blob: a4f84f9145b2297f59f497cebe6a6cc9ac0b124d [file] [log] [blame]
/*
* Copyright © 2018, VideoLAN and dav1d authors
* Copyright © 2018, Two Orioles, LLC
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* 1. Redistributions of source code must retain the above copyright notice, this
* list of conditions and the following disclaimer.
*
* 2. Redistributions in binary form must reproduce the above copyright notice,
* this list of conditions and the following disclaimer in the documentation
* and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
* WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
* ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
* LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
* ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
* SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#include "config.h"
#include <errno.h>
#include <limits.h>
#include <string.h>
#include <stdio.h>
#include <inttypes.h>
#include "dav1d/data.h"
#include "common/intops.h"
#include "common/mem.h"
#include "src/ctx.h"
#include "src/decode.h"
#include "src/dequant_tables.h"
#include "src/env.h"
#include "src/film_grain.h"
#include "src/log.h"
#include "src/qm.h"
#include "src/recon.h"
#include "src/ref.h"
#include "src/tables.h"
#include "src/thread_task.h"
#include "src/warpmv.h"
#include "common/starboard_memory.h"
static void init_quant_tables(const Dav1dSequenceHeader *const seq_hdr,
const Dav1dFrameHeader *const frame_hdr,
const int qidx, uint16_t (*dq)[3][2])
{
for (int i = 0; i < (frame_hdr->segmentation.enabled ? 8 : 1); i++) {
const int yac = frame_hdr->segmentation.enabled ?
iclip_u8(qidx + frame_hdr->segmentation.seg_data.d[i].delta_q) : qidx;
const int ydc = iclip_u8(yac + frame_hdr->quant.ydc_delta);
const int uac = iclip_u8(yac + frame_hdr->quant.uac_delta);
const int udc = iclip_u8(yac + frame_hdr->quant.udc_delta);
const int vac = iclip_u8(yac + frame_hdr->quant.vac_delta);
const int vdc = iclip_u8(yac + frame_hdr->quant.vdc_delta);
dq[i][0][0] = dav1d_dq_tbl[seq_hdr->hbd][ydc][0];
dq[i][0][1] = dav1d_dq_tbl[seq_hdr->hbd][yac][1];
dq[i][1][0] = dav1d_dq_tbl[seq_hdr->hbd][udc][0];
dq[i][1][1] = dav1d_dq_tbl[seq_hdr->hbd][uac][1];
dq[i][2][0] = dav1d_dq_tbl[seq_hdr->hbd][vdc][0];
dq[i][2][1] = dav1d_dq_tbl[seq_hdr->hbd][vac][1];
}
}
static int read_mv_component_diff(Dav1dTileContext *const t,
CdfMvComponent *const mv_comp,
const int have_fp)
{
Dav1dTileState *const ts = t->ts;
const Dav1dFrameContext *const f = t->f;
const int have_hp = f->frame_hdr->hp;
const int sign = dav1d_msac_decode_bool_adapt(&ts->msac, mv_comp->sign);
const int cl = dav1d_msac_decode_symbol_adapt16(&ts->msac,
mv_comp->classes, 10);
int up, fp, hp;
if (!cl) {
up = dav1d_msac_decode_bool_adapt(&ts->msac, mv_comp->class0);
if (have_fp) {
fp = dav1d_msac_decode_symbol_adapt4(&ts->msac,
mv_comp->class0_fp[up], 3);
hp = have_hp ? dav1d_msac_decode_bool_adapt(&ts->msac,
mv_comp->class0_hp) : 1;
} else {
fp = 3;
hp = 1;
}
} else {
up = 1 << cl;
for (int n = 0; n < cl; n++)
up |= dav1d_msac_decode_bool_adapt(&ts->msac,
mv_comp->classN[n]) << n;
if (have_fp) {
fp = dav1d_msac_decode_symbol_adapt4(&ts->msac,
mv_comp->classN_fp, 3);
hp = have_hp ? dav1d_msac_decode_bool_adapt(&ts->msac,
mv_comp->classN_hp) : 1;
} else {
fp = 3;
hp = 1;
}
}
const int diff = ((up << 3) | (fp << 1) | hp) + 1;
return sign ? -diff : diff;
}
static void read_mv_residual(Dav1dTileContext *const t, mv *const ref_mv,
CdfMvContext *const mv_cdf, const int have_fp)
{
switch (dav1d_msac_decode_symbol_adapt4(&t->ts->msac, t->ts->cdf.mv.joint,
N_MV_JOINTS - 1))
{
case MV_JOINT_HV:
ref_mv->y += read_mv_component_diff(t, &mv_cdf->comp[0], have_fp);
ref_mv->x += read_mv_component_diff(t, &mv_cdf->comp[1], have_fp);
break;
case MV_JOINT_H:
ref_mv->x += read_mv_component_diff(t, &mv_cdf->comp[1], have_fp);
break;
case MV_JOINT_V:
ref_mv->y += read_mv_component_diff(t, &mv_cdf->comp[0], have_fp);
break;
default:
break;
}
}
static void read_tx_tree(Dav1dTileContext *const t,
const enum RectTxfmSize from,
const int depth, uint16_t *const masks,
const int x_off, const int y_off)
{
const Dav1dFrameContext *const f = t->f;
const int bx4 = t->bx & 31, by4 = t->by & 31;
const TxfmInfo *const t_dim = &dav1d_txfm_dimensions[from];
const int txw = t_dim->lw, txh = t_dim->lh;
int is_split;
if (depth < 2 && from > (int) TX_4X4) {
const int cat = 2 * (TX_64X64 - t_dim->max) - depth;
const int a = t->a->tx[bx4] < txw;
const int l = t->l.tx[by4] < txh;
is_split = dav1d_msac_decode_bool_adapt(&t->ts->msac,
t->ts->cdf.m.txpart[cat][a + l]);
if (is_split)
masks[depth] |= 1 << (y_off * 4 + x_off);
} else {
is_split = 0;
}
if (is_split && t_dim->max > TX_8X8) {
const enum RectTxfmSize sub = t_dim->sub;
const TxfmInfo *const sub_t_dim = &dav1d_txfm_dimensions[sub];
const int txsw = sub_t_dim->w, txsh = sub_t_dim->h;
read_tx_tree(t, sub, depth + 1, masks, x_off * 2 + 0, y_off * 2 + 0);
t->bx += txsw;
if (txw >= txh && t->bx < f->bw)
read_tx_tree(t, sub, depth + 1, masks, x_off * 2 + 1, y_off * 2 + 0);
t->bx -= txsw;
t->by += txsh;
if (txh >= txw && t->by < f->bh) {
read_tx_tree(t, sub, depth + 1, masks, x_off * 2 + 0, y_off * 2 + 1);
t->bx += txsw;
if (txw >= txh && t->bx < f->bw)
read_tx_tree(t, sub, depth + 1, masks,
x_off * 2 + 1, y_off * 2 + 1);
t->bx -= txsw;
}
t->by -= txsh;
} else {
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
rep_macro(type, t->dir tx, off, is_split ? TX_4X4 : mul * txh)
case_set_upto16(t_dim->h, l., 1, by4);
#undef set_ctx
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
rep_macro(type, t->dir tx, off, is_split ? TX_4X4 : mul * txw)
case_set_upto16(t_dim->w, a->, 0, bx4);
#undef set_ctx
}
}
static int neg_deinterleave(int diff, int ref, int max) {
if (!ref) return diff;
if (ref >= (max - 1)) return max - diff - 1;
if (2 * ref < max) {
if (diff <= 2 * ref) {
if (diff & 1)
return ref + ((diff + 1) >> 1);
else
return ref - (diff >> 1);
}
return diff;
} else {
if (diff <= 2 * (max - ref - 1)) {
if (diff & 1)
return ref + ((diff + 1) >> 1);
else
return ref - (diff >> 1);
}
return max - (diff + 1);
}
}
static void find_matching_ref(const Dav1dTileContext *const t,
const enum EdgeFlags intra_edge_flags,
const int bw4, const int bh4,
const int w4, const int h4,
const int have_left, const int have_top,
const int ref, uint64_t masks[2])
{
const Dav1dFrameContext *const f = t->f;
const ptrdiff_t b4_stride = f->b4_stride;
const refmvs *const r = &f->mvs[t->by * b4_stride + t->bx];
int count = 0;
int have_topleft = have_top && have_left;
int have_topright = imax(bw4, bh4) < 32 &&
have_top && t->bx + bw4 < t->ts->tiling.col_end &&
(intra_edge_flags & EDGE_I444_TOP_HAS_RIGHT);
#define bs(rp) dav1d_block_dimensions[dav1d_sbtype_to_bs[(rp)->sb_type]]
#define matches(rp) ((rp)->ref[0] == ref + 1 && (rp)->ref[1] == -1)
if (have_top) {
const refmvs *r2 = &r[-b4_stride];
if (matches(r2)) {
masks[0] |= 1;
count = 1;
}
int aw4 = bs(r2)[0];
if (aw4 >= bw4) {
const int off = t->bx & (aw4 - 1);
if (off) have_topleft = 0;
if (aw4 - off > bw4) have_topright = 0;
} else {
unsigned mask = 1 << aw4;
for (int x = aw4; x < w4; x += aw4) {
r2 += aw4;
if (matches(r2)) {
masks[0] |= mask;
if (++count >= 8) return;
}
aw4 = bs(r2)[0];
mask <<= aw4;
}
}
}
if (have_left) {
const refmvs *r2 = &r[-1];
if (matches(r2)) {
masks[1] |= 1;
if (++count >= 8) return;
