1
0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2024-12-12 19:18:44 +02:00
FFmpeg/libavcodec/vp9block.c

1750 lines
74 KiB
C
Raw Normal View History

/*
* VP9 compatible video decoder
*
* Copyright (C) 2013 Ronald S. Bultje <rsbultje gmail com>
* Copyright (C) 2013 Clément Bœsch <u pkh me>
*
* This file is part of Libav.
*
* Libav is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* Libav is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with Libav; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "libavutil/avassert.h"
#include "avcodec.h"
#include "internal.h"
#include "videodsp.h"
#include "vp56.h"
#include "vp9.h"
#include "vp9data.h"
static const uint8_t bwh_tab[2][N_BS_SIZES][2] = {
{
{ 16, 16 }, { 16, 8 }, { 8, 16 }, { 8, 8 }, { 8, 4 }, { 4, 8 },
{ 4, 4 }, { 4, 2 }, { 2, 4 }, { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 },
}, {
{ 8, 8 }, { 8, 4 }, { 4, 8 }, { 4, 4 }, { 4, 2 }, { 2, 4 },
{ 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 }, { 1, 1 }, { 1, 1 }, { 1, 1 },
}
};
// differential forward probability updates
static void decode_mode(VP9Context *s, VP9Block *const b)
{
static const uint8_t left_ctx[N_BS_SIZES] = {
0x0, 0x8, 0x0, 0x8, 0xc, 0x8, 0xc, 0xe, 0xc, 0xe, 0xf, 0xe, 0xf
};
static const uint8_t above_ctx[N_BS_SIZES] = {
0x0, 0x0, 0x8, 0x8, 0x8, 0xc, 0xc, 0xc, 0xe, 0xe, 0xe, 0xf, 0xf
};
static const uint8_t max_tx_for_bl_bp[N_BS_SIZES] = {
TX_32X32, TX_32X32, TX_32X32, TX_32X32, TX_16X16, TX_16X16,
TX_16X16, TX_8X8, TX_8X8, TX_8X8, TX_4X4, TX_4X4, TX_4X4
};
int row = b->row, col = b->col, row7 = b->row7;
enum TxfmMode max_tx = max_tx_for_bl_bp[b->bs];
int w4 = FFMIN(s->cols - col, bwh_tab[1][b->bs][0]);
int h4 = FFMIN(s->rows - row, bwh_tab[1][b->bs][1]);
int have_a = row > 0, have_l = col > s->tiling.tile_col_start;
int y;
if (!s->segmentation.enabled) {
b->seg_id = 0;
} else if (s->keyframe || s->intraonly) {
b->seg_id = s->segmentation.update_map ?
vp8_rac_get_tree(&s->c, ff_vp9_segmentation_tree, s->prob.seg) : 0;
} else if (!s->segmentation.update_map ||
(s->segmentation.temporal &&
vp56_rac_get_prob_branchy(&s->c,
s->prob.segpred[s->above_segpred_ctx[col] +
s->left_segpred_ctx[row7]]))) {
if (!s->errorres) {
uint8_t *refsegmap = s->frames[LAST_FRAME].segmentation_map;
int pred = MAX_SEGMENT - 1;
int x;
if (!s->last_uses_2pass)
ff_thread_await_progress(&s->frames[LAST_FRAME].tf, row >> 3, 0);
for (y = 0; y < h4; y++)
for (x = 0; x < w4; x++)
pred = FFMIN(pred,
refsegmap[(y + row) * 8 * s->sb_cols + x + col]);
b->seg_id = pred;
} else {
b->seg_id = 0;
}
memset(&s->above_segpred_ctx[col], 1, w4);
memset(&s->left_segpred_ctx[row7], 1, h4);
} else {
b->seg_id = vp8_rac_get_tree(&s->c, ff_vp9_segmentation_tree,
s->prob.seg);
memset(&s->above_segpred_ctx[col], 0, w4);
memset(&s->left_segpred_ctx[row7], 0, h4);
}
if ((s->segmentation.enabled && s->segmentation.update_map) || s->keyframe) {
uint8_t *segmap = s->frames[CUR_FRAME].segmentation_map;
for (y = 0; y < h4; y++)
memset(&segmap[(y + row) * 8 * s->sb_cols + col],
b->seg_id, w4);
}
b->skip = s->segmentation.enabled &&
s->segmentation.feat[b->seg_id].skip_enabled;
if (!b->skip) {
int c = s->left_skip_ctx[row7] + s->above_skip_ctx[col];
b->skip = vp56_rac_get_prob(&s->c, s->prob.p.skip[c]);
s->counts.skip[c][b->skip]++;
}
if (s->keyframe || s->intraonly) {
b->intra = 1;
} else if (s->segmentation.feat[b->seg_id].ref_enabled) {
b->intra = !s->segmentation.feat[b->seg_id].ref_val;
} else {
int c, bit;
if (have_a && have_l) {
c = s->above_intra_ctx[col] + s->left_intra_ctx[row7];
c += (c == 2);
} else {
c = have_a ? 2 * s->above_intra_ctx[col] :
have_l ? 2 * s->left_intra_ctx[row7] : 0;
}
bit = vp56_rac_get_prob(&s->c, s->prob.p.intra[c]);
s->counts.intra[c][bit]++;
b->intra = !bit;
}
if ((b->intra || !b->skip) && s->txfmmode == TX_SWITCHABLE) {
int c;
if (have_a) {
if (have_l) {
c = (s->above_skip_ctx[col] ? max_tx :
s->above_txfm_ctx[col]) +
(s->left_skip_ctx[row7] ? max_tx :
s->left_txfm_ctx[row7]) > max_tx;
} else {
c = s->above_skip_ctx[col] ? 1 :
(s->above_txfm_ctx[col] * 2 > max_tx);
}
} else if (have_l) {
c = s->left_skip_ctx[row7] ? 1 :
(s->left_txfm_ctx[row7] * 2 > max_tx);
} else {
c = 1;
}
switch (max_tx) {
case TX_32X32:
b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][0]);
if (b->tx) {
b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][1]);
if (b->tx == 2)
b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][2]);
}
s->counts.tx32p[c][b->tx]++;
break;
case TX_16X16:
b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][0]);
if (b->tx)
b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][1]);
s->counts.tx16p[c][b->tx]++;
break;
case TX_8X8:
b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx8p[c]);
s->counts.tx8p[c][b->tx]++;
break;
case TX_4X4:
b->tx = TX_4X4;
break;
}
} else {
b->tx = FFMIN(max_tx, s->txfmmode);
}
if (s->keyframe || s->intraonly) {
uint8_t *a = &s->above_mode_ctx[col * 2];
uint8_t *l = &s->left_mode_ctx[(row7) << 1];
b->comp = 0;
if (b->bs > BS_8x8) {
// FIXME the memory storage intermediates here aren't really
// necessary, they're just there to make the code slightly
// simpler for now
b->mode[0] =
a[0] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
ff_vp9_default_kf_ymode_probs[a[0]][l[0]]);
if (b->bs != BS_8x4) {
b->mode[1] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
ff_vp9_default_kf_ymode_probs[a[1]][b->mode[0]]);
l[0] =
a[1] = b->mode[1];
} else {
l[0] =
a[1] =
b->mode[1] = b->mode[0];
}
if (b->bs != BS_4x8) {
b->mode[2] =
a[0] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
ff_vp9_default_kf_ymode_probs[a[0]][l[1]]);
if (b->bs != BS_8x4) {
b->mode[3] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
ff_vp9_default_kf_ymode_probs[a[1]][b->mode[2]]);
l[1] =
a[1] = b->mode[3];
} else {
l[1] =
a[1] =
b->mode[3] = b->mode[2];
}
} else {
b->mode[2] = b->mode[0];
l[1] =
a[1] =
b->mode[3] = b->mode[1];
}
} else {
b->mode[0] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
ff_vp9_default_kf_ymode_probs[*a][*l]);
b->mode[3] =
b->mode[2] =
b->mode[1] = b->mode[0];
// FIXME this can probably be optimized
memset(a, b->mode[0], bwh_tab[0][b->bs][0]);
memset(l, b->mode[0], bwh_tab[0][b->bs][1]);
}
b->uvmode = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
ff_vp9_default_kf_uvmode_probs[b->mode[3]]);
} else if (b->intra) {
b->comp = 0;
if (b->bs > BS_8x8) {
b->mode[0] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
s->prob.p.y_mode[0]);
s->counts.y_mode[0][b->mode[0]]++;
if (b->bs != BS_8x4) {
b->mode[1] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
s->prob.p.y_mode[0]);
s->counts.y_mode[0][b->mode[1]]++;
} else {
b->mode[1] = b->mode[0];
}
if (b->bs != BS_4x8) {
b->mode[2] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
s->prob.p.y_mode[0]);
s->counts.y_mode[0][b->mode[2]]++;
if (b->bs != BS_8x4) {
b->mode[3] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
s->prob.p.y_mode[0]);
s->counts.y_mode[0][b->mode[3]]++;
} else {
b->mode[3] = b->mode[2];
}
} else {
b->mode[2] = b->mode[0];
b->mode[3] = b->mode[1];
}
} else {
static const uint8_t size_group[10] = {
3, 3, 3, 3, 2, 2, 2, 1, 1, 1
};
int sz = size_group[b->bs];
b->mode[0] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
s->prob.p.y_mode[sz]);
b->mode[1] =
b->mode[2] =
b->mode[3] = b->mode[0];
s->counts.y_mode[sz][b->mode[3]]++;
}
b->uvmode = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
s->prob.p.uv_mode[b->mode[3]]);
s->counts.uv_mode[b->mode[3]][b->uvmode]++;
} else {
static const uint8_t inter_mode_ctx_lut[14][14] = {
{ 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
{ 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
{ 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
{ 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
{ 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
{ 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
{ 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
{ 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
{ 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
{ 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
{ 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },
{ 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },
{ 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 1, 1, 0, 3 },
{ 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 3, 3, 3, 4 },
};
if (s->segmentation.feat[b->seg_id].ref_enabled) {
av_assert2(s->segmentation.feat[b->seg_id].ref_val != 0);
b->comp = 0;
b->ref[0] = s->segmentation.feat[b->seg_id].ref_val - 1;
} else {
// read comp_pred flag
if (s->comppredmode != PRED_SWITCHABLE) {
b->comp = s->comppredmode == PRED_COMPREF;
} else {
int c;
// FIXME add intra as ref=0xff (or -1) to make these easier?
if (have_a) {
if (have_l) {
if (s->above_comp_ctx[col] && s->left_comp_ctx[row7]) {
c = 4;
} else if (s->above_comp_ctx[col]) {
c = 2 + (s->left_intra_ctx[row7] ||
s->left_ref_ctx[row7] == s->fixcompref);
} else if (s->left_comp_ctx[row7]) {
c = 2 + (s->above_intra_ctx[col] ||
s->above_ref_ctx[col] == s->fixcompref);
} else {
c = (!s->above_intra_ctx[col] &&
s->above_ref_ctx[col] == s->fixcompref) ^
(!s->left_intra_ctx[row7] &&
s->left_ref_ctx[row & 7] == s->fixcompref);
}
} else {
c = s->above_comp_ctx[col] ? 3 :
(!s->above_intra_ctx[col] && s->above_ref_ctx[col] == s->fixcompref);
}
} else if (have_l) {
c = s->left_comp_ctx[row7] ? 3 :
(!s->left_intra_ctx[row7] && s->left_ref_ctx[row7] == s->fixcompref);
} else {
c = 1;
}
b->comp = vp56_rac_get_prob(&s->c, s->prob.p.comp[c]);
s->counts.comp[c][b->comp]++;
}
// read actual references
// FIXME probably cache a few variables here to prevent repetitive
// memory accesses below
if (b->comp) { /* two references */
int fix_idx = s->signbias[s->fixcompref], var_idx = !fix_idx, c, bit;
b->ref[fix_idx] = s->fixcompref;
// FIXME can this codeblob be replaced by some sort of LUT?
