-
Paul Wilkins authored
Also tweaks to other features and experiments with what is on and off at different speed settings. Change-Id: I3e1d0be0d195216bf17c2ac5df67f34ce0b306b2
9f3ab834
/* * 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 <stdio.h>#include <limits.h>#include <math.h>#include "vp9/encoder/vp9_onyx_int.h"#include "vp9/encoder/vp9_mcomp.h"#include "vpx_mem/vpx_mem.h"#include "./vpx_config.h"#include "vp9/common/vp9_findnearmv.h"#include "vp9/common/vp9_common.h"void vp9_clamp_mv_min_max(MACROBLOCK *x, int_mv *ref_mv) { int col_min = (ref_mv->as_mv.col >> 3) - MAX_FULL_PEL_VAL + ((ref_mv->as_mv.col & 7) ? 1 : 0); int row_min = (ref_mv->as_mv.row >> 3) - MAX_FULL_PEL_VAL + ((ref_mv->as_mv.row & 7) ? 1 : 0); int col_max = (ref_mv->as_mv.col >> 3) + MAX_FULL_PEL_VAL; int row_max = (ref_mv->as_mv.row >> 3) + MAX_FULL_PEL_VAL; /* Get intersection of UMV window and valid MV window to reduce # of checks in diamond search. */ if (x->mv_col_min < col_min) x->mv_col_min = col_min; if (x->mv_col_max > col_max) x->mv_col_max = col_max; if (x->mv_row_min < row_min) x->mv_row_min = row_min; if (x->mv_row_max > row_max) x->mv_row_max = row_max;}int vp9_init_search_range(VP9_COMP *cpi, int size) { int sr = 0; // Minimum search size no matter what the passed in value. size = MAX(16, size); while ((size << sr) < MAX_FULL_PEL_VAL) sr++; if (sr) sr--; sr += cpi->sf.reduce_first_step_size; sr = MIN(sr, (cpi->sf.max_step_search_steps - 2)); return sr;}int vp9_mv_bit_cost(int_mv *mv, int_mv *ref, int *mvjcost, int *mvcost[2], int weight, int ishp) { MV v; v.row = mv->as_mv.row - ref->as_mv.row; v.col = mv->as_mv.col - ref->as_mv.col; return ROUND_POWER_OF_TWO((mvjcost[vp9_get_mv_joint(&v)] + mvcost[0][v.row] + mvcost[1][v.col]) * weight, 7);}static int mv_err_cost(int_mv *mv, int_mv *ref, int *mvjcost, int *mvcost[2], int error_per_bit, int ishp) { if (mvcost) {7172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140
MV v;
v.row = mv->as_mv.row - ref->as_mv.row;
v.col = mv->as_mv.col - ref->as_mv.col;
return ROUND_POWER_OF_TWO((mvjcost[vp9_get_mv_joint(&v)] +
mvcost[0][v.row] +
mvcost[1][v.col]) * error_per_bit, 13);
}
return 0;
}
static int mvsad_err_cost(int_mv *mv, int_mv *ref, int *mvjsadcost,
int *mvsadcost[2], int error_per_bit) {
if (mvsadcost) {
MV v;
v.row = mv->as_mv.row - ref->as_mv.row;
v.col = mv->as_mv.col - ref->as_mv.col;
return ROUND_POWER_OF_TWO((mvjsadcost[vp9_get_mv_joint(&v)] +
mvsadcost[0][v.row] +
mvsadcost[1][v.col]) * error_per_bit, 8);
}
return 0;
}
void vp9_init_dsmotion_compensation(MACROBLOCK *x, int stride) {
int len;
int search_site_count = 0;
// Generate offsets for 4 search sites per 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++;
for (len = MAX_FIRST_STEP; len > 0; len /= 2) {
// 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++;
}
x->ss_count = search_site_count;
x->searches_per_step = 4;
}
void vp9_init3smotion_compensation(MACROBLOCK *x, int stride) {
int len;
int search_site_count = 0;
// Generate offsets for 8 search sites per step.
x->ss[search_site_count].mv.col = 0;
x->ss[search_site_count].mv.row = 0;
141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210
x->ss[search_site_count].offset = 0;
search_site_count++;
for (len = MAX_FIRST_STEP; len > 0; len /= 2) {
// 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++;
// 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++;
}
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.