}
int lh4 = bs(r2)[1];
if (lh4 >= bh4) {
if (t->by & (lh4 - 1)) have_topleft = 0;
} else {
unsigned mask = 1 << lh4;
for (int y = lh4; y < h4; y += lh4) {
r2 += lh4 * b4_stride;
if (matches(r2)) {
masks[1] |= mask;
if (++count >= 8) return;
}
lh4 = bs(r2)[1];
mask <<= lh4;
}
}
}
if (have_topleft && matches(&r[-(1 + b4_stride)])) {
masks[1] |= 1ULL << 32;
if (++count >= 8) return;
}
if (have_topright && matches(&r[bw4 - b4_stride])) {
masks[0] |= 1ULL << 32;
}
#undef matches
}
static void derive_warpmv(const Dav1dTileContext *const t,
const int bw4, const int bh4,
const uint64_t masks[2], const struct mv mv,
Dav1dWarpedMotionParams *const wmp)
{
int pts[8][2 /* in, out */][2 /* x, y */], np = 0;
const Dav1dFrameContext *const f = t->f;
const ptrdiff_t b4_stride = f->b4_stride;
const refmvs *const r = &f->mvs[t->by * b4_stride + t->bx];
#define add_sample(dx, dy, sx, sy, rp) do { \
pts[np][0][0] = 16 * (2 * dx + sx * bs(rp)[0]) - 8; \
pts[np][0][1] = 16 * (2 * dy + sy * bs(rp)[1]) - 8; \
pts[np][1][0] = pts[np][0][0] + (rp)->mv[0].x; \
pts[np][1][1] = pts[np][0][1] + (rp)->mv[0].y; \
np++; \
} while (0)
// use masks[] to find the projectable motion vectors in the edges
if ((unsigned) masks[0] == 1 && !(masks[1] >> 32)) {
const int off = t->bx & (bs(&r[-b4_stride])[0] - 1);
add_sample(-off, 0, 1, -1, &r[-b4_stride]);
} else for (unsigned off = 0, xmask = (uint32_t) masks[0]; np < 8 && xmask;) { // top
const int tz = ctz(xmask);
off += tz;
xmask >>= tz;
add_sample(off, 0, 1, -1, &r[off - b4_stride]);
xmask &= ~1;
}
if (np < 8 && masks[1] == 1) {
const int off = t->by & (bs(&r[-1])[1] - 1);
add_sample(0, -off, -1, 1, &r[-1 - off * b4_stride]);
} else for (unsigned off = 0, ymask = (uint32_t) masks[1]; np < 8 && ymask;) { // left
const int tz = ctz(ymask);
off += tz;
ymask >>= tz;
add_sample(0, off, -1, 1, &r[off * b4_stride - 1]);
ymask &= ~1;
}
if (np < 8 && masks[1] >> 32) // top/left
add_sample(0, 0, -1, -1, &r[-(1 + b4_stride)]);
if (np < 8 && masks[0] >> 32) // top/right
add_sample(bw4, 0, 1, -1, &r[bw4 - b4_stride]);
assert(np > 0 && np <= 8);
#undef bs
// select according to motion vector difference against a threshold
int mvd[8], ret = 0;
const int thresh = 4 * iclip(imax(bw4, bh4), 4, 28);
for (int i = 0; i < np; i++) {
mvd[i] = abs(pts[i][1][0] - pts[i][0][0] - mv.x) +
abs(pts[i][1][1] - pts[i][0][1] - mv.y);
if (mvd[i] > thresh)
mvd[i] = -1;
else
ret++;
}
if (!ret) {
ret = 1;
} else for (int i = 0, j = np - 1, k = 0; k < np - ret; k++, i++, j--) {
while (mvd[i] != -1) i++;
while (mvd[j] == -1) j--;
assert(i != j);
if (i > j) break;
// replace the discarded samples;
mvd[i] = mvd[j];
memcpy(pts[i], pts[j], sizeof(*pts));
}
if (!dav1d_find_affine_int(pts, ret, bw4, bh4, mv, wmp, t->bx, t->by) &&
!dav1d_get_shear_params(wmp))
{
wmp->type = DAV1D_WM_TYPE_AFFINE;
} else
wmp->type = DAV1D_WM_TYPE_IDENTITY;
}
static inline int findoddzero(const uint8_t *buf, int len) {
for (int n = 0; n < len; n++)
if (!buf[n * 2]) return 1;
return 0;
}
static void read_pal_plane(Dav1dTileContext *const t, Av1Block *const b,
const int pl, const int sz_ctx,
const int bx4, const int by4)
{
Dav1dTileState *const ts = t->ts;
const Dav1dFrameContext *const f = t->f;
const int pal_sz = b->pal_sz[pl] = dav1d_msac_decode_symbol_adapt8(&ts->msac,
ts->cdf.m.pal_sz[pl][sz_ctx], 6) + 2;
uint16_t cache[16], used_cache[8];
int l_cache = pl ? t->pal_sz_uv[1][by4] : t->l.pal_sz[by4];
int n_cache = 0;
// don't reuse above palette outside SB64 boundaries
int a_cache = by4 & 15 ? pl ? t->pal_sz_uv[0][bx4] : t->a->pal_sz[bx4] : 0;
const uint16_t *l = t->al_pal[1][by4][pl], *a = t->al_pal[0][bx4][pl];
// fill/sort cache
while (l_cache && a_cache) {
if (*l < *a) {
if (!n_cache || cache[n_cache - 1] != *l)
cache[n_cache++] = *l;
l++;
l_cache--;
} else {
if (*a == *l) {
l++;
l_cache--;
}
if (!n_cache || cache[n_cache - 1] != *a)
cache[n_cache++] = *a;
a++;
a_cache--;
}
}
if (l_cache) {
do {
if (!n_cache || cache[n_cache - 1] != *l)
cache[n_cache++] = *l;
l++;
} while (--l_cache > 0);
} else if (a_cache) {
do {
if (!n_cache || cache[n_cache - 1] != *a)
cache[n_cache++] = *a;
a++;
} while (--a_cache > 0);
}
// find reused cache entries
int i = 0;
for (int n = 0; n < n_cache && i < pal_sz; n++)
if (dav1d_msac_decode_bool_equi(&ts->msac))
used_cache[i++] = cache[n];
const int n_used_cache = i;
// parse new entries
uint16_t *const pal = f->frame_thread.pass ?
f->frame_thread.pal[((t->by >> 1) + (t->bx & 1)) * (f->b4_stride >> 1) +
((t->bx >> 1) + (t->by & 1))][pl] : t->scratch.pal[pl];
if (i < pal_sz) {
int prev = pal[i++] = dav1d_msac_decode_bools(&ts->msac, f->cur.p.bpc);
if (i < pal_sz) {
int bits = f->cur.p.bpc - 3 + dav1d_msac_decode_bools(&ts->msac, 2);
const int max = (1 << f->cur.p.bpc) - 1;
do {
const int delta = dav1d_msac_decode_bools(&ts->msac, bits);
prev = pal[i++] = imin(prev + delta + !pl, max);
if (prev + !pl >= max) {
for (; i < pal_sz; i++)
pal[i] = max;
break;
}
bits = imin(bits, 1 + ulog2(max - prev - !pl));
} while (i < pal_sz);
}
// merge cache+new entries
int n = 0, m = n_used_cache;
for (i = 0; i < pal_sz; i++) {
if (n < n_used_cache && (m >= pal_sz || used_cache[n] <= pal[m])) {
pal[i] = used_cache[n++];
} else {
assert(m < pal_sz);
pal[i] = pal[m++];
}
}
} else {
memcpy(pal, used_cache, n_used_cache * sizeof(*used_cache));
}
if (DEBUG_BLOCK_INFO) {
printf("Post-pal[pl=%d,sz=%d,cache_size=%d,used_cache=%d]: r=%d, cache=",
pl, pal_sz, n_cache, n_used_cache, ts->msac.rng);
for (int n = 0; n < n_cache; n++)
printf("%c%02x", n ? ' ' : '[', cache[n]);
printf("%s, pal=", n_cache ? "]" : "[]");
for (int n = 0; n < pal_sz; n++)
printf("%c%02x", n ? ' ' : '[', pal[n]);
printf("]\n");
}
}
static void read_pal_uv(Dav1dTileContext *const t, Av1Block *const b,
const int sz_ctx, const int bx4, const int by4)
{
read_pal_plane(t, b, 1, sz_ctx, bx4, by4);
// V pal coding
Dav1dTileState *const ts = t->ts;
const Dav1dFrameContext *const f = t->f;
uint16_t *const pal = f->frame_thread.pass ?
f->frame_thread.pal[((t->by >> 1) + (t->bx & 1)) * (f->b4_stride >> 1) +
((t->bx >> 1) + (t->by & 1))][2] : t->scratch.pal[2];
if (dav1d_msac_decode_bool_equi(&ts->msac)) {
const int bits = f->cur.p.bpc - 4 +
dav1d_msac_decode_bools(&ts->msac, 2);
int prev = pal[0] = dav1d_msac_decode_bools(&ts->msac, f->cur.p.bpc);
const int max = (1 << f->cur.p.bpc) - 1;
for (int i = 1; i < b->pal_sz[1]; i++) {
int delta = dav1d_msac_decode_bools(&ts->msac, bits);
if (delta && dav1d_msac_decode_bool_equi(&ts->msac)) delta = -delta;
prev = pal[i] = (prev + delta) & max;
}
} else {
for (int i = 0; i < b->pal_sz[1]; i++)
pal[i] = dav1d_msac_decode_bools(&ts->msac, f->cur.p.bpc);
}
if (DEBUG_BLOCK_INFO) {
printf("Post-pal[pl=2]: r=%d ", ts->msac.rng);
for (int n = 0; n < b->pal_sz[1]; n++)
printf("%c%02x", n ? ' ' : '[', pal[n]);
printf("]\n");
}
}
// meant to be SIMD'able, so that theoretical complexity of this function
// times block size goes from w4*h4 to w4+h4-1
// a and b are previous two lines containing (a) top/left entries or (b)
// top/left entries, with a[0] being either the first top or first left entry,
// depending on top_offset being 1 or 0, and b being the first top/left entry
// for whichever has one. left_offset indicates whether the (len-1)th entry
// has a left neighbour.
// output is order[] and ctx for each member of this diagonal.