if (have_a) {
if (have_l) {
if (s->above_intra_ctx[col]) {
if (s->left_intra_ctx[row7]) {
c = 2;
} else {
c = 1 + 2 * (s->left_ref_ctx[row7] != s->varcompref[1]);
}
} else if (s->left_intra_ctx[row7]) {
c = 1 + 2 * (s->above_ref_ctx[col] != s->varcompref[1]);
} else {
int refl = s->left_ref_ctx[row7], refa = s->above_ref_ctx[col];
if (refl == refa && refa == s->varcompref[1]) {
c = 0;
} else if (!s->left_comp_ctx[row7] && !s->above_comp_ctx[col]) {
if ((refa == s->fixcompref && refl == s->varcompref[0]) ||
(refl == s->fixcompref && refa == s->varcompref[0])) {
c = 4;
} else {
c = (refa == refl) ? 3 : 1;
}
} else if (!s->left_comp_ctx[row7]) {
if (refa == s->varcompref[1] && refl != s->varcompref[1]) {
c = 1;
} else {
c = (refl == s->varcompref[1] &&
refa != s->varcompref[1]) ? 2 : 4;
}
} else if (!s->above_comp_ctx[col]) {
if (refl == s->varcompref[1] && refa != s->varcompref[1]) {
c = 1;
} else {
c = (refa == s->varcompref[1] &&
refl != s->varcompref[1]) ? 2 : 4;
}
} else {
c = (refl == refa) ? 4 : 2;
}
}
} else {
if (s->above_intra_ctx[col]) {
c = 2;
} else if (s->above_comp_ctx[col]) {
c = 4 * (s->above_ref_ctx[col] != s->varcompref[1]);
} else {
c = 3 * (s->above_ref_ctx[col] != s->varcompref[1]);
}
}
} else if (have_l) {
if (s->left_intra_ctx[row7]) {
c = 2;
} else if (s->left_comp_ctx[row7]) {
c = 4 * (s->left_ref_ctx[row7] != s->varcompref[1]);
} else {
c = 3 * (s->left_ref_ctx[row7] != s->varcompref[1]);
}
} else {
c = 2;
}
bit = vp56_rac_get_prob(&s->c, s->prob.p.comp_ref[c]);
b->ref[var_idx] = s->varcompref[bit];
s->counts.comp_ref[c][bit]++;
} else { /* single reference */
int bit, c;
if (have_a && !s->above_intra_ctx[col]) {
if (have_l && !s->left_intra_ctx[row7]) {
if (s->left_comp_ctx[row7]) {
if (s->above_comp_ctx[col]) {
c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7] ||
!s->above_ref_ctx[col]);
} else {
c = (3 * !s->above_ref_ctx[col]) +
(!s->fixcompref || !s->left_ref_ctx[row7]);
}
} else if (s->above_comp_ctx[col]) {
c = (3 * !s->left_ref_ctx[row7]) +
(!s->fixcompref || !s->above_ref_ctx[col]);
} else {
c = 2 * !s->left_ref_ctx[row7] + 2 * !s->above_ref_ctx[col];
}
} else if (s->above_intra_ctx[col]) {
c = 2;
} else if (s->above_comp_ctx[col]) {
c = 1 + (!s->fixcompref || !s->above_ref_ctx[col]);
} else {
c = 4 * (!s->above_ref_ctx[col]);
}
} else if (have_l && !s->left_intra_ctx[row7]) {
if (s->left_intra_ctx[row7]) {
c = 2;
} else if (s->left_comp_ctx[row7]) {
c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7]);
} else {
c = 4 * (!s->left_ref_ctx[row7]);
}
} else {
c = 2;
}
bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][0]);
s->counts.single_ref[c][0][bit]++;
if (!bit) {
b->ref[0] = 0;
} else {
// FIXME can this codeblob be replaced by some sort of LUT?
if (have_a) {
if (have_l) {
if (s->left_intra_ctx[row7]) {
if (s->above_intra_ctx[col]) {
c = 2;
} else if (s->above_comp_ctx[col]) {
c = 1 + 2 * (s->fixcompref == 1 ||
s->above_ref_ctx[col] == 1);
} else if (!s->above_ref_ctx[col]) {
c = 3;
} else {
c = 4 * (s->above_ref_ctx[col] == 1);
}
} else if (s->above_intra_ctx[col]) {
if (s->left_intra_ctx[row7]) {
c = 2;
} else if (s->left_comp_ctx[row7]) {
c = 1 + 2 * (s->fixcompref == 1 ||
s->left_ref_ctx[row7] == 1);
} else if (!s->left_ref_ctx[row7]) {
c = 3;
} else {
c = 4 * (s->left_ref_ctx[row7] == 1);
}
} else if (s->above_comp_ctx[col]) {
if (s->left_comp_ctx[row7]) {
if (s->left_ref_ctx[row7] == s->above_ref_ctx[col]) {
c = 3 * (s->fixcompref == 1 ||
s->left_ref_ctx[row7] == 1);
} else {
c = 2;
}
} else if (!s->left_ref_ctx[row7]) {
c = 1 + 2 * (s->fixcompref == 1 ||
s->above_ref_ctx[col] == 1);
} else {
c = 3 * (s->left_ref_ctx[row7] == 1) +
(s->fixcompref == 1 || s->above_ref_ctx[col] == 1);
}
} else if (s->left_comp_ctx[row7]) {
if (!s->above_ref_ctx[col]) {
c = 1 + 2 * (s->fixcompref == 1 ||
s->left_ref_ctx[row7] == 1);
} else {
c = 3 * (s->above_ref_ctx[col] == 1) +
(s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);
}
} else if (!s->above_ref_ctx[col]) {
if (!s->left_ref_ctx[row7]) {
c = 3;
} else {
c = 4 * (s->left_ref_ctx[row7] == 1);
}
} else if (!s->left_ref_ctx[row7]) {
c = 4 * (s->above_ref_ctx[col] == 1);
} else {
c = 2 * (s->left_ref_ctx[row7] == 1) +
2 * (s->above_ref_ctx[col] == 1);
}
} else {
if (s->above_intra_ctx[col] ||
(!s->above_comp_ctx[col] && !s->above_ref_ctx[col])) {
c = 2;
} else if (s->above_comp_ctx[col]) {
c = 3 * (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);
} else {
c = 4 * (s->above_ref_ctx[col] == 1);
}
}
} else if (have_l) {
if (s->left_intra_ctx[row7] ||
(!s->left_comp_ctx[row7] && !s->left_ref_ctx[row7])) {
c = 2;
} else if (s->left_comp_ctx[row7]) {
c = 3 * (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);
} else {
c = 4 * (s->left_ref_ctx[row7] == 1);
}
} else {
c = 2;
}
bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][1]);
s->counts.single_ref[c][1][bit]++;
b->ref[0] = 1 + bit;
}
}
}
if (b->bs <= BS_8x8) {
if (s->segmentation.feat[b->seg_id].skip_enabled) {
b->mode[0] =
b->mode[1] =
b->mode[2] =
b->mode[3] = ZEROMV;
} else {
static const uint8_t off[10] = {
3, 0, 0, 1, 0, 0, 0, 0, 0, 0
};
// FIXME this needs to use the LUT tables from find_ref_mvs
// because not all are -1,0/0,-1
int c = inter_mode_ctx_lut[s->above_mode_ctx[col + off[b->bs]]]
[s->left_mode_ctx[row7 + off[b->bs]]];
b->mode[0] = vp8_rac_get_tree(&s->c, ff_vp9_inter_mode_tree,
s->prob.p.mv_mode[c]);
b->mode[1] =
b->mode[2] =
b->mode[3] = b->mode[0];
s->counts.mv_mode[c][b->mode[0] - 10]++;
}
}
if (s->filtermode == FILTER_SWITCHABLE) {
int c;
if (have_a && s->above_mode_ctx[col] >= NEARESTMV) {
if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {
c = s->above_filter_ctx[col] == s->left_filter_ctx[row7] ?
