vpx/vp9/encoder/vp9_bitstream.c

2406 lines
80 KiB
C
Raw Normal View History

2010-05-18 17:58:33 +02:00
/*
* Copyright (c) 2010 The WebM project authors. All Rights Reserved.
2010-05-18 17:58:33 +02:00
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
2010-05-18 17:58:33 +02:00
*/
#include <assert.h>
#include <stdio.h>
#include <limits.h>
2010-05-18 17:58:33 +02:00
#include "vp9/common/vp9_header.h"
#include "vp9/encoder/vp9_encodemv.h"
#include "vp9/common/vp9_entropymode.h"
#include "vp9/common/vp9_entropymv.h"
#include "vp9/common/vp9_findnearmv.h"
#include "vp9/common/vp9_tile_common.h"
#include "vp9/encoder/vp9_mcomp.h"
#include "vp9/common/vp9_systemdependent.h"
#include "vp9/common/vp9_pragmas.h"
#include "vpx/vpx_encoder.h"
2010-05-18 17:58:33 +02:00
#include "vpx_mem/vpx_mem.h"
#include "vp9/encoder/vp9_bitstream.h"
#include "vp9/encoder/vp9_segmentation.h"
#include "vp9/common/vp9_seg_common.h"
#include "vp9/common/vp9_pred_common.h"
#include "vp9/common/vp9_entropy.h"
#include "vp9/encoder/vp9_encodemv.h"
#include "vp9/common/vp9_entropymv.h"
#include "vp9/common/vp9_mvref_common.h"
#include "vp9/common/vp9_treecoder.h"
2010-05-18 17:58:33 +02:00
#if defined(SECTIONBITS_OUTPUT)
unsigned __int64 Sectionbits[500];
#endif
#ifdef ENTROPY_STATS
int intra_mode_stats[VP9_KF_BINTRAMODES]
[VP9_KF_BINTRAMODES]
[VP9_KF_BINTRAMODES];
vp9_coeff_stats tree_update_hist_4x4[BLOCK_TYPES];
vp9_coeff_stats tree_update_hist_8x8[BLOCK_TYPES];
vp9_coeff_stats tree_update_hist_16x16[BLOCK_TYPES];
vp9_coeff_stats tree_update_hist_32x32[BLOCK_TYPES];
2010-05-18 17:58:33 +02:00
extern unsigned int active_section;
#endif
#if CONFIG_CODE_ZEROGROUP
#ifdef ZPC_STATS
vp9_zpc_count zpc_stats_4x4;
vp9_zpc_count zpc_stats_8x8;
vp9_zpc_count zpc_stats_16x16;
vp9_zpc_count zpc_stats_32x32;
void init_zpcstats();
void update_zpcstats(VP9_COMMON *const cm);
void print_zpcstats();
#endif
#endif
2010-05-18 17:58:33 +02:00
#ifdef MODE_STATS
int count_mb_seg[4] = { 0, 0, 0, 0 };
#endif
#define vp9_cost_upd ((int)(vp9_cost_one(upd) - vp9_cost_zero(upd)) >> 8)
#define vp9_cost_upd256 ((int)(vp9_cost_one(upd) - vp9_cost_zero(upd)))
#define SEARCH_NEWP
static int update_bits[255];
static INLINE void write_le16(uint8_t *p, int value) {
p[0] = value;
p[1] = value >> 8;
}
static INLINE void write_le32(uint8_t *p, int value) {
p[0] = value;
p[1] = value >> 8;
p[2] = value >> 16;
p[3] = value >> 24;
}
static void compute_update_table() {
int i;
for (i = 0; i < 255; i++)
update_bits[i] = vp9_count_term_subexp(i, SUBEXP_PARAM, 255);
}
static int split_index(int i, int n, int modulus) {
int max1 = (n - 1 - modulus / 2) / modulus + 1;
if (i % modulus == modulus / 2) i = i / modulus;
else i = max1 + i - (i + modulus - modulus / 2) / modulus;
return i;
}
static int remap_prob(int v, int m) {
const int n = 256;
const int modulus = MODULUS_PARAM;
int i;
if ((m << 1) <= n)
i = vp9_recenter_nonneg(v, m) - 1;
else
i = vp9_recenter_nonneg(n - 1 - v, n - 1 - m) - 1;
i = split_index(i, n - 1, modulus);
return i;
}
static void write_prob_diff_update(vp9_writer *const bc,
vp9_prob newp, vp9_prob oldp) {
int delp = remap_prob(newp, oldp);
vp9_encode_term_subexp(bc, delp, SUBEXP_PARAM, 255);
}
static int prob_diff_update_cost(vp9_prob newp, vp9_prob oldp) {
int delp = remap_prob(newp, oldp);
return update_bits[delp] * 256;
}
2010-05-18 17:58:33 +02:00
static void update_mode(
vp9_writer *const bc,
int n,
const struct vp9_token tok[/* n */],
vp9_tree tree,
