ss[i].mv.row)> x->mv_row_min);
all_in &= ((best_mv->as_mv.row + ss[i+1].mv.row) < x->mv_row_max);
all_in &= ((best_mv->as_mv.col + ss[i+2].mv.col) > x->mv_col_min);
all_in &= ((best_mv->as_mv.col + ss[i+3].mv.col) < x->mv_col_max);
if (all_in)
{
unsigned int sad_array[4];
for (j = 0 ; j < x->searches_per_step ; j += 4)
{
unsigned char *block_offset[4];
for (t = 0; t < 4; t++)
block_offset[t] = ss[i+t].offset + best_address;
fn_ptr->sdx4df(what, what_stride, block_offset, in_what_stride, sad_array);
for (t = 0; t < 4; t++, i++)
{
if (sad_array[t] < bestsad)
{
this_mv.as_mv.row = best_mv->as_mv.row + ss[i].mv.row;
this_mv.as_mv.col = best_mv->as_mv.col + ss[i].mv.col;
sad_array[t] += mvsad_err_cost(&this_mv, &fcenter_mv,
mvsadcost, sad_per_bit);
if (sad_array[t] < bestsad)
{
bestsad = sad_array[t];
best_site = i;
}
}
}
}
}
else
{
for (j = 0 ; j < x->searches_per_step ; j++)
{
// Trap illegal vectors
this_row_offset = best_mv->as_mv.row + ss[i].mv.row;
this_col_offset = best_mv->as_mv.col + ss[i].mv.col;
if ((this_col_offset > x->mv_col_min) && (this_col_offset < x->mv_col_max) &&
(this_row_offset > x->mv_row_min) && (this_row_offset < x->mv_row_max))
{
check_here = ss[i].offset + best_address;
thissad = fn_ptr->sdf(what, what_stride, check_here , in_what_stride, bestsad);
if (thissad < bestsad)
{
this_mv.as_mv.row = this_row_offset;
this_mv.as_mv.col = this_col_offset;
thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
mvsadcost, sad_per_bit);
if (thissad < bestsad)
{
bestsad = thissad;
best_site = i;
}
}
}
i++;
}
}
1261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330
if (best_site != last_site)
{
best_mv->as_mv.row += ss[best_site].mv.row;
best_mv->as_mv.col += ss[best_site].mv.col;
best_address += ss[best_site].offset;
last_site = best_site;
}
else if (best_address == in_what)
(*num00)++;
}
this_mv.as_mv.row = best_mv->as_mv.row << 3;
this_mv.as_mv.col = best_mv->as_mv.col << 3;
if (bestsad == INT_MAX)
return INT_MAX;
return fn_ptr->vf(what, what_stride, best_address, in_what_stride, (unsigned int *)(&thissad))
+ mv_err_cost(&this_mv, center_mv, mvcost, x->errorperbit);
}
int vp8_full_search_sad(MACROBLOCK *x, BLOCK *b, BLOCKD *d, int_mv *ref_mv,
int sad_per_bit, int distance,
vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2],
int_mv *center_mv)
{
unsigned char *what = (*(b->base_src) + b->src);
int what_stride = b->src_stride;
unsigned char *in_what;
int in_what_stride = d->pre_stride;
int mv_stride = d->pre_stride;
unsigned char *bestaddress;
int_mv *best_mv = &d->bmi.mv;
int_mv this_mv;
int bestsad = INT_MAX;
int r, c;
unsigned char *check_here;
int thissad;
int ref_row = ref_mv->as_mv.row;
int ref_col = ref_mv->as_mv.col;
int row_min = ref_row - distance;
int row_max = ref_row + distance;
int col_min = ref_col - distance;
int col_max = ref_col + distance;
int *mvsadcost[2] = {x->mvsadcost[0], x->mvsadcost[1]};
int_mv fcenter_mv;
fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
// Work out the mid point for the search
in_what = *(d->base_pre) + d->pre;
bestaddress = in_what + (ref_row * d->pre_stride) + ref_col;
