vpx/vp8/decoder/decodemv.c
Scott LaVarnway ff1d170e69 Improved read_mb_modes_mv()
Interleaved vp8_find_near_mvs and vp8_mv_ref_probs.
2.5% to 4% performance improvement for the HD clips used.

Change-Id: Id888b667cf5ae2f0e19da18743140f055ff7de8d
2011-10-26 10:46:36 -04:00

621 lines
19 KiB
C

/*
* 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 "treereader.h"
#include "vp8/common/entropymv.h"
#include "vp8/common/entropymode.h"
#include "onyxd_int.h"
#include "vp8/common/findnearmv.h"
#if CONFIG_DEBUG
#include <assert.h>
#endif
static B_PREDICTION_MODE read_bmode(vp8_reader *bc, const vp8_prob *p)
{
const int i = vp8_treed_read(bc, vp8_bmode_tree, p);
return (B_PREDICTION_MODE)i;
}
static MB_PREDICTION_MODE read_ymode(vp8_reader *bc, const vp8_prob *p)
{
const int i = vp8_treed_read(bc, vp8_ymode_tree, p);
return (MB_PREDICTION_MODE)i;
}
static MB_PREDICTION_MODE read_kf_ymode(vp8_reader *bc, const vp8_prob *p)
{
const int i = vp8_treed_read(bc, vp8_kf_ymode_tree, p);
return (MB_PREDICTION_MODE)i;
}
static MB_PREDICTION_MODE read_uv_mode(vp8_reader *bc, const vp8_prob *p)
{
const int i = vp8_treed_read(bc, vp8_uv_mode_tree, p);
return (MB_PREDICTION_MODE)i;
}
static void read_kf_modes(VP8D_COMP *pbi, MODE_INFO *mi)
{
vp8_reader *const bc = & pbi->bc;
const int mis = pbi->common.mode_info_stride;
mi->mbmi.ref_frame = INTRA_FRAME;
mi->mbmi.mode = read_kf_ymode(bc, pbi->common.kf_ymode_prob);
if (mi->mbmi.mode == B_PRED)
{
int i = 0;
do
{
const B_PREDICTION_MODE A = above_block_mode(mi, i, mis);
const B_PREDICTION_MODE L = left_block_mode(mi, i);
mi->bmi[i].as_mode =
read_bmode(bc, pbi->common.kf_bmode_prob [A] [L]);
}
while (++i < 16);
}
mi->mbmi.uv_mode = read_uv_mode(bc, pbi->common.kf_uv_mode_prob);
}
static int read_mvcomponent(vp8_reader *r, const MV_CONTEXT *mvc)
{
const vp8_prob *const p = (const vp8_prob *) mvc;
int x = 0;
if (vp8_read(r, p [mvpis_short])) /* Large */
{
int i = 0;
do
{
x += vp8_read(r, p [MVPbits + i]) << i;
}
while (++i < 3);
i = mvlong_width - 1; /* Skip bit 3, which is sometimes implicit */
do
{
x += vp8_read(r, p [MVPbits + i]) << i;
}
while (--i > 3);
if (!(x & 0xFFF0) || vp8_read(r, p [MVPbits + 3]))
x += 8;
}
else /* small */
x = vp8_treed_read(r, vp8_small_mvtree, p + MVPshort);
if (x && vp8_read(r, p [MVPsign]))
x = -x;
return x;
}
static void read_mv(vp8_reader *r, MV *mv, const MV_CONTEXT *mvc)
{
mv->row = (short)(read_mvcomponent(r, mvc) << 1);
mv->col = (short)(read_mvcomponent(r, ++mvc) << 1);
}
static void read_mvcontexts(vp8_reader *bc, MV_CONTEXT *mvc)
{
int i = 0;
do
{
const vp8_prob *up = vp8_mv_update_probs[i].prob;
vp8_prob *p = (vp8_prob *)(mvc + i);
vp8_prob *const pstop = p + MVPcount;
do
{
if (vp8_read(bc, *up++))
{
const vp8_prob x = (vp8_prob)vp8_read_literal(bc, 7);