vp9_prob Pnew [/* n-1 */],
vp9_prob Pcur [/* n-1 */],
unsigned int bct [/* n-1 */] [2],
const unsigned int num_events[/* n */]
) {
unsigned int new_b = 0, old_b = 0;
int i = 0;
vp9_tree_probs_from_distribution(tree, Pnew, bct, num_events, 0);
n--;
do {
new_b += cost_branch(bct[i], Pnew[i]);
old_b += cost_branch(bct[i], Pcur[i]);
} while (++i < n);
if (new_b + (n << 8) < old_b) {
2010-05-18 17:58:33 +02:00
int i = 0;
vp9_write_bit(bc, 1);
2010-05-18 17:58:33 +02:00
do {
const vp9_prob p = Pnew[i];
2010-05-18 17:58:33 +02:00
vp9_write_literal(bc, Pcur[i] = p ? p : 1, 8);
} while (++i < n);
} else
vp9_write_bit(bc, 0);
2010-05-18 17:58:33 +02:00
}
static void update_mbintra_mode_probs(VP9_COMP* const cpi,
vp9_writer* const bc) {
VP9_COMMON *const cm = &cpi->common;
2010-05-18 17:58:33 +02:00
{
vp9_prob Pnew [VP9_YMODES - 1];
unsigned int bct [VP9_YMODES - 1] [2];
2010-05-18 17:58:33 +02:00
update_mode(
bc, VP9_YMODES, vp9_ymode_encodings, vp9_ymode_tree,
Pnew, cm->fc.ymode_prob, bct, (unsigned int *)cpi->ymode_count
);
update_mode(bc, VP9_I32X32_MODES, vp9_sb_ymode_encodings,
vp9_sb_ymode_tree, Pnew, cm->fc.sb_ymode_prob, bct,
(unsigned int *)cpi->sb_ymode_count);
}
2010-05-18 17:58:33 +02:00
}
void vp9_update_skip_probs(VP9_COMP *cpi) {
VP9_COMMON *const pc = &cpi->common;
int k;
for (k = 0; k < MBSKIP_CONTEXTS; ++k) {
pc->mbskip_pred_probs[k] = get_binary_prob(cpi->skip_false_count[k],
cpi->skip_true_count[k]);
}
}
static void update_switchable_interp_probs(VP9_COMP *cpi,
vp9_writer* const bc) {
VP9_COMMON *const pc = &cpi->common;
unsigned int branch_ct[32][2];
int i, j;
for (j = 0; j <= VP9_SWITCHABLE_FILTERS; ++j) {
vp9_tree_probs_from_distribution(
vp9_switchable_interp_tree,
pc->fc.switchable_interp_prob[j], branch_ct,
cpi->switchable_interp_count[j], 0);
for (i = 0; i < VP9_SWITCHABLE_FILTERS - 1; ++i) {
if (pc->fc.switchable_interp_prob[j][i] < 1)
pc->fc.switchable_interp_prob[j][i] = 1;
vp9_write_prob(bc, pc->fc.switchable_interp_prob[j][i]);
}
}
}
// This function updates the reference frame prediction stats
static void update_refpred_stats(VP9_COMP *cpi) {
VP9_COMMON *const cm = &cpi->common;
int i;
vp9_prob new_pred_probs[PREDICTION_PROBS];
int old_cost, new_cost;
// Set the prediction probability structures to defaults
if (cm->frame_type != KEY_FRAME) {
// From the prediction counts set the probabilities for each context
for (i = 0; i < PREDICTION_PROBS; i++) {
new_pred_probs[i] = get_binary_prob(cpi->ref_pred_count[i][0],
cpi->ref_pred_count[i][1]);
// Decide whether or not to update the reference frame probs.
// Returned costs are in 1/256 bit units.
old_cost =
(cpi->ref_pred_count[i][0] * vp9_cost_zero(cm->ref_pred_probs[i])) +
(cpi->ref_pred_count[i][1] * vp9_cost_one(cm->ref_pred_probs[i]));
new_cost =
(cpi->ref_pred_count[i][0] * vp9_cost_zero(new_pred_probs[i])) +
(cpi->ref_pred_count[i][1] * vp9_cost_one(new_pred_probs[i]));
// Cost saving must be >= 8 bits (2048 in these units)
if ((old_cost - new_cost) >= 2048) {
cpi->ref_pred_probs_update[i] = 1;
cm->ref_pred_probs[i] = new_pred_probs[i];
} else
cpi->ref_pred_probs_update[i] = 0;
}
}
}
// This function is called to update the mode probability context used to encode
// inter modes. It assumes the branch counts table has already been populated
// prior to the actual packing of the bitstream (in rd stage or dummy pack)
//
// The branch counts table is re-populated during the actual pack stage and in
// the decoder to facilitate backwards update of the context.