best_mv->as_mv.row = ref_row;
best_mv->as_mv.col = ref_col;
// Baseline value at the centre
bestsad = fn_ptr->sdf(what, what_stride, bestaddress,
in_what_stride, 0x7fffffff)
+ mvsad_err_cost(best_mv, &fcenter_mv, mvsadcost, sad_per_bit);
// Apply further limits to prevent us looking using vectors that stretch beyiond the UMV border
if (col_min < x->mv_col_min)
col_min = x->mv_col_min;
1331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400
if (col_max > x->mv_col_max)
col_max = x->mv_col_max;
if (row_min < x->mv_row_min)
row_min = x->mv_row_min;
if (row_max > x->mv_row_max)
row_max = x->mv_row_max;
for (r = row_min; r < row_max ; r++)
{
this_mv.as_mv.row = r;
check_here = r * mv_stride + in_what + col_min;
for (c = col_min; c < col_max; c++)
{
thissad = fn_ptr->sdf(what, what_stride, check_here , in_what_stride, bestsad);
this_mv.as_mv.col = c;
thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
mvsadcost, sad_per_bit);
if (thissad < bestsad)
{
bestsad = thissad;
best_mv->as_mv.row = r;
best_mv->as_mv.col = c;
bestaddress = check_here;
}
check_here++;
}
}
this_mv.as_mv.row = best_mv->as_mv.row << 3;
this_mv.as_mv.col = best_mv->as_mv.col << 3;
if (bestsad < INT_MAX)
return fn_ptr->vf(what, what_stride, bestaddress, in_what_stride, (unsigned int *)(&thissad))
+ mv_err_cost(&this_mv, center_mv, mvcost, x->errorperbit);
else
return INT_MAX;
}
int vp8_full_search_sadx3(MACROBLOCK *x, BLOCK *b, BLOCKD *d, int_mv *ref_mv,
int sad_per_bit, int distance,
vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2],
int_mv *center_mv)
{
unsigned char *what = (*(b->base_src) + b->src);
int what_stride = b->src_stride;
unsigned char *in_what;
int in_what_stride = d->pre_stride;
int mv_stride = d->pre_stride;
unsigned char *bestaddress;
int_mv *best_mv = &d->bmi.mv;
int_mv this_mv;
int bestsad = INT_MAX;
int r, c;
unsigned char *check_here;
unsigned int thissad;
int ref_row = ref_mv->as_mv.row;
int ref_col = ref_mv->as_mv.col;
int row_min = ref_row - distance;
int row_max = ref_row + distance;
int col_min = ref_col - distance;
int col_max = ref_col + distance;
1401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470
unsigned int sad_array[3];
int *mvsadcost[2] = {x->mvsadcost[0], x->mvsadcost[1]};
int_mv fcenter_mv;
fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
// Work out the mid point for the search
in_what = *(d->base_pre) + d->pre;
bestaddress = in_what + (ref_row * d->pre_stride) + ref_col;
best_mv->as_mv.row = ref_row;
best_mv->as_mv.col = ref_col;
// Baseline value at the centre
bestsad = fn_ptr->sdf(what, what_stride,
bestaddress, in_what_stride, 0x7fffffff)
+ mvsad_err_cost(best_mv, &fcenter_mv, mvsadcost, sad_per_bit);
// Apply further limits to prevent us looking using vectors that stretch beyiond the UMV border
if (col_min < x->mv_col_min)
col_min = x->mv_col_min;
if (col_max > x->mv_col_max)
col_max = x->mv_col_max;
if (row_min < x->mv_row_min)
row_min = x->mv_row_min;
if (row_max > x->mv_row_max)
row_max = x->mv_row_max;
for (r = row_min; r < row_max ; r++)
{
this_mv.as_mv.row = r;