*/
/* estimated cost of a motion vector (r,c) */
#define MVC(r, c) \
(mvcost ? \
211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280
((mvjcost[((r) != rr) * 2 + ((c) != rc)] + \
mvcost[0][((r) - rr)] + mvcost[1][((c) - rc)]) * \
error_per_bit + 4096) >> 13 : 0)
#define SP(x) (((x) & 7) << 1) // convert motion vector component to offset
// for svf calc
#define IFMVCV(r, c, s, e) \
if (c >= minc && c <= maxc && r >= minr && r <= maxr) \
s \
else \
e;
/* pointer to predictor base of a motionvector */
#define PRE(r, c) (y + (((r) >> 3) * y_stride + ((c) >> 3) -(offset)))
/* returns subpixel variance error function */
#define DIST(r, c) \
vfp->svf(PRE(r, c), y_stride, SP(c), SP(r), z, src_stride, &sse)
/* checks if (r, c) has better score than previous best */
#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;)
int vp9_find_best_sub_pixel_step_iteratively(MACROBLOCK *x,
int_mv *bestmv, int_mv *ref_mv,
int error_per_bit,
const vp9_variance_fn_ptr_t *vfp,
int *mvjcost, int *mvcost[2],
int *distortion,
unsigned int *sse1) {
uint8_t *z = x->plane[0].src.buf;
int src_stride = x->plane[0].src.stride;
MACROBLOCKD *xd = &x->e_mbd;
int rr, rc, br, bc, hstep;
int tr, tc;
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;
unsigned int eighthiters = 4;
int thismse;
int maxc, minc, maxr, minr;
int y_stride;
int offset;
int usehp = xd->allow_high_precision_mv;
uint8_t *y = xd->plane[0].pre[0].buf +
(bestmv->as_mv.row) * xd->plane[0].pre[0].stride +
bestmv->as_mv.col;
y_stride = xd->plane[0].pre[0].stride;
rr = ref_mv->as_mv.row;
rc = ref_mv->as_mv.col;
br = bestmv->as_mv.row << 3;
281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350
bc = bestmv->as_mv.col << 3;
hstep = 4;
minc = MAX(x->mv_col_min << 3, (ref_mv->as_mv.col) - ((1 << MV_MAX_BITS) - 1));
maxc = MIN(x->mv_col_max << 3, (ref_mv->as_mv.col) + ((1 << MV_MAX_BITS) - 1));
minr = MAX(x->mv_row_min << 3, (ref_mv->as_mv.row) - ((1 << MV_MAX_BITS) - 1));
maxr = MIN(x->mv_row_max << 3, (ref_mv->as_mv.row) + ((1 << MV_MAX_BITS) - 1));
tr = br;
tc = bc;
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, src_stride, sse1);
*distortion = besterr;
besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost,
error_per_bit, xd->allow_high_precision_mv);
// 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 - hstep);
CHECK_BETTER(right, tr, tc + hstep);
CHECK_BETTER(up, tr - hstep, tc);
CHECK_BETTER(down, tr + hstep, tc);
whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
switch (whichdir) {
case 0:
CHECK_BETTER(diag, tr - hstep, tc - hstep);
break;
case 1:
CHECK_BETTER(diag, tr - hstep, tc + hstep);
break;
case 2:
CHECK_BETTER(diag, tr + hstep, tc - hstep);
break;
case 3:
CHECK_BETTER(diag, tr + hstep, tc + hstep);
break;
}
// no reason to check the same one again.
if (tr == br && tc == bc)
break;
tr = br;
tc = bc;
}
// TODO: Each subsequent iteration checks at least one point in common with
// the last iteration could be 2 ( if diag selected) 1/4 pel
hstep >>= 1;
while (--quarteriters) {
CHECK_BETTER(left, tr, tc - hstep);
CHECK_BETTER(right, tr, tc + hstep);
CHECK_BETTER(up, tr - hstep, tc);
CHECK_BETTER(down, tr + hstep, tc);
whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
switch (whichdir) {
case 0:
351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420
CHECK_BETTER(diag, tr - hstep, tc - hstep);
break;
case 1:
CHECK_BETTER(diag, tr - hstep, tc + hstep);
break;
case 2:
CHECK_BETTER(diag, tr + hstep, tc - hstep);
break;
case 3:
CHECK_BETTER(diag, tr + hstep, tc + hstep);
break;
}
// no reason to check the same one again.
if (tr == br && tc == bc)
break;
tr = br;
tc = bc;
}
if (xd->allow_high_precision_mv) {
usehp = vp9_use_mv_hp(&ref_mv->as_mv);
} else {
usehp = 0;
}
if (usehp) {
hstep >>= 1;
while (--eighthiters) {
CHECK_BETTER(left, tr, tc - hstep);
CHECK_BETTER(right, tr, tc + hstep);
CHECK_BETTER(up, tr - hstep, tc);
CHECK_BETTER(down, tr + hstep, tc);
whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
switch (whichdir) {
case 0:
CHECK_BETTER(diag, tr - hstep, tc - hstep);
break;
case 1:
CHECK_BETTER(diag, tr - hstep, tc + hstep);
break;
case 2:
CHECK_BETTER(diag, tr + hstep, tc - hstep);
break;
case 3:
CHECK_BETTER(diag, tr + hstep, tc + hstep);
break;
}
// no reason to check the same one again.
if (tr == br && tc == bc)
break;
tr = br;
tc = bc;
}
}
bestmv->as_mv.row = br;
bestmv->as_mv.col = bc;
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;
}
421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490
#undef DIST
/* returns subpixel variance error function */
#define DIST(r, c) \
vfp->svaf(PRE(r, c), y_stride, SP(c), SP(r), \
z, src_stride, &sse, second_pred)
int vp9_find_best_sub_pixel_comp(MACROBLOCK *x,
int_mv *bestmv, int_mv *ref_mv,
int error_per_bit,
const vp9_variance_fn_ptr_t *vfp,
int *mvjcost, int *mvcost[2],
int *distortion,
unsigned int *sse1,
const uint8_t *second_pred, int w, int h) {
uint8_t *z = x->plane[0].src.buf;
int src_stride = x->plane[0].src.stride;
MACROBLOCKD *xd = &x->e_mbd;
int rr, rc, br, bc, hstep;
int tr, tc;
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;
unsigned int eighthiters = 4;
int thismse;
int maxc, minc, maxr, minr;
int y_stride;
int offset;
int usehp = xd->allow_high_precision_mv;
DECLARE_ALIGNED_ARRAY(16, uint8_t, comp_pred, 64 * 64);
uint8_t *y = xd->plane[0].pre[0].buf +
(bestmv->as_mv.row) * xd->plane[0].pre[0].stride +
bestmv->as_mv.col;
y_stride = xd->plane[0].pre[0].stride;
rr = ref_mv->as_mv.row;
rc = ref_mv->as_mv.col;
br = bestmv->as_mv.row << 3;
bc = bestmv->as_mv.col << 3;
hstep = 4;
minc = MAX(x->mv_col_min << 3, (ref_mv->as_mv.col) -
((1 << MV_MAX_BITS) - 1));
maxc = MIN(x->mv_col_max << 3, (ref_mv->as_mv.col) +
((1 << MV_MAX_BITS) - 1));
minr = MAX(x->mv_row_min << 3, (ref_mv->as_mv.row) -
((1 << MV_MAX_BITS) - 1));
maxr = MIN(x->mv_row_max << 3, (ref_mv->as_mv.row) +
((1 << MV_MAX_BITS) - 1));
tr = br;
tc = bc;
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
// TODO(yunqingwang): central pointer error was already calculated in full-
// pixel search, and can be passed in this function.