static void update_inter_mode_probs(VP9_COMMON *cm,
int mode_context[INTER_MODE_CONTEXTS][4]) {
int i, j;
unsigned int (*mv_ref_ct)[4][2];
vpx_memcpy(mode_context, cm->fc.vp9_mode_contexts,
sizeof(cm->fc.vp9_mode_contexts));
mv_ref_ct = cm->fc.mv_ref_ct;
for (i = 0; i < INTER_MODE_CONTEXTS; i++) {
for (j = 0; j < 4; j++) {
int new_prob, old_cost, new_cost;
// Work out cost of coding branches with the old and optimal probability
old_cost = cost_branch256(mv_ref_ct[i][j], mode_context[i][j]);
new_prob = get_binary_prob(mv_ref_ct[i][j][0], mv_ref_ct[i][j][1]);
new_cost = cost_branch256(mv_ref_ct[i][j], new_prob);
// If cost saving is >= 14 bits then update the mode probability.
// This is the approximate net cost of updating one probability given
// that the no update case ismuch more common than the update case.
if (new_cost <= (old_cost - (14 << 8))) {
mode_context[i][j] = new_prob;
}
}
}
}
static void write_ymode(vp9_writer *bc, int m, const vp9_prob *p) {
write_token(bc, vp9_ymode_tree, p, vp9_ymode_encodings + m);
2010-05-18 17:58:33 +02:00
}
static void kfwrite_ymode(vp9_writer *bc, int m, const vp9_prob *p) {
write_token(bc, vp9_kf_ymode_tree, p, vp9_kf_ymode_encodings + m);
2010-05-18 17:58:33 +02:00
}
static void write_sb_ymode(vp9_writer *bc, int m, const vp9_prob *p) {
write_token(bc, vp9_sb_ymode_tree, p, vp9_sb_ymode_encodings + m);
}
static void sb_kfwrite_ymode(vp9_writer *bc, int m, const vp9_prob *p) {
write_token(bc, vp9_uv_mode_tree, p, vp9_sb_kf_ymode_encodings + m);
}
static void write_i8x8_mode(vp9_writer *bc, int m, const vp9_prob *p) {
write_token(bc, vp9_i8x8_mode_tree, p, vp9_i8x8_mode_encodings + m);
}
static void write_uv_mode(vp9_writer *bc, int m, const vp9_prob *p) {
write_token(bc, vp9_uv_mode_tree, p, vp9_uv_mode_encodings + m);
2010-05-18 17:58:33 +02:00
}
static void write_bmode(vp9_writer *bc, int m, const vp9_prob *p) {
#if CONFIG_NEWBINTRAMODES
assert(m < B_CONTEXT_PRED - CONTEXT_PRED_REPLACEMENTS || m == B_CONTEXT_PRED);
if (m == B_CONTEXT_PRED) m -= CONTEXT_PRED_REPLACEMENTS;
#endif
write_token(bc, vp9_bmode_tree, p, vp9_bmode_encodings + m);
2010-05-18 17:58:33 +02:00
}
static void write_kf_bmode(vp9_writer *bc, int m, const vp9_prob *p) {
write_token(bc, vp9_kf_bmode_tree, p, vp9_kf_bmode_encodings + m);
}
static void write_split(vp9_writer *bc, int x, const vp9_prob *p) {
write_token(
bc, vp9_mbsplit_tree, p, vp9_mbsplit_encodings + x);
2010-05-18 17:58:33 +02:00
}
static int prob_update_savings(const unsigned int *ct,
const vp9_prob oldp, const vp9_prob newp,
const vp9_prob upd) {
const int old_b = cost_branch256(ct, oldp);
const int new_b = cost_branch256(ct, newp);
const int update_b = 2048 + vp9_cost_upd256;
return (old_b - new_b - update_b);
}
static int prob_diff_update_savings(const unsigned int *ct,
const vp9_prob oldp, const vp9_prob newp,
const vp9_prob upd) {
const int old_b = cost_branch256(ct, oldp);
const int new_b = cost_branch256(ct, newp);
const int update_b = (newp == oldp ? 0 :
prob_diff_update_cost(newp, oldp) + vp9_cost_upd256);
return (old_b - new_b - update_b);
}
static int prob_diff_update_savings_search(const unsigned int *ct,
const vp9_prob oldp, vp9_prob *bestp,
const vp9_prob upd) {
const int old_b = cost_branch256(ct, oldp);
int new_b, update_b, savings, bestsavings, step;
vp9_prob newp, bestnewp;
bestsavings = 0;
bestnewp = oldp;
step = (*bestp > oldp ? -1 : 1);
for (newp = *bestp; newp != oldp; newp += step) {
new_b = cost_branch256(ct, newp);
update_b = prob_diff_update_cost(newp, oldp) + vp9_cost_upd256;
savings = old_b - new_b - update_b;
if (savings > bestsavings) {
bestsavings = savings;
bestnewp = newp;
}
}
*bestp = bestnewp;
return bestsavings;
}