*p = x ? x << 1 : 1;
}
}
while (++p < pstop);
}
while (++i < 2);
}
static B_PREDICTION_MODE sub_mv_ref(vp8_reader *bc, const vp8_prob *p)
{
const int i = vp8_treed_read(bc, vp8_sub_mv_ref_tree, p);
return (B_PREDICTION_MODE)i;
}
static const unsigned char mbsplit_fill_count[4] = {8, 8, 4, 1};
static const unsigned char mbsplit_fill_offset[4][16] = {
{ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15},
{ 0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15},
{ 0, 1, 4, 5, 2, 3, 6, 7, 8, 9, 12, 13, 10, 11, 14, 15},
{ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15}
};
static void mb_mode_mv_init(VP8D_COMP *pbi)
{
vp8_reader *const bc = & pbi->bc;
MV_CONTEXT *const mvc = pbi->common.fc.mvc;
#if CONFIG_ERROR_CONCEALMENT
/* Default is that no macroblock is corrupt, therefore we initialize
* mvs_corrupt_from_mb to something very big, which we can be sure is
* outside the frame. */
pbi->mvs_corrupt_from_mb = UINT_MAX;
#endif
pbi->prob_skip_false = 0;
if (pbi->common.mb_no_coeff_skip)
pbi->prob_skip_false = (vp8_prob)vp8_read_literal(bc, 8);
if(pbi->common.frame_type != KEY_FRAME)
{
pbi->prob_intra = (vp8_prob)vp8_read_literal(bc, 8);
pbi->prob_last = (vp8_prob)vp8_read_literal(bc, 8);
pbi->prob_gf = (vp8_prob)vp8_read_literal(bc, 8);
if (vp8_read_bit(bc))
{
int i = 0;
do
{
pbi->common.fc.ymode_prob[i] =
(vp8_prob) vp8_read_literal(bc, 8);
}
while (++i < 4);
}
if (vp8_read_bit(bc))
{
int i = 0;
do
{
pbi->common.fc.uv_mode_prob[i] =
(vp8_prob) vp8_read_literal(bc, 8);
}
while (++i < 3);
}
read_mvcontexts(bc, mvc);
}
}
static void decode_split_mv(vp8_reader *const bc, MODE_INFO *mi,
MB_MODE_INFO *mbmi, int mis, int_mv best_mv,
MV_CONTEXT *const mvc, int mb_to_left_edge,
int mb_to_right_edge, int mb_to_top_edge,
int mb_to_bottom_edge)
{
const int s = mbmi->partitioning =
vp8_treed_read(bc, vp8_mbsplit_tree, vp8_mbsplit_probs);
const int num_p = vp8_mbsplit_count [s];
int j = 0;
do /* for each subset j */
{
int_mv leftmv, abovemv;
int_mv blockmv;
int k; /* first block in subset j */
int mv_contz;
k = vp8_mbsplit_offset[s][j];
leftmv.as_int = left_block_mv(mi, k);
abovemv.as_int = above_block_mv(mi, k, mis);
mv_contz = vp8_mv_cont(&leftmv, &abovemv);
switch (sub_mv_ref(bc, vp8_sub_mv_ref_prob2 [mv_contz]))
{
case NEW4X4:
read_mv(bc, &blockmv.as_mv, (const MV_CONTEXT *) mvc);
blockmv.as_mv.row += best_mv.as_mv.row;
blockmv.as_mv.col += best_mv.as_mv.col;
break;
case LEFT4X4:
blockmv.as_int = leftmv.as_int;
break;
case ABOVE4X4:
blockmv.as_int = abovemv.as_int;
break;
case ZERO4X4:
blockmv.as_int = 0;
break;
default:
break;
}
mbmi->need_to_clamp_mvs = vp8_check_mv_bounds(&blockmv,
mb_to_left_edge,
mb_to_right_edge,
mb_to_top_edge,
mb_to_bottom_edge);
{
/* Fill (uniform) modes, mvs of jth subset.