comp_avg_pred(comp_pred, second_pred, w, h, y, y_stride);
besterr = vfp->vf(comp_pred, w, z, src_stride, sse1);
*distortion = besterr;
491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560
besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost,
error_per_bit, xd->allow_high_precision_mv);
// 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 - hstep);
CHECK_BETTER(right, tr, tc + hstep);
CHECK_BETTER(up, tr - hstep, tc);
CHECK_BETTER(down, tr + hstep, tc);
whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
switch (whichdir) {
case 0:
CHECK_BETTER(diag, tr - hstep, tc - hstep);
break;
case 1:
CHECK_BETTER(diag, tr - hstep, tc + hstep);
break;
case 2:
CHECK_BETTER(diag, tr + hstep, tc - hstep);
break;
case 3:
CHECK_BETTER(diag, tr + hstep, tc + hstep);
break;
}
// no reason to check the same one again.
if (tr == br && tc == bc)
break;
tr = br;
tc = bc;
}
// Each subsequent iteration checks at least one point in common with
// the last iteration could be 2 ( if diag selected) 1/4 pel
hstep >>= 1;
while (--quarteriters) {
CHECK_BETTER(left, tr, tc - hstep);
CHECK_BETTER(right, tr, tc + hstep);
CHECK_BETTER(up, tr - hstep, tc);
CHECK_BETTER(down, tr + hstep, tc);
whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
switch (whichdir) {
case 0:
CHECK_BETTER(diag, tr - hstep, tc - hstep);
break;
case 1:
CHECK_BETTER(diag, tr - hstep, tc + hstep);
break;
case 2:
CHECK_BETTER(diag, tr + hstep, tc - hstep);
break;
case 3:
CHECK_BETTER(diag, tr + hstep, tc + hstep);
break;
}
// no reason to check the same one again.
if (tr == br && tc == bc)
break;
tr = br;
tc = bc;
}
561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630
if (xd->allow_high_precision_mv) {
usehp = vp9_use_mv_hp(&ref_mv->as_mv);
} else {
usehp = 0;
}
if (usehp) {
hstep >>= 1;
while (--eighthiters) {
CHECK_BETTER(left, tr, tc - hstep);
CHECK_BETTER(right, tr, tc + hstep);
CHECK_BETTER(up, tr - hstep, tc);
CHECK_BETTER(down, tr + hstep, tc);
whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
switch (whichdir) {
case 0:
CHECK_BETTER(diag, tr - hstep, tc - hstep);
break;
case 1:
CHECK_BETTER(diag, tr - hstep, tc + hstep);
break;
case 2:
CHECK_BETTER(diag, tr + hstep, tc - hstep);
break;
case 3:
CHECK_BETTER(diag, tr + hstep, tc + hstep);
break;
}
// no reason to check the same one again.
if (tr == br && tc == bc)
break;
tr = br;
tc = bc;
}
}
bestmv->as_mv.row = br;
bestmv->as_mv.col = bc;
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 DIST
#undef IFMVCV
#undef CHECK_BETTER
#undef MIN
#undef MAX
int vp9_find_best_sub_pixel_step(MACROBLOCK *x,
int_mv *bestmv, int_mv *ref_mv,
int error_per_bit,
const vp9_variance_fn_ptr_t *vfp,
int *mvjcost, int *mvcost[2], int *distortion,
unsigned int *sse1) {
int bestmse = INT_MAX;
int_mv startmv;
int_mv this_mv;
int_mv orig_mv;
int yrow_movedback = 0, ycol_movedback = 0;
631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700
uint8_t *z = x->plane[0].src.buf;
int src_stride = x->plane[0].src.stride;
int left, right, up, down, diag;
unsigned int sse;
int whichdir;
int thismse;
int y_stride;
MACROBLOCKD *xd = &x->e_mbd;
int usehp = xd->allow_high_precision_mv;
uint8_t *y = xd->plane[0].pre[0].buf +
(bestmv->as_mv.row) * xd->plane[0].pre[0].stride +
bestmv->as_mv.col;
y_stride = xd->plane[0].pre[0].stride;
// central mv
bestmv->as_mv.row <<= 3;
bestmv->as_mv.col <<= 3;
startmv = *bestmv;
orig_mv = *bestmv;
// calculate central point error
bestmse = vfp->vf(y, y_stride, z, src_stride, sse1);
*distortion = bestmse;
bestmse += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit,
xd->allow_high_precision_mv);
// 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, src_stride, &sse);
left = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit,
xd->allow_high_precision_mv);
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, src_stride, &sse);
right = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost,
error_per_bit, xd->allow_high_precision_mv);
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, src_stride, &sse);
up = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit,
xd->allow_high_precision_mv);
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, src_stride, &sse);
down = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit,
701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770
xd->allow_high_precision_mv);
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, 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, 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, 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, src_stride, &sse);
break;
}
diag = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit,
xd->allow_high_precision_mv);
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;
yrow_movedback = 1;
}
if (bestmv->as_mv.col < startmv.as_mv.col) {
y--;
ycol_movedback = 1;
}
startmv = *bestmv;
771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840
// 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;
thismse = vfp->svf(y, y_stride,
SP(this_mv.as_mv.col), SP(this_mv.as_mv.row),
z, src_stride, &sse);
} else {
this_mv.as_mv.col = (startmv.as_mv.col - 8) | 6;
thismse = vfp->svf(y - 1, y_stride, SP(6), SP(this_mv.as_mv.row), z,
src_stride, &sse);
}
left = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit,
xd->allow_high_precision_mv);
if (left < bestmse) {
*bestmv = this_mv;
bestmse = left;
*distortion = thismse;
*sse1 = sse;
}
this_mv.as_mv.col += 4;
thismse = vfp->svf(y, y_stride,
SP(this_mv.as_mv.col), SP(this_mv.as_mv.row),
z, src_stride, &sse);
right = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost,
error_per_bit, xd->allow_high_precision_mv);
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,
SP(this_mv.as_mv.col), SP(this_mv.as_mv.row),
z, src_stride, &sse);
} else {
this_mv.as_mv.row = (startmv.as_mv.row - 8) | 6;
thismse = vfp->svf(y - y_stride, y_stride, SP(this_mv.as_mv.col), SP(6),
z, src_stride, &sse);
}
up = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit,
xd->allow_high_precision_mv);
if (up < bestmse) {
*bestmv = this_mv;
bestmse = up;
*distortion = thismse;
*sse1 = sse;
}
this_mv.as_mv.row += 4;
thismse = vfp->svf(y, y_stride, SP(this_mv.as_mv.col), SP(this_mv.as_mv.row),
z, src_stride, &sse);
down = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit,
xd->allow_high_precision_mv);
if (down < bestmse) {
*bestmv = this_mv;
841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910
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:
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,
SP(this_mv.as_mv.col), SP(this_mv.as_mv.row),
z, src_stride, &sse);
} else {
this_mv.as_mv.col = (startmv.as_mv.col - 8) | 6;
thismse = vfp->svf(y - 1, y_stride,
SP(6), SP(this_mv.as_mv.row), z, 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,
SP(this_mv.as_mv.col), SP(6), z, src_stride, &sse);
} else {
this_mv.as_mv.col = (startmv.as_mv.col - 8) | 6;
thismse = vfp->svf(y - y_stride - 1, y_stride,
SP(6), SP(6), z, 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,
SP(this_mv.as_mv.col), SP(this_mv.as_mv.row),
z, src_stride, &sse);
} else {
this_mv.as_mv.row = (startmv.as_mv.row - 8) | 6;
thismse = vfp->svf(y - y_stride, y_stride,
SP(this_mv.as_mv.col), SP(6), z, 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,
SP(this_mv.as_mv.col), SP(this_mv.as_mv.row),
z, src_stride, &sse);
} else {
this_mv.as_mv.col = (startmv.as_mv.col - 8) | 6;
thismse = vfp->svf(y - 1, y_stride, SP(6), SP(this_mv.as_mv.row), z,
911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980
src_stride, &sse);
}
break;
case 3:
this_mv.as_mv.col += 2;
this_mv.as_mv.row += 2;
thismse = vfp->svf(y, y_stride,
SP(this_mv.as_mv.col), SP(this_mv.as_mv.row),
z, src_stride, &sse);
break;
}
diag = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit,
xd->allow_high_precision_mv);
if (diag < bestmse) {
*bestmv = this_mv;
bestmse = diag;
*distortion = thismse;
*sse1 = sse;
}
if (x->e_mbd.allow_high_precision_mv) {
usehp = vp9_use_mv_hp(&ref_mv->as_mv);
} else {
usehp = 0;
}
if (!usehp)
return bestmse;
/* Now do 1/8th pixel */
if (bestmv->as_mv.row < orig_mv.as_mv.row && !yrow_movedback) {
y -= y_stride;
yrow_movedback = 1;
}
if (bestmv->as_mv.col < orig_mv.as_mv.col && !ycol_movedback) {
y--;
ycol_movedback = 1;
}
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 - 1;
thismse = vfp->svf(y, y_stride,
SP(this_mv.as_mv.col), SP(this_mv.as_mv.row),
z, src_stride, &sse);
} else {
this_mv.as_mv.col = (startmv.as_mv.col - 8) | 7;
thismse = vfp->svf(y - 1, y_stride, SP(7), SP(this_mv.as_mv.row),
z, src_stride, &sse);
}
left = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit,
xd->allow_high_precision_mv);
if (left < bestmse) {
*bestmv = this_mv;
bestmse = left;
*distortion = thismse;
*sse1 = sse;
}
this_mv.as_mv.col += 2;
thismse = vfp->svf(y, y_stride, SP(this_mv.as_mv.col), SP(this_mv.as_mv.row),
981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050
z, src_stride, &sse);
right = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost,
error_per_bit, xd->allow_high_precision_mv);
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 - 1;
thismse = vfp->svf(y, y_stride,
SP(this_mv.as_mv.col), SP(this_mv.as_mv.row),
z, src_stride, &sse);
} else {
this_mv.as_mv.row = (startmv.as_mv.row - 8) | 7;
thismse = vfp->svf(y - y_stride, y_stride,
SP(this_mv.as_mv.col), SP(7), z, src_stride, &sse);
}
up = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit,
xd->allow_high_precision_mv);
if (up < bestmse) {
*bestmv = this_mv;
bestmse = up;
*distortion = thismse;
*sse1 = sse;
}
this_mv.as_mv.row += 2;
thismse = vfp->svf(y, y_stride,
SP(this_mv.as_mv.col), SP(this_mv.as_mv.row),
z, src_stride, &sse);
down = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit,
xd->allow_high_precision_mv);
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:
if (startmv.as_mv.row & 7) {
this_mv.as_mv.row -= 1;
if (startmv.as_mv.col & 7) {
this_mv.as_mv.col -= 1;
thismse = vfp->svf(y, y_stride,
SP(this_mv.as_mv.col), SP(this_mv.as_mv.row),
z, src_stride, &sse);
} else {
this_mv.as_mv.col = (startmv.as_mv.col - 8) | 7;
thismse = vfp->svf(y - 1, y_stride,
1051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120
SP(7), SP(this_mv.as_mv.row),
z, src_stride, &sse);
}
} else {
this_mv.as_mv.row = (startmv.as_mv.row - 8) | 7;