s->left_filter_ctx[row7] : 3;
} else {
c = s->above_filter_ctx[col];
}
} else if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {
c = s->left_filter_ctx[row7];
} else {
c = 3;
}
b->filter = vp8_rac_get_tree(&s->c, ff_vp9_filter_tree,
s->prob.p.filter[c]);
s->counts.filter[c][b->filter]++;
} else {
b->filter = s->filtermode;
}
if (b->bs > BS_8x8) {
int c = inter_mode_ctx_lut[s->above_mode_ctx[col]][s->left_mode_ctx[row7]];
b->mode[0] = vp8_rac_get_tree(&s->c, ff_vp9_inter_mode_tree,
s->prob.p.mv_mode[c]);
s->counts.mv_mode[c][b->mode[0] - 10]++;
ff_vp9_fill_mv(s, b->mv[0], b->mode[0], 0);
if (b->bs != BS_8x4) {
b->mode[1] = vp8_rac_get_tree(&s->c, ff_vp9_inter_mode_tree,
s->prob.p.mv_mode[c]);
s->counts.mv_mode[c][b->mode[1] - 10]++;
ff_vp9_fill_mv(s, b->mv[1], b->mode[1], 1);
} else {
b->mode[1] = b->mode[0];
AV_COPY32(&b->mv[1][0], &b->mv[0][0]);
AV_COPY32(&b->mv[1][1], &b->mv[0][1]);
}
if (b->bs != BS_4x8) {
b->mode[2] = vp8_rac_get_tree(&s->c, ff_vp9_inter_mode_tree,
s->prob.p.mv_mode[c]);
s->counts.mv_mode[c][b->mode[2] - 10]++;
ff_vp9_fill_mv(s, b->mv[2], b->mode[2], 2);
if (b->bs != BS_8x4) {
b->mode[3] = vp8_rac_get_tree(&s->c, ff_vp9_inter_mode_tree,
s->prob.p.mv_mode[c]);
s->counts.mv_mode[c][b->mode[3] - 10]++;
ff_vp9_fill_mv(s, b->mv[3], b->mode[3], 3);
} else {
b->mode[3] = b->mode[2];
AV_COPY32(&b->mv[3][0], &b->mv[2][0]);
AV_COPY32(&b->mv[3][1], &b->mv[2][1]);
}
} else {
b->mode[2] = b->mode[0];
AV_COPY32(&b->mv[2][0], &b->mv[0][0]);
AV_COPY32(&b->mv[2][1], &b->mv[0][1]);
b->mode[3] = b->mode[1];
AV_COPY32(&b->mv[3][0], &b->mv[1][0]);
AV_COPY32(&b->mv[3][1], &b->mv[1][1]);
}
} else {
ff_vp9_fill_mv(s, b->mv[0], b->mode[0], -1);
AV_COPY32(&b->mv[1][0], &b->mv[0][0]);
AV_COPY32(&b->mv[2][0], &b->mv[0][0]);
AV_COPY32(&b->mv[3][0], &b->mv[0][0]);
AV_COPY32(&b->mv[1][1], &b->mv[0][1]);
AV_COPY32(&b->mv[2][1], &b->mv[0][1]);
AV_COPY32(&b->mv[3][1], &b->mv[0][1]);
}
}
// FIXME this can probably be optimized
memset(&s->above_skip_ctx[col], b->skip, w4);
memset(&s->left_skip_ctx[row7], b->skip, h4);
memset(&s->above_txfm_ctx[col], b->tx, w4);
memset(&s->left_txfm_ctx[row7], b->tx, h4);
memset(&s->above_partition_ctx[col], above_ctx[b->bs], w4);
memset(&s->left_partition_ctx[row7], left_ctx[b->bs], h4);
if (!s->keyframe && !s->intraonly) {
memset(&s->above_intra_ctx[col], b->intra, w4);
memset(&s->left_intra_ctx[row7], b->intra, h4);
memset(&s->above_comp_ctx[col], b->comp, w4);
memset(&s->left_comp_ctx[row7], b->comp, h4);
memset(&s->above_mode_ctx[col], b->mode[3], w4);
memset(&s->left_mode_ctx[row7], b->mode[3], h4);
if (s->filtermode == FILTER_SWITCHABLE && !b->intra) {
memset(&s->above_filter_ctx[col], b->filter, w4);
memset(&s->left_filter_ctx[row7], b->filter, h4);
b->filter = ff_vp9_filter_lut[b->filter];
}
if (b->bs > BS_8x8) {
int mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);
AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][0], &b->mv[1][0]);
AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][1], &b->mv[1][1]);
AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][0], mv0);
AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][1], mv1);
AV_COPY32(&s->above_mv_ctx[col * 2 + 0][0], &b->mv[2][0]);
AV_COPY32(&s->above_mv_ctx[col * 2 + 0][1], &b->mv[2][1]);
AV_WN32A(&s->above_mv_ctx[col * 2 + 1][0], mv0);
AV_WN32A(&s->above_mv_ctx[col * 2 + 1][1], mv1);
} else {
int n, mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);
for (n = 0; n < w4 * 2; n++) {
AV_WN32A(&s->above_mv_ctx[col * 2 + n][0], mv0);
AV_WN32A(&s->above_mv_ctx[col * 2 + n][1], mv1);
}
for (n = 0; n < h4 * 2; n++) {
AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][0], mv0);
AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][1], mv1);
}
}
if (!b->intra) { // FIXME write 0xff or -1 if intra, so we can use this
// as a direct check in above branches
int vref = b->ref[b->comp ? s->signbias[s->varcompref[0]] : 0];
memset(&s->above_ref_ctx[col], vref, w4);
memset(&s->left_ref_ctx[row7], vref, h4);
}
}
// FIXME kinda ugly
for (y = 0; y < h4; y++) {
int x, o = (row + y) * s->sb_cols * 8 + col;
VP9MVRefPair *mv = &s->frames[CUR_FRAME].mv[o];
if (b->intra) {
for (x = 0; x < w4; x++) {
mv[x].ref[0] =
mv[x].ref[1] = -1;
}
} else if (b->comp) {
for (x = 0; x < w4; x++) {
mv[x].ref[0] = b->ref[0];
mv[x].ref[1] = b->ref[1];
AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);
AV_COPY32(&mv[x].mv[1], &b->mv[3][1]);
}
} else {
for (x = 0; x < w4; x++) {
mv[x].ref[0] = b->ref[0];
mv[x].ref[1] = -1;
AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);
}
}
}
}
// FIXME remove tx argument, and merge cnt/eob arguments?
static int decode_block_coeffs(VP56RangeCoder *c, int16_t *coef, int n_coeffs,
enum TxfmMode tx, unsigned (*cnt)[6][3],
unsigned (*eob)[6][2], uint8_t(*p)[6][11],
int nnz, const int16_t *scan,
const int16_t(*nb)[2],
const int16_t *band_counts, const int16_t *qmul)
{
int i = 0, band = 0, band_left = band_counts[band];
uint8_t *tp = p[0][nnz];
uint8_t cache[1024];
do {
int val, rc;
val = vp56_rac_get_prob_branchy(c, tp[0]); // eob
eob[band][nnz][val]++;
if (!val)
break;
skip_eob:
if (!vp56_rac_get_prob_branchy(c, tp[1])) { // zero
cnt[band][nnz][0]++;
if (!--band_left)
band_left = band_counts[++band];
cache[scan[i]] = 0;
nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;
tp = p[band][nnz];
if (++i == n_coeffs)
break; //invalid input; blocks should end with EOB
goto skip_eob;
}
rc = scan[i];
if (!vp56_rac_get_prob_branchy(c, tp[2])) { // one
cnt[band][nnz][1]++;
val = 1;
cache[rc] = 1;
} else {
// fill in p[3-10] (model fill) - only once per frame for each pos
if (!tp[3])
memcpy(&tp[3], ff_vp9_model_pareto8[tp[2]], 8);
cnt[band][nnz][2]++;
if (!vp56_rac_get_prob_branchy(c, tp[3])) { // 2, 3, 4
if (!vp56_rac_get_prob_branchy(c, tp[4])) {
cache[rc] = val = 2;
} else {
val = 3 + vp56_rac_get_prob(c, tp[5]);
cache[rc] = 3;
}
} else if (!vp56_rac_get_prob_branchy(c, tp[6])) { // cat1/2
cache[rc] = 4;
if (!vp56_rac_get_prob_branchy(c, tp[7])) {
val = vp56_rac_get_prob(c, 159) + 5;
} else {
val = (vp56_rac_get_prob(c, 165) << 1) + 7;
val += vp56_rac_get_prob(c, 145);
}
} else { // cat 3-6
cache[rc] = 5;
if (!vp56_rac_get_prob_branchy(c, tp[8])) {
if (!vp56_rac_get_prob_branchy(c, tp[9])) {
val = (vp56_rac_get_prob(c, 173) << 2) + 11;
val += (vp56_rac_get_prob(c, 148) << 1);
val += vp56_rac_get_prob(c, 140);
} else {
val = (vp56_rac_get_prob(c, 176) << 3) + 19;
val += (vp56_rac_get_prob(c, 155) << 2);
val += (vp56_rac_get_prob(c, 140) << 1);
val += vp56_rac_get_prob(c, 135);
}
} else if (!vp56_rac_get_prob_branchy(c, tp[10])) {
val = (vp56_rac_get_prob(c, 180) << 4) + 35;
val += (vp56_rac_get_prob(c, 157) << 3);
val += (vp56_rac_get_prob(c, 141) << 2);
val += (vp56_rac_get_prob(c, 134) << 1);
val += vp56_rac_get_prob(c, 130);
} else {
val = (vp56_rac_get_prob(c, 254) << 13) + 67;
val += (vp56_rac_get_prob(c, 254) << 12);
val += (vp56_rac_get_prob(c, 254) << 11);
val += (vp56_rac_get_prob(c, 252) << 10);
val += (vp56_rac_get_prob(c, 249) << 9);
val += (vp56_rac_get_prob(c, 243) << 8);
val += (vp56_rac_get_prob(c, 230) << 7);
val += (vp56_rac_get_prob(c, 196) << 6);
val += (vp56_rac_get_prob(c, 177) << 5);
val += (vp56_rac_get_prob(c, 153) << 4);
val += (vp56_rac_get_prob(c, 140) << 3);
val += (vp56_rac_get_prob(c, 133) << 2);
val += (vp56_rac_get_prob(c, 130) << 1);
val += vp56_rac_get_prob(c, 129);
}
}
}
if (!--band_left)
band_left = band_counts[++band];
if (tx == TX_32X32) // FIXME slow
coef[rc] = ((vp8_rac_get(c) ? -val : val) * qmul[!!i]) / 2;
else
coef[rc] = (vp8_rac_get(c) ? -val : val) * qmul[!!i];
nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;
tp = p[band][nnz];
} while (++i < n_coeffs);
return i;
}
static int decode_coeffs(AVCodecContext *avctx)
{
VP9Context *s = avctx->priv_data;
VP9Block *b = s->b;
int row = b->row, col = b->col;
uint8_t (*p)[6][11] = s->prob.coef[b->tx][0 /* y */][!b->intra];
unsigned (*c)[6][3] = s->counts.coef[b->tx][0 /* y */][!b->intra];
unsigned (*e)[6][2] = s->counts.eob[b->tx][0 /* y */][!b->intra];
int w4 = bwh_tab[1][b->bs][0] << 1, h4 = bwh_tab[1][b->bs][1] << 1;
int end_x = FFMIN(2 * (s->cols - col), w4);
int end_y = FFMIN(2 * (s->rows - row), h4);
int n, pl, x, y, step1d = 1 << b->tx, step = 1 << (b->tx * 2);
int uvstep1d = 1 << b->uvtx, uvstep = 1 << (b->uvtx * 2), ret;
int16_t (*qmul)[2] = s->segmentation.feat[b->seg_id].qmul;
int tx = 4 * s->lossless + b->tx;
const int16_t * const *yscans = ff_vp9_scans[tx];
const int16_t (* const * ynbs)[2] = ff_vp9_scans_nb[tx];
const int16_t *uvscan = ff_vp9_scans[b->uvtx][DCT_DCT];
const int16_t (*uvnb)[2] = ff_vp9_scans_nb[b->uvtx][DCT_DCT];
uint8_t *a = &s->above_y_nnz_ctx[col * 2];
uint8_t *l = &s->left_y_nnz_ctx[(row & 7) << 1];
static const int16_t band_counts[4][8] = {
{ 1, 2, 3, 4, 3, 16 - 13, 0 },
{ 1, 2, 3, 4, 11, 64 - 21, 0 },
{ 1, 2, 3, 4, 11, 256 - 21, 0 },
{ 1, 2, 3, 4, 11, 1024 - 21, 0 },
};
const int16_t *y_band_counts = band_counts[b->tx];
const int16_t *uv_band_counts = band_counts[b->uvtx];
/* y tokens */
if (b->tx > TX_4X4) { // FIXME slow
for (y = 0; y < end_y; y += step1d)
for (x = 1; x < step1d; x++)
l[y] |= l[y + x];
for (x = 0; x < end_x; x += step1d)
for (y = 1; y < step1d; y++)
a[x] |= a[x + y];
}
for (n = 0, y = 0; y < end_y; y += step1d) {
for (x = 0; x < end_x; x += step1d, n += step) {
enum TxfmType txtp = ff_vp9_intra_txfm_type[b->mode[b->tx == TX_4X4 &&
b->bs > BS_8x8 ?