Modeling default coef probs with distribution Replaces the default tables for single coefficient magnitudes with those obtained from an appropriate distribution. The EOB node is left unchanged. The model is represeted as a 256-size codebook where the index corresponds to the probability of the Zero or the One node. Two variations are implemented corresponding to whether the Zero node or the One-node is used as the peg. The main advantage is that the default prob tables will become considerably smaller and manageable. Besides there is substantially less risk of over-fitting for a training set. Various distributions are tried and the one that gives the best results is the family of Generalized Gaussian distributions with shape parameter 0.75. The results are within about 0.2% of fully trained tables for the Zero peg variant, and within 0.1% of the One peg variant. The forward updates are optionally (controlled by a macro) model-based, i.e. restricted to only convey probabilities from the codebook. Backward updates can also be optionally (controlled by another macro) model-based, but is turned off by default. Currently model-based forward updates work about the same as unconstrained updates, but there is a drop in performance with backward-updates being model based. The model based approach also allows the probabilities for the key frames to be adjusted from the defaults based on the base_qindex of the frame. Currently the adjustment function is a placeholder that adjusts the prob of EOB and Zero node from the nominal one at higher quality (lower qindex) or lower quality (higher qindex) ends of the range. The rest of the probabilities are then derived based on the model from the adjusted prob of zero. Change-Id: Iae050f3cbcc6d8b3f204e8dc395ae47b3b2192c9
2013-03-13 19:03:17 +01:00
#if CONFIG_MODELCOEFPROB && MODEL_BASED_UPDATE
static int prob_diff_update_savings_search_model(const unsigned int *ct,
const vp9_prob *oldp,
vp9_prob *bestp,
const vp9_prob upd,
int b, int r, int q) {
Modeling default coef probs with distribution Replaces the default tables for single coefficient magnitudes with those obtained from an appropriate distribution. The EOB node is left unchanged. The model is represeted as a 256-size codebook where the index corresponds to the probability of the Zero or the One node. Two variations are implemented corresponding to whether the Zero node or the One-node is used as the peg. The main advantage is that the default prob tables will become considerably smaller and manageable. Besides there is substantially less risk of over-fitting for a training set. Various distributions are tried and the one that gives the best results is the family of Generalized Gaussian distributions with shape parameter 0.75. The results are within about 0.2% of fully trained tables for the Zero peg variant, and within 0.1% of the One peg variant. The forward updates are optionally (controlled by a macro) model-based, i.e. restricted to only convey probabilities from the codebook. Backward updates can also be optionally (controlled by another macro) model-based, but is turned off by default. Currently model-based forward updates work about the same as unconstrained updates, but there is a drop in performance with backward-updates being model based. The model based approach also allows the probabilities for the key frames to be adjusted from the defaults based on the base_qindex of the frame. Currently the adjustment function is a placeholder that adjusts the prob of EOB and Zero node from the nominal one at higher quality (lower qindex) or lower quality (higher qindex) ends of the range. The rest of the probabilities are then derived based on the model from the adjusted prob of zero. Change-Id: Iae050f3cbcc6d8b3f204e8dc395ae47b3b2192c9
2013-03-13 19:03:17 +01:00
int i, old_b, new_b, update_b, savings, bestsavings, step;
int newp;
vp9_prob bestnewp, newplist[ENTROPY_NODES];
for (i = UNCONSTRAINED_NODES - 1, old_b = 0; i < ENTROPY_NODES; ++i)
old_b += cost_branch256(ct + 2 * i, oldp[i]);
bestsavings = 0;
bestnewp = oldp[UNCONSTRAINED_NODES - 1];
step = (*bestp > oldp[UNCONSTRAINED_NODES - 1] ? -1 : 1);