if (startmv.as_mv.col & 7) {
this_mv.as_mv.col -= 1;
thismse = vfp->svf(y - y_stride, y_stride,
SP(this_mv.as_mv.col), SP(7), z, src_stride, &sse);
} else {
this_mv.as_mv.col = (startmv.as_mv.col - 8) | 7;
thismse = vfp->svf(y - y_stride - 1, y_stride,
SP(7), SP(7), z, src_stride, &sse);
}
}
break;
case 1:
this_mv.as_mv.col += 1;
if (startmv.as_mv.row & 7) {
this_mv.as_mv.row -= 1;
thismse = vfp->svf(y, y_stride,
SP(this_mv.as_mv.col), SP(this_mv.as_mv.row),
z, src_stride, &sse);
} else {
this_mv.as_mv.row = (startmv.as_mv.row - 8) | 7;
thismse = vfp->svf(y - y_stride, y_stride,
SP(this_mv.as_mv.col), SP(7), z, src_stride, &sse);
}
break;
case 2:
this_mv.as_mv.row += 1;
if (startmv.as_mv.col & 7) {
this_mv.as_mv.col -= 1;
thismse = vfp->svf(y, y_stride,
SP(this_mv.as_mv.col), SP(this_mv.as_mv.row),
z, src_stride, &sse);
} else {
this_mv.as_mv.col = (startmv.as_mv.col - 8) | 7;
thismse = vfp->svf(y - 1, y_stride,
SP(7), SP(this_mv.as_mv.row), z, src_stride, &sse);
}
break;
case 3:
this_mv.as_mv.col += 1;
this_mv.as_mv.row += 1;
thismse = vfp->svf(y, y_stride,
SP(this_mv.as_mv.col), SP(this_mv.as_mv.row),
z, src_stride, &sse);
break;
}
diag = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit,
xd->allow_high_precision_mv);
if (diag < bestmse) {
*bestmv = this_mv;
bestmse = diag;
*distortion = thismse;
*sse1 = sse;
}
return bestmse;
}
1121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190
#undef SP
int vp9_find_best_half_pixel_step(MACROBLOCK *x,
int_mv *bestmv, int_mv *ref_mv,
int error_per_bit,
const vp9_variance_fn_ptr_t *vfp,
int *mvjcost, int *mvcost[2],
int *distortion,
unsigned int *sse1) {
int bestmse = INT_MAX;
int_mv startmv;
int_mv this_mv;
uint8_t *z = x->plane[0].src.buf;
int src_stride = x->plane[0].src.stride;
int left, right, up, down, diag;
unsigned int sse;
int whichdir;
int thismse;
int y_stride;
MACROBLOCKD *xd = &x->e_mbd;
uint8_t *y = xd->plane[0].pre[0].buf +
(bestmv->as_mv.row) * xd->plane[0].pre[0].stride + bestmv->as_mv.col;
y_stride = xd->plane[0].pre[0].stride;
// 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, src_stride, sse1);
*distortion = bestmse;
bestmse += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit,
xd->allow_high_precision_mv);
// 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, src_stride, &sse);
left = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit,
xd->allow_high_precision_mv);
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, src_stride, &sse);
right = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost,
error_per_bit, xd->allow_high_precision_mv);
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, src_stride, &sse);
up = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit,
xd->allow_high_precision_mv);
if (up < bestmse) {
1191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260
*bestmv = this_mv;
bestmse = up;
*distortion = thismse;
*sse1 = sse;
}
this_mv.as_mv.row += 8;
thismse = vfp->svf_halfpix_v(y, y_stride, z, src_stride, &sse);
down = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit,
xd->allow_high_precision_mv);
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, 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, 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, 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, src_stride, &sse);
break;
}
diag = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit,
xd->allow_high_precision_mv);
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);\
}
1261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330
#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, mvjsadcost, 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 vp9_hex_search
(
MACROBLOCK *x,
int_mv *ref_mv,
int_mv *best_mv,
int search_param,
int sad_per_bit,
const vp9_variance_fn_ptr_t *vfp,
int *mvjsadcost, int *mvsadcost[2],
int *mvjcost, int *mvcost[2],
int_mv *center_mv
) {
const MACROBLOCKD* const xd = &x->e_mbd;
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;
uint8_t *what = x->plane[0].src.buf;
int what_stride = x->plane[0].src.stride;
int in_what_stride = xd->plane[0].pre[0].stride;
int br, bc;
int_mv this_mv;
unsigned int bestsad = 0x7fffffff;
unsigned int thissad;
uint8_t *base_offset;
uint8_t *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
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;
1331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400
// Work out the start point for the search
base_offset = (uint8_t *)(xd->plane[0].pre[0].buf);
this_offset = base_offset + (br * (xd->plane[0].pre[0].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, mvjsadcost, 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;
}
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;
1401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470
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
}
}
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 vp9_diamond_search_sad_c(MACROBLOCK *x,
int_mv *ref_mv, int_mv *best_mv,
int search_param, int sad_per_bit, int *num00,
vp9_variance_fn_ptr_t *fn_ptr, int *mvjcost,
int *mvcost[2], int_mv *center_mv) {
int i, j, step;
const MACROBLOCKD* const xd = &x->e_mbd;