static void order_palette(const uint8_t *pal_idx, const ptrdiff_t stride,
const int i, const int first, const int last,
uint8_t (*const order)[8], uint8_t *const ctx)
{
int have_top = i > first;
assert(pal_idx);
pal_idx += first + (i - first) * stride;
for (int j = first, n = 0; j >= last; have_top = 1, j--, n++, pal_idx += stride - 1) {
const int have_left = j > 0;
assert(have_left || have_top);
#define add(v_in) do { \
const int v = v_in; \
assert((unsigned)v < 8U); \
order[n][o_idx++] = v; \
mask |= 1 << v; \
} while (0)
unsigned mask = 0;
int o_idx = 0;
if (!have_left) {
ctx[n] = 0;
add(pal_idx[-stride]);
} else if (!have_top) {
ctx[n] = 0;
add(pal_idx[-1]);
} else {
const int l = pal_idx[-1], t = pal_idx[-stride], tl = pal_idx[-(stride + 1)];
const int same_t_l = t == l;
const int same_t_tl = t == tl;
const int same_l_tl = l == tl;
const int same_all = same_t_l & same_t_tl & same_l_tl;
if (same_all) {
ctx[n] = 4;
add(t);
} else if (same_t_l) {
ctx[n] = 3;
add(t);
add(tl);
} else if (same_t_tl | same_l_tl) {
ctx[n] = 2;
add(tl);
add(same_t_tl ? l : t);
} else {
ctx[n] = 1;
add(imin(t, l));
add(imax(t, l));
add(tl);
}
}
for (unsigned m = 1, bit = 0; m < 0x100; m <<= 1, bit++)
if (!(mask & m))
order[n][o_idx++] = bit;
assert(o_idx == 8);
#undef add
}
}
static void read_pal_indices(Dav1dTileContext *const t,
uint8_t *const pal_idx,
const Av1Block *const b, const int pl,
const int w4, const int h4,
const int bw4, const int bh4)
{
Dav1dTileState *const ts = t->ts;
const ptrdiff_t stride = bw4 * 4;
assert(pal_idx);
pal_idx[0] = dav1d_msac_decode_uniform(&ts->msac, b->pal_sz[pl]);
uint16_t (*const color_map_cdf)[8] =
ts->cdf.m.color_map[pl][b->pal_sz[pl] - 2];
uint8_t (*const order)[8] = t->scratch.pal_order;
uint8_t *const ctx = t->scratch.pal_ctx;
for (int i = 1; i < 4 * (w4 + h4) - 1; i++) {
// top/left-to-bottom/right diagonals ("wave-front")
const int first = imin(i, w4 * 4 - 1);
const int last = imax(0, i - h4 * 4 + 1);
order_palette(pal_idx, stride, i, first, last, order, ctx);
for (int j = first, m = 0; j >= last; j--, m++) {
const int color_idx = dav1d_msac_decode_symbol_adapt8(&ts->msac,
color_map_cdf[ctx[m]], b->pal_sz[pl] - 1);
pal_idx[(i - j) * stride + j] = order[m][color_idx];
}
}
// fill invisible edges
if (bw4 > w4)
for (int y = 0; y < 4 * h4; y++)
memset(&pal_idx[y * stride + 4 * w4],
pal_idx[y * stride + 4 * w4 - 1], 4 * (bw4 - w4));
if (h4 < bh4) {
const uint8_t *const src = &pal_idx[stride * (4 * h4 - 1)];
for (int y = h4 * 4; y < bh4 * 4; y++)
memcpy(&pal_idx[y * stride], src, bw4 * 4);
}
}
static void read_vartx_tree(Dav1dTileContext *const t,
Av1Block *const b, const enum BlockSize bs,
const int bx4, const int by4)
{
const Dav1dFrameContext *const f = t->f;
const uint8_t *const b_dim = dav1d_block_dimensions[bs];
const int bw4 = b_dim[0], bh4 = b_dim[1];
// var-tx tree coding
b->tx_split[0] = b->tx_split[1] = 0;
b->max_ytx = dav1d_max_txfm_size_for_bs[bs][0];
if (f->frame_hdr->segmentation.lossless[b->seg_id] ||
b->max_ytx == TX_4X4)
{
b->max_ytx = b->uvtx = TX_4X4;
if (f->frame_hdr->txfm_mode == DAV1D_TX_SWITCHABLE) {
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
rep_macro(type, t->dir tx, off, TX_4X4)
case_set(bh4, l., 1, by4);
case_set(bw4, a->, 0, bx4);
#undef set_ctx
}
} else if (f->frame_hdr->txfm_mode != DAV1D_TX_SWITCHABLE || b->skip) {
if (f->frame_hdr->txfm_mode == DAV1D_TX_SWITCHABLE) {
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
rep_macro(type, t->dir tx, off, mul * b_dim[2 + diridx])
case_set(bh4, l., 1, by4);
case_set(bw4, a->, 0, bx4);
#undef set_ctx
} else {
assert(f->frame_hdr->txfm_mode == DAV1D_TX_LARGEST);
}
b->uvtx = dav1d_max_txfm_size_for_bs[bs][f->cur.p.layout];
} else {
assert(bw4 <= 16 || bh4 <= 16 || b->max_ytx == TX_64X64);
int y, x, y_off, x_off;
const TxfmInfo *const ytx = &dav1d_txfm_dimensions[b->max_ytx];
for (y = 0, y_off = 0; y < bh4; y += ytx->h, y_off++) {
for (x = 0, x_off = 0; x < bw4; x += ytx->w, x_off++) {
read_tx_tree(t, b->max_ytx, 0, b->tx_split, x_off, y_off);
// contexts are updated inside read_tx_tree()
t->bx += ytx->w;
}
t->bx -= x;
t->by += ytx->h;
}
t->by -= y;
if (DEBUG_BLOCK_INFO)
printf("Post-vartxtree[%x/%x]: r=%d\n",
b->tx_split[0], b->tx_split[1], t->ts->msac.rng);
b->uvtx = dav1d_max_txfm_size_for_bs[bs][f->cur.p.layout];
}
}
static inline unsigned get_prev_frame_segid(const Dav1dFrameContext *const f,
const int by, const int bx,
const int w4, int h4,
const uint8_t *ref_seg_map,
const ptrdiff_t stride)
{
assert(f->frame_hdr->primary_ref_frame != DAV1D_PRIMARY_REF_NONE);
if (dav1d_thread_picture_wait(&f->refp[f->frame_hdr->primary_ref_frame],
(by + h4) * 4, PLANE_TYPE_BLOCK))
{
return 8;
}
unsigned seg_id = 8;
ref_seg_map += by * stride + bx;
do {
for (int x = 0; x < w4; x++)
seg_id = imin(seg_id, ref_seg_map[x]);
ref_seg_map += stride;
} while (--h4 > 0 && seg_id);
assert(seg_id < 8);
return seg_id;
}
static int decode_b(Dav1dTileContext *const t,
const enum BlockLevel bl,
const enum BlockSize bs,
const enum BlockPartition bp,
const enum EdgeFlags intra_edge_flags)
{
Dav1dTileState *const ts = t->ts;
const Dav1dFrameContext *const f = t->f;
Av1Block b_mem, *const b = f->frame_thread.pass ?
&f->frame_thread.b[t->by * f->b4_stride + t->bx] : &b_mem;
const uint8_t *const b_dim = dav1d_block_dimensions[bs];
const int bx4 = t->bx & 31, by4 = t->by & 31;
const int ss_ver = f->cur.p.layout == DAV1D_PIXEL_LAYOUT_I420;
const int ss_hor = f->cur.p.layout != DAV1D_PIXEL_LAYOUT_I444;
const int cbx4 = bx4 >> ss_hor, cby4 = by4 >> ss_ver;
const int bw4 = b_dim[0], bh4 = b_dim[1];
const int w4 = imin(bw4, f->bw - t->bx), h4 = imin(bh4, f->bh - t->by);
const int cbw4 = (bw4 + ss_hor) >> ss_hor, cbh4 = (bh4 + ss_ver) >> ss_ver;
const int have_left = t->bx > ts->tiling.col_start;
const int have_top = t->by > ts->tiling.row_start;
const int has_chroma = f->cur.p.layout != DAV1D_PIXEL_LAYOUT_I400 &&
(bw4 > ss_hor || t->bx & 1) &&
(bh4 > ss_ver || t->by & 1);
if (f->frame_thread.pass == 2) {
if (b->intra) {
f->bd_fn.recon_b_intra(t, bs, intra_edge_flags, b);
const enum IntraPredMode y_mode_nofilt =
b->y_mode == FILTER_PRED ? DC_PRED : b->y_mode;
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
rep_macro(type, t->dir mode, off, mul * y_mode_nofilt); \
rep_macro(type, t->dir intra, off, mul)
case_set(bh4, l., 1, by4);
case_set(bw4, a->, 0, bx4);
#undef set_ctx
if (has_chroma) {
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
rep_macro(type, t->dir uvmode, off, mul * b->uv_mode)
case_set(cbh4, l., 1, cby4);
case_set(cbw4, a->, 0, cbx4);
#undef set_ctx
}
} else {
if (f->frame_hdr->frame_type & 1 /* not intrabc */ &&
b->comp_type == COMP_INTER_NONE && b->motion_mode == MM_WARP)
{
uint64_t mask[2] = { 0, 0 };
find_matching_ref(t, intra_edge_flags, bw4, bh4, w4, h4,
have_left, have_top, b->ref[0], mask);
derive_warpmv(t, bw4, bh4, mask, b->mv[0], &t->warpmv);
}
if (f->bd_fn.recon_b_inter(t, bs, b)) return -1;
const uint8_t *const filter = dav1d_filter_dir[b->filter2d];
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
rep_macro(type, t->dir filter[0], off, mul * filter[0]); \
rep_macro(type, t->dir filter[1], off, mul * filter[1]); \
rep_macro(type, t->dir intra, off, 0)
case_set(bh4, l., 1, by4);
case_set(bw4, a->, 0, bx4);
#undef set_ctx
if (has_chroma) {
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
rep_macro(type, t->dir uvmode, off, mul * DC_PRED)
case_set(cbh4, l., 1, cby4);
case_set(cbw4, a->, 0, cbx4);
#undef set_ctx
}
}
return 0;
}
const int cw4 = (w4 + ss_hor) >> ss_hor, ch4 = (h4 + ss_ver) >> ss_ver;
b->bl = bl;
b->bp = bp;
b->bs = bs;
const Dav1dSegmentationData *seg = NULL;
// segment_id (if seg_feature for skip/ref/gmv is enabled)
int seg_pred = 0;
if (f->frame_hdr->segmentation.enabled) {
if (!f->frame_hdr->segmentation.update_map) {
if (f->prev_segmap) {
unsigned seg_id = get_prev_frame_segid(f, t->by, t->bx, w4, h4,
f->prev_segmap,
f->b4_stride);
if (seg_id >= 8) return -1;
b->seg_id = seg_id;
} else {
b->seg_id = 0;
}
seg = &f->frame_hdr->segmentation.seg_data.d[b->seg_id];
} else if (f->frame_hdr->segmentation.seg_data.preskip) {
if (f->frame_hdr->segmentation.temporal &&
(seg_pred = dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.seg_pred[t->a->seg_pred[bx4] +
t->l.seg_pred[by4]])))
{
// temporal predicted seg_id
if (f->prev_segmap) {
unsigned seg_id = get_prev_frame_segid(f, t->by, t->bx,
w4, h4,
f->prev_segmap,
f->b4_stride);
if (seg_id >= 8) return -1;
b->seg_id = seg_id;
} else {
b->seg_id = 0;
}
} else {
int seg_ctx;
const unsigned pred_seg_id =
get_cur_frame_segid(t->by, t->bx, have_top, have_left,
&seg_ctx, f->cur_segmap, f->b4_stride);
const unsigned diff = dav1d_msac_decode_symbol_adapt8(&ts->msac,
ts->cdf.m.seg_id[seg_ctx],
DAV1D_MAX_SEGMENTS - 1);
const unsigned last_active_seg_id =
f->frame_hdr->segmentation.seg_data.last_active_segid;
b->seg_id = neg_deinterleave(diff, pred_seg_id,
last_active_seg_id + 1);
if (b->seg_id > last_active_seg_id) b->seg_id = 0; // error?
if (b->seg_id >= DAV1D_MAX_SEGMENTS) b->seg_id = 0; // error?
}
if (DEBUG_BLOCK_INFO)
printf("Post-segid[preskip;%d]: r=%d\n",
b->seg_id, ts->msac.rng);
seg = &f->frame_hdr->segmentation.seg_data.d[b->seg_id];
}
} else {
b->seg_id = 0;
}
// skip_mode
if ((!seg || (!seg->globalmv && seg->ref == -1 && !seg->skip)) &&
f->frame_hdr->skip_mode_enabled && imin(bw4, bh4) > 1)
{
const int smctx = t->a->skip_mode[bx4] + t->l.skip_mode[by4];
b->skip_mode = dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.skip_mode[smctx]);
if (DEBUG_BLOCK_INFO)
printf("Post-skipmode[%d]: r=%d\n", b->skip_mode, ts->msac.rng);
} else {
b->skip_mode = 0;
}
// skip
if (b->skip_mode || (seg && seg->skip)) {
b->skip = 1;
} else {
const int sctx = t->a->skip[bx4] + t->l.skip[by4];
b->skip = dav1d_msac_decode_bool_adapt(&ts->msac, ts->cdf.m.skip[sctx]);
if (DEBUG_BLOCK_INFO)
printf("Post-skip[%d]: r=%d\n", b->skip, ts->msac.rng);
}
// segment_id
if (f->frame_hdr->segmentation.enabled &&
f->frame_hdr->segmentation.update_map &&
!f->frame_hdr->segmentation.seg_data.preskip)
{
if (!b->skip && f->frame_hdr->segmentation.temporal &&
(seg_pred = dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.seg_pred[t->a->seg_pred[bx4] +
t->l.seg_pred[by4]])))
{
// temporal predicted seg_id
if (f->prev_segmap) {
unsigned seg_id = get_prev_frame_segid(f, t->by, t->bx, w4, h4,
f->prev_segmap,
f->b4_stride);
if (seg_id >= 8) return -1;
b->seg_id = seg_id;
} else {
b->seg_id = 0;
}
} else {
int seg_ctx;
const unsigned pred_seg_id =
get_cur_frame_segid(t->by, t->bx, have_top, have_left,
&seg_ctx, f->cur_segmap, f->b4_stride);
if (b->skip) {
b->seg_id = pred_seg_id;
} else {
const unsigned diff = dav1d_msac_decode_symbol_adapt8(&ts->msac,
ts->cdf.m.seg_id[seg_ctx],
DAV1D_MAX_SEGMENTS - 1);
const unsigned last_active_seg_id =
f->frame_hdr->segmentation.seg_data.last_active_segid;
b->seg_id = neg_deinterleave(diff, pred_seg_id,
last_active_seg_id + 1);
if (b->seg_id > last_active_seg_id) b->seg_id = 0; // error?
}
if (b->seg_id >= DAV1D_MAX_SEGMENTS) b->seg_id = 0; // error?
}
seg = &f->frame_hdr->segmentation.seg_data.d[b->seg_id];
if (DEBUG_BLOCK_INFO)
printf("Post-segid[postskip;%d]: r=%d\n",
b->seg_id, ts->msac.rng);
}
// cdef index
if (!b->skip) {
const int idx = f->seq_hdr->sb128 ? ((t->bx & 16) >> 4) +
((t->by & 16) >> 3) : 0;
if (t->cur_sb_cdef_idx_ptr[idx] == -1) {
const int v = dav1d_msac_decode_bools(&ts->msac,
f->frame_hdr->cdef.n_bits);
t->cur_sb_cdef_idx_ptr[idx] = v;
if (bw4 > 16) t->cur_sb_cdef_idx_ptr[idx + 1] = v;
if (bh4 > 16) t->cur_sb_cdef_idx_ptr[idx + 2] = v;
if (bw4 == 32 && bh4 == 32) t->cur_sb_cdef_idx_ptr[idx + 3] = v;
if (DEBUG_BLOCK_INFO)
printf("Post-cdef_idx[%d]: r=%d\n",
*t->cur_sb_cdef_idx_ptr, ts->msac.rng);
}
}
// delta-q/lf
if (!(t->bx & (31 >> !f->seq_hdr->sb128)) &&
!(t->by & (31 >> !f->seq_hdr->sb128)))
{
const int prev_qidx = ts->last_qidx;
const int have_delta_q = f->frame_hdr->delta.q.present &&
(bs != (f->seq_hdr->sb128 ? BS_128x128 : BS_64x64) || !b->skip);
int8_t prev_delta_lf[4];
memcpy(prev_delta_lf, ts->last_delta_lf, 4);
if (have_delta_q) {
int delta_q = dav1d_msac_decode_symbol_adapt4(&ts->msac,
ts->cdf.m.delta_q, 3);
if (delta_q == 3) {
const int n_bits = 1 + dav1d_msac_decode_bools(&ts->msac, 3);
delta_q = dav1d_msac_decode_bools(&ts->msac, n_bits) +
1 + (1 << n_bits);
}
if (delta_q) {
if (dav1d_msac_decode_bool_equi(&ts->msac)) delta_q = -delta_q;
delta_q *= 1 << f->frame_hdr->delta.q.res_log2;
}
ts->last_qidx = iclip(ts->last_qidx + delta_q, 1, 255);
if (have_delta_q && DEBUG_BLOCK_INFO)
printf("Post-delta_q[%d->%d]: r=%d\n",
delta_q, ts->last_qidx, ts->msac.rng);
if (f->frame_hdr->delta.lf.present) {
const int n_lfs = f->frame_hdr->delta.lf.multi ?
f->cur.p.layout != DAV1D_PIXEL_LAYOUT_I400 ? 4 : 2 : 1;
for (int i = 0; i < n_lfs; i++) {
int delta_lf = dav1d_msac_decode_symbol_adapt4(&ts->msac,
ts->cdf.m.delta_lf[i + f->frame_hdr->delta.lf.multi], 3);
if (delta_lf == 3) {
const int n_bits = 1 + dav1d_msac_decode_bools(&ts->msac, 3);
delta_lf = dav1d_msac_decode_bools(&ts->msac, n_bits) +
1 + (1 << n_bits);
}
if (delta_lf) {
if (dav1d_msac_decode_bool_equi(&ts->msac))
delta_lf = -delta_lf;
delta_lf *= 1 << f->frame_hdr->delta.lf.res_log2;
}
ts->last_delta_lf[i] =
iclip(ts->last_delta_lf[i] + delta_lf, -63, 63);
if (have_delta_q && DEBUG_BLOCK_INFO)
printf("Post-delta_lf[%d:%d]: r=%d\n", i, delta_lf,
ts->msac.rng);
}
}
}
if (ts->last_qidx == f->frame_hdr->quant.yac) {
// assign frame-wide q values to this sb
ts->dq = f->dq;
} else if (ts->last_qidx != prev_qidx) {
// find sb-specific quant parameters
init_quant_tables(f->seq_hdr, f->frame_hdr, ts->last_qidx, ts->dqmem);
ts->dq = ts->dqmem;
}
if (!memcmp(ts->last_delta_lf, (int8_t[4]) { 0, 0, 0, 0 }, 4)) {
// assign frame-wide lf values to this sb
ts->lflvl = f->lf.lvl;
} else if (memcmp(ts->last_delta_lf, prev_delta_lf, 4)) {
// find sb-specific lf lvl parameters
dav1d_calc_lf_values(ts->lflvlmem, f->frame_hdr, ts->last_delta_lf);
ts->lflvl = ts->lflvlmem;
}
}
if (b->skip_mode) {
b->intra = 0;
} else if (f->frame_hdr->frame_type & 1) {
if (seg && (seg->ref >= 0 || seg->globalmv)) {
b->intra = !seg->ref;
} else {
const int ictx = get_intra_ctx(t->a, &t->l, by4, bx4,
have_top, have_left);
b->intra = !dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.intra[ictx]);
if (DEBUG_BLOCK_INFO)
printf("Post-intra[%d]: r=%d\n", b->intra, ts->msac.rng);
}
} else if (f->frame_hdr->allow_intrabc) {
b->intra = !dav1d_msac_decode_bool_adapt(&ts->msac, ts->cdf.m.intrabc);
if (DEBUG_BLOCK_INFO)
printf("Post-intrabcflag[%d]: r=%d\n", b->intra, ts->msac.rng);
} else {
b->intra = 1;
}
// intra/inter-specific stuff
if (b->intra) {
uint16_t *const ymode_cdf = f->frame_hdr->frame_type & 1 ?
ts->cdf.m.y_mode[dav1d_ymode_size_context[bs]] :
ts->cdf.kfym[dav1d_intra_mode_context[t->a->mode[bx4]]]
[dav1d_intra_mode_context[t->l.mode[by4]]];
b->y_mode = dav1d_msac_decode_symbol_adapt16(&ts->msac, ymode_cdf,
N_INTRA_PRED_MODES - 1);
if (DEBUG_BLOCK_INFO)
printf("Post-ymode[%d]: r=%d\n", b->y_mode, ts->msac.rng);
// angle delta
if (b_dim[2] + b_dim[3] >= 2 && b->y_mode >= VERT_PRED &&
b->y_mode <= VERT_LEFT_PRED)
{
uint16_t *const acdf = ts->cdf.m.angle_delta[b->y_mode - VERT_PRED];
const int angle = dav1d_msac_decode_symbol_adapt8(&ts->msac, acdf, 6);
b->y_angle = angle - 3;
} else {
b->y_angle = 0;
}
if (has_chroma) {
const int cfl_allowed = f->frame_hdr->segmentation.lossless[b->seg_id] ?
cbw4 == 1 && cbh4 == 1 : !!(cfl_allowed_mask & (1 << bs));
uint16_t *const uvmode_cdf = ts->cdf.m.uv_mode[cfl_allowed][b->y_mode];
b->uv_mode = dav1d_msac_decode_symbol_adapt16(&ts->msac, uvmode_cdf,
N_UV_INTRA_PRED_MODES - 1 - !cfl_allowed);
if (DEBUG_BLOCK_INFO)
printf("Post-uvmode[%d]: r=%d\n", b->uv_mode, ts->msac.rng);
if (b->uv_mode == CFL_PRED) {
#define SIGN(a) (!!(a) + ((a) > 0))
const int sign = dav1d_msac_decode_symbol_adapt8(&ts->msac,
ts->cdf.m.cfl_sign, 7) + 1;
const int sign_u = sign * 0x56 >> 8, sign_v = sign - sign_u * 3;
assert(sign_u == sign / 3);
if (sign_u) {
const int ctx = (sign_u == 2) * 3 + sign_v;
b->cfl_alpha[0] = dav1d_msac_decode_symbol_adapt16(&ts->msac,
ts->cdf.m.cfl_alpha[ctx], 15) + 1;
if (sign_u == 1) b->cfl_alpha[0] = -b->cfl_alpha[0];
} else {
b->cfl_alpha[0] = 0;
}
if (sign_v) {
const int ctx = (sign_v == 2) * 3 + sign_u;
b->cfl_alpha[1] = dav1d_msac_decode_symbol_adapt16(&ts->msac,
ts->cdf.m.cfl_alpha[ctx], 15) + 1;
if (sign_v == 1) b->cfl_alpha[1] = -b->cfl_alpha[1];
} else {
b->cfl_alpha[1] = 0;
}
#undef SIGN
if (DEBUG_BLOCK_INFO)
printf("Post-uvalphas[%d/%d]: r=%d\n",
b->cfl_alpha[0], b->cfl_alpha[1], ts->msac.rng);
} else if (b_dim[2] + b_dim[3] >= 2 && b->uv_mode >= VERT_PRED &&
b->uv_mode <= VERT_LEFT_PRED)
{
uint16_t *const acdf = ts->cdf.m.angle_delta[b->uv_mode - VERT_PRED];
const int angle = dav1d_msac_decode_symbol_adapt8(&ts->msac, acdf, 6);
b->uv_angle = angle - 3;
} else {
b->uv_angle = 0;
}
}
b->pal_sz[0] = b->pal_sz[1] = 0;
if (f->frame_hdr->allow_screen_content_tools &&
imax(bw4, bh4) <= 16 && bw4 + bh4 >= 4)
{
const int sz_ctx = b_dim[2] + b_dim[3] - 2;
if (b->y_mode == DC_PRED) {
const int pal_ctx = (t->a->pal_sz[bx4] > 0) + (t->l.pal_sz[by4] > 0);
const int use_y_pal = dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.pal_y[sz_ctx][pal_ctx]);
if (DEBUG_BLOCK_INFO)
printf("Post-y_pal[%d]: r=%d\n", use_y_pal, ts->msac.rng);
if (use_y_pal)
read_pal_plane(t, b, 0, sz_ctx, bx4, by4);
}
if (has_chroma && b->uv_mode == DC_PRED) {
const int pal_ctx = b->pal_sz[0] > 0;
const int use_uv_pal = dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.pal_uv[pal_ctx]);
if (DEBUG_BLOCK_INFO)
printf("Post-uv_pal[%d]: r=%d\n", use_uv_pal, ts->msac.rng);
if (use_uv_pal) // see aomedia bug 2183 for why we use luma coordinates
read_pal_uv(t, b, sz_ctx, bx4, by4);
}
}
if (b->y_mode == DC_PRED && !b->pal_sz[0] &&
imax(b_dim[2], b_dim[3]) <= 3 && f->seq_hdr->filter_intra)
{
const int is_filter = dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.use_filter_intra[bs]);
if (is_filter) {
b->y_mode = FILTER_PRED;
b->y_angle = dav1d_msac_decode_symbol_adapt4(&ts->msac,
ts->cdf.m.filter_intra, 4);
}
if (DEBUG_BLOCK_INFO)
printf("Post-filterintramode[%d/%d]: r=%d\n",
b->y_mode, b->y_angle, ts->msac.rng);
}
if (b->pal_sz[0]) {
uint8_t *pal_idx;
if (f->frame_thread.pass) {
assert(ts->frame_thread.pal_idx);
pal_idx = ts->frame_thread.pal_idx;
ts->frame_thread.pal_idx += bw4 * bh4 * 16;
} else
pal_idx = t->scratch.pal_idx;
read_pal_indices(t, pal_idx, b, 0, w4, h4, bw4, bh4);
if (DEBUG_BLOCK_INFO)
printf("Post-y-pal-indices: r=%d\n", ts->msac.rng);
}
if (has_chroma && b->pal_sz[1]) {
uint8_t *pal_idx;
if (f->frame_thread.pass) {
assert(ts->frame_thread.pal_idx);
pal_idx = ts->frame_thread.pal_idx;
ts->frame_thread.pal_idx += cbw4 * cbh4 * 16;
} else
pal_idx = &t->scratch.pal_idx[bw4 * bh4 * 16];
read_pal_indices(t, pal_idx, b, 1, cw4, ch4, cbw4, cbh4);
if (DEBUG_BLOCK_INFO)
printf("Post-uv-pal-indices: r=%d\n", ts->msac.rng);
}
const TxfmInfo *t_dim;
if (f->frame_hdr->segmentation.lossless[b->seg_id]) {
b->tx = b->uvtx = (int) TX_4X4;
t_dim = &dav1d_txfm_dimensions[TX_4X4];
} else {
b->tx = dav1d_max_txfm_size_for_bs[bs][0];
b->uvtx = dav1d_max_txfm_size_for_bs[bs][f->cur.p.layout];
t_dim = &dav1d_txfm_dimensions[b->tx];
if (f->frame_hdr->txfm_mode == DAV1D_TX_SWITCHABLE && t_dim->max > TX_4X4) {
const int tctx = get_tx_ctx(t->a, &t->l, t_dim, by4, bx4);
uint16_t *const tx_cdf = ts->cdf.m.txsz[t_dim->max - 1][tctx];
int depth = dav1d_msac_decode_symbol_adapt4(&ts->msac, tx_cdf,
imin(t_dim->max, 2));
while (depth--) {
b->tx = t_dim->sub;
t_dim = &dav1d_txfm_dimensions[b->tx];
}
}
if (DEBUG_BLOCK_INFO)
printf("Post-tx[%d]: r=%d\n", b->tx, ts->msac.rng);
}
// reconstruction
if (f->frame_thread.pass == 1) {
f->bd_fn.read_coef_blocks(t, bs, b);
} else {
f->bd_fn.recon_b_intra(t, bs, intra_edge_flags, b);
}
if (f->frame_hdr->loopfilter.level_y[0] ||
f->frame_hdr->loopfilter.level_y[1])
{
dav1d_create_lf_mask_intra(t->lf_mask, f->lf.level, f->b4_stride,
(const uint8_t (*)[8][2])
&ts->lflvl[b->seg_id][0][0][0],
t->bx, t->by, f->w4, f->h4, bs,
b->tx, b->uvtx, f->cur.p.layout,
&t->a->tx_lpf_y[bx4], &t->l.tx_lpf_y[by4],
has_chroma ? &t->a->tx_lpf_uv[cbx4] : NULL,
has_chroma ? &t->l.tx_lpf_uv[cby4] : NULL);
}
// update contexts
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
rep_macro(type, t->dir tx_intra, off, mul * (((uint8_t *) &t_dim->lw)[diridx])); \
rep_macro(type, t->dir tx, off, mul * (((uint8_t *) &t_dim->lw)[diridx])); \
rep_macro(type, t->dir mode, off, mul * y_mode_nofilt); \
rep_macro(type, t->dir pal_sz, off, mul * b->pal_sz[0]); \
rep_macro(type, t->dir seg_pred, off, mul * seg_pred); \
rep_macro(type, t->dir skip_mode, off, 0); \
rep_macro(type, t->dir intra, off, mul); \
rep_macro(type, t->dir skip, off, mul * b->skip); \
/* see aomedia bug 2183 for why we use luma coordinates here */ \
rep_macro(type, t->pal_sz_uv[diridx], off, mul * (has_chroma ? b->pal_sz[1] : 0)); \
if (f->frame_hdr->frame_type & 1) { \
rep_macro(type, t->dir comp_type, off, mul * COMP_INTER_NONE); \
rep_macro(type, t->dir ref[0], off, mul * ((uint8_t) -1)); \
rep_macro(type, t->dir ref[1], off, mul * ((uint8_t) -1)); \
rep_macro(type, t->dir filter[0], off, mul * DAV1D_N_SWITCHABLE_FILTERS); \
rep_macro(type, t->dir filter[1], off, mul * DAV1D_N_SWITCHABLE_FILTERS); \
}
const enum IntraPredMode y_mode_nofilt =
b->y_mode == FILTER_PRED ? DC_PRED : b->y_mode;
case_set(bh4, l., 1, by4);
case_set(bw4, a->, 0, bx4);
#undef set_ctx
if (b->pal_sz[0]) {
uint16_t *const pal = f->frame_thread.pass ?
f->frame_thread.pal[((t->by >> 1) + (t->bx & 1)) * (f->b4_stride >> 1) +
((t->bx >> 1) + (t->by & 1))][0] : t->scratch.pal[0];
for (int x = 0; x < bw4; x++)
memcpy(t->al_pal[0][bx4 + x][0], pal, 16);
for (int y = 0; y < bh4; y++)
memcpy(t->al_pal[1][by4 + y][0], pal, 16);
}
if (has_chroma) {
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
rep_macro(type, t->dir uvmode, off, mul * b->uv_mode)
case_set(cbh4, l., 1, cby4);
case_set(cbw4, a->, 0, cbx4);
#undef set_ctx
if (b->pal_sz[1]) {
const uint16_t (*const pal)[8] = f->frame_thread.pass ?