newp = *bestp;
// newp = *bestp - step * (abs(*bestp - oldp[UNCONSTRAINED_NODES - 1]) >> 1);
for (; newp != oldp[UNCONSTRAINED_NODES - 1]; newp += step) {
if (newp < 1 || newp > 255) continue;
newplist[UNCONSTRAINED_NODES - 1] = newp;
vp9_get_model_distribution(newp, newplist, b, r);
for (i = UNCONSTRAINED_NODES - 1, new_b = 0; i < ENTROPY_NODES; ++i)
new_b += cost_branch256(ct + 2 * i, newplist[i]);
update_b = prob_diff_update_cost(newp, oldp[UNCONSTRAINED_NODES - 1]) +
vp9_cost_upd256;
savings = old_b - new_b - update_b;
if (savings > bestsavings) {
bestsavings = savings;
bestnewp = newp;
}
}
*bestp = bestnewp;
return bestsavings;
}
#endif
static void vp9_cond_prob_update(vp9_writer *bc, vp9_prob *oldp, vp9_prob upd,
unsigned int *ct) {
vp9_prob newp;
int savings;
newp = get_binary_prob(ct[0], ct[1]);
savings = prob_update_savings(ct, *oldp, newp, upd);
if (savings > 0) {
vp9_write(bc, 1, upd);
vp9_write_prob(bc, newp);
*oldp = newp;
} else {
vp9_write(bc, 0, upd);
}
}
static void pack_mb_tokens(vp9_writer* const bc,
TOKENEXTRA **tp,
const TOKENEXTRA *const stop) {
TOKENEXTRA *p = *tp;
2010-05-18 17:58:33 +02:00
while (p < stop) {
const int t = p->token;
const struct vp9_token *const a = vp9_coef_encodings + t;
const vp9_extra_bit *const b = vp9_extra_bits + t;
int i = 0;
const unsigned char *pp = p->context_tree;
int v = a->value;
int n = a->len;
int ncount = n;
if (t == EOSB_TOKEN)
{
++p;
break;
}
assert(pp != 0);
#if CONFIG_CODE_ZEROGROUP
if (t == ZPC_ISOLATED || t == ZPC_EOORIENT) {
assert((p - 1)->token == ZERO_TOKEN);
encode_bool(bc, t == ZPC_ISOLATED, *pp);
++p;
continue;
} else if (p->skip_coef_val) {
assert(p->skip_eob_node == 0);
assert(t == DCT_EOB_TOKEN || t == ZERO_TOKEN);
encode_bool(bc, t == ZERO_TOKEN, *pp);
++p;
continue;
}
#endif
/* skip one or two nodes */
if (p->skip_eob_node) {
n -= p->skip_eob_node;
i = 2 * p->skip_eob_node;
ncount -= p->skip_eob_node;
}
2010-05-18 17:58:33 +02:00
do {
const int bb = (v >> --n) & 1;
vp9_write(bc, bb, pp[i >> 1]);
i = vp9_coef_tree[i + bb];
ncount--;
} while (n && ncount);
2010-05-18 17:58:33 +02:00
if (b->base_val) {
const int e = p->extra, l = b->len;
2010-05-18 17:58:33 +02:00
if (l) {
const unsigned char *pp = b->prob;
int v = e >> 1;
int n = l; /* number of bits in v, assumed nonzero */
int i = 0;
2010-05-18 17:58:33 +02:00
do {
const int bb = (v >> --n) & 1;
vp9_write(bc, bb, pp[i >> 1]);
i = b->tree[i + bb];
} while (n);
}
2010-05-18 17:58:33 +02:00
vp9_write_bit(bc, e & 1);
2010-05-18 17:58:33 +02:00
}
++p;
}
*tp = p;
2010-05-18 17:58:33 +02:00
}
static void write_mv_ref
(
vp9_writer *bc, MB_PREDICTION_MODE m, const vp9_prob *p
) {
#if CONFIG_DEBUG
assert(NEARESTMV <= m && m <= SPLITMV);
#endif
write_token(bc, vp9_mv_ref_tree, p,
vp9_mv_ref_encoding_array - NEARESTMV + m);
2010-05-18 17:58:33 +02:00
}
static void write_sb_mv_ref(vp9_writer *bc, MB_PREDICTION_MODE m,
const vp9_prob *p) {
#if CONFIG_DEBUG
assert(NEARESTMV <= m && m < SPLITMV);
#endif
write_token(bc, vp9_sb_mv_ref_tree, p,
vp9_sb_mv_ref_encoding_array - NEARESTMV + m);
}
2010-05-18 17:58:33 +02:00
static void write_sub_mv_ref
(
vp9_writer *bc, B_PREDICTION_MODE m, const vp9_prob *p
) {
#if CONFIG_DEBUG
assert(LEFT4X4 <= m && m <= NEW4X4);
#endif
write_token(bc, vp9_sub_mv_ref_tree, p,
vp9_sub_mv_ref_encoding_array - LEFT4X4 + m);
2010-05-18 17:58:33 +02:00
}
static void write_nmv(VP9_COMP *cpi, vp9_writer *bc,
const MV *mv, const int_mv *ref,
const nmv_context *nmvc, int usehp) {
MV e;
e.row = mv->row - ref->as_mv.row;
e.col = mv->col - ref->as_mv.col;
vp9_encode_nmv(bc, &e, &ref->as_mv, nmvc);
vp9_encode_nmv_fp(bc, &e, &ref->as_mv, nmvc, usehp);
}
// This function writes the current macro block's segnment id to the bitstream
// It should only be called if a segment map update is indicated.
static void write_mb_segid(vp9_writer *bc,
const MB_MODE_INFO *mi, const MACROBLOCKD *xd) {
// Encode the MB segment id.