uint8_t *what = x->plane[0].src.buf;
int what_stride = x->plane[0].src.stride;
uint8_t *in_what;
int in_what_stride = xd->plane[0].pre[0].stride;
uint8_t *best_address;
int tot_steps;
int_mv this_mv;
int bestsad = INT_MAX;
int best_site = 0;
int last_site = 0;
int ref_row, ref_col;
int this_row_offset, this_col_offset;
1471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540
search_site *ss;
uint8_t *check_here;
int thissad;
int_mv fcenter_mv;
int *mvjsadcost = x->nmvjointsadcost;
int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]};
fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
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 = (uint8_t *)(xd->plane[0].pre[0].buf +
(ref_row * (xd->plane[0].pre[0].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, mvjsadcost, 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,
mvjsadcost, 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;
1541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610
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, mvjcost, mvcost, x->errorperbit,
xd->allow_high_precision_mv);
}
int vp9_diamond_search_sadx4(MACROBLOCK *x,
int_mv *ref_mv, int_mv *best_mv, int search_param,
int sad_per_bit, int *num00,
vp9_variance_fn_ptr_t *fn_ptr,
int *mvjcost, int *mvcost[2], int_mv *center_mv) {
int i, j, step;
const MACROBLOCKD* const xd = &x->e_mbd;
uint8_t *what = x->plane[0].src.buf;
int what_stride = x->plane[0].src.stride;
uint8_t *in_what;
int in_what_stride = xd->plane[0].pre[0].stride;
uint8_t *best_address;
int tot_steps;
int_mv this_mv;
unsigned 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;
uint8_t *check_here;
unsigned int thissad;
int_mv fcenter_mv;
int *mvjsadcost = x->nmvjointsadcost;
int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]};
fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
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 = (uint8_t *)(xd->plane[0].pre[0].buf +
(ref_row * (xd->plane[0].pre[0].stride)) + ref_col);
best_address = in_what;
// Check the starting position
bestsad = fn_ptr->sdf(what, what_stride,
in_what, in_what_stride, 0x7fffffff)
1611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680
+ mvsad_err_cost(best_mv, &fcenter_mv, mvjsadcost, 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;
// 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 const *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,
mvjsadcost, 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,
mvjsadcost, mvsadcost, sad_per_bit);
if (thissad < bestsad) {
bestsad = thissad;
best_site = i;
}
}
}
i++;
1681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750
}
}
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, mvjcost, mvcost, x->errorperbit,
xd->allow_high_precision_mv);
}
/* do_refine: If last step (1-away) of n-step search doesn't pick the center
point as the best match, we will do a final 1-away diamond
refining search */
int vp9_full_pixel_diamond(VP9_COMP *cpi, MACROBLOCK *x,
int_mv *mvp_full, int step_param,
int sadpb, int further_steps,
int do_refine, vp9_variance_fn_ptr_t *fn_ptr,
int_mv *ref_mv, int_mv *dst_mv) {
int_mv temp_mv;
int thissme, n, num00;
int bestsme = cpi->diamond_search_sad(x, mvp_full, &temp_mv,
step_param, sadpb, &num00,
fn_ptr, x->nmvjointcost,
x->mvcost, ref_mv);
dst_mv->as_int = temp_mv.as_int;
n = num00;
num00 = 0;
/* If there won't be more n-step search, check to see if refining search is needed. */
if (n > further_steps)
do_refine = 0;
while (n < further_steps) {
n++;
if (num00)
num00--;
else {
thissme = cpi->diamond_search_sad(x, mvp_full, &temp_mv,
step_param + n, sadpb, &num00,
fn_ptr, x->nmvjointcost, x->mvcost,
ref_mv);
/* check to see if refining search is needed. */
if (num00 > (further_steps - n))
do_refine = 0;
if (thissme < bestsme) {
bestsme = thissme;
dst_mv->as_int = temp_mv.as_int;
}
}
}
1751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820
/* final 1-away diamond refining search */
if (do_refine == 1) {
int search_range = 8;
int_mv best_mv;
best_mv.as_int = dst_mv->as_int;
thissme = cpi->refining_search_sad(x, &best_mv, sadpb, search_range,
fn_ptr, x->nmvjointcost, x->mvcost,
ref_mv);
if (thissme < bestsme) {
bestsme = thissme;
dst_mv->as_int = best_mv.as_int;
}
}
return bestsme;
}
int vp9_full_search_sad_c(MACROBLOCK *x, int_mv *ref_mv,
int sad_per_bit, int distance,
vp9_variance_fn_ptr_t *fn_ptr, int *mvjcost,
int *mvcost[2],
int_mv *center_mv, int n) {
const MACROBLOCKD* const xd = &x->e_mbd;
uint8_t *what = x->plane[0].src.buf;
int what_stride = x->plane[0].src.stride;
uint8_t *in_what;
int in_what_stride = xd->plane[0].pre[0].stride;
int mv_stride = xd->plane[0].pre[0].stride;
uint8_t *bestaddress;
int_mv *best_mv = &x->e_mbd.mode_info_context->bmi[n].as_mv[0];
int_mv this_mv;
int bestsad = INT_MAX;
int r, c;
uint8_t *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_mv fcenter_mv;
int *mvjsadcost = x->nmvjointsadcost;
int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]};
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 = xd->plane[0].pre[0].buf;
bestaddress = in_what + (ref_row * xd->plane[0].pre[0].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, mvjsadcost, 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)