check_here = r * mv_stride + in_what + col_min;
c = col_min;
while ((c + 2) < col_max)
{
int i;
fn_ptr->sdx3f(what, what_stride, check_here , in_what_stride, sad_array);
for (i = 0; i < 3; i++)
{
thissad = sad_array[i];
if (thissad < bestsad)
{
this_mv.as_mv.col = c;
thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
mvsadcost, sad_per_bit);
if (thissad < bestsad)
{
bestsad = thissad;
best_mv->as_mv.row = r;
best_mv->as_mv.col = c;
bestaddress = check_here;
}
}
check_here++;
c++;
}
}
while (c < col_max)
1471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540
{
thissad = fn_ptr->sdf(what, what_stride, check_here , in_what_stride, bestsad);
if (thissad < bestsad)
{
this_mv.as_mv.col = c;
thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
mvsadcost, sad_per_bit);
if (thissad < bestsad)
{
bestsad = thissad;
best_mv->as_mv.row = r;
best_mv->as_mv.col = c;
bestaddress = check_here;
}
}
check_here ++;
c ++;
}
}
this_mv.as_mv.row = best_mv->as_mv.row << 3;
this_mv.as_mv.col = best_mv->as_mv.col << 3;
if (bestsad < INT_MAX)
return fn_ptr->vf(what, what_stride, bestaddress, in_what_stride, (unsigned int *)(&thissad))
+ mv_err_cost(&this_mv, center_mv, mvcost, x->errorperbit);
else
return INT_MAX;
}
int vp8_full_search_sadx8(MACROBLOCK *x, BLOCK *b, BLOCKD *d, int_mv *ref_mv,
int sad_per_bit, int distance,
vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2],
int_mv *center_mv)
{
unsigned char *what = (*(b->base_src) + b->src);
int what_stride = b->src_stride;
unsigned char *in_what;
int in_what_stride = d->pre_stride;
int mv_stride = d->pre_stride;
unsigned char *bestaddress;
int_mv *best_mv = &d->bmi.mv;
int_mv this_mv;
int bestsad = INT_MAX;
int r, c;
unsigned char *check_here;
unsigned int thissad;
int ref_row = ref_mv->as_mv.row;
int ref_col = ref_mv->as_mv.col;
int row_min = ref_row - distance;
int row_max = ref_row + distance;
int col_min = ref_col - distance;
int col_max = ref_col + distance;
DECLARE_ALIGNED_ARRAY(16, unsigned short, sad_array8, 8);
unsigned int sad_array[3];
int *mvsadcost[2] = {x->mvsadcost[0], x->mvsadcost[1]};
int_mv fcenter_mv;
fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
// Work out the mid point for the search
1541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610
in_what = *(d->base_pre) + d->pre;
bestaddress = in_what + (ref_row * d->pre_stride) + ref_col;
best_mv->as_mv.row = ref_row;
best_mv->as_mv.col = ref_col;
// Baseline value at the centre
bestsad = fn_ptr->sdf(what, what_stride,
bestaddress, in_what_stride, 0x7fffffff)
+ mvsad_err_cost(best_mv, &fcenter_mv, mvsadcost, sad_per_bit);
// Apply further limits to prevent us looking using vectors that stretch beyiond the UMV border
if (col_min < x->mv_col_min)
col_min = x->mv_col_min;
if (col_max > x->mv_col_max)
col_max = x->mv_col_max;
if (row_min < x->mv_row_min)
row_min = x->mv_row_min;
if (row_max > x->mv_row_max)
row_max = x->mv_row_max;
for (r = row_min; r < row_max ; r++)
{
this_mv.as_mv.row = r;
check_here = r * mv_stride + in_what + col_min;
c = col_min;
while ((c + 7) < col_max)
{
int i;