int seg_id = mi->segment_id;
if (xd->segmentation_enabled && xd->update_mb_segmentation_map) {
switch (seg_id) {
case 0:
vp9_write(bc, 0, xd->mb_segment_tree_probs[0]);
vp9_write(bc, 0, xd->mb_segment_tree_probs[1]);
vp9_write(bc, 0, xd->mb_segment_tree_probs[2]);
break;
case 1:
vp9_write(bc, 0, xd->mb_segment_tree_probs[0]);
vp9_write(bc, 0, xd->mb_segment_tree_probs[1]);
vp9_write(bc, 1, xd->mb_segment_tree_probs[2]);
break;
case 2:
vp9_write(bc, 0, xd->mb_segment_tree_probs[0]);
vp9_write(bc, 1, xd->mb_segment_tree_probs[1]);
vp9_write(bc, 0, xd->mb_segment_tree_probs[3]);
break;
case 3:
vp9_write(bc, 0, xd->mb_segment_tree_probs[0]);
vp9_write(bc, 1, xd->mb_segment_tree_probs[1]);
vp9_write(bc, 1, xd->mb_segment_tree_probs[3]);
break;
case 4:
vp9_write(bc, 1, xd->mb_segment_tree_probs[0]);
vp9_write(bc, 0, xd->mb_segment_tree_probs[4]);
vp9_write(bc, 0, xd->mb_segment_tree_probs[5]);
break;
case 5:
vp9_write(bc, 1, xd->mb_segment_tree_probs[0]);
vp9_write(bc, 0, xd->mb_segment_tree_probs[4]);
vp9_write(bc, 1, xd->mb_segment_tree_probs[5]);
break;
case 6:
vp9_write(bc, 1, xd->mb_segment_tree_probs[0]);
vp9_write(bc, 1, xd->mb_segment_tree_probs[4]);
vp9_write(bc, 0, xd->mb_segment_tree_probs[6]);
break;
case 7:
vp9_write(bc, 1, xd->mb_segment_tree_probs[0]);
vp9_write(bc, 1, xd->mb_segment_tree_probs[4]);
vp9_write(bc, 1, xd->mb_segment_tree_probs[6]);
break;
// TRAP.. This should not happen
default:
vp9_write(bc, 0, xd->mb_segment_tree_probs[0]);
vp9_write(bc, 0, xd->mb_segment_tree_probs[1]);
vp9_write(bc, 0, xd->mb_segment_tree_probs[2]);
break;
2010-05-18 17:58:33 +02:00
}
}
2010-05-18 17:58:33 +02:00
}
// This function encodes the reference frame
static void encode_ref_frame(vp9_writer *const bc,
VP9_COMMON *const cm,
MACROBLOCKD *xd,
int segment_id,
MV_REFERENCE_FRAME rf) {
int seg_ref_active;
int seg_ref_count = 0;
seg_ref_active = vp9_segfeature_active(xd,
segment_id,
SEG_LVL_REF_FRAME);
if (seg_ref_active) {
seg_ref_count = vp9_check_segref(xd, segment_id, INTRA_FRAME) +
vp9_check_segref(xd, segment_id, LAST_FRAME) +
vp9_check_segref(xd, segment_id, GOLDEN_FRAME) +
vp9_check_segref(xd, segment_id, ALTREF_FRAME);
}
// If segment level coding of this signal is disabled...
// or the segment allows multiple reference frame options
if (!seg_ref_active || (seg_ref_count > 1)) {
// Values used in prediction model coding
unsigned char prediction_flag;
vp9_prob pred_prob;
MV_REFERENCE_FRAME pred_rf;
// Get the context probability the prediction flag
pred_prob = vp9_get_pred_prob(cm, xd, PRED_REF);
// Get the predicted value.
pred_rf = vp9_get_pred_ref(cm, xd);
// Did the chosen reference frame match its predicted value.
prediction_flag =
(xd->mode_info_context->mbmi.ref_frame == pred_rf);
vp9_set_pred_flag(xd, PRED_REF, prediction_flag);
vp9_write(bc, prediction_flag, pred_prob);
// If not predicted correctly then code value explicitly
if (!prediction_flag) {
vp9_prob mod_refprobs[PREDICTION_PROBS];
vpx_memcpy(mod_refprobs,
cm->mod_refprobs[pred_rf], sizeof(mod_refprobs));
// If segment coding enabled blank out options that cant occur by
// setting the branch probability to 0.