1821182218231824182518261827182818291830183118321833183418351836183718381839184018411842184318441845184618471848184918501851185218531854185518561857185818591860186118621863186418651866186718681869187018711872187318741875187618771878187918801881188218831884188518861887188818891890
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,
mvjsadcost, 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, mvjcost, mvcost, x->errorperbit,
xd->allow_high_precision_mv);
else
return INT_MAX;
}
int vp9_full_search_sadx3(MACROBLOCK *x, int_mv *ref_mv,
int sad_per_bit, int distance,
vp9_variance_fn_ptr_t *fn_ptr, int *mvjcost,
int *mvcost[2], int_mv *center_mv, int n) {
const MACROBLOCKD* const xd = &x->e_mbd;
uint8_t *what = x->plane[0].src.buf;
int what_stride = x->plane[0].src.stride;
uint8_t *in_what;
int in_what_stride = xd->plane[0].pre[0].stride;
int mv_stride = xd->plane[0].pre[0].stride;
uint8_t *bestaddress;
int_mv *best_mv = &x->e_mbd.mode_info_context->bmi[n].as_mv[0];
int_mv this_mv;
unsigned int bestsad = INT_MAX;
int r, c;
uint8_t *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;
1891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927192819291930193119321933193419351936193719381939194019411942194319441945194619471948194919501951195219531954195519561957195819591960
unsigned int sad_array[3];
int_mv fcenter_mv;
int *mvjsadcost = x->nmvjointsadcost;
int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]};
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 = xd->plane[0].pre[0].buf;
bestaddress = in_what + (ref_row * xd->plane[0].pre[0].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, mvjsadcost, 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,
mvjsadcost, 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) {
1961196219631964196519661967196819691970197119721973197419751976197719781979198019811982198319841985198619871988198919901991199219931994199519961997199819992000200120022003200420052006200720082009201020112012201320142015201620172018201920202021202220232024202520262027202820292030
this_mv.as_mv.col = c;
thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
mvjsadcost, 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, mvjcost, mvcost, x->errorperbit,
xd->allow_high_precision_mv);
else
return INT_MAX;
}
int vp9_full_search_sadx8(MACROBLOCK *x, int_mv *ref_mv,
int sad_per_bit, int distance,
vp9_variance_fn_ptr_t *fn_ptr,
int *mvjcost, int *mvcost[2],
int_mv *center_mv, int n) {
const MACROBLOCKD* const xd = &x->e_mbd;
uint8_t *what = x->plane[0].src.buf;
int what_stride = x->plane[0].src.stride;
uint8_t *in_what;
int in_what_stride = xd->plane[0].pre[0].stride;
int mv_stride = xd->plane[0].pre[0].stride;
uint8_t *bestaddress;
int_mv *best_mv = &x->e_mbd.mode_info_context->bmi[n].as_mv[0];
int_mv this_mv;
unsigned int bestsad = INT_MAX;
int r, c;
uint8_t *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, uint32_t, sad_array8, 8);
unsigned int sad_array[3];
int_mv fcenter_mv;
int *mvjsadcost = x->nmvjointsadcost;
int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]};
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
2031203220332034203520362037203820392040204120422043204420452046204720482049205020512052205320542055205620572058205920602061206220632064206520662067206820692070207120722073207420752076207720782079208020812082208320842085208620872088208920902091209220932094209520962097209820992100
in_what = xd->plane[0].pre[0].buf;
bestaddress = in_what + (ref_row * xd->plane[0].pre[0].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, mvjsadcost, 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,
mvjsadcost, 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];
if (thissad < bestsad) {
this_mv.as_mv.col = c;
thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
mvjsadcost, mvsadcost, sad_per_bit);
if (thissad < bestsad) {
2101210221032104210521062107210821092110211121122113211421152116211721182119212021212122212321242125212621272128212921302131213221332134213521362137213821392140214121422143214421452146214721482149215021512152215321542155215621572158215921602161216221632164216521662167216821692170
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,
mvjsadcost, 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, mvjcost, mvcost, x->errorperbit,
xd->allow_high_precision_mv);
else
return INT_MAX;
}
int vp9_refining_search_sad_c(MACROBLOCK *x,
int_mv *ref_mv, int error_per_bit,
int search_range, vp9_variance_fn_ptr_t *fn_ptr,
int *mvjcost, int *mvcost[2], int_mv *center_mv) {
const MACROBLOCKD* const xd = &x->e_mbd;
MV neighbors[4] = {{ -1, 0}, {0, -1}, {0, 1}, {1, 0}};
int i, j;
int this_row_offset, this_col_offset;
int what_stride = x->plane[0].src.stride;
int in_what_stride = xd->plane[0].pre[0].stride;
uint8_t *what = x->plane[0].src.buf;
uint8_t *best_address = xd->plane[0].pre[0].buf +
(ref_mv->as_mv.row * xd->plane[0].pre[0].stride) +
ref_mv->as_mv.col;
uint8_t *check_here;
unsigned int thissad;
int_mv this_mv;
unsigned int bestsad = INT_MAX;
int_mv fcenter_mv;
int *mvjsadcost = x->nmvjointsadcost;