Must do it here because ensuing subsets can
refer back to us via "left" or "above". */
const unsigned char *fill_offset;
unsigned int fill_count = mbsplit_fill_count[s];
fill_offset = &mbsplit_fill_offset[s]
[(unsigned char)j * mbsplit_fill_count[s]];
do {
mi->bmi[ *fill_offset].mv.as_int = blockmv.as_int;
fill_offset++;
}while (--fill_count);
}
}
while (++j < num_p);
}
static void read_mb_modes_mv(VP8D_COMP *pbi, MODE_INFO *mi, MB_MODE_INFO *mbmi,
int mb_row, int mb_col)
{
vp8_reader *const bc = & pbi->bc;
mbmi->ref_frame = (MV_REFERENCE_FRAME) vp8_read(bc, pbi->prob_intra);
if (mbmi->ref_frame) /* inter MB */
{
enum {CNT_INTRA, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV};
vp8_prob mv_ref_p [VP8_MVREFS-1];
int cnt[4];
int *cntx = cnt;
int_mv near_mvs[4];
int_mv *nmv = near_mvs;
const int mis = pbi->mb.mode_info_stride;
const MODE_INFO *above = mi - mis;
const MODE_INFO *left = mi - 1;
const MODE_INFO *aboveleft = above - 1;
MV_CONTEXT *const mvc = pbi->common.fc.mvc;
int *ref_frame_sign_bias = pbi->common.ref_frame_sign_bias;
int propogate_mv_for_ec = 0;
mbmi->need_to_clamp_mvs = 0;
if (vp8_read(bc, pbi->prob_last))
{
mbmi->ref_frame =
(MV_REFERENCE_FRAME)((int)(2 + vp8_read(bc, pbi->prob_gf)));
}
/* Zero accumulators */
nmv[0].as_int = nmv[1].as_int = nmv[2].as_int = 0;
cnt[0] = cnt[1] = cnt[2] = cnt[3] = 0;
/* Process above */
if (above->mbmi.ref_frame != INTRA_FRAME)
{
if (above->mbmi.mv.as_int)
{
(++nmv)->as_int = above->mbmi.mv.as_int;
mv_bias(ref_frame_sign_bias[above->mbmi.ref_frame],
mbmi->ref_frame, nmv, ref_frame_sign_bias);
++cntx;
}
*cntx += 2;
}
/* Process left */
if (left->mbmi.ref_frame != INTRA_FRAME)
{
if (left->mbmi.mv.as_int)
{
int_mv this_mv;
this_mv.as_int = left->mbmi.mv.as_int;
mv_bias(ref_frame_sign_bias[left->mbmi.ref_frame],
mbmi->ref_frame, &this_mv, ref_frame_sign_bias);
if (this_mv.as_int != nmv->as_int)
{
(++nmv)->as_int = this_mv.as_int;
++cntx;
}
*cntx += 2;
}
else
cnt[CNT_INTRA] += 2;
}
/* Process above left */
if (aboveleft->mbmi.ref_frame != INTRA_FRAME)
{
if (aboveleft->mbmi.mv.as_int)
{
int_mv this_mv;
this_mv.as_int = aboveleft->mbmi.mv.as_int;
mv_bias(ref_frame_sign_bias[aboveleft->mbmi.ref_frame],
mbmi->ref_frame, &this_mv, ref_frame_sign_bias);
if (this_mv.as_int != nmv->as_int)
{
(++nmv)->as_int = this_mv.as_int;
++cntx;
}
*cntx += 1;
}
else
cnt[CNT_INTRA] += 1;
}
mv_ref_p[0] = vp8_mode_contexts [cnt[CNT_INTRA]] [0];
if( vp8_read(bc, mv_ref_p[0]) )
{
int mb_to_left_edge;
int mb_to_right_edge;
/* Distance of Mb to the various image edges.