fn_ptr->sdx8f(what, what_stride, check_here , in_what_stride, sad_array8);
for (i = 0; i < 8; i++)
{
thissad = (unsigned int)sad_array8[i];
if (thissad < bestsad)
{
this_mv.as_mv.col = c;
thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
mvsadcost, sad_per_bit);
if (thissad < bestsad)
{
bestsad = thissad;
best_mv->as_mv.row = r;
best_mv->as_mv.col = c;
bestaddress = check_here;
}
}
check_here++;
c++;
}
}
while ((c + 2) < col_max)
{
int i;
fn_ptr->sdx3f(what, what_stride, check_here , in_what_stride, sad_array);
for (i = 0; i < 3; i++)
{
thissad = sad_array[i];
1611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680
if (thissad < bestsad)
{
this_mv.as_mv.col = c;
thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
mvsadcost, sad_per_bit);
if (thissad < bestsad)
{
bestsad = thissad;
best_mv->as_mv.row = r;
best_mv->as_mv.col = c;
bestaddress = check_here;
}
}
check_here++;
c++;
}
}
while (c < col_max)
{
thissad = fn_ptr->sdf(what, what_stride, check_here , in_what_stride, bestsad);
if (thissad < bestsad)
{
this_mv.as_mv.col = c;
thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
mvsadcost, sad_per_bit);
if (thissad < bestsad)
{
bestsad = thissad;
best_mv->as_mv.row = r;
best_mv->as_mv.col = c;
bestaddress = check_here;
}
}
check_here ++;
c ++;
}
}
this_mv.as_mv.row = best_mv->as_mv.row << 3;
this_mv.as_mv.col = best_mv->as_mv.col << 3;
if (bestsad < INT_MAX)
return fn_ptr->vf(what, what_stride, bestaddress, in_what_stride, (unsigned int *)(&thissad))
+ mv_err_cost(&this_mv, center_mv, mvcost, x->errorperbit);
else
return INT_MAX;
}
int vp8_refining_search_sad(MACROBLOCK *x, BLOCK *b, BLOCKD *d, int_mv *ref_mv,
int error_per_bit, int search_range,
vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2],
int_mv *center_mv)
{
MV neighbors[4] = {{-1, 0}, {0, -1}, {0, 1}, {1, 0}};
int i, j;
short this_row_offset, this_col_offset;
int what_stride = b->src_stride;
int in_what_stride = d->pre_stride;
unsigned char *what = (*(b->base_src) + b->src);
unsigned char *best_address = (unsigned char *)(*(d->base_pre) + d->pre +
(ref_mv->as_mv.row * (d->pre_stride)) + ref_mv->as_mv.col);
unsigned char *check_here;
unsigned int thissad;
1681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750
int_mv this_mv;
unsigned int bestsad = INT_MAX;
int *mvsadcost[2] = {x->mvsadcost[0], x->mvsadcost[1]};
int_mv fcenter_mv;
fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
bestsad = fn_ptr->sdf(what, what_stride, best_address, in_what_stride, 0x7fffffff) + mvsad_err_cost(ref_mv, &fcenter_mv, mvsadcost, error_per_bit);
for (i=0; i<search_range; i++)
{
int best_site = -1;
for (j = 0 ; j < 4 ; j++)
{
this_row_offset = ref_mv->as_mv.row + neighbors[j].row;
this_col_offset = ref_mv->as_mv.col + neighbors[j].col;
if ((this_col_offset > x->mv_col_min) && (this_col_offset < x->mv_col_max) &&
(this_row_offset > x->mv_row_min) && (this_row_offset < x->mv_row_max))
{
check_here = (neighbors[j].row)*in_what_stride + neighbors[j].col + best_address;
thissad = fn_ptr->sdf(what, what_stride, check_here , in_what_stride, bestsad);
if (thissad < bestsad)
{
this_mv.as_mv.row = this_row_offset;