n : 0]];
int nnz = a[x] + l[y];
if ((ret = decode_block_coeffs(&s->c, s->block + 16 * n, 16 * step,
b->tx, c, e, p, nnz, yscans[txtp],
ynbs[txtp], y_band_counts,
qmul[0])) < 0)
return ret;
a[x] = l[y] = !!ret;
if (b->tx > TX_8X8)
AV_WN16A(&s->eob[n], ret);
else
s->eob[n] = ret;
}
}
if (b->tx > TX_4X4) { // FIXME slow
for (y = 0; y < end_y; y += step1d)
memset(&l[y + 1], l[y], FFMIN(end_y - y - 1, step1d - 1));
for (x = 0; x < end_x; x += step1d)
memset(&a[x + 1], a[x], FFMIN(end_x - x - 1, step1d - 1));
}
p = s->prob.coef[b->uvtx][1 /* uv */][!b->intra];
c = s->counts.coef[b->uvtx][1 /* uv */][!b->intra];
e = s->counts.eob[b->uvtx][1 /* uv */][!b->intra];
w4 >>= 1;
h4 >>= 1;
end_x >>= 1;
end_y >>= 1;
for (pl = 0; pl < 2; pl++) {
a = &s->above_uv_nnz_ctx[pl][col];
l = &s->left_uv_nnz_ctx[pl][row & 7];
if (b->uvtx > TX_4X4) { // FIXME slow
for (y = 0; y < end_y; y += uvstep1d)
for (x = 1; x < uvstep1d; x++)
l[y] |= l[y + x];
for (x = 0; x < end_x; x += uvstep1d)
for (y = 1; y < uvstep1d; y++)
a[x] |= a[x + y];
}
for (n = 0, y = 0; y < end_y; y += uvstep1d) {
for (x = 0; x < end_x; x += uvstep1d, n += uvstep) {
int nnz = a[x] + l[y];
if ((ret = decode_block_coeffs(&s->c, s->uvblock[pl] + 16 * n,
16 * uvstep, b->uvtx, c, e, p,
nnz, uvscan, uvnb,
uv_band_counts, qmul[1])) < 0)
return ret;
a[x] = l[y] = !!ret;
if (b->uvtx > TX_8X8)
AV_WN16A(&s->uveob[pl][n], ret);
else
s->uveob[pl][n] = ret;
}
}
if (b->uvtx > TX_4X4) { // FIXME slow
for (y = 0; y < end_y; y += uvstep1d)
memset(&l[y + 1], l[y], FFMIN(end_y - y - 1, uvstep1d - 1));
for (x = 0; x < end_x; x += uvstep1d)
memset(&a[x + 1], a[x], FFMIN(end_x - x - 1, uvstep1d - 1));
}
}
return 0;
}
static av_always_inline int check_intra_mode(VP9Context *s, int mode,
uint8_t **a,
uint8_t *dst_edge,
ptrdiff_t stride_edge,
uint8_t *dst_inner,
ptrdiff_t stride_inner,
uint8_t *l, int col, int x, int w,
int row, int y, enum TxfmMode tx,
int p)
{
int have_top = row > 0 || y > 0;
int have_left = col > s->tiling.tile_col_start || x > 0;
int have_right = x < w - 1;
static const uint8_t mode_conv[10][2 /* have_left */][2 /* have_top */] = {
[VERT_PRED] = { { DC_127_PRED, VERT_PRED },
{ DC_127_PRED, VERT_PRED } },
[HOR_PRED] = { { DC_129_PRED, DC_129_PRED },
{ HOR_PRED, HOR_PRED } },
[DC_PRED] = { { DC_128_PRED, TOP_DC_PRED },
{ LEFT_DC_PRED, DC_PRED } },
[DIAG_DOWN_LEFT_PRED] = { { DC_127_PRED, DIAG_DOWN_LEFT_PRED },
{ DC_127_PRED, DIAG_DOWN_LEFT_PRED } },
[DIAG_DOWN_RIGHT_PRED] = { { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED },
{ DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED } },
[VERT_RIGHT_PRED] = { { VERT_RIGHT_PRED, VERT_RIGHT_PRED },
{ VERT_RIGHT_PRED, VERT_RIGHT_PRED } },
[HOR_DOWN_PRED] = { { HOR_DOWN_PRED, HOR_DOWN_PRED },
{ HOR_DOWN_PRED, HOR_DOWN_PRED } },
[VERT_LEFT_PRED] = { { DC_127_PRED, VERT_LEFT_PRED },
{ DC_127_PRED, VERT_LEFT_PRED } },
[HOR_UP_PRED] = { { DC_129_PRED, DC_129_PRED },
{ HOR_UP_PRED, HOR_UP_PRED } },
[TM_VP8_PRED] = { { DC_129_PRED, VERT_PRED },
{ HOR_PRED, TM_VP8_PRED } },
};
static const struct {
uint8_t needs_left:1;
uint8_t needs_top:1;
uint8_t needs_topleft:1;
uint8_t needs_topright:1;
} edges[N_INTRA_PRED_MODES] = {
[VERT_PRED] = { .needs_top = 1 },
[HOR_PRED] = { .needs_left = 1 },
[DC_PRED] = { .needs_top = 1, .needs_left = 1 },
[DIAG_DOWN_LEFT_PRED] = { .needs_top = 1, .needs_topright = 1 },
[DIAG_DOWN_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1,
.needs_topleft = 1 },
[VERT_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1,
.needs_topleft = 1 },
[HOR_DOWN_PRED] = { .needs_left = 1, .needs_top = 1,
.needs_topleft = 1 },
[VERT_LEFT_PRED] = { .needs_top = 1, .needs_topright = 1 },
[HOR_UP_PRED] = { .needs_left = 1 },
[TM_VP8_PRED] = { .needs_left = 1, .needs_top = 1,
.needs_topleft = 1 },
[LEFT_DC_PRED] = { .needs_left = 1 },
[TOP_DC_PRED] = { .needs_top = 1 },
[DC_128_PRED] = { 0 },
[DC_127_PRED] = { 0 },
[DC_129_PRED] = { 0 }
};
av_assert2(mode >= 0 && mode < 10);
mode = mode_conv[mode][have_left][have_top];
if (edges[mode].needs_top) {
uint8_t *top = NULL, *topleft = NULL;
int n_px_need = 4 << tx, n_px_have = (((s->cols - col) << !p) - x) * 4;
int n_px_need_tr = 0;
if (tx == TX_4X4 && edges[mode].needs_topright && have_right)
n_px_need_tr = 4;
// if top of sb64-row, use s->intra_pred_data[] instead of
// dst[-stride] for intra prediction (it contains pre- instead of
// post-loopfilter data)
if (have_top) {
top = !(row & 7) && !y ?
s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :
y == 0 ? &dst_edge[-stride_edge] : &dst_inner[-stride_inner];
if (have_left)
topleft = !(row & 7) && !y ?
s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :
y == 0 || x == 0 ? &dst_edge[-stride_edge] :
&dst_inner[-stride_inner];
}
if (have_top &&
(!edges[mode].needs_topleft || (have_left && top == topleft)) &&
(tx != TX_4X4 || !edges[mode].needs_topright || have_right) &&
n_px_need + n_px_need_tr <= n_px_have) {
*a = top;
} else {
if (have_top) {
if (n_px_need <= n_px_have) {
memcpy(*a, top, n_px_need);
} else {
memcpy(*a, top, n_px_have);
memset(&(*a)[n_px_have], (*a)[n_px_have - 1],
n_px_need - n_px_have);
}
} else {
memset(*a, 127, n_px_need);
}
if (edges[mode].needs_topleft) {
if (have_left && have_top)
(*a)[-1] = topleft[-1];
else
(*a)[-1] = have_top ? 129 : 127;
}
if (tx == TX_4X4 && edges[mode].needs_topright) {
if (have_top && have_right &&
n_px_need + n_px_need_tr <= n_px_have) {
memcpy(&(*a)[4], &top[4], 4);
} else {
memset(&(*a)[4], (*a)[3], 4);
}
}
}
}
if (edges[mode].needs_left) {
if (have_left) {
int i;
int n_px_need = 4 << tx;
int n_px_have = (((s->rows - row) << !p) - y) * 4;
uint8_t *dst = x == 0 ? dst_edge : dst_inner;
ptrdiff_t stride = x == 0 ? stride_edge : stride_inner;
if (n_px_need <= n_px_have) {
for (i = 0; i < n_px_need; i++)
l[i] = dst[i * stride - 1];
} else {
for (i = 0; i < n_px_have; i++)
l[i] = dst[i * stride - 1];
memset(&l[i], l[i - 1], n_px_need - n_px_have);
}
} else {
memset(l, 129, 4 << tx);
}
}
return mode;
}
static void intra_recon(AVCodecContext *avctx, ptrdiff_t y_off, ptrdiff_t uv_off)
{
VP9Context *s = avctx->priv_data;
VP9Block *b = s->b;
AVFrame *f = s->frames[CUR_FRAME].tf.f;
int row = b->row, col = b->col;
int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;
int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);
int end_x = FFMIN(2 * (s->cols - col), w4);
int end_y = FFMIN(2 * (s->rows - row), h4);
int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;
int uvstep1d = 1 << b->uvtx, p;
uint8_t *dst = b->dst[0], *dst_r = f->data[0] + y_off;
for (n = 0, y = 0; y < end_y; y += step1d) {
uint8_t *ptr = dst, *ptr_r = dst_r;
for (x = 0; x < end_x;
x += step1d, ptr += 4 * step1d, ptr_r += 4 * step1d, n += step) {
int mode = b->mode[b->bs > BS_8x8 && b->tx == TX_4X4 ?