if (seg_ref_active) {
mod_refprobs[INTRA_FRAME] *=
vp9_check_segref(xd, segment_id, INTRA_FRAME);
mod_refprobs[LAST_FRAME] *=
vp9_check_segref(xd, segment_id, LAST_FRAME);
mod_refprobs[GOLDEN_FRAME] *=
(vp9_check_segref(xd, segment_id, GOLDEN_FRAME) *
vp9_check_segref(xd, segment_id, ALTREF_FRAME));
}
if (mod_refprobs[0]) {
vp9_write(bc, (rf != INTRA_FRAME), mod_refprobs[0]);
}
// Inter coded
if (rf != INTRA_FRAME) {
if (mod_refprobs[1]) {
vp9_write(bc, (rf != LAST_FRAME), mod_refprobs[1]);
}
if (rf != LAST_FRAME) {
if (mod_refprobs[2]) {
vp9_write(bc, (rf != GOLDEN_FRAME), mod_refprobs[2]);
}
}
}
}
}
// if using the prediction mdoel we have nothing further to do because
// the reference frame is fully coded by the segment
}
2010-05-18 17:58:33 +02:00
// Update the probabilities used to encode reference frame data
static void update_ref_probs(VP9_COMP *const cpi) {
VP9_COMMON *const cm = &cpi->common;
const int *const rfct = cpi->count_mb_ref_frame_usage;
const int rf_intra = rfct[INTRA_FRAME];
const int rf_inter = rfct[LAST_FRAME] +
rfct[GOLDEN_FRAME] + rfct[ALTREF_FRAME];
cm->prob_intra_coded = get_binary_prob(rf_intra, rf_inter);
cm->prob_last_coded = get_prob(rfct[LAST_FRAME], rf_inter);
cm->prob_gf_coded = get_binary_prob(rfct[GOLDEN_FRAME], rfct[ALTREF_FRAME]);
// Compute a modified set of probabilities to use when prediction of the
// reference frame fails
vp9_compute_mod_refprobs(cm);
}
static void pack_inter_mode_mvs(VP9_COMP *cpi, MODE_INFO *m,
vp9_writer *bc, int mi_row, int mi_col) {
VP9_COMMON *const pc = &cpi->common;
const nmv_context *nmvc = &pc->fc.nmvc;
MACROBLOCK *const x = &cpi->mb;
MACROBLOCKD *const xd = &x->e_mbd;
const int mis = pc->mode_info_stride;
MB_MODE_INFO *const mi = &m->mbmi;
const MV_REFERENCE_FRAME rf = mi->ref_frame;
const MB_PREDICTION_MODE mode = mi->mode;
const int segment_id = mi->segment_id;
int skip_coeff;
xd->prev_mode_info_context = pc->prev_mi + (m - pc->mi);
x->partition_info = x->pi + (m - pc->mi);
#ifdef ENTROPY_STATS
active_section = 9;
#endif
if (cpi->mb.e_mbd.update_mb_segmentation_map) {
// Is temporal coding of the segment map enabled
if (pc->temporal_update) {
unsigned char prediction_flag = vp9_get_pred_flag(xd, PRED_SEG_ID);
vp9_prob pred_prob = vp9_get_pred_prob(pc, xd, PRED_SEG_ID);
// Code the segment id prediction flag for this mb
vp9_write(bc, prediction_flag, pred_prob);
// If the mb segment id wasn't predicted code explicitly
if (!prediction_flag)
write_mb_segid(bc, mi, &cpi->mb.e_mbd);
} else {
// Normal unpredicted coding
write_mb_segid(bc, mi, &cpi->mb.e_mbd);
}
}
if (vp9_segfeature_active(xd, segment_id, SEG_LVL_SKIP)) {
skip_coeff = 1;
} else {
skip_coeff = m->mbmi.mb_skip_coeff;
vp9_write(bc, skip_coeff,
vp9_get_pred_prob(pc, xd, PRED_MBSKIP));
}
2010-05-18 17:58:33 +02:00
// Encode the reference frame.
encode_ref_frame(bc, pc, xd, segment_id, rf);
if (rf == INTRA_FRAME) {
#ifdef ENTROPY_STATS
active_section = 6;
#endif
if (m->mbmi.sb_type > BLOCK_SIZE_MB16X16)
write_sb_ymode(bc, mode, pc->fc.sb_ymode_prob);
else
write_ymode(bc, mode, pc->fc.ymode_prob);
if (mode == I4X4_PRED) {
int j = 0;
do {
write_bmode(bc, m->bmi[j].as_mode.first,
pc->fc.bmode_prob);
} while (++j < 16);
}
if (mode == I8X8_PRED) {
write_i8x8_mode(bc, m->bmi[0].as_mode.first,
pc->fc.i8x8_mode_prob);
write_i8x8_mode(bc, m->bmi[2].as_mode.first,
pc->fc.i8x8_mode_prob);
write_i8x8_mode(bc, m->bmi[8].as_mode.first,
pc->fc.i8x8_mode_prob);
write_i8x8_mode(bc, m->bmi[10].as_mode.first,
pc->fc.i8x8_mode_prob);
} else {
write_uv_mode(bc, mi->uv_mode,
pc->fc.uv_mode_prob[mode]);
}
} else {
vp9_prob mv_ref_p[VP9_MVREFS - 1];
vp9_mv_ref_probs(&cpi->common, mv_ref_p, mi->mb_mode_context[rf]);
2010-05-18 17:58:33 +02:00
#ifdef ENTROPY_STATS
active_section = 3;
2010-05-18 17:58:33 +02:00
#endif
// If segment skip is not enabled code the mode.