this_mv.as_mv.col = this_col_offset;
thissad += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, error_per_bit);
if (thissad < bestsad)
{
bestsad = thissad;
best_site = j;
}
}
}
}
if (best_site == -1)
break;
else
{
ref_mv->as_mv.row += neighbors[best_site].row;
ref_mv->as_mv.col += neighbors[best_site].col;
best_address += (neighbors[best_site].row)*in_what_stride + neighbors[best_site].col;
}
}
this_mv.as_mv.row = ref_mv->as_mv.row << 3;
this_mv.as_mv.col = ref_mv->as_mv.col << 3;
if (bestsad < INT_MAX)
return fn_ptr->vf(what, what_stride, best_address, in_what_stride, (unsigned int *)(&thissad))
+ mv_err_cost(&this_mv, center_mv, mvcost, x->errorperbit);
else
return INT_MAX;
}
int vp8_refining_search_sadx4(MACROBLOCK *x, BLOCK *b, BLOCKD *d,
int_mv *ref_mv, int error_per_bit,
int search_range, vp8_variance_fn_ptr_t *fn_ptr,
int *mvcost[2], int_mv *center_mv)
{
MV neighbors[4] = {{-1, 0}, {0, -1}, {0, 1}, {1, 0}};
int i, j;
short this_row_offset, this_col_offset;
1751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820
int what_stride = b->src_stride;
int in_what_stride = d->pre_stride;
unsigned char *what = (*(b->base_src) + b->src);
unsigned char *best_address = (unsigned char *)(*(d->base_pre) + d->pre +
(ref_mv->as_mv.row * (d->pre_stride)) + ref_mv->as_mv.col);
unsigned char *check_here;
unsigned int thissad;
int_mv this_mv;
unsigned int bestsad = INT_MAX;
int *mvsadcost[2] = {x->mvsadcost[0], x->mvsadcost[1]};
int_mv fcenter_mv;
fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
bestsad = fn_ptr->sdf(what, what_stride, best_address, in_what_stride, 0x7fffffff) + mvsad_err_cost(ref_mv, &fcenter_mv, mvsadcost, error_per_bit);
for (i=0; i<search_range; i++)
{
int best_site = -1;
int all_in = 1;
all_in &= ((ref_mv->as_mv.row - 1) > x->mv_row_min);
all_in &= ((ref_mv->as_mv.row + 1) < x->mv_row_max);
all_in &= ((ref_mv->as_mv.col - 1) > x->mv_col_min);
all_in &= ((ref_mv->as_mv.col + 1) < x->mv_col_max);
if(all_in)
{
unsigned int sad_array[4];
unsigned char *block_offset[4];
block_offset[0] = best_address - in_what_stride;
block_offset[1] = best_address - 1;
block_offset[2] = best_address + 1;
block_offset[3] = best_address + in_what_stride;
fn_ptr->sdx4df(what, what_stride, block_offset, in_what_stride, sad_array);
for (j = 0; j < 4; j++)
{
if (sad_array[j] < bestsad)
{
this_mv.as_mv.row = ref_mv->as_mv.row + neighbors[j].row;
this_mv.as_mv.col = ref_mv->as_mv.col + neighbors[j].col;
sad_array[j] += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, error_per_bit);
if (sad_array[j] < bestsad)
{
bestsad = sad_array[j];
best_site = j;
}
}
}
}
else
{
for (j = 0 ; j < 4 ; j++)
{
this_row_offset = ref_mv->as_mv.row + neighbors[j].row;
this_col_offset = ref_mv->as_mv.col + neighbors[j].col;
if ((this_col_offset > x->mv_col_min) && (this_col_offset < x->mv_col_max) &&
(this_row_offset > x->mv_row_min) && (this_row_offset < x->mv_row_max))
{
check_here = (neighbors[j].row)*in_what_stride + neighbors[j].col + best_address;
thissad = fn_ptr->sdf(what, what_stride, check_here , in_what_stride, bestsad);