y * 2 + x : 0];
LOCAL_ALIGNED_16(uint8_t, a_buf, [48]);
uint8_t *a = &a_buf[16], l[32];
enum TxfmType txtp = ff_vp9_intra_txfm_type[mode];
int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];
mode = check_intra_mode(s, mode, &a, ptr_r,
f->linesize[0],
ptr, b->y_stride, l,
col, x, w4, row, y, b->tx, 0);
s->dsp.intra_pred[b->tx][mode](ptr, b->y_stride, l, a);
if (eob)
s->dsp.itxfm_add[tx][txtp](ptr, b->y_stride,
s->block + 16 * n, eob);
}
dst_r += 4 * f->linesize[0] * step1d;
dst += 4 * b->y_stride * step1d;
}
// U/V
h4 >>= 1;
w4 >>= 1;
end_x >>= 1;
end_y >>= 1;
step = 1 << (b->uvtx * 2);
for (p = 0; p < 2; p++) {
dst = b->dst[1 + p];
dst_r = f->data[1 + p] + uv_off;
for (n = 0, y = 0; y < end_y; y += uvstep1d) {
uint8_t *ptr = dst, *ptr_r = dst_r;
for (x = 0; x < end_x;
x += uvstep1d, ptr += 4 * uvstep1d,
ptr_r += 4 * uvstep1d, n += step) {
int mode = b->uvmode;
LOCAL_ALIGNED_16(uint8_t, a_buf, [48]);
uint8_t *a = &a_buf[16], l[32];
int eob = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n])
: s->uveob[p][n];
mode = check_intra_mode(s, mode, &a, ptr_r,
f->linesize[1],
ptr, b->uv_stride, l,
col, x, w4, row, y, b->uvtx, p + 1);
s->dsp.intra_pred[b->uvtx][mode](ptr, b->uv_stride, l, a);
if (eob)
s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, b->uv_stride,
s->uvblock[p] + 16 * n,
eob);
}
dst_r += 4 * uvstep1d * f->linesize[1];
dst += 4 * uvstep1d * b->uv_stride;
}
}
}
static av_always_inline void mc_luma_dir(VP9Context *s, vp9_mc_func(*mc)[2],
uint8_t *dst, ptrdiff_t dst_stride,
const uint8_t *ref,
ptrdiff_t ref_stride,
ThreadFrame *ref_frame,
ptrdiff_t y, ptrdiff_t x,
const VP56mv *mv,
int bw, int bh, int w, int h)
{
int mx = mv->x, my = mv->y;
int th;
y += my >> 3;
x += mx >> 3;
ref += y * ref_stride + x;
mx &= 7;
my &= 7;
// we use +7 because the last 7 pixels of each sbrow can be changed in
// the longest loopfilter of the next sbrow
th = (y + bh + 4 * !!my + 7) >> 6;
ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);
// FIXME bilinear filter only needs 0/1 pixels, not 3/4
aarch64: vp9: Add NEON optimizations of VP9 MC functions This work is sponsored by, and copyright, Google. These are ported from the ARM version; it is essentially a 1:1 port with no extra added features, but with some hand tuning (especially for the plain copy/avg functions). The ARM version isn't very register starved to begin with, so there's not much to be gained from having more spare registers here - we only avoid having to clobber callee-saved registers. Examples of runtimes vs the 32 bit version, on a Cortex A53: ARM AArch64 vp9_avg4_neon: 27.2 23.7 vp9_avg8_neon: 56.5 54.7 vp9_avg16_neon: 169.9 167.4 vp9_avg32_neon: 585.8 585.2 vp9_avg64_neon: 2460.3 2294.7 vp9_avg_8tap_smooth_4h_neon: 132.7 125.2 vp9_avg_8tap_smooth_4hv_neon: 478.8 442.0 vp9_avg_8tap_smooth_4v_neon: 126.0 93.7 vp9_avg_8tap_smooth_8h_neon: 241.7 234.2 vp9_avg_8tap_smooth_8hv_neon: 690.9 646.5 vp9_avg_8tap_smooth_8v_neon: 245.0 205.5 vp9_avg_8tap_smooth_64h_neon: 11273.2 11280.1 vp9_avg_8tap_smooth_64hv_neon: 22980.6 22184.1 vp9_avg_8tap_smooth_64v_neon: 11549.7 10781.1 vp9_put4_neon: 18.0 17.2 vp9_put8_neon: 40.2 37.7 vp9_put16_neon: 97.4 99.5 vp9_put32_neon/armv8: 346.0 307.4 vp9_put64_neon/armv8: 1319.0 1107.5 vp9_put_8tap_smooth_4h_neon: 126.7 118.2 vp9_put_8tap_smooth_4hv_neon: 465.7 434.0 vp9_put_8tap_smooth_4v_neon: 113.0 86.5 vp9_put_8tap_smooth_8h_neon: 229.7 221.6 vp9_put_8tap_smooth_8hv_neon: 658.9 621.3 vp9_put_8tap_smooth_8v_neon: 215.0 187.5 vp9_put_8tap_smooth_64h_neon: 10636.7 10627.8 vp9_put_8tap_smooth_64hv_neon: 21076.8 21026.9 vp9_put_8tap_smooth_64v_neon: 9635.0 9632.4 These are generally about as fast as the corresponding ARM routines on the same CPU (at least on the A53), in most cases marginally faster. The speedup vs C code is pretty much the same as for the 32 bit case; on the A53 it's around 6-13x for ther larger 8tap filters. The exact speedup varies a little, since the C versions generally don't end up exactly as slow/fast as on 32 bit. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-10 11:06:26 +02:00
// The arm/aarch64 _hv filters read one more row than what actually is
arm: vp9: Add NEON optimizations of VP9 MC functions This work is sponsored by, and copyright, Google. The filter coefficients are signed values, where the product of the multiplication with one individual filter coefficient doesn't overflow a 16 bit signed value (the largest filter coefficient is 127). But when the products are accumulated, the resulting sum can overflow the 16 bit signed range. Instead of accumulating in 32 bit, we accumulate the largest product (either index 3 or 4) last with a saturated addition. (The VP8 MC asm does something similar, but slightly simpler, by accumulating each half of the filter separately. In the VP9 MC filters, each half of the filter can also overflow though, so the largest component has to be handled individually.) Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_avg4_neon: 1.71 1.15 1.42 1.49 vp9_avg8_neon: 2.51 3.63 3.14 2.58 vp9_avg16_neon: 2.95 6.76 3.01 2.84 vp9_avg32_neon: 3.29 6.64 2.85 3.00 vp9_avg64_neon: 3.47 6.67 3.14 2.80 vp9_avg_8tap_smooth_4h_neon: 3.22 4.73 2.76 4.67 vp9_avg_8tap_smooth_4hv_neon: 3.67 4.76 3.28 4.71 vp9_avg_8tap_smooth_4v_neon: 5.52 7.60 4.60 6.31 vp9_avg_8tap_smooth_8h_neon: 6.22 9.04 5.12 9.32 vp9_avg_8tap_smooth_8hv_neon: 6.38 8.21 5.72 8.17 vp9_avg_8tap_smooth_8v_neon: 9.22 12.66 8.15 11.10 vp9_avg_8tap_smooth_64h_neon: 7.02 10.23 5.54 11.58 vp9_avg_8tap_smooth_64hv_neon: 6.76 9.46 5.93 9.40 vp9_avg_8tap_smooth_64v_neon: 10.76 14.13 9.46 13.37 vp9_put4_neon: 1.11 1.47 1.00 1.21 vp9_put8_neon: 1.23 2.17 1.94 1.48 vp9_put16_neon: 1.63 4.02 1.73 1.97 vp9_put32_neon: 1.56 4.92 2.00 1.96 vp9_put64_neon: 2.10 5.28 2.03 2.35 vp9_put_8tap_smooth_4h_neon: 3.11 4.35 2.63 4.35 vp9_put_8tap_smooth_4hv_neon: 3.67 4.69 3.25 4.71 vp9_put_8tap_smooth_4v_neon: 5.45 7.27 4.49 6.52 vp9_put_8tap_smooth_8h_neon: 5.97 8.18 4.81 8.56 vp9_put_8tap_smooth_8hv_neon: 6.39 7.90 5.64 8.15 vp9_put_8tap_smooth_8v_neon: 9.03 11.84 8.07 11.51 vp9_put_8tap_smooth_64h_neon: 6.78 9.48 4.88 10.89 vp9_put_8tap_smooth_64hv_neon: 6.99 8.87 5.94 9.56 vp9_put_8tap_smooth_64v_neon: 10.69 13.30 9.43 14.34 For the larger 8tap filters, the speedup vs C code is around 5-14x. This is significantly faster than libvpx's implementation of the same functions, at least when comparing the put_8tap_smooth_64 functions (compared to vpx_convolve8_horiz_neon and vpx_convolve8_vert_neon from libvpx). Absolute runtimes from checkasm: Cortex A7 A8 A9 A53 vp9_put_8tap_smooth_64h_neon: 20150.3 14489.4 19733.6 10863.7 libvpx vpx_convolve8_horiz_neon: 52623.3 19736.4 21907.7 25027.7 vp9_put_8tap_smooth_64v_neon: 14455.0 12303.9 13746.4 9628.9 libvpx vpx_convolve8_vert_neon: 42090.0 17706.2 17659.9 16941.2 Thus, on the A9, the horizontal filter is only marginally faster than libvpx, while our version is significantly faster on the other cores, and the vertical filter is significantly faster on all cores. The difference is especially large on the A7. The libvpx implementation does the accumulation in 32 bit, which probably explains most of the differences. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-03 09:12:08 +02:00
// needed, so switch to emulated edge one pixel sooner vertically
// (!!my * 5) than horizontally (!!mx * 4).
if (x < !!mx * 3 || y < !!my * 3 ||