f->frame_thread.pal[((t->by >> 1) + (t->bx & 1)) *
(f->b4_stride >> 1) + ((t->bx >> 1) + (t->by & 1))] :
t->scratch.pal;
// see aomedia bug 2183 for why we use luma coordinates here
for (int pl = 1; pl <= 2; pl++) {
for (int x = 0; x < bw4; x++)
memcpy(t->al_pal[0][bx4 + x][pl], pal[pl], 16);
for (int y = 0; y < bh4; y++)
memcpy(t->al_pal[1][by4 + y][pl], pal[pl], 16);
}
}
}
if ((f->frame_hdr->frame_type & 1) || f->frame_hdr->allow_intrabc) {
splat_intraref(f->mvs, f->b4_stride, t->by, t->bx, bs,
y_mode_nofilt);
}
} else if (!(f->frame_hdr->frame_type & 1)) {
// intra block copy
candidate_mv mvstack[8];
int n_mvs;
mv mvlist[2][2];
dav1d_find_ref_mvs(mvstack, &n_mvs, mvlist, NULL,
(int[2]) { -1, -1 }, f->bw, f->bh,
bs, bp, t->by, t->bx, ts->tiling.col_start,
ts->tiling.col_end, ts->tiling.row_start,
ts->tiling.row_end, f->libaom_cm);
if (mvlist[0][0].y | mvlist[0][0].x)
b->mv[0] = mvlist[0][0];
else if (mvlist[0][1].y | mvlist[0][1].x)
b->mv[0] = mvlist[0][1];
else {
if (t->by - (16 << f->seq_hdr->sb128) < ts->tiling.row_start) {
b->mv[0].y = 0;
b->mv[0].x = -(512 << f->seq_hdr->sb128) - 2048;
} else {
b->mv[0].y = -(512 << f->seq_hdr->sb128);
b->mv[0].x = 0;
}
}
const struct mv ref = b->mv[0];
read_mv_residual(t, &b->mv[0], &ts->cdf.dmv, 0);
// clip intrabc motion vector to decoded parts of current tile
int border_left = ts->tiling.col_start * 4;
int border_top = ts->tiling.row_start * 4;
if (has_chroma) {
if (bw4 < 2 && ss_hor)
border_left += 4;
if (bh4 < 2 && ss_ver)
border_top += 4;
}
int src_left = t->bx * 4 + (b->mv[0].x >> 3);
int src_top = t->by * 4 + (b->mv[0].y >> 3);
int src_right = src_left + bw4 * 4;
int src_bottom = src_top + bh4 * 4;
const int border_right = ((ts->tiling.col_end + (bw4 - 1)) & ~(bw4 - 1)) * 4;
// check against left or right tile boundary and adjust if necessary
if (src_left < border_left) {
src_right += border_left - src_left;
src_left += border_left - src_left;
} else if (src_right > border_right) {
src_left -= src_right - border_right;
src_right -= src_right - border_right;
}
// check against top tile boundary and adjust if necessary
if (src_top < border_top) {
src_bottom += border_top - src_top;
src_top += border_top - src_top;
}
const int sbx = (t->bx >> (4 + f->seq_hdr->sb128)) << (6 + f->seq_hdr->sb128);
const int sby = (t->by >> (4 + f->seq_hdr->sb128)) << (6 + f->seq_hdr->sb128);
const int sb_size = 1 << (6 + f->seq_hdr->sb128);
// check for overlap with current superblock
if (src_bottom > sby && src_right > sbx) {
if (src_top - border_top >= src_bottom - sby) {
// if possible move src up into the previous suberblock row
src_top -= src_bottom - sby;
src_bottom -= src_bottom - sby;
} else if (src_left - border_left >= src_right - sbx) {
// if possible move src left into the previous suberblock
src_left -= src_right - sbx;
src_right -= src_right - sbx;
}
}
// move src up if it is below current superblock row
if (src_bottom > sby + sb_size) {
src_top -= src_bottom - (sby + sb_size);
src_bottom -= src_bottom - (sby + sb_size);
}
// error out if mv still overlaps with the current superblock
if (src_bottom > sby && src_right > sbx)
return -1;
b->mv[0].x = (src_left - t->bx * 4) * 8;
b->mv[0].y = (src_top - t->by * 4) * 8;
if (DEBUG_BLOCK_INFO)
printf("Post-dmv[%d/%d,ref=%d/%d|%d/%d]: r=%d\n",
b->mv[0].y, b->mv[0].x, ref.y, ref.x,
mvlist[0][0].y, mvlist[0][0].x, ts->msac.rng);
read_vartx_tree(t, b, bs, bx4, by4);
// reconstruction
if (f->frame_thread.pass == 1) {
f->bd_fn.read_coef_blocks(t, bs, b);
b->filter2d = FILTER_2D_BILINEAR;
} else {
if (f->bd_fn.recon_b_inter(t, bs, b)) return -1;
}
splat_intrabc_mv(f->mvs, f->b4_stride, t->by, t->bx, bs, b->mv[0]);
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
rep_macro(type, t->dir tx_intra, off, mul * b_dim[2 + diridx]); \
rep_macro(type, t->dir mode, off, mul * DC_PRED); \
rep_macro(type, t->dir pal_sz, off, 0); \
/* see aomedia bug 2183 for why this is outside if (has_chroma) */ \
rep_macro(type, t->pal_sz_uv[diridx], off, 0); \
rep_macro(type, t->dir seg_pred, off, mul * seg_pred); \
rep_macro(type, t->dir skip_mode, off, 0); \
rep_macro(type, t->dir intra, off, 0); \
rep_macro(type, t->dir skip, off, mul * b->skip)
case_set(bh4, l., 1, by4);
case_set(bw4, a->, 0, bx4);
#undef set_ctx
if (has_chroma) {
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
rep_macro(type, t->dir uvmode, off, mul * DC_PRED)
case_set(cbh4, l., 1, cby4);
case_set(cbw4, a->, 0, cbx4);
#undef set_ctx
}
} else {
// inter-specific mode/mv coding
int is_comp, has_subpel_filter;
if (b->skip_mode) {
is_comp = 1;
} else if ((!seg || (seg->ref == -1 && !seg->globalmv && !seg->skip)) &&
f->frame_hdr->switchable_comp_refs && imin(bw4, bh4) > 1)
{
const int ctx = get_comp_ctx(t->a, &t->l, by4, bx4,
have_top, have_left);
is_comp = dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.comp[ctx]);
if (DEBUG_BLOCK_INFO)
printf("Post-compflag[%d]: r=%d\n", is_comp, ts->msac.rng);
} else {
is_comp = 0;
}
if (b->skip_mode) {
b->ref[0] = f->frame_hdr->skip_mode_refs[0];
b->ref[1] = f->frame_hdr->skip_mode_refs[1];
b->comp_type = COMP_INTER_AVG;
b->inter_mode = NEARESTMV_NEARESTMV;
b->drl_idx = NEAREST_DRL;
has_subpel_filter = 0;
candidate_mv mvstack[8];
int n_mvs, ctx;
mv mvlist[2][2];
dav1d_find_ref_mvs(mvstack, &n_mvs, mvlist, &ctx,
(int[2]) { b->ref[0], b->ref[1] }, f->bw, f->bh,
bs, bp, t->by, t->bx, ts->tiling.col_start,
ts->tiling.col_end, ts->tiling.row_start,
ts->tiling.row_end, f->libaom_cm);
b->mv[0] = mvstack[0].this_mv;
b->mv[1] = mvstack[0].comp_mv;
fix_mv_precision(f->frame_hdr, &b->mv[0]);
fix_mv_precision(f->frame_hdr, &b->mv[1]);
if (DEBUG_BLOCK_INFO)
printf("Post-skipmodeblock[mv=1:y=%d,x=%d,2:y=%d,x=%d,refs=%d+%d\n",
b->mv[0].y, b->mv[0].x, b->mv[1].y, b->mv[1].x,
b->ref[0], b->ref[1]);
} else if (is_comp) {
const int dir_ctx = get_comp_dir_ctx(t->a, &t->l, by4, bx4,
have_top, have_left);
if (dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.comp_dir[dir_ctx]))
{
// bidir - first reference (fw)
const int ctx1 = av1_get_fwd_ref_ctx(t->a, &t->l, by4, bx4,
have_top, have_left);
if (dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.comp_fwd_ref[0][ctx1]))
{
const int ctx2 = av1_get_fwd_ref_2_ctx(t->a, &t->l, by4, bx4,
have_top, have_left);
b->ref[0] = 2 + dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.comp_fwd_ref[2][ctx2]);
} else {
const int ctx2 = av1_get_fwd_ref_1_ctx(t->a, &t->l, by4, bx4,
have_top, have_left);
b->ref[0] = dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.comp_fwd_ref[1][ctx2]);
}
// second reference (bw)
const int ctx3 = av1_get_bwd_ref_ctx(t->a, &t->l, by4, bx4,
have_top, have_left);
if (dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.comp_bwd_ref[0][ctx3]))
{
b->ref[1] = 6;
} else {
const int ctx4 = av1_get_bwd_ref_1_ctx(t->a, &t->l, by4, bx4,
have_top, have_left);
b->ref[1] = 4 + dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.comp_bwd_ref[1][ctx4]);
}
} else {
// unidir
const int uctx_p = av1_get_uni_p_ctx(t->a, &t->l, by4, bx4,
have_top, have_left);
if (dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.comp_uni_ref[0][uctx_p]))
{
b->ref[0] = 4;
b->ref[1] = 6;
} else {
const int uctx_p1 = av1_get_uni_p1_ctx(t->a, &t->l, by4, bx4,
have_top, have_left);
b->ref[0] = 0;
b->ref[1] = 1 + dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.comp_uni_ref[1][uctx_p1]);
if (b->ref[1] == 2) {
const int uctx_p2 = av1_get_uni_p2_ctx(t->a, &t->l, by4, bx4,
have_top, have_left);
b->ref[1] += dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.comp_uni_ref[2][uctx_p2]);
}
}
}
if (DEBUG_BLOCK_INFO)
printf("Post-refs[%d/%d]: r=%d\n",
b->ref[0], b->ref[1], ts->msac.rng);
candidate_mv mvstack[8];
int n_mvs, ctx;
mv mvlist[2][2];
dav1d_find_ref_mvs(mvstack, &n_mvs, mvlist, &ctx,
(int[2]) { b->ref[0], b->ref[1] }, f->bw, f->bh,
bs, bp, t->by, t->bx, ts->tiling.col_start,
ts->tiling.col_end, ts->tiling.row_start,
ts->tiling.row_end, f->libaom_cm);
b->inter_mode = dav1d_msac_decode_symbol_adapt8(&ts->msac,
ts->cdf.m.comp_inter_mode[ctx],
N_COMP_INTER_PRED_MODES - 1);
if (DEBUG_BLOCK_INFO)
printf("Post-compintermode[%d,ctx=%d,n_mvs=%d]: r=%d\n",
b->inter_mode, ctx, n_mvs, ts->msac.rng);
const uint8_t *const im = dav1d_comp_inter_pred_modes[b->inter_mode];
b->drl_idx = NEAREST_DRL;
if (b->inter_mode == NEWMV_NEWMV) {
if (n_mvs > 1) { // NEARER, NEAR or NEARISH
const int drl_ctx_v1 = get_drl_context(mvstack, 0);
b->drl_idx += dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.drl_bit[drl_ctx_v1]);
if (b->drl_idx == NEARER_DRL && n_mvs > 2) {
const int drl_ctx_v2 = get_drl_context(mvstack, 1);
b->drl_idx += dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.drl_bit[drl_ctx_v2]);
}
if (DEBUG_BLOCK_INFO)
printf("Post-drlidx[%d,n_mvs=%d]: r=%d\n",
b->drl_idx, n_mvs, ts->msac.rng);
}
} else if (im[0] == NEARMV || im[1] == NEARMV) {
b->drl_idx = NEARER_DRL;
if (n_mvs > 2) { // NEAR or NEARISH
const int drl_ctx_v2 = get_drl_context(mvstack, 1);
b->drl_idx += dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.drl_bit[drl_ctx_v2]);
if (b->drl_idx == NEAR_DRL && n_mvs > 3) {
const int drl_ctx_v3 = get_drl_context(mvstack, 2);
b->drl_idx += dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.drl_bit[drl_ctx_v3]);
}
if (DEBUG_BLOCK_INFO)
printf("Post-drlidx[%d,n_mvs=%d]: r=%d\n",
b->drl_idx, n_mvs, ts->msac.rng);
}
}
assert(b->drl_idx >= NEAREST_DRL && b->drl_idx <= NEARISH_DRL);
#define assign_comp_mv(idx, pfx) \
switch (im[idx]) { \
case NEARMV: \
case NEARESTMV: \
b->mv[idx] = mvstack[b->drl_idx].pfx##_mv; \
fix_mv_precision(f->frame_hdr, &b->mv[idx]); \
break; \
case GLOBALMV: \
has_subpel_filter |= \
f->frame_hdr->gmv[b->ref[idx]].type == DAV1D_WM_TYPE_TRANSLATION; \
b->mv[idx] = get_gmv_2d(&f->frame_hdr->gmv[b->ref[idx]], \
t->bx, t->by, bw4, bh4, f->frame_hdr); \
fix_mv_precision(f->frame_hdr, &b->mv[idx]); \
break; \
case NEWMV: \
b->mv[idx] = mvstack[b->drl_idx].pfx##_mv; \
read_mv_residual(t, &b->mv[idx], &ts->cdf.mv, \
!f->frame_hdr->force_integer_mv); \
break; \
}
has_subpel_filter = imin(bw4, bh4) == 1 ||
b->inter_mode != GLOBALMV_GLOBALMV;
assign_comp_mv(0, this);
assign_comp_mv(1, comp);
#undef assign_comp_mv
if (DEBUG_BLOCK_INFO)
printf("Post-residual_mv[1:y=%d,x=%d,2:y=%d,x=%d]: r=%d\n",
b->mv[0].y, b->mv[0].x, b->mv[1].y, b->mv[1].x,
ts->msac.rng);
// jnt_comp vs. seg vs. wedge
int is_segwedge = 0;
if (f->seq_hdr->masked_compound) {
const int mask_ctx = get_mask_comp_ctx(t->a, &t->l, by4, bx4);
is_segwedge = dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.mask_comp[mask_ctx]);
if (DEBUG_BLOCK_INFO)
printf("Post-segwedge_vs_jntavg[%d,ctx=%d]: r=%d\n",
is_segwedge, mask_ctx, ts->msac.rng);
}
if (!is_segwedge) {
if (f->seq_hdr->jnt_comp) {
const int jnt_ctx =
get_jnt_comp_ctx(f->seq_hdr->order_hint_n_bits,
f->cur.frame_hdr->frame_offset,
f->refp[b->ref[0]].p.frame_hdr->frame_offset,
f->refp[b->ref[1]].p.frame_hdr->frame_offset,
t->a, &t->l, by4, bx4);
b->comp_type = COMP_INTER_WEIGHTED_AVG +
dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.jnt_comp[jnt_ctx]);
if (DEBUG_BLOCK_INFO)
printf("Post-jnt_comp[%d,ctx=%d[ac:%d,ar:%d,lc:%d,lr:%d]]: r=%d\n",
b->comp_type == COMP_INTER_AVG,
jnt_ctx, t->a->comp_type[bx4], t->a->ref[0][bx4],
t->l.comp_type[by4], t->l.ref[0][by4],
ts->msac.rng);
} else {
b->comp_type = COMP_INTER_AVG;
}
} else {
if (wedge_allowed_mask & (1 << bs)) {
const int ctx = dav1d_wedge_ctx_lut[bs];
b->comp_type = COMP_INTER_WEDGE -
dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.wedge_comp[ctx]);
if (b->comp_type == COMP_INTER_WEDGE)
b->wedge_idx = dav1d_msac_decode_symbol_adapt16(&ts->msac,
ts->cdf.m.wedge_idx[ctx], 15);
} else {
b->comp_type = COMP_INTER_SEG;
}
b->mask_sign = dav1d_msac_decode_bool_equi(&ts->msac);
if (DEBUG_BLOCK_INFO)
printf("Post-seg/wedge[%d,wedge_idx=%d,sign=%d]: r=%d\n",
b->comp_type == COMP_INTER_WEDGE,
b->wedge_idx, b->mask_sign, ts->msac.rng);
}
} else {
b->comp_type = COMP_INTER_NONE;
// ref
if (seg && seg->ref > 0) {
b->ref[0] = seg->ref - 1;
} else if (seg && (seg->globalmv || seg->skip)) {
b->ref[0] = 0;
} else {
const int ctx1 = av1_get_ref_ctx(t->a, &t->l, by4, bx4,
have_top, have_left);
if (dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.ref[0][ctx1]))
{
const int ctx2 = av1_get_ref_2_ctx(t->a, &t->l, by4, bx4,
have_top, have_left);
if (dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.ref[1][ctx2]))
{
b->ref[0] = 6;
} else {
const int ctx3 = av1_get_ref_6_ctx(t->a, &t->l, by4, bx4,
have_top, have_left);
b->ref[0] = 4 + dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.ref[5][ctx3]);
}
} else {
const int ctx2 = av1_get_ref_3_ctx(t->a, &t->l, by4, bx4,
have_top, have_left);
if (dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.ref[2][ctx2]))
{
const int ctx3 = av1_get_ref_5_ctx(t->a, &t->l, by4, bx4,
have_top, have_left);
b->ref[0] = 2 + dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.ref[4][ctx3]);
} else {
const int ctx3 = av1_get_ref_4_ctx(t->a, &t->l, by4, bx4,
have_top, have_left);
b->ref[0] = dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.ref[3][ctx3]);
}
}
if (DEBUG_BLOCK_INFO)
printf("Post-ref[%d]: r=%d\n", b->ref[0], ts->msac.rng);
}
b->ref[1] = -1;
candidate_mv mvstack[8];
int n_mvs, ctx;
mv mvlist[2][2];
dav1d_find_ref_mvs(mvstack, &n_mvs, mvlist, &ctx,
(int[2]) { b->ref[0], -1 }, f->bw, f->bh, bs, bp,
t->by, t->bx, ts->tiling.col_start,
ts->tiling.col_end, ts->tiling.row_start,
ts->tiling.row_end, f->libaom_cm);
// mode parsing and mv derivation from ref_mvs
if ((seg && (seg->skip || seg->globalmv)) ||
dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.newmv_mode[ctx & 7]))
{
if ((seg && (seg->skip || seg->globalmv)) ||
!dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.globalmv_mode[(ctx >> 3) & 1]))
{
b->inter_mode = GLOBALMV;
b->mv[0] = get_gmv_2d(&f->frame_hdr->gmv[b->ref[0]],
t->bx, t->by, bw4, bh4, f->frame_hdr);
fix_mv_precision(f->frame_hdr, &b->mv[0]);
has_subpel_filter = imin(bw4, bh4) == 1 ||
f->frame_hdr->gmv[b->ref[0]].type == DAV1D_WM_TYPE_TRANSLATION;
} else {
has_subpel_filter = 1;
if (dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.refmv_mode[(ctx >> 4) & 15]))
{ // NEAREST, NEARER, NEAR or NEARISH
b->inter_mode = NEARMV;
b->drl_idx = NEARER_DRL;
if (n_mvs > 2) { // NEARER, NEAR or NEARISH
const int drl_ctx_v2 = get_drl_context(mvstack, 1);
b->drl_idx += dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.drl_bit[drl_ctx_v2]);
if (b->drl_idx == NEAR_DRL && n_mvs > 3) { // NEAR or NEARISH
const int drl_ctx_v3 =
get_drl_context(mvstack, 2);
b->drl_idx += dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.drl_bit[drl_ctx_v3]);
}
}
} else {
b->inter_mode = NEARESTMV;
b->drl_idx = NEAREST_DRL;
}
assert(b->drl_idx >= NEAREST_DRL && b->drl_idx <= NEARISH_DRL);
if (b->drl_idx >= NEAR_DRL) {
b->mv[0] = mvstack[b->drl_idx].this_mv;
} else {
b->mv[0] = mvlist[0][b->drl_idx];
fix_mv_precision(f->frame_hdr, &b->mv[0]);
}
}
if (DEBUG_BLOCK_INFO)
printf("Post-intermode[%d,drl=%d,mv=y:%d,x:%d,n_mvs=%d]: r=%d\n",
b->inter_mode, b->drl_idx, b->mv[0].y, b->mv[0].x, n_mvs,
ts->msac.rng);
} else {
has_subpel_filter = 1;
b->inter_mode = NEWMV;
b->drl_idx = NEAREST_DRL;
if (n_mvs > 1) { // NEARER, NEAR or NEARISH
const int drl_ctx_v1 = get_drl_context(mvstack, 0);
b->drl_idx += dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.drl_bit[drl_ctx_v1]);
if (b->drl_idx == NEARER_DRL && n_mvs > 2) { // NEAR or NEARISH
const int drl_ctx_v2 = get_drl_context(mvstack, 1);
b->drl_idx += dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.drl_bit[drl_ctx_v2]);
}
}
assert(b->drl_idx >= NEAREST_DRL && b->drl_idx <= NEARISH_DRL);
if (n_mvs > 1) {
b->mv[0] = mvstack[b->drl_idx].this_mv;
} else {
assert(!b->drl_idx);
b->mv[0] = mvlist[0][0];
fix_mv_precision(f->frame_hdr, &b->mv[0]);
}
if (DEBUG_BLOCK_INFO)
printf("Post-intermode[%d,drl=%d]: r=%d\n",
b->inter_mode, b->drl_idx, ts->msac.rng);
read_mv_residual(t, &b->mv[0], &ts->cdf.mv,
!f->frame_hdr->force_integer_mv);
if (DEBUG_BLOCK_INFO)
printf("Post-residualmv[mv=y:%d,x:%d]: r=%d\n",
b->mv[0].y, b->mv[0].x, ts->msac.rng);
}
// interintra flags
const int ii_sz_grp = dav1d_ymode_size_context[bs];
if (f->seq_hdr->inter_intra &&
interintra_allowed_mask & (1 << bs) &&
dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.interintra[ii_sz_grp]))
{
b->interintra_mode = dav1d_msac_decode_symbol_adapt4(&ts->msac,
ts->cdf.m.interintra_mode[ii_sz_grp],
N_INTER_INTRA_PRED_MODES - 1);
const int wedge_ctx = dav1d_wedge_ctx_lut[bs];
b->interintra_type = INTER_INTRA_BLEND +
dav1d_msac_decode_bool_adapt(&ts->msac,
ts->cdf.m.interintra_wedge[wedge_ctx]);
if (b->interintra_type == INTER_INTRA_WEDGE)
b->wedge_idx = dav1d_msac_decode_symbol_adapt16(&ts->msac,
ts->cdf.m.wedge_idx[wedge_ctx], 15);
} else {
b->interintra_type = INTER_INTRA_NONE;
}
if (DEBUG_BLOCK_INFO && f->seq_hdr->inter_intra &&
interintra_allowed_mask & (1 << bs))
{
printf("Post-interintra[t=%d,m=%d,w=%d]: r=%d\n",
b->interintra_type, b->interintra_mode,
b->wedge_idx, ts->msac.rng);
}
// motion variation
if (f->frame_hdr->switchable_motion_mode &&
b->interintra_type == INTER_INTRA_NONE && imin(bw4, bh4) >= 2 &&
// is not warped global motion
!(!f->frame_hdr->force_integer_mv && b->inter_mode == GLOBALMV &&
f->frame_hdr->gmv[b->ref[0]].type > DAV1D_WM_TYPE_TRANSLATION) &&
// has overlappable neighbours
((have_left && findoddzero(&t->l.intra[by4 + 1], h4 >> 1)) ||
(have_top && findoddzero(&t->a->intra[bx4 + 1], w4 >> 1))))
{
// reaching here means the block allows obmc - check warp by
// finding matching-ref blocks in top/left edges
uint64_t mask[2] = { 0, 0 };
find_matching_ref(t, intra_edge_flags, bw4, bh4, w4, h4,
have_left, have_top, b->ref[0], mask);
const int allow_warp = !f->svc[b->ref[0]][0].scale &&
!f->frame_hdr->force_integer_mv &&
f->frame_hdr->warp_motion && (mask[0] | mask[1]);
b->motion_mode = allow_warp ?
dav1d_msac_decode_symbol_adapt4(&ts->msac,
ts->cdf.m.motion_mode[bs], 2) :
dav1d_msac_decode_bool_adapt(&ts->msac, ts->cdf.m.obmc[bs]);
if (b->motion_mode == MM_WARP) {
has_subpel_filter = 0;
derive_warpmv(t, bw4, bh4, mask, b->mv[0], &t->warpmv);
#define signabs(v) v < 0 ? '-' : ' ', abs(v)
if (DEBUG_BLOCK_INFO)
printf("[ %c%x %c%x %c%x\n %c%x %c%x %c%x ]\n"
"alpha=%c%x, beta=%c%x, gamma=%c%x, delta=%c%x\n",
signabs(t->warpmv.matrix[0]),
signabs(t->warpmv.matrix[1]),
signabs(t->warpmv.matrix[2]),
signabs(t->warpmv.matrix[3]),
signabs(t->warpmv.matrix[4]),
signabs(t->warpmv.matrix[5]),
signabs(t->warpmv.alpha),
signabs(t->warpmv.beta),
signabs(t->warpmv.gamma),
signabs(t->warpmv.delta));
#undef signabs
}
if (DEBUG_BLOCK_INFO)
printf("Post-motionmode[%d]: r=%d [mask: 0x%" PRIu64 "x/0x%"
PRIu64 "x]\n", b->motion_mode, ts->msac.rng, mask[0],
mask[1]);
} else {
b->motion_mode = MM_TRANSLATION;
}
}
// subpel filter
enum Dav1dFilterMode filter[2];
if (f->frame_hdr->subpel_filter_mode == DAV1D_FILTER_SWITCHABLE) {
if (has_subpel_filter) {
const int comp = b->comp_type != COMP_INTER_NONE;
const int ctx1 = get_filter_ctx(t->a, &t->l, comp, 0, b->ref[0],
by4, bx4);
filter[0] = dav1d_msac_decode_symbol_adapt4(&ts->msac,
ts->cdf.m.filter[0][ctx1],
DAV1D_N_SWITCHABLE_FILTERS - 1);
if (f->seq_hdr->dual_filter) {
const int ctx2 = get_filter_ctx(t->a, &t->l, comp, 1,
b->ref[0], by4, bx4);
if (DEBUG_BLOCK_INFO)
printf("Post-subpel_filter1[%d,ctx=%d]: r=%d\n",
filter[0], ctx1, ts->msac.rng);
filter[1] = dav1d_msac_decode_symbol_adapt4(&ts->msac,
ts->cdf.m.filter[1][ctx2],
DAV1D_N_SWITCHABLE_FILTERS - 1);
if (DEBUG_BLOCK_INFO)
printf("Post-subpel_filter2[%d,ctx=%d]: r=%d\n",
filter[1], ctx2, ts->msac.rng);
} else {
filter[1] = filter[0];
if (DEBUG_BLOCK_INFO)
printf("Post-subpel_filter[%d,ctx=%d]: r=%d\n",
filter[0], ctx1, ts->msac.rng);
}
} else {
filter[0] = filter[1] = DAV1D_FILTER_8TAP_REGULAR;
}
} else {
filter[0] = filter[1] = f->frame_hdr->subpel_filter_mode;
}
b->filter2d = dav1d_filter_2d[filter[1]][filter[0]];
read_vartx_tree(t, b, bs, bx4, by4);
// reconstruction
if (f->frame_thread.pass == 1) {
f->bd_fn.read_coef_blocks(t, bs, b);
} else {
if (f->bd_fn.recon_b_inter(t, bs, b)) return -1;
}
if (f->frame_hdr->loopfilter.level_y[0] ||
f->frame_hdr->loopfilter.level_y[1])
{
const int is_globalmv =
b->inter_mode == (is_comp ? GLOBALMV_GLOBALMV : GLOBALMV);
const uint8_t (*const lf_lvls)[8][2] = (const uint8_t (*)[8][2])
&ts->lflvl[b->seg_id][0][b->ref[0] + 1][!is_globalmv];
dav1d_create_lf_mask_inter(t->lf_mask, f->lf.level, f->b4_stride,
lf_lvls, t->bx, t->by, f->w4, f->h4,
b->skip, bs, b->tx_split, b->uvtx,
f->cur.p.layout,
&t->a->tx_lpf_y[bx4], &t->l.tx_lpf_y[by4],
has_chroma ? &t->a->tx_lpf_uv[cbx4] : NULL,
has_chroma ? &t->l.tx_lpf_uv[cby4] : NULL);
}
// context updates
if (is_comp) {
splat_tworef_mv(f->mvs, f->b4_stride, t->by, t->bx, bs,
b->inter_mode, b->ref[0], b->ref[1],
b->mv[0], b->mv[1]);
} else {
splat_oneref_mv(f->mvs, f->b4_stride, t->by, t->bx, bs,
b->inter_mode, b->ref[0], b->mv[0],
b->interintra_type);
}
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
rep_macro(type, t->dir seg_pred, off, mul * seg_pred); \
rep_macro(type, t->dir skip_mode, off, mul * b->skip_mode); \
rep_macro(type, t->dir intra, off, 0); \
rep_macro(type, t->dir skip, off, mul * b->skip); \
rep_macro(type, t->dir pal_sz, off, 0); \
/* see aomedia bug 2183 for why this is outside if (has_chroma) */ \
rep_macro(type, t->pal_sz_uv[diridx], off, 0); \
rep_macro(type, t->dir tx_intra, off, mul * b_dim[2 + diridx]); \
rep_macro(type, t->dir comp_type, off, mul * b->comp_type); \
rep_macro(type, t->dir filter[0], off, mul * filter[0]); \
rep_macro(type, t->dir filter[1], off, mul * filter[1]); \
rep_macro(type, t->dir mode, off, mul * b->inter_mode); \
rep_macro(type, t->dir ref[0], off, mul * b->ref[0]); \
rep_macro(type, t->dir ref[1], off, mul * ((uint8_t) b->ref[1]))
case_set(bh4, l., 1, by4);
case_set(bw4, a->, 0, bx4);
#undef set_ctx
if (has_chroma) {
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
rep_macro(type, t->dir uvmode, off, mul * DC_PRED)
case_set(cbh4, l., 1, cby4);
case_set(cbw4, a->, 0, cbx4);
#undef set_ctx
}
}
// update contexts
if (f->frame_hdr->segmentation.enabled &&
f->frame_hdr->segmentation.update_map)
{
uint8_t *seg_ptr = &f->cur_segmap[t->by * f->b4_stride + t->bx];
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
for (int y = 0; y < bh4; y++) { \
rep_macro(type, seg_ptr, 0, mul * b->seg_id); \
seg_ptr += f->b4_stride; \
}
case_set(bw4, NULL, 0, 0);
#undef set_ctx
}
if (!b->skip) {
uint16_t (*noskip_mask)[2] = &t->lf_mask->noskip_mask[by4];
const unsigned mask = (~0U >> (32 - bw4)) << (bx4 & 15);
const int bx_idx = (bx4 & 16) >> 4;
for (int y = 0; y < bh4; y++, noskip_mask++) {
(*noskip_mask)[bx_idx] |= mask;
if (bw4 == 32) // this should be mask >> 16, but it's 0xffffffff anyway
(*noskip_mask)[1] |= mask;
}
}
return 0;
}
#if defined(__has_feature)
#if __has_feature(memory_sanitizer)
#include <sanitizer/msan_interface.h>
static int checked_decode_b(Dav1dTileContext *const t,
const enum BlockLevel bl,
const enum BlockSize bs,
const enum BlockPartition bp,
const enum EdgeFlags intra_edge_flags)
{
const Dav1dFrameContext *const f = t->f;
const int err = decode_b(t, bl, bs, bp, intra_edge_flags);
if (err == 0 && !(f->frame_thread.pass & 1)) {
const int ss_ver = f->cur.p.layout == DAV1D_PIXEL_LAYOUT_I420;
const int ss_hor = f->cur.p.layout != DAV1D_PIXEL_LAYOUT_I444;
const uint8_t *const b_dim = dav1d_block_dimensions[bs];
const int bw4 = b_dim[0], bh4 = b_dim[1];
const int w4 = imin(bw4, f->bw - t->bx), h4 = imin(bh4, f->bh - t->by);
const int has_chroma = f->seq_hdr->layout != DAV1D_PIXEL_LAYOUT_I400 &&
(bw4 > ss_hor || t->bx & 1) &&
(bh4 > ss_ver || t->by & 1);
for (int p = 0; p < 1 + 2 * has_chroma; p++) {
const int ss_ver = p && f->cur.p.layout == DAV1D_PIXEL_LAYOUT_I420;
const int ss_hor = p && f->cur.p.layout != DAV1D_PIXEL_LAYOUT_I444;
const ptrdiff_t stride = f->cur.stride[!!p];
const int bx = t->bx & ~ss_hor;
const int by = t->by & ~ss_ver;
const int width = w4 << (2 - ss_hor + (bw4 == ss_hor));
const int height = h4 << (2 - ss_ver + (bh4 == ss_ver));
const uint8_t *data = f->cur.data[p] + (by << (2 - ss_ver)) * stride +
(bx << (2 - ss_hor + !!f->seq_hdr->hbd));
for (int y = 0; y < height; data += stride, y++) {
const size_t line_sz = width << !!f->seq_hdr->hbd;
if (__msan_test_shadow(data, line_sz) != -1) {
fprintf(stderr, "B[%d](%d, %d) w4:%d, h4:%d, row:%d\n",
p, bx, by, w4, h4, y);
__msan_check_mem_is_initialized(data, line_sz);
}
}
}
}
return err;
}
#define decode_b checked_decode_b
#endif /* defined(__has_feature) */
#endif /* __has_feature(memory_sanitizer) */
static int decode_sb(Dav1dTileContext *const t, const enum BlockLevel bl,
const EdgeNode *const node)
{
const Dav1dFrameContext *const f = t->f;
const int hsz = 16 >> bl;
const int have_h_split = f->bw > t->bx + hsz;
const int have_v_split = f->bh > t->by + hsz;