* These specified to 8th pel as they are always compared to MV
* values that are in 1/8th pel units
*/
pbi->mb.mb_to_left_edge =
mb_to_left_edge = -((mb_col * 16) << 3);
mb_to_left_edge -= LEFT_TOP_MARGIN;
pbi->mb.mb_to_right_edge =
mb_to_right_edge = ((pbi->common.mb_cols - 1 - mb_col) * 16) << 3;
mb_to_right_edge += RIGHT_BOTTOM_MARGIN;
/* If we have three distinct MV's ... */
if (cnt[CNT_SPLITMV])
{
/* See if above-left MV can be merged with NEAREST */
if (nmv->as_int == near_mvs[CNT_NEAREST].as_int)
cnt[CNT_NEAREST] += 1;
}
cnt[CNT_SPLITMV] = ((above->mbmi.mode == SPLITMV)
+ (left->mbmi.mode == SPLITMV)) * 2
+ (aboveleft->mbmi.mode == SPLITMV);
/* Swap near and nearest if necessary */
if (cnt[CNT_NEAR] > cnt[CNT_NEAREST])
{
int tmp;
tmp = cnt[CNT_NEAREST];
cnt[CNT_NEAREST] = cnt[CNT_NEAR];
cnt[CNT_NEAR] = tmp;
tmp = near_mvs[CNT_NEAREST].as_int;
near_mvs[CNT_NEAREST].as_int = near_mvs[CNT_NEAR].as_int;
near_mvs[CNT_NEAR].as_int = tmp;
}
mv_ref_p[1] = vp8_mode_contexts [cnt[CNT_NEAREST]] [1];
if( vp8_read(bc, mv_ref_p[1]) )
{
mv_ref_p[2] = vp8_mode_contexts [cnt[CNT_NEAR]] [2];
if( vp8_read(bc, mv_ref_p[2]) )
{
int mb_to_top_edge;
int mb_to_bottom_edge;
mb_to_top_edge = pbi->mb.mb_to_top_edge;
mb_to_bottom_edge = pbi->mb.mb_to_bottom_edge;
mb_to_top_edge -= LEFT_TOP_MARGIN;
mb_to_bottom_edge += RIGHT_BOTTOM_MARGIN;
/* Use near_mvs[0] to store the "best" MV */
if (cnt[CNT_NEAREST] >= cnt[CNT_INTRA])
near_mvs[CNT_INTRA] = near_mvs[CNT_NEAREST];
mv_ref_p[3] = vp8_mode_contexts [cnt[CNT_SPLITMV]] [3];
vp8_clamp_mv2(&near_mvs[CNT_INTRA], &pbi->mb);
if( vp8_read(bc, mv_ref_p[3]) )
{
decode_split_mv(bc, mi,
mbmi, mis,
near_mvs[CNT_INTRA],
mvc, mb_to_left_edge,
mb_to_right_edge,
mb_to_top_edge,
mb_to_bottom_edge);
mbmi->mv.as_int = mi->bmi[15].mv.as_int;
mbmi->mode = SPLITMV;
}
else
{
int_mv *const mbmi_mv = & mbmi->mv;
read_mv(bc, &mbmi_mv->as_mv, (const MV_CONTEXT *) mvc);
mbmi_mv->as_mv.row += near_mvs[CNT_INTRA].as_mv.row;
mbmi_mv->as_mv.col += near_mvs[CNT_INTRA].as_mv.col;
/* Don't need to check this on NEARMV and NEARESTMV
* modes since those modes clamp the MV. The NEWMV mode
* does not, so signal to the prediction stage whether
* special handling may be required.
*/
mbmi->need_to_clamp_mvs =
vp8_check_mv_bounds(mbmi_mv, mb_to_left_edge,
mb_to_right_edge,
mb_to_top_edge,
mb_to_bottom_edge);
mbmi->mode = NEWMV;
propogate_mv_for_ec = 1;
}
}
else
{
mbmi->mode = NEARMV;
vp8_clamp_mv2(&near_mvs[CNT_NEAR], &pbi->mb);
mbmi->mv.as_int = near_mvs[CNT_NEAR].as_int;
propogate_mv_for_ec = 1;
}
}
else
{
mbmi->mode = NEARESTMV;
vp8_clamp_mv2(&near_mvs[CNT_NEAREST], &pbi->mb);
mbmi->mv.as_int = near_mvs[CNT_NEAREST].as_int;
propogate_mv_for_ec = 1;
}
}
else {
mbmi->mode = ZEROMV;
mbmi->mv.as_int = 0;
propogate_mv_for_ec = 1;
}
mbmi->uv_mode = DC_PRED;
#if CONFIG_ERROR_CONCEALMENT
if(pbi->ec_enabled && propogate_mv_for_ec)
{