arm: vp9: Add NEON optimizations of VP9 MC functions This work is sponsored by, and copyright, Google. The filter coefficients are signed values, where the product of the multiplication with one individual filter coefficient doesn't overflow a 16 bit signed value (the largest filter coefficient is 127). But when the products are accumulated, the resulting sum can overflow the 16 bit signed range. Instead of accumulating in 32 bit, we accumulate the largest product (either index 3 or 4) last with a saturated addition. (The VP8 MC asm does something similar, but slightly simpler, by accumulating each half of the filter separately. In the VP9 MC filters, each half of the filter can also overflow though, so the largest component has to be handled individually.) Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_avg4_neon: 1.71 1.15 1.42 1.49 vp9_avg8_neon: 2.51 3.63 3.14 2.58 vp9_avg16_neon: 2.95 6.76 3.01 2.84 vp9_avg32_neon: 3.29 6.64 2.85 3.00 vp9_avg64_neon: 3.47 6.67 3.14 2.80 vp9_avg_8tap_smooth_4h_neon: 3.22 4.73 2.76 4.67 vp9_avg_8tap_smooth_4hv_neon: 3.67 4.76 3.28 4.71 vp9_avg_8tap_smooth_4v_neon: 5.52 7.60 4.60 6.31 vp9_avg_8tap_smooth_8h_neon: 6.22 9.04 5.12 9.32 vp9_avg_8tap_smooth_8hv_neon: 6.38 8.21 5.72 8.17 vp9_avg_8tap_smooth_8v_neon: 9.22 12.66 8.15 11.10 vp9_avg_8tap_smooth_64h_neon: 7.02 10.23 5.54 11.58 vp9_avg_8tap_smooth_64hv_neon: 6.76 9.46 5.93 9.40 vp9_avg_8tap_smooth_64v_neon: 10.76 14.13 9.46 13.37 vp9_put4_neon: 1.11 1.47 1.00 1.21 vp9_put8_neon: 1.23 2.17 1.94 1.48 vp9_put16_neon: 1.63 4.02 1.73 1.97 vp9_put32_neon: 1.56 4.92 2.00 1.96 vp9_put64_neon: 2.10 5.28 2.03 2.35 vp9_put_8tap_smooth_4h_neon: 3.11 4.35 2.63 4.35 vp9_put_8tap_smooth_4hv_neon: 3.67 4.69 3.25 4.71 vp9_put_8tap_smooth_4v_neon: 5.45 7.27 4.49 6.52 vp9_put_8tap_smooth_8h_neon: 5.97 8.18 4.81 8.56 vp9_put_8tap_smooth_8hv_neon: 6.39 7.90 5.64 8.15 vp9_put_8tap_smooth_8v_neon: 9.03 11.84 8.07 11.51 vp9_put_8tap_smooth_64h_neon: 6.78 9.48 4.88 10.89 vp9_put_8tap_smooth_64hv_neon: 6.99 8.87 5.94 9.56 vp9_put_8tap_smooth_64v_neon: 10.69 13.30 9.43 14.34 For the larger 8tap filters, the speedup vs C code is around 5-14x. This is significantly faster than libvpx's implementation of the same functions, at least when comparing the put_8tap_smooth_64 functions (compared to vpx_convolve8_horiz_neon and vpx_convolve8_vert_neon from libvpx). Absolute runtimes from checkasm: Cortex A7 A8 A9 A53 vp9_put_8tap_smooth_64h_neon: 20150.3 14489.4 19733.6 10863.7 libvpx vpx_convolve8_horiz_neon: 52623.3 19736.4 21907.7 25027.7 vp9_put_8tap_smooth_64v_neon: 14455.0 12303.9 13746.4 9628.9 libvpx vpx_convolve8_vert_neon: 42090.0 17706.2 17659.9 16941.2 Thus, on the A9, the horizontal filter is only marginally faster than libvpx, while our version is significantly faster on the other cores, and the vertical filter is significantly faster on all cores. The difference is especially large on the A7. The libvpx implementation does the accumulation in 32 bit, which probably explains most of the differences. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-03 09:12:08 +02:00
x + !!mx * 4 > w - bw || y + !!my * 5 > h - bh) {
s->vdsp.emulated_edge_mc(s->edge_emu_buffer,
ref - !!my * 3 * ref_stride - !!mx * 3,
80,
ref_stride,
bw + !!mx * 7, bh + !!my * 7,
x - !!mx * 3, y - !!my * 3, w, h);
ref = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;
ref_stride = 80;
}
mc[!!mx][!!my](dst, dst_stride, ref, ref_stride, bh, mx << 1, my << 1);
}
static av_always_inline void mc_chroma_dir(VP9Context *s, vp9_mc_func(*mc)[2],
uint8_t *dst_u, uint8_t *dst_v,
ptrdiff_t dst_stride,
const uint8_t *ref_u,
ptrdiff_t src_stride_u,
const uint8_t *ref_v,
ptrdiff_t src_stride_v,
ThreadFrame *ref_frame,
ptrdiff_t y, ptrdiff_t x,
const VP56mv *mv,
int bw, int bh, int w, int h)
{
int mx = mv->x, my = mv->y;
int th;
y += my >> 4;
x += mx >> 4;
ref_u += y * src_stride_u + x;
ref_v += y * src_stride_v + x;
mx &= 15;
my &= 15;
// we use +7 because the last 7 pixels of each sbrow can be changed in
// the longest loopfilter of the next sbrow
th = (y + bh + 4 * !!my + 7) >> 5;
ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);
// FIXME bilinear filter only needs 0/1 pixels, not 3/4
aarch64: vp9: Add NEON optimizations of VP9 MC functions This work is sponsored by, and copyright, Google. These are ported from the ARM version; it is essentially a 1:1 port with no extra added features, but with some hand tuning (especially for the plain copy/avg functions). The ARM version isn't very register starved to begin with, so there's not much to be gained from having more spare registers here - we only avoid having to clobber callee-saved registers. Examples of runtimes vs the 32 bit version, on a Cortex A53: ARM AArch64 vp9_avg4_neon: 27.2 23.7 vp9_avg8_neon: 56.5 54.7 vp9_avg16_neon: 169.9 167.4 vp9_avg32_neon: 585.8 585.2 vp9_avg64_neon: 2460.3 2294.7 vp9_avg_8tap_smooth_4h_neon: 132.7 125.2 vp9_avg_8tap_smooth_4hv_neon: 478.8 442.0 vp9_avg_8tap_smooth_4v_neon: 126.0 93.7 vp9_avg_8tap_smooth_8h_neon: 241.7 234.2 vp9_avg_8tap_smooth_8hv_neon: 690.9 646.5 vp9_avg_8tap_smooth_8v_neon: 245.0 205.5 vp9_avg_8tap_smooth_64h_neon: 11273.2 11280.1 vp9_avg_8tap_smooth_64hv_neon: 22980.6 22184.1 vp9_avg_8tap_smooth_64v_neon: 11549.7 10781.1 vp9_put4_neon: 18.0 17.2 vp9_put8_neon: 40.2 37.7 vp9_put16_neon: 97.4 99.5 vp9_put32_neon/armv8: 346.0 307.4 vp9_put64_neon/armv8: 1319.0 1107.5 vp9_put_8tap_smooth_4h_neon: 126.7 118.2 vp9_put_8tap_smooth_4hv_neon: 465.7 434.0 vp9_put_8tap_smooth_4v_neon: 113.0 86.5 vp9_put_8tap_smooth_8h_neon: 229.7 221.6 vp9_put_8tap_smooth_8hv_neon: 658.9 621.3 vp9_put_8tap_smooth_8v_neon: 215.0 187.5 vp9_put_8tap_smooth_64h_neon: 10636.7 10627.8 vp9_put_8tap_smooth_64hv_neon: 21076.8 21026.9 vp9_put_8tap_smooth_64v_neon: 9635.0 9632.4 These are generally about as fast as the corresponding ARM routines on the same CPU (at least on the A53), in most cases marginally faster. The speedup vs C code is pretty much the same as for the 32 bit case; on the A53 it's around 6-13x for ther larger 8tap filters. The exact speedup varies a little, since the C versions generally don't end up exactly as slow/fast as on 32 bit. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-10 11:06:26 +02:00
// The arm/aarch64 _hv filters read one more row than what actually is
arm: vp9: Add NEON optimizations of VP9 MC functions This work is sponsored by, and copyright, Google. The filter coefficients are signed values, where the product of the multiplication with one individual filter coefficient doesn't overflow a 16 bit signed value (the largest filter coefficient is 127). But when the products are accumulated, the resulting sum can overflow the 16 bit signed range. Instead of accumulating in 32 bit, we accumulate the largest product (either index 3 or 4) last with a saturated addition. (The VP8 MC asm does something similar, but slightly simpler, by accumulating each half of the filter separately. In the VP9 MC filters, each half of the filter can also overflow though, so the largest component has to be handled individually.) Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_avg4_neon: 1.71 1.15 1.42 1.49 vp9_avg8_neon: 2.51 3.63 3.14 2.58 vp9_avg16_neon: 2.95 6.76 3.01 2.84 vp9_avg32_neon: 3.29 6.64 2.85 3.00 vp9_avg64_neon: 3.47 6.67 3.14 2.80 vp9_avg_8tap_smooth_4h_neon: 3.22 4.73 2.76 4.67 vp9_avg_8tap_smooth_4hv_neon: 3.67 4.76 3.28 4.71 vp9_avg_8tap_smooth_4v_neon: 5.52 7.60 4.60 6.31 vp9_avg_8tap_smooth_8h_neon: 6.22 9.04 5.12 9.32 vp9_avg_8tap_smooth_8hv_neon: 6.38 8.21 5.72 8.17 vp9_avg_8tap_smooth_8v_neon: 9.22 12.66 8.15 11.10 vp9_avg_8tap_smooth_64h_neon: 7.02 10.23 5.54 11.58 vp9_avg_8tap_smooth_64hv_neon: 6.76 9.46 5.93 9.40 vp9_avg_8tap_smooth_64v_neon: 10.76 14.13 9.46 13.37 vp9_put4_neon: 1.11 1.47 1.00 1.21 vp9_put8_neon: 1.23 2.17 1.94 1.48 vp9_put16_neon: 1.63 4.02 1.73 1.97 vp9_put32_neon: 1.56 4.92 2.00 1.96 vp9_put64_neon: 2.10 5.28 2.03 2.35 vp9_put_8tap_smooth_4h_neon: 3.11 4.35 2.63 4.35 vp9_put_8tap_smooth_4hv_neon: 3.67 4.69 3.25 4.71 vp9_put_8tap_smooth_4v_neon: 5.45 7.27 4.49 6.52 vp9_put_8tap_smooth_8h_neon: 5.97 8.18 4.81 8.56 vp9_put_8tap_smooth_8hv_neon: 6.39 7.90 5.64 8.15 vp9_put_8tap_smooth_8v_neon: 9.03 11.84 8.07 11.51 vp9_put_8tap_smooth_64h_neon: 6.78 9.48 4.88 10.89 vp9_put_8tap_smooth_64hv_neon: 6.99 8.87 5.94 9.56 vp9_put_8tap_smooth_64v_neon: 10.69 13.30 9.43 14.34 For the larger 8tap filters, the speedup vs C code is around 5-14x. This is significantly faster than libvpx's implementation of the same functions, at least when comparing the put_8tap_smooth_64 functions (compared to vpx_convolve8_horiz_neon and vpx_convolve8_vert_neon from libvpx). Absolute runtimes from checkasm: Cortex A7 A8 A9 A53 vp9_put_8tap_smooth_64h_neon: 20150.3 14489.4 19733.6 10863.7 libvpx vpx_convolve8_horiz_neon: 52623.3 19736.4 21907.7 25027.7 vp9_put_8tap_smooth_64v_neon: 14455.0 12303.9 13746.4 9628.9 libvpx vpx_convolve8_vert_neon: 42090.0 17706.2 17659.9 16941.2 Thus, on the A9, the horizontal filter is only marginally faster than libvpx, while our version is significantly faster on the other cores, and the vertical filter is significantly faster on all cores. The difference is especially large on the A7. The libvpx implementation does the accumulation in 32 bit, which probably explains most of the differences. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-03 09:12:08 +02:00
// needed, so switch to emulated edge one pixel sooner vertically
// (!!my * 5) than horizontally (!!mx * 4).