int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]};
fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
2171217221732174217521762177217821792180218121822183218421852186218721882189219021912192219321942195219621972198219922002201220222032204220522062207220822092210221122122213221422152216221722182219222022212222222322242225222622272228222922302231223222332234223522362237223822392240
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, mvjsadcost, 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, mvjsadcost,
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, mvjcost, mvcost, x->errorperbit,
xd->allow_high_precision_mv);
else
return INT_MAX;
}
int vp9_refining_search_sadx4(MACROBLOCK *x,
int_mv *ref_mv, int error_per_bit,
int search_range, vp9_variance_fn_ptr_t *fn_ptr,
int *mvjcost, int *mvcost[2], int_mv *center_mv) {
const MACROBLOCKD* const xd = &x->e_mbd;
MV neighbors[4] = {{ -1, 0}, {0, -1}, {0, 1}, {1, 0}};
int i, j;
int this_row_offset, this_col_offset;
int what_stride = x->plane[0].src.stride;
int in_what_stride = xd->plane[0].pre[0].stride;
uint8_t *what = x->plane[0].src.buf;
uint8_t *best_address = xd->plane[0].pre[0].buf +
(ref_mv->as_mv.row * xd->plane[0].pre[0].stride) +
ref_mv->as_mv.col;
uint8_t *check_here;
unsigned int thissad;
2241224222432244224522462247224822492250225122522253225422552256225722582259226022612262226322642265226622672268226922702271227222732274227522762277227822792280228122822283228422852286228722882289229022912292229322942295229622972298229923002301230223032304230523062307230823092310
int_mv this_mv;
unsigned int bestsad = INT_MAX;
int_mv fcenter_mv;
int *mvjsadcost = x->nmvjointsadcost;
int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]};
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, mvjsadcost, mvsadcost, error_per_bit);
for (i = 0; i < search_range; i++) {
int best_site = -1;
int all_in = ((ref_mv->as_mv.row - 1) > x->mv_row_min) &
((ref_mv->as_mv.row + 1) < x->mv_row_max) &
((ref_mv->as_mv.col - 1) > x->mv_col_min) &
((ref_mv->as_mv.col + 1) < x->mv_col_max);
if (all_in) {
unsigned int sad_array[4];
unsigned char const *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, mvjsadcost,
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) {
this_mv.as_mv.row = this_row_offset;
this_mv.as_mv.col = this_col_offset;
thissad += mvsad_err_cost(&this_mv, &fcenter_mv, mvjsadcost,
mvsadcost, error_per_bit);
if (thissad < bestsad) {
bestsad = thissad;
best_site = j;
}
}
}
}
}
if (best_site == -1)
break;
2311231223132314231523162317231823192320232123222323232423252326232723282329233023312332233323342335233623372338233923402341234223432344234523462347234823492350235123522353235423552356235723582359236023612362236323642365236623672368236923702371237223732374237523762377237823792380
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, mvjcost, mvcost, x->errorperbit,
xd->allow_high_precision_mv);
else
return INT_MAX;
}
/* This function is called when we do joint motion search in comp_inter_inter
* mode.
*/
int vp9_refining_search_8p_c(MACROBLOCK *x,
int_mv *ref_mv, int error_per_bit,
int search_range, vp9_variance_fn_ptr_t *fn_ptr,
int *mvjcost, int *mvcost[2], int_mv *center_mv,
const uint8_t *second_pred, int w, int h) {
const MACROBLOCKD* const xd = &x->e_mbd;
MV neighbors[8] = {{-1, 0}, {0, -1}, {0, 1}, {1, 0},
{-1, -1}, {1, -1}, {-1, 1}, {1, 1}};
int i, j;
int this_row_offset, this_col_offset;
int what_stride = x->plane[0].src.stride;
int in_what_stride = xd->plane[0].pre[0].stride;
uint8_t *what = x->plane[0].src.buf;
uint8_t *best_address = xd->plane[0].pre[0].buf +
(ref_mv->as_mv.row * xd->plane[0].pre[0].stride) +
ref_mv->as_mv.col;
uint8_t *check_here;
unsigned int thissad;
int_mv this_mv;
unsigned int bestsad = INT_MAX;
int_mv fcenter_mv;
int *mvjsadcost = x->nmvjointsadcost;
int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]};
fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
/* Get compound pred by averaging two pred blocks. */
bestsad = fn_ptr->sdaf(what, what_stride, best_address, in_what_stride,
second_pred, 0x7fffffff) +
mvsad_err_cost(ref_mv, &fcenter_mv, mvjsadcost, mvsadcost, error_per_bit);
for (i = 0; i < search_range; i++) {
int best_site = -1;
for (j = 0; j < 8; 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;
2381238223832384238523862387238823892390239123922393239423952396239723982399240024012402240324042405240624072408240924102411241224132414241524162417241824192420242124222423242424252426
/* Get compound block and use it to calculate SAD. */
thissad = fn_ptr->sdaf(what, what_stride, check_here, in_what_stride,
second_pred, 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, mvjsadcost,
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) {
// FIXME(rbultje, yunqing): add full-pixel averaging variance functions
// so we don't have to use the subpixel with xoff=0,yoff=0 here.
int besterr = fn_ptr->svaf(best_address, in_what_stride, 0, 0,
what, what_stride, (unsigned int *)(&thissad),
second_pred) +
mv_err_cost(&this_mv, center_mv, mvjcost, mvcost, x->errorperbit,
xd->allow_high_precision_mv);
return besterr;
} else {
return INT_MAX;
}
}