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]->mbmi);
vp10_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 = pd->plane_type;
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);
TX_TYPE tx_type = get_tx_type(type, xd, block);
const scan_order *const so = get_scan(tx_size, tx_type);
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 int16_t *cat6_high_cost = vp10_get_high_cost_table(xd->bd);
#else
const int16_t *cat6_high_cost = vp10_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]] =
vp10_pt_energy_class[vp10_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;
vp10_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 = vp10_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 {
vp10_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 = vp10_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 blk_row, int blk_col,
BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
struct tokenize_b_args *const args = arg;
AV1_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];
MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
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 = pd->plane_type;
const tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
const int segment_id = mbmi->segment_id;
const int16_t *scan, *nb;
const TX_TYPE tx_type = get_tx_type(type, xd, block);
const scan_order *const so = get_scan(tx_size, tx_type);
const int ref = is_inter_block(mbmi);
unsigned int (*const counts)[COEFF_CONTEXTS][ENTROPY_TOKENS] =
td->rd_counts.coef_counts[tx_size][type][ref];
aom_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 seg_eob = get_tx_eob(&cpi->common.seg, segment_id, tx_size);
int16_t token;
EXTRABIT extra;
pt = get_entropy_context(tx_size, pd->above_context + blk_col,
pd->left_context + blk_row);
scan = so->scan;
nb = so->neighbors;
c = 0;
while (c < eob) {
int v = 0;
int skip_eob = 0;
v = qcoeff[scan[c]];
while (!v) {
add_token_no_extra(&t, coef_probs[band[c]][pt], ZERO_TOKEN, skip_eob,
counts[band[c]][pt]);
eob_branch[band[c]][pt] += !skip_eob;
skip_eob = 1;
token_cache[scan[c]] = 0;
++c;
pt = get_coef_context(nb, token_cache, c);
v = qcoeff[scan[c]];
}
av1_get_token_extra(v, &token, &extra);
add_token(&t, coef_probs[band[c]][pt], extra, (uint8_t)token,
(uint8_t)skip_eob, counts[band[c]][pt]);
eob_branch[band[c]][pt] += !skip_eob;
token_cache[scan[c]] = av1_pt_energy_class[token];
++c;
pt = get_coef_context(nb, token_cache, c);
}
if (c < seg_eob) {
add_token_no_extra(&t, coef_probs[band[c]][pt], EOB_TOKEN, 0,
counts[band[c]][pt]);
++eob_branch[band[c]][pt];
}
*tp = t;
av1_set_contexts(xd, pd, plane_bsize, tx_size, c > 0, blk_col, blk_row);
}
static void encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
TX_SIZE tx_size, void *arg) {
struct encode_b_args* const args = arg;
MACROBLOCK *const x = args->x;
MACROBLOCKD *const xd = &x->e_mbd;
MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
struct macroblock_plane *const p = &x->plane[plane];
struct macroblockd_plane *const pd = &xd->plane[plane];
int16_t *coeff = BLOCK_OFFSET(p->coeff, block);
int16_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
const scan_order *scan_order;
TX_TYPE tx_type;
PREDICTION_MODE mode;
const int bwl = b_width_log2(plane_bsize);
const int diff_stride = 4 * (1 << bwl);
uint8_t *src, *dst;
int16_t *src_diff;
uint16_t *eob = &p->eobs[block];
const int src_stride = p->src.stride;
const int dst_stride = pd->dst.stride;
int i, j;
txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
dst = &pd->dst.buf[4 * (j * dst_stride + i)];
src = &p->src.buf[4 * (j * src_stride + i)];
src_diff = &p->src_diff[4 * (j * diff_stride + i)];
switch (tx_size) {
case TX_32X32:
scan_order = &vp9_default_scan_orders[TX_32X32];
mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
vp9_predict_intra_block(xd, block >> 6, bwl, TX_32X32, mode,
x->skip_encode ? src : dst,
x->skip_encode ? src_stride : dst_stride,
dst, dst_stride, i, j, plane);
if (!x->skip_recode) {
vp9_subtract_block(32, 32, src_diff, diff_stride,
src, src_stride, dst, dst_stride);
fdct32x32(x->use_lp32x32fdct, src_diff, coeff, diff_stride);
vp9_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, p->round,
p->quant, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, scan_order->scan,
scan_order->iscan);
}
if (!x->skip_encode && *eob)
vp9_idct32x32_add(dqcoeff, dst, dst_stride, *eob);
break;
case TX_16X16:
tx_type = get_tx_type(pd->plane_type, xd);
scan_order = &vp9_scan_orders[TX_16X16][tx_type];
mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
vp9_predict_intra_block(xd, block >> 4, bwl, TX_16X16, mode,
x->skip_encode ? src : dst,
x->skip_encode ? src_stride : dst_stride,
dst, dst_stride, i, j, plane);
if (!x->skip_recode) {
vp9_subtract_block(16, 16, src_diff, diff_stride,
src, src_stride, dst, dst_stride);
vp9_fht16x16(src_diff, coeff, diff_stride, tx_type);
vp9_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
p->quant, p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, scan_order->scan,
scan_order->iscan);
}
if (!x->skip_encode && *eob)
vp9_iht16x16_add(tx_type, dqcoeff, dst, dst_stride, *eob);
break;
case TX_8X8:
tx_type = get_tx_type(pd->plane_type, xd);
scan_order = &vp9_scan_orders[TX_8X8][tx_type];
mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
vp9_predict_intra_block(xd, block >> 2, bwl, TX_8X8, mode,
x->skip_encode ? src : dst,
x->skip_encode ? src_stride : dst_stride,
dst, dst_stride, i, j, plane);
if (!x->skip_recode) {
vp9_subtract_block(8, 8, src_diff, diff_stride,
src, src_stride, dst, dst_stride);
vp9_fht8x8(src_diff, coeff, diff_stride, tx_type);
vp9_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round, p->quant,
p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, scan_order->scan,
scan_order->iscan);
}
if (!x->skip_encode && *eob)
vp9_iht8x8_add(tx_type, dqcoeff, dst, dst_stride, *eob);
break;
case TX_4X4:
tx_type = get_tx_type_4x4(pd->plane_type, xd, block);
scan_order = &vp9_scan_orders[TX_4X4][tx_type];
mode = plane == 0 ? get_y_mode(xd->mi[0], block) : mbmi->uv_mode;
vp9_predict_intra_block(xd, block, bwl, TX_4X4, mode,
x->skip_encode ? src : dst,
x->skip_encode ? src_stride : dst_stride,
dst, dst_stride, i, j, plane);
if (!x->skip_recode) {
vp9_subtract_block(4, 4, src_diff, diff_stride,
src, src_stride, dst, dst_stride);
if (tx_type != DCT_DCT)
vp9_fht4x4(src_diff, coeff, diff_stride, tx_type);
else
x->fwd_txm4x4(src_diff, coeff, diff_stride);
vp9_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round, p->quant,
p->quant_shift, qcoeff, dqcoeff,
pd->dequant, p->zbin_extra, eob, scan_order->scan,
scan_order->iscan);
}
if (!x->skip_encode && *eob) {
if (tx_type == DCT_DCT)
// this is like vp9_short_idct4x4 but has a special case around eob<=1
// which is significant (not just an optimization) for the lossless
// case.
xd->itxm_add(dqcoeff, dst, dst_stride, *eob);
else
vp9_iht4x4_16_add(dqcoeff, dst, dst_stride, tx_type);
}
break;
default:
assert(0);
}
if (*eob)
*(args->skip) = 0;
}