if (x < !!mx * 3 || y < !!my * 3 ||
arm: vp9: Add NEON optimizations of VP9 MC functions This work is sponsored by, and copyright, Google. The filter coefficients are signed values, where the product of the multiplication with one individual filter coefficient doesn't overflow a 16 bit signed value (the largest filter coefficient is 127). But when the products are accumulated, the resulting sum can overflow the 16 bit signed range. Instead of accumulating in 32 bit, we accumulate the largest product (either index 3 or 4) last with a saturated addition. (The VP8 MC asm does something similar, but slightly simpler, by accumulating each half of the filter separately. In the VP9 MC filters, each half of the filter can also overflow though, so the largest component has to be handled individually.) Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_avg4_neon: 1.71 1.15 1.42 1.49 vp9_avg8_neon: 2.51 3.63 3.14 2.58 vp9_avg16_neon: 2.95 6.76 3.01 2.84 vp9_avg32_neon: 3.29 6.64 2.85 3.00 vp9_avg64_neon: 3.47 6.67 3.14 2.80 vp9_avg_8tap_smooth_4h_neon: 3.22 4.73 2.76 4.67 vp9_avg_8tap_smooth_4hv_neon: 3.67 4.76 3.28 4.71 vp9_avg_8tap_smooth_4v_neon: 5.52 7.60 4.60 6.31 vp9_avg_8tap_smooth_8h_neon: 6.22 9.04 5.12 9.32 vp9_avg_8tap_smooth_8hv_neon: 6.38 8.21 5.72 8.17 vp9_avg_8tap_smooth_8v_neon: 9.22 12.66 8.15 11.10 vp9_avg_8tap_smooth_64h_neon: 7.02 10.23 5.54 11.58 vp9_avg_8tap_smooth_64hv_neon: 6.76 9.46 5.93 9.40 vp9_avg_8tap_smooth_64v_neon: 10.76 14.13 9.46 13.37 vp9_put4_neon: 1.11 1.47 1.00 1.21 vp9_put8_neon: 1.23 2.17 1.94 1.48 vp9_put16_neon: 1.63 4.02 1.73 1.97 vp9_put32_neon: 1.56 4.92 2.00 1.96 vp9_put64_neon: 2.10 5.28 2.03 2.35 vp9_put_8tap_smooth_4h_neon: 3.11 4.35 2.63 4.35 vp9_put_8tap_smooth_4hv_neon: 3.67 4.69 3.25 4.71 vp9_put_8tap_smooth_4v_neon: 5.45 7.27 4.49 6.52 vp9_put_8tap_smooth_8h_neon: 5.97 8.18 4.81 8.56 vp9_put_8tap_smooth_8hv_neon: 6.39 7.90 5.64 8.15 vp9_put_8tap_smooth_8v_neon: 9.03 11.84 8.07 11.51 vp9_put_8tap_smooth_64h_neon: 6.78 9.48 4.88 10.89 vp9_put_8tap_smooth_64hv_neon: 6.99 8.87 5.94 9.56 vp9_put_8tap_smooth_64v_neon: 10.69 13.30 9.43 14.34 For the larger 8tap filters, the speedup vs C code is around 5-14x. This is significantly faster than libvpx's implementation of the same functions, at least when comparing the put_8tap_smooth_64 functions (compared to vpx_convolve8_horiz_neon and vpx_convolve8_vert_neon from libvpx). Absolute runtimes from checkasm: Cortex A7 A8 A9 A53 vp9_put_8tap_smooth_64h_neon: 20150.3 14489.4 19733.6 10863.7 libvpx vpx_convolve8_horiz_neon: 52623.3 19736.4 21907.7 25027.7 vp9_put_8tap_smooth_64v_neon: 14455.0 12303.9 13746.4 9628.9 libvpx vpx_convolve8_vert_neon: 42090.0 17706.2 17659.9 16941.2 Thus, on the A9, the horizontal filter is only marginally faster than libvpx, while our version is significantly faster on the other cores, and the vertical filter is significantly faster on all cores. The difference is especially large on the A7. The libvpx implementation does the accumulation in 32 bit, which probably explains most of the differences. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-03 09:12:08 +02:00
x + !!mx * 4 > w - bw || y + !!my * 5 > h - bh) {
s->vdsp.emulated_edge_mc(s->edge_emu_buffer,
ref_u - !!my * 3 * src_stride_u - !!mx * 3,
80,
src_stride_u,
bw + !!mx * 7, bh + !!my * 7,
x - !!mx * 3, y - !!my * 3, w, h);
ref_u = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;
mc[!!mx][!!my](dst_u, dst_stride, ref_u, 80, bh, mx, my);
s->vdsp.emulated_edge_mc(s->edge_emu_buffer,
ref_v - !!my * 3 * src_stride_v - !!mx * 3,
80,
src_stride_v,
bw + !!mx * 7, bh + !!my * 7,
x - !!mx * 3, y - !!my * 3, w, h);
ref_v = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;
mc[!!mx][!!my](dst_v, dst_stride, ref_v, 80, bh, mx, my);
} else {
mc[!!mx][!!my](dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my);
mc[!!mx][!!my](dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my);
}
}
static int inter_recon(AVCodecContext *avctx)
{
static const uint8_t bwlog_tab[2][N_BS_SIZES] = {
{ 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4 },
{ 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 4, 4 },
};
VP9Context *s = avctx->priv_data;
VP9Block *b = s->b;
int row = b->row, col = b->col;
ThreadFrame *tref1 = &s->refs[s->refidx[b->ref[0]]];
ThreadFrame *tref2 = b->comp ? &s->refs[s->refidx[b->ref[1]]] : NULL;
AVFrame *ref1 = tref1->f;
AVFrame *ref2 = tref2 ? tref2->f : NULL;
int w = avctx->width, h = avctx->height;
ptrdiff_t ls_y = b->y_stride, ls_uv = b->uv_stride;
if (!ref1->data[0] || (b->comp && !ref2->data[0]))
return AVERROR_INVALIDDATA;
// y inter pred
if (b->bs > BS_8x8) {
if (b->bs == BS_8x4) {
mc_luma_dir(s, s->dsp.mc[3][b->filter][0], b->dst[0], ls_y,
ref1->data[0], ref1->linesize[0], tref1,
row << 3, col << 3, &b->mv[0][0], 8, 4, w, h);
mc_luma_dir(s, s->dsp.mc[3][b->filter][0],
b->dst[0] + 4 * ls_y, ls_y,
ref1->data[0], ref1->linesize[0], tref1,
(row << 3) + 4, col << 3, &b->mv[2][0], 8, 4, w, h);
if (b->comp) {
mc_luma_dir(s, s->dsp.mc[3][b->filter][1], b->dst[0], ls_y,
ref2->data[0], ref2->linesize[0], tref2,
row << 3, col << 3, &b->mv[0][1], 8, 4, w, h);
mc_luma_dir(s, s->dsp.mc[3][b->filter][1],
b->dst[0] + 4 * ls_y, ls_y,
ref2->data[0], ref2->linesize[0], tref2,
(row << 3) + 4, col << 3, &b->mv[2][1], 8, 4, w, h);
}
} else if (b->bs == BS_4x8) {
mc_luma_dir(s, s->dsp.mc[4][b->filter][0], b->dst[0], ls_y,
ref1->data[0], ref1->linesize[0], tref1,
row << 3, col << 3, &b->mv[0][0], 4, 8, w, h);
mc_luma_dir(s, s->dsp.mc[4][b->filter][0], b->dst[0] + 4, ls_y,
ref1->data[0], ref1->linesize[0], tref1,
row << 3, (col << 3) + 4, &b->mv[1][0], 4, 8, w, h);
if (b->comp) {
mc_luma_dir(s, s->dsp.mc[4][b->filter][1], b->dst[0], ls_y,
ref2->data[0], ref2->linesize[0], tref2,
row << 3, col << 3, &b->mv[0][1], 4, 8, w, h);
mc_luma_dir(s, s->dsp.mc[4][b->filter][1], b->dst[0] + 4, ls_y,
ref2->data[0], ref2->linesize[0], tref2,
row << 3, (col << 3) + 4, &b->mv[1][1], 4, 8, w, h);
}
} else {
av_assert2(b->bs == BS_4x4);
// FIXME if two horizontally adjacent blocks have the same MV,
// do a w8 instead of a w4 call
mc_luma_dir(s, s->dsp.mc[4][b->filter][0], b->dst[0], ls_y,
ref1->data[0], ref1->linesize[0], tref1,
row << 3, col << 3, &b->mv[0][0], 4, 4, w, h);
mc_luma_dir(s, s->dsp.mc[4][b->filter][0], b->dst[0] + 4, ls_y,
ref1->data[0], ref1->linesize[0], tref1,
row << 3, (col << 3) + 4, &b->mv[1][0], 4, 4, w, h);
mc_luma_dir(s, s->dsp.mc[4][b->filter][0],
b->dst[0] + 4 * ls_y, ls_y,
ref1->data[0], ref1->linesize[0], tref1,
(row << 3) + 4, col << 3, &b->mv[2][0], 4, 4, w, h);
mc_luma_dir(s, s->dsp.mc[4][b->filter][0],
b->dst[0] + 4 * ls_y + 4, ls_y,
ref1->data[0], ref1->linesize[0], tref1,
(row << 3) + 4, (col << 3) + 4, &b->mv[3][0], 4, 4, w, h);
if (b->comp) {
mc_luma_dir(s, s->dsp.mc[4][b->filter][1], b->dst[0], ls_y,
ref2->data[0], ref2->linesize[0], tref2,
row << 3, col << 3, &b->mv[0][1], 4, 4, w, h);
mc_luma_dir(s, s->dsp.mc[4][b->filter][1], b->dst[0] + 4, ls_y,
ref2->data[0], ref2->linesize[0], tref2,
row << 3, (col << 3) + 4, &b->mv[1][1], 4, 4, w, h);
mc_luma_dir(s, s->dsp.mc[4][b->filter][1],
b->dst[0] + 4 * ls_y, ls_y,
ref2->data[0], ref2->linesize[0], tref2,
(row << 3) + 4, col << 3, &b->mv[2][1], 4, 4, w, h);
mc_luma_dir(s, s->dsp.mc[4][b->filter][1],
b->dst[0] + 4 * ls_y + 4, ls_y,
ref2->data[0], ref2->linesize[0], tref2,
(row << 3) + 4, (col << 3) + 4, &b->mv[3][1], 4, 4, w, h);
}
}
} else {
int bwl = bwlog_tab[0][b->bs];
int bw = bwh_tab[0][b->bs][0] * 4;
int bh = bwh_tab[0][b->bs][1] * 4;
mc_luma_dir(s, s->dsp.mc[bwl][b->filter][0], b->dst[0], ls_y,
ref1->data[0], ref1->linesize[0], tref1,
row << 3, col << 3, &b->mv[0][0], bw, bh, w, h);
if (b->comp)
mc_luma_dir(s, s->dsp.mc[bwl][b->filter][1], b->dst[0], ls_y,
ref2->data[0], ref2->linesize[0], tref2,
row << 3, col << 3, &b->mv[0][1], bw, bh, w, h);
}
// uv inter pred
{
int bwl = bwlog_tab[1][b->bs];
int bw = bwh_tab[1][b->bs][0] * 4, bh = bwh_tab[1][b->bs][1] * 4;
VP56mv mvuv;
w = (w + 1) >> 1;
h = (h + 1) >> 1;
if (b->bs > BS_8x8) {
mvuv.x = ROUNDED_DIV(b->mv[0][0].x + b->mv[1][0].x +
b->mv[2][0].x + b->mv[3][0].x, 4);
mvuv.y = ROUNDED_DIV(b->mv[0][0].y + b->mv[1][0].y +
b->mv[2][0].y + b->mv[3][0].y, 4);
} else {
mvuv = b->mv[0][0];
}
mc_chroma_dir(s, s->dsp.mc[bwl][b->filter][0],
b->dst[1], b->dst[2], ls_uv,
ref1->data[1], ref1->linesize[1],
ref1->data[2], ref1->linesize[2], tref1,
row << 2, col << 2, &mvuv, bw, bh, w, h);
if (b->comp) {
if (b->bs > BS_8x8) {
mvuv.x = ROUNDED_DIV(b->mv[0][1].x + b->mv[1][1].x +
b->mv[2][1].x + b->mv[3][1].x, 4);
mvuv.y = ROUNDED_DIV(b->mv[0][1].y + b->mv[1][1].y +
b->mv[2][1].y + b->mv[3][1].y, 4);
} else {
mvuv = b->mv[0][1];
}
mc_chroma_dir(s, s->dsp.mc[bwl][b->filter][1],
b->dst[1], b->dst[2], ls_uv,
ref2->data[1], ref2->linesize[1],
ref2->data[2], ref2->linesize[2], tref2,
row << 2, col << 2, &mvuv, bw, bh, w, h);
}
}
if (!b->skip) {
/* mostly copied intra_reconn() */
int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;
int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);
int end_x = FFMIN(2 * (s->cols - col), w4);
int end_y = FFMIN(2 * (s->rows - row), h4);
int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;
int uvstep1d = 1 << b->uvtx, p;
uint8_t *dst = b->dst[0];
// y itxfm add
for (n = 0, y = 0; y < end_y; y += step1d) {
uint8_t *ptr = dst;
for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d, n += step) {
int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];
if (eob)
s->dsp.itxfm_add[tx][DCT_DCT](ptr, b->y_stride,
s->block + 16 * n, eob);
}
dst += 4 * b->y_stride * step1d;
}
// uv itxfm add
h4 >>= 1;
w4 >>= 1;
end_x >>= 1;
end_y >>= 1;
step = 1 << (b->uvtx * 2);
for (p = 0; p < 2; p++) {
dst = b->dst[p + 1];
for (n = 0, y = 0; y < end_y; y += uvstep1d) {
uint8_t *ptr = dst;
for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d, n += step) {
int eob = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n])
: s->uveob[p][n];
if (eob)
s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, b->uv_stride,
s->uvblock[p] + 16 * n, eob);
}
dst += 4 * uvstep1d * b->uv_stride;
}
}
}
return 0;
}
static av_always_inline void mask_edges(VP9Filter *lflvl, int is_uv,
int row_and_7, int col_and_7,
int w, int h, int col_end, int row_end,
enum TxfmMode tx, int skip_inter)
{
// FIXME I'm pretty sure all loops can be replaced by a single LUT if
// we make VP9Filter.mask uint64_t (i.e. row/col all single variable)
// and make the LUT 5-indexed (bl, bp, is_uv, tx and row/col), and then
// use row_and_7/col_and_7 as shifts (1*col_and_7+8*row_and_7)
// the intended behaviour of the vp9 loopfilter is to work on 8-pixel
// edges. This means that for UV, we work on two subsampled blocks at
// a time, and we only use the topleft block's mode information to set
// things like block strength. Thus, for any block size smaller than
// 16x16, ignore the odd portion of the block.
if (tx == TX_4X4 && is_uv) {
if (h == 1) {
if (row_and_7 & 1)
return;
if (!row_end)
h += 1;
}
if (w == 1) {
if (col_and_7 & 1)
return;
if (!col_end)
w += 1;
}
}
if (tx == TX_4X4 && !skip_inter) {
int t = 1 << col_and_7, m_col = (t << w) - t, y;
int m_col_odd = (t << (w - 1)) - t;
// on 32-px edges, use the 8-px wide loopfilter; else, use 4-px wide
if (is_uv) {
int m_row_8 = m_col & 0x01, m_row_4 = m_col - m_row_8;
for (y = row_and_7; y < h + row_and_7; y++) {
int col_mask_id = 2 - !(y & 7);
lflvl->mask[is_uv][0][y][1] |= m_row_8;
lflvl->mask[is_uv][0][y][2] |= m_row_4;
// for odd lines, if the odd col is not being filtered,
// skip odd row also:
// .---. <-- a
// | |
// |___| <-- b
// ^ ^
// c d
//
// if a/c are even row/col and b/d are odd, and d is skipped,
// e.g. right edge of size-66x66.webm, then skip b also (bug)
if ((col_end & 1) && (y & 1)) {
lflvl->mask[is_uv][1][y][col_mask_id] |= m_col_odd;
} else {
lflvl->mask[is_uv][1][y][col_mask_id] |= m_col;
}
}
} else {
int m_row_8 = m_col & 0x11, m_row_4 = m_col - m_row_8;
for (y = row_and_7; y < h + row_and_7; y++) {
int col_mask_id = 2 - !(y & 3);
lflvl->mask[is_uv][0][y][1] |= m_row_8; // row edge
lflvl->mask[is_uv][0][y][2] |= m_row_4;
lflvl->mask[is_uv][1][y][col_mask_id] |= m_col; // col edge
lflvl->mask[is_uv][0][y][3] |= m_col;
lflvl->mask[is_uv][1][y][3] |= m_col;
}
}
} else {
int y, t = 1 << col_and_7, m_col = (t << w) - t;
if (!skip_inter) {
int mask_id = (tx == TX_8X8);
int l2 = tx + is_uv - 1, step1d = 1 << l2;
static const unsigned masks[4] = { 0xff, 0x55, 0x11, 0x01 };
int m_row = m_col & masks[l2];
// at odd UV col/row edges tx16/tx32 loopfilter edges, force
// 8wd loopfilter to prevent going off the visible edge.
if (is_uv && tx > TX_8X8 && (w ^ (w - 1)) == 1) {
int m_row_16 = ((t << (w - 1)) - t) & masks[l2];
int m_row_8 = m_row - m_row_16;
for (y = row_and_7; y < h + row_and_7; y++) {
lflvl->mask[is_uv][0][y][0] |= m_row_16;
lflvl->mask[is_uv][0][y][1] |= m_row_8;
}
} else {
for (y = row_and_7; y < h + row_and_7; y++)
lflvl->mask[is_uv][0][y][mask_id] |= m_row;
}
if (is_uv && tx > TX_8X8 && (h ^ (h - 1)) == 1) {
for (y = row_and_7; y < h + row_and_7 - 1; y += step1d)
lflvl->mask[is_uv][1][y][0] |= m_col;
if (y - row_and_7 == h - 1)
lflvl->mask[is_uv][1][y][1] |= m_col;
} else {
for (y = row_and_7; y < h + row_and_7; y += step1d)
lflvl->mask[is_uv][1][y][mask_id] |= m_col;
}
} else if (tx != TX_4X4) {
int mask_id;
mask_id = (tx == TX_8X8) || (is_uv && h == 1);
lflvl->mask[is_uv][1][row_and_7][mask_id] |= m_col;
mask_id = (tx == TX_8X8) || (is_uv && w == 1);
for (y = row_and_7; y < h + row_and_7; y++)
lflvl->mask[is_uv][0][y][mask_id] |= t;
} else if (is_uv) {
int t8 = t & 0x01, t4 = t - t8;
for (y = row_and_7; y < h + row_and_7; y++) {
lflvl->mask[is_uv][0][y][2] |= t4;
lflvl->mask[is_uv][0][y][1] |= t8;
}
lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 7)] |= m_col;
} else {
int t8 = t & 0x11, t4 = t - t8;
for (y = row_and_7; y < h + row_and_7; y++) {
lflvl->mask[is_uv][0][y][2] |= t4;
lflvl->mask[is_uv][0][y][1] |= t8;
}
lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 3)] |= m_col;
}
}
}
int ff_vp9_decode_block(AVCodecContext *avctx, int row, int col,
VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,
enum BlockLevel bl, enum BlockPartition bp)
{
VP9Context *s = avctx->priv_data;
VP9Block *b = s->b;
AVFrame *f = s->frames[CUR_FRAME].tf.f;
enum BlockSize bs = bl * 3 + bp;
int ret, y, w4 = bwh_tab[1][bs][0], h4 = bwh_tab[1][bs][1], lvl;
int emu[2];
b->row = row;
b->row7 = row & 7;
b->col = col;
b->col7 = col & 7;
s->min_mv.x = -(128 + col * 64);
s->min_mv.y = -(128 + row * 64);
s->max_mv.x = 128 + (s->cols - col - w4) * 64;
s->max_mv.y = 128 + (s->rows - row - h4) * 64;
if (s->pass < 2) {
b->bs = bs;
b->bl = bl;
b->bp = bp;
decode_mode(s, b);
b->uvtx = b->tx - (w4 * 2 == (1 << b->tx) || h4 * 2 == (1 << b->tx));
if (!b->skip) {
if ((ret = decode_coeffs(avctx)) < 0)
return ret;
} else {
int pl;
memset(&s->above_y_nnz_ctx[col * 2], 0, w4 * 2);
memset(&s->left_y_nnz_ctx[(row & 7) << 1], 0, h4 * 2);
for (pl = 0; pl < 2; pl++) {
memset(&s->above_uv_nnz_ctx[pl][col], 0, w4);
memset(&s->left_uv_nnz_ctx[pl][row & 7], 0, h4);
}
}
if (s->pass == 1) {
s->b++;
s->block += w4 * h4 * 64;
s->uvblock[0] += w4 * h4 * 16;
s->uvblock[1] += w4 * h4 * 16;
s->eob += w4 * h4 * 4;
s->uveob[0] += w4 * h4;
s->uveob[1] += w4 * h4;
return 0;
}
}
/* Emulated overhangs if the stride of the target buffer can't hold.
* This allows to support emu-edge and so on even if we have large
* block overhangs. */
emu[0] = (col + w4) * 8 > f->linesize[0] ||
(row + h4) > s->rows;
emu[1] = (col + w4) * 4 > f->linesize[1] ||
(row + h4) > s->rows;
if (emu[0]) {
b->dst[0] = s->tmp_y;
b->y_stride = 64;
} else {
b->dst[0] = f->data[0] + yoff;
b->y_stride = f->linesize[0];
}
if (emu[1]) {
b->dst[1] = s->tmp_uv[0];
b->dst[2] = s->tmp_uv[1];
b->uv_stride = 32;
} else {
b->dst[1] = f->data[1] + uvoff;
b->dst[2] = f->data[2] + uvoff;
b->uv_stride = f->linesize[1];
}
if (b->intra) {
intra_recon(avctx, yoff, uvoff);
} else {
if ((ret = inter_recon(avctx)) < 0)
return ret;
}
if (emu[0]) {
int w = FFMIN(s->cols - col, w4) * 8;
int h = FFMIN(s->rows - row, h4) * 8;
int n, o = 0;
for (n = 0; o < w; n++) {
int bw = 64 >> n;
av_assert2(n <= 4);
if (w & bw) {
s->dsp.mc[n][0][0][0][0](f->data[0] + yoff + o,
f->linesize[0],
s->tmp_y + o,
64, h, 0, 0);
o += bw;
}
}
}
if (emu[1]) {
int w = FFMIN(s->cols - col, w4) * 4;
int h = FFMIN(s->rows - row, h4) * 4;
int n, o = 0;
for (n = 1; o < w; n++) {
int bw = 64 >> n;
av_assert2(n <= 4);
if (w & bw) {
s->dsp.mc[n][0][0][0][0](f->data[1] + uvoff + o,
f->linesize[1],
s->tmp_uv[0] + o,
32, h, 0, 0);
s->dsp.mc[n][0][0][0][0](f->data[2] + uvoff + o,
f->linesize[2],
s->tmp_uv[1] + o,
32, h, 0, 0);
o += bw;
}
}
}
// pick filter level and find edges to apply filter to
if (s->filter.level &&
(lvl = s->segmentation.feat[b->seg_id].lflvl[b->intra ? 0 : b->ref[0] + 1]
[b->mode[3] != ZEROMV]) > 0) {
int x_end = FFMIN(s->cols - col, w4);
int y_end = FFMIN(s->rows - row, h4);
int skip_inter = !b->intra && b->skip;
for (y = 0; y < h4; y++)
memset(&lflvl->level[((row & 7) + y) * 8 + (col & 7)], lvl, w4);
mask_edges(lflvl, 0, row & 7, col & 7, x_end, y_end, 0, 0, b->tx, skip_inter);
mask_edges(lflvl, 1, row & 7, col & 7, x_end, y_end,
s->cols & 1 && col + w4 >= s->cols ? s->cols & 7 : 0,
s->rows & 1 && row + h4 >= s->rows ? s->rows & 7 : 0,
b->uvtx, skip_inter);
if (!s->filter.lim_lut[lvl]) {
int sharp = s->filter.sharpness;
int limit = lvl;
if (sharp > 0) {
limit >>= (sharp + 3) >> 2;
limit = FFMIN(limit, 9 - sharp);
}
limit = FFMAX(limit, 1);
s->filter.lim_lut[lvl] = limit;
s->filter.mblim_lut[lvl] = 2 * (lvl + 2) + limit;
}
}
if (s->pass == 2) {
s->b++;
s->block += w4 * h4 * 64;
s->uvblock[0] += w4 * h4 * 16;
s->uvblock[1] += w4 * h4 * 16;
s->eob += w4 * h4 * 4;
s->uveob[0] += w4 * h4;
s->uveob[1] += w4 * h4;
}
return 0;
}