if (thissad < bestsad)
{
1821182218231824182518261827182818291830183118321833183418351836183718381839184018411842184318441845184618471848184918501851185218531854185518561857185818591860186118621863186418651866186718681869187018711872187318741875187618771878187918801881188218831884188518861887188818891890
this_mv.as_mv.row = this_row_offset;
this_mv.as_mv.col = this_col_offset;
thissad += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, error_per_bit);
if (thissad < bestsad)
{
bestsad = thissad;
best_site = j;
}
}
}
}
}
if (best_site == -1)
break;
else
{
ref_mv->as_mv.row += neighbors[best_site].row;
ref_mv->as_mv.col += neighbors[best_site].col;
best_address += (neighbors[best_site].row)*in_what_stride + neighbors[best_site].col;
}
}
this_mv.as_mv.row = ref_mv->as_mv.row << 3;
this_mv.as_mv.col = ref_mv->as_mv.col << 3;
if (bestsad < INT_MAX)
return fn_ptr->vf(what, what_stride, best_address, in_what_stride, (unsigned int *)(&thissad))
+ mv_err_cost(&this_mv, center_mv, mvcost, x->errorperbit);
else
return INT_MAX;
}
#ifdef ENTROPY_STATS
void print_mode_context(void)
{
FILE *f = fopen("modecont.c", "w");
int i, j;
fprintf(f, "#include \"entropy.h\"\n");
fprintf(f, "const int vp8_mode_contexts[6][4] =\n");
fprintf(f, "{\n");
for (j = 0; j < 6; j++)
{
fprintf(f, " { // %d \n", j);
fprintf(f, " ");
for (i = 0; i < 4; i++)
{
int overal_prob;
int this_prob;
int count; // = mv_ref_ct[j][i][0]+mv_ref_ct[j][i][1];
// Overall probs
count = mv_mode_cts[i][0] + mv_mode_cts[i][1];
if (count)
overal_prob = 256 * mv_mode_cts[i][0] / count;
else
overal_prob = 128;
if (overal_prob == 0)
overal_prob = 1;
// context probs
count = mv_ref_ct[j][i][0] + mv_ref_ct[j][i][1];
if (count)
1891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927192819291930193119321933193419351936193719381939194019411942194319441945194619471948194919501951195219531954195519561957195819591960
this_prob = 256 * mv_ref_ct[j][i][0] / count;
else
this_prob = 128;
if (this_prob == 0)
this_prob = 1;
fprintf(f, "%5d, ", this_prob);
//fprintf(f,"%5d, %5d, %8d,", this_prob, overal_prob, (this_prob << 10)/overal_prob);
//fprintf(f,"%8d, ", (this_prob << 10)/overal_prob);
}
fprintf(f, " },\n");
}
fprintf(f, "};\n");
fclose(f);
}
/* MV ref count ENTROPY_STATS stats code */
#ifdef ENTROPY_STATS
void init_mv_ref_counts()
{
vpx_memset(mv_ref_ct, 0, sizeof(mv_ref_ct));
vpx_memset(mv_mode_cts, 0, sizeof(mv_mode_cts));
}
void accum_mv_refs(MB_PREDICTION_MODE m, const int ct[4])
{
if (m == ZEROMV)
{
++mv_ref_ct [ct[0]] [0] [0];
++mv_mode_cts[0][0];
}
else
{
++mv_ref_ct [ct[0]] [0] [1];
++mv_mode_cts[0][1];
if (m == NEARESTMV)
{
++mv_ref_ct [ct[1]] [1] [0];
++mv_mode_cts[1][0];
}
else
{
++mv_ref_ct [ct[1]] [1] [1];
++mv_mode_cts[1][1];
if (m == NEARMV)
{
++mv_ref_ct [ct[2]] [2] [0];
++mv_mode_cts[2][0];
}
else
{
++mv_ref_ct [ct[2]] [2] [1];
++mv_mode_cts[2][1];
if (m == NEWMV)
{
++mv_ref_ct [ct[3]] [3] [0];
++mv_mode_cts[3][0];
}
else
{
++mv_ref_ct [ct[3]] [3] [1];
++mv_mode_cts[3][1];
}
}
19611962196319641965196619671968
}
}
}
#endif/* END MV ref count ENTROPY_STATS stats code */
#endif