mi->bmi[ 0].mv.as_int =
mi->bmi[ 1].mv.as_int =
mi->bmi[ 2].mv.as_int =
mi->bmi[ 3].mv.as_int =
mi->bmi[ 4].mv.as_int =
mi->bmi[ 5].mv.as_int =
mi->bmi[ 6].mv.as_int =
mi->bmi[ 7].mv.as_int =
mi->bmi[ 8].mv.as_int =
mi->bmi[ 9].mv.as_int =
mi->bmi[10].mv.as_int =
mi->bmi[11].mv.as_int =
mi->bmi[12].mv.as_int =
mi->bmi[13].mv.as_int =
mi->bmi[14].mv.as_int =
mi->bmi[15].mv.as_int = mbmi->mv.as_int;
}
#endif
}
else
{
/* required for left and above block mv */
mbmi->mv.as_int = 0;
/* MB is intra coded */
if ((mbmi->mode = read_ymode(bc, pbi->common.fc.ymode_prob)) == B_PRED)
{
int j = 0;
do
{
mi->bmi[j].as_mode = read_bmode(bc, pbi->common.fc.bmode_prob);
}
while (++j < 16);
}
mbmi->uv_mode = read_uv_mode(bc, pbi->common.fc.uv_mode_prob);
}
}
static void read_mb_features(vp8_reader *r, MB_MODE_INFO *mi, MACROBLOCKD *x)
{
/* Is segmentation enabled */
if (x->segmentation_enabled && x->update_mb_segmentation_map)
{
/* If so then read the segment id. */
if (vp8_read(r, x->mb_segment_tree_probs[0]))
mi->segment_id =
(unsigned char)(2 + vp8_read(r, x->mb_segment_tree_probs[2]));
else
mi->segment_id =
(unsigned char)(vp8_read(r, x->mb_segment_tree_probs[1]));
}
}
static void decode_mb_mode_mvs(VP8D_COMP *pbi, MODE_INFO *mi,
MB_MODE_INFO *mbmi, int mb_row, int mb_col)
{
/* Read the Macroblock segmentation map if it is being updated explicitly
* this frame (reset to 0 above by default)
* By default on a key frame reset all MBs to segment 0
*/
if (pbi->mb.update_mb_segmentation_map)
read_mb_features(&pbi->bc, &mi->mbmi, &pbi->mb);
else if(pbi->common.frame_type == KEY_FRAME)
mi->mbmi.segment_id = 0;
/* Read the macroblock coeff skip flag if this feature is in use,
* else default to 0 */
if (pbi->common.mb_no_coeff_skip)
mi->mbmi.mb_skip_coeff = vp8_read(&pbi->bc, pbi->prob_skip_false);
else
mi->mbmi.mb_skip_coeff = 0;
if(pbi->common.frame_type == KEY_FRAME)
read_kf_modes(pbi, mi);
else
read_mb_modes_mv(pbi, mi, &mi->mbmi, mb_row, mb_col);
}
void vp8_decode_mode_mvs(VP8D_COMP *pbi)
{
MODE_INFO *mi = pbi->common.mi;
int mb_row = -1;
mb_mode_mv_init(pbi);
while (++mb_row < pbi->common.mb_rows)
{
int mb_col = -1;
pbi->mb.mb_to_top_edge = -((mb_row * 16)) << 3;
pbi->mb.mb_to_bottom_edge =
((pbi->common.mb_rows - 1 - mb_row) * 16) << 3;
while (++mb_col < pbi->common.mb_cols)
{
#if CONFIG_ERROR_CONCEALMENT
int mb_num = mb_row * pbi->common.mb_cols + mb_col;
#endif
decode_mb_mode_mvs(pbi, mi, &mi->mbmi, mb_row, mb_col);
#if CONFIG_ERROR_CONCEALMENT
/* look for corruption. set mvs_corrupt_from_mb to the current
* mb_num if the frame is corrupt from this macroblock. */
if (vp8dx_bool_error(&pbi->bc) && mb_num < pbi->mvs_corrupt_from_mb)
{
pbi->mvs_corrupt_from_mb = mb_num;
/* no need to continue since the partition is corrupt from
* here on.
*/
return;
}
#endif
mi++; /* next macroblock */
}
mi++; /* skip left predictor each row */
}
}