if (!vp9_segfeature_active(xd, segment_id, SEG_LVL_SKIP)) {
if (mi->sb_type > BLOCK_SIZE_MB16X16) {
write_sb_mv_ref(bc, mode, mv_ref_p);
} else {
write_mv_ref(bc, mode, mv_ref_p);
}
vp9_accum_mv_refs(&cpi->common, mode, mi->mb_mode_context[rf]);
}
if (is_inter_mode(mode)) {
if (cpi->common.mcomp_filter_type == SWITCHABLE) {
write_token(bc, vp9_switchable_interp_tree,
vp9_get_pred_probs(&cpi->common, xd,
PRED_SWITCHABLE_INTERP),
vp9_switchable_interp_encodings +
vp9_switchable_interp_map[mi->interp_filter]);
} else {
assert(mi->interp_filter == cpi->common.mcomp_filter_type);
}
}
// does the feature use compound prediction or not
// (if not specified at the frame/segment level)
if (cpi->common.comp_pred_mode == HYBRID_PREDICTION) {
vp9_write(bc, mi->second_ref_frame > INTRA_FRAME,
vp9_get_pred_prob(pc, xd, PRED_COMP));
}
#if CONFIG_COMP_INTERINTRA_PRED
if (cpi->common.use_interintra &&
mode >= NEARESTMV && mode < SPLITMV &&
mi->second_ref_frame <= INTRA_FRAME) {
vp9_write(bc, mi->second_ref_frame == INTRA_FRAME,
pc->fc.interintra_prob);
// if (!cpi->dummy_packing)
// printf("-- %d (%d)\n", mi->second_ref_frame == INTRA_FRAME,
// pc->fc.interintra_prob);
if (mi->second_ref_frame == INTRA_FRAME) {
// if (!cpi->dummy_packing)
// printf("** %d %d\n", mi->interintra_mode,
// mi->interintra_uv_mode);
write_ymode(bc, mi->interintra_mode, pc->fc.ymode_prob);
#if SEPARATE_INTERINTRA_UV
write_uv_mode(bc, mi->interintra_uv_mode,
pc->fc.uv_mode_prob[mi->interintra_mode]);
#endif
}
}
#endif
switch (mode) { /* new, split require MVs */
case NEWMV:
#ifdef ENTROPY_STATS
active_section = 5;
#endif
write_nmv(cpi, bc, &mi->mv[0].as_mv, &mi->best_mv,
(const nmv_context*) nmvc,
xd->allow_high_precision_mv);
if (mi->second_ref_frame > 0) {
write_nmv(cpi, bc, &mi->mv[1].as_mv, &mi->best_second_mv,
(const nmv_context*) nmvc,
xd->allow_high_precision_mv);
}
break;
case SPLITMV: {
int j = 0;
2010-05-18 17:58:33 +02:00
#ifdef MODE_STATS
++count_mb_seg[mi->partitioning];
#endif
2010-05-18 17:58:33 +02:00
write_split(bc, mi->partitioning, cpi->common.fc.mbsplit_prob);
cpi->mbsplit_count[mi->partitioning]++;
do {
B_PREDICTION_MODE blockmode;
int_mv blockmv;
const int *const L = vp9_mbsplits[mi->partitioning];
int k = -1; /* first block in subset j */
int mv_contz;
int_mv leftmv, abovemv;
blockmode = cpi->mb.partition_info->bmi[j].mode;
blockmv = cpi->mb.partition_info->bmi[j].mv;
#if CONFIG_DEBUG
while (j != L[++k])
if (k >= 16)
assert(0);
#else
while (j != L[++k]);
#endif
[WIP] Add column-based tiling. This patch adds column-based tiling. The idea is to make each tile independently decodable (after reading the common frame header) and also independendly encodable (minus within-frame cost adjustments in the RD loop) to speed-up hardware & software en/decoders if they used multi-threading. Column-based tiling has the added advantage (over other tiling methods) that it minimizes realtime use-case latency, since all threads can start encoding data as soon as the first SB-row worth of data is available to the encoder. There is some test code that does random tile ordering in the decoder, to confirm that each tile is indeed independently decodable from other tiles in the same frame. At tile edges, all contexts assume default values (i.e. 0, 0 motion vector, no coefficients, DC intra4x4 mode), and motion vector search and ordering do not cross tiles in the same frame. t log Tile independence is not maintained between frames ATM, i.e. tile 0 of frame 1 is free to use motion vectors that point into any tile of frame 0. We support 1 (i.e. no tiling), 2 or 4 column-tiles. The loopfilter crosses tile boundaries. I discussed this briefly with Aki and he says that's OK. An in-loop loopfilter would need to do some sync between tile threads, but that shouldn't be a big issue. Resuls: with tiling disabled, we go up slightