int vp9_decode_block_tokens(MACROBLOCKD *xd, int plane, const scan_order *sc,
int x, int y, TX_SIZE tx_size, vpx_reader *r,
int seg_id) {
struct macroblockd_plane *const pd = &xd->plane[plane];
const int16_t *const dequant = pd->seg_dequant[seg_id];
int eob;
ENTROPY_CONTEXT *a = pd->above_context + x;
ENTROPY_CONTEXT *l = pd->left_context + y;
int ctx;
int ctx_shift_a = 0;
int ctx_shift_l = 0;
switch (tx_size) {
case TX_4X4:
ctx = a[0] != 0;
ctx += l[0] != 0;
eob = decode_coefs(xd, get_plane_type(plane), pd->dqcoeff, tx_size,
dequant, ctx, sc->scan, sc->neighbors, r);
a[0] = l[0] = (eob > 0);
break;
case TX_8X8:
get_ctx_shift(xd, &ctx_shift_a, &ctx_shift_l, x, y, 1 << TX_8X8);
ctx = !!*(const uint16_t *)a;
ctx += !!*(const uint16_t *)l;
eob = decode_coefs(xd, get_plane_type(plane), pd->dqcoeff, tx_size,
dequant, ctx, sc->scan, sc->neighbors, r);
*(uint16_t *)a = ((eob > 0) * 0x0101) >> ctx_shift_a;
*(uint16_t *)l = ((eob > 0) * 0x0101) >> ctx_shift_l;
break;
case TX_16X16:
get_ctx_shift(xd, &ctx_shift_a, &ctx_shift_l, x, y, 1 << TX_16X16);
ctx = !!*(const uint32_t *)a;
ctx += !!*(const uint32_t *)l;
eob = decode_coefs(xd, get_plane_type(plane), pd->dqcoeff, tx_size,
dequant, ctx, sc->scan, sc->neighbors, r);
*(uint32_t *)a = ((eob > 0) * 0x01010101) >> ctx_shift_a;
*(uint32_t *)l = ((eob > 0) * 0x01010101) >> ctx_shift_l;
break;
case TX_32X32:
get_ctx_shift(xd, &ctx_shift_a, &ctx_shift_l, x, y, 1 << TX_32X32);
// NOTE: casting to uint64_t here is safe because the default memory
// alignment is at least 8 bytes and the TX_32X32 is aligned on 8 byte
// boundaries.
ctx = !!*(const uint64_t *)a;
ctx += !!*(const uint64_t *)l;
eob = decode_coefs(xd, get_plane_type(plane), pd->dqcoeff, tx_size,
dequant, ctx, sc->scan, sc->neighbors, r);
*(uint64_t *)a = ((eob > 0) * 0x0101010101010101ULL) >> ctx_shift_a;
*(uint64_t *)l = ((eob > 0) * 0x0101010101010101ULL) >> ctx_shift_l;
break;
default:
assert(0 && "Invalid transform size.");
eob = 0;
break;
}
return eob;
}
int vp9_optimize_b(MACROBLOCK *mb, int plane, int block, TX_SIZE tx_size,
int ctx) {
MACROBLOCKD *const xd = &mb->e_mbd;
struct macroblock_plane *const p = &mb->plane[plane];
struct macroblockd_plane *const pd = &xd->plane[plane];
const int ref = is_inter_block(xd->mi[0]);
vp9_token_state tokens[1025][2];
uint8_t token_cache[1024];
const tran_low_t *const coeff = BLOCK_OFFSET(mb->plane[plane].coeff, block);
tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
const int eob = p->eobs[block];
const PLANE_TYPE type = get_plane_type(plane);
const int default_eob = 16 << (tx_size << 1);
const int shift = (tx_size == TX_32X32);
const int16_t *const dequant_ptr = pd->dequant;
const uint8_t *const band_translate = get_band_translate(tx_size);
const scan_order *const so = get_scan(xd, tx_size, type, block);
const int16_t *const scan = so->scan;
const int16_t *const nb = so->neighbors;
const int dq_step[2] = { dequant_ptr[0] >> shift, dequant_ptr[1] >> shift };
static int optimize_b(MACROBLOCK *mb, int plane, int block,
TX_SIZE tx_size, int ctx) {
MACROBLOCKD *const xd = &mb->e_mbd;
struct macroblock_plane *const p = &mb->plane[plane];
struct macroblockd_plane *const pd = &xd->plane[plane];
const int ref = is_inter_block(xd->mi[0]);
vp9_token_state tokens[1025][2];
unsigned best_index[1025][2];
uint8_t token_cache[1024];
const tran_low_t *const coeff = BLOCK_OFFSET(mb->plane[plane].coeff, block);
tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
const int eob = p->eobs[block];
const PLANE_TYPE type = get_plane_type(plane);
const int default_eob = 16 << (tx_size << 1);
const int mul = 1 + (tx_size == TX_32X32);
const int16_t *dequant_ptr = pd->dequant;
const uint8_t *const band_translate = get_band_translate(tx_size);
const scan_order *const so = get_scan(xd, tx_size, type, block);
const int16_t *const scan = so->scan;
const int16_t *const nb = so->neighbors;
int next = eob, sz = 0;
int64_t rdmult = mb->rdmult * plane_rd_mult[type], rddiv = mb->rddiv;
int64_t rd_cost0, rd_cost1;
int rate0, rate1, error0, error1;
int16_t t0, t1;
EXTRABIT e0;
int best, band, pt, i, final_eob;
#if CONFIG_VP9_HIGHBITDEPTH
const int *cat6_high_cost = vp9_get_high_cost_table(xd->bd);
#else
const int *cat6_high_cost = vp9_get_high_cost_table(8);
#endif
assert((!type && !plane) || (type && plane));
assert(eob <= default_eob);
/* Now set up a Viterbi trellis to evaluate alternative roundings. */
if (!ref)
rdmult = (rdmult * 9) >> 4;
/* Initialize the sentinel node of the trellis. */
tokens[eob][0].rate = 0;
tokens[eob][0].error = 0;
tokens[eob][0].next = default_eob;
tokens[eob][0].token = EOB_TOKEN;
tokens[eob][0].qc = 0;
tokens[eob][1] = tokens[eob][0];
for (i = 0; i < eob; i++)
token_cache[scan[i]] =
vp9_pt_energy_class[vp9_get_token(qcoeff[scan[i]])];
for (i = eob; i-- > 0;) {
int base_bits, d2, dx;
const int rc = scan[i];
int x = qcoeff[rc];
/* Only add a trellis state for non-zero coefficients. */
if (x) {
int shortcut = 0;
error0 = tokens[next][0].error;
error1 = tokens[next][1].error;
/* Evaluate the first possibility for this state. */
rate0 = tokens[next][0].rate;
rate1 = tokens[next][1].rate;
vp9_get_token_extra(x, &t0, &e0);
/* Consider both possible successor states. */
if (next < default_eob) {
band = band_translate[i + 1];
pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
rate0 += mb->token_costs[tx_size][type][ref][band][0][pt]
[tokens[next][0].token];
rate1 += mb->token_costs[tx_size][type][ref][band][0][pt]
[tokens[next][1].token];
}
UPDATE_RD_COST();
/* And pick the best. */
best = rd_cost1 < rd_cost0;
base_bits = vp9_get_cost(t0, e0, cat6_high_cost);
dx = mul * (dqcoeff[rc] - coeff[rc]);
#if CONFIG_VP9_HIGHBITDEPTH
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
dx >>= xd->bd - 8;
}
#endif // CONFIG_VP9_HIGHBITDEPTH
d2 = dx * dx;
tokens[i][0].rate = base_bits + (best ? rate1 : rate0);
tokens[i][0].error = d2 + (best ? error1 : error0);
tokens[i][0].next = next;
tokens[i][0].token = t0;
tokens[i][0].qc = x;
best_index[i][0] = best;
/* Evaluate the second possibility for this state. */
rate0 = tokens[next][0].rate;
rate1 = tokens[next][1].rate;
if ((abs(x) * dequant_ptr[rc != 0] > abs(coeff[rc]) * mul) &&
(abs(x) * dequant_ptr[rc != 0] < abs(coeff[rc]) * mul +
dequant_ptr[rc != 0]))
shortcut = 1;
else
shortcut = 0;
if (shortcut) {
sz = -(x < 0);
x -= 2 * sz + 1;
}
/* Consider both possible successor states. */
if (!x) {
/* If we reduced this coefficient to zero, check to see if
* we need to move the EOB back here.
*/
t0 = tokens[next][0].token == EOB_TOKEN ? EOB_TOKEN : ZERO_TOKEN;
t1 = tokens[next][1].token == EOB_TOKEN ? EOB_TOKEN : ZERO_TOKEN;
e0 = 0;
} else {
vp9_get_token_extra(x, &t0, &e0);
t1 = t0;
}
if (next < default_eob) {
band = band_translate[i + 1];
if (t0 != EOB_TOKEN) {
pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
rate0 += mb->token_costs[tx_size][type][ref][band][!x][pt]
[tokens[next][0].token];
}
if (t1 != EOB_TOKEN) {
pt = trellis_get_coeff_context(scan, nb, i, t1, token_cache);
rate1 += mb->token_costs[tx_size][type][ref][band][!x][pt]
[tokens[next][1].token];
}
}
UPDATE_RD_COST();
/* And pick the best. */
best = rd_cost1 < rd_cost0;
base_bits = vp9_get_cost(t0, e0, cat6_high_cost);
if (shortcut) {
#if CONFIG_VP9_HIGHBITDEPTH
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
dx -= ((dequant_ptr[rc != 0] >> (xd->bd - 8)) + sz) ^ sz;
} else {
dx -= (dequant_ptr[rc != 0] + sz) ^ sz;
}
#else
dx -= (dequant_ptr[rc != 0] + sz) ^ sz;
#endif // CONFIG_VP9_HIGHBITDEPTH
d2 = dx * dx;
}
static void tokenize_b(int plane, int block, int row, int col,
BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
struct tokenize_b_args *const args = arg;
VP9_COMP *cpi = args->cpi;
ThreadData *const td = args->td;
MACROBLOCK *const x = &td->mb;
MACROBLOCKD *const xd = &x->e_mbd;
TOKENEXTRA **tp = args->tp;
uint8_t token_cache[32 * 32];
struct macroblock_plane *p = &x->plane[plane];
struct macroblockd_plane *pd = &xd->plane[plane];
MODE_INFO *mi = xd->mi[0];
int pt; /* near block/prev token context index */
int c;
TOKENEXTRA *t = *tp; /* store tokens starting here */
int eob = p->eobs[block];
const PLANE_TYPE type = get_plane_type(plane);
const tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
const int16_t *scan, *nb;
const scan_order *so;
const int ref = is_inter_block(mi);
unsigned int (*const counts)[COEFF_CONTEXTS][ENTROPY_TOKENS] =
td->rd_counts.coef_counts[tx_size][type][ref];
vpx_prob (*const coef_probs)[COEFF_CONTEXTS][UNCONSTRAINED_NODES] =
cpi->common.fc->coef_probs[tx_size][type][ref];
unsigned int (*const eob_branch)[COEFF_CONTEXTS] =
td->counts->eob_branch[tx_size][type][ref];
const uint8_t *const band = get_band_translate(tx_size);
const int tx_eob = 16 << (tx_size << 1);
int16_t token;
EXTRABIT extra;
pt = get_entropy_context(tx_size, pd->above_context + col,
pd->left_context + row);
so = get_scan(xd, tx_size, type, block);
scan = so->scan;
nb = so->neighbors;
c = 0;
while (c < eob) {
int v = 0;
v = qcoeff[scan[c]];
++eob_branch[band[c]][pt];
while (!v) {
add_token_no_extra(&t, coef_probs[band[c]][pt], ZERO_TOKEN,
counts[band[c]][pt]);
token_cache[scan[c]] = 0;
++c;
pt = get_coef_context(nb, token_cache, c);
v = qcoeff[scan[c]];
}
vp9_get_token_extra(v, &token, &extra);
add_token(&t, coef_probs[band[c]][pt], token, extra, counts[band[c]][pt]);
token_cache[scan[c]] = vp9_pt_energy_class[token];
++c;
pt = get_coef_context(nb, token_cache, c);
}
if (c < tx_eob) {
++eob_branch[band[c]][pt];
add_token_no_extra(&t, coef_probs[band[c]][pt], EOB_TOKEN,
counts[band[c]][pt]);
}
*tp = t;
vp9_set_contexts(xd, pd, plane_bsize, tx_size, c > 0, col, row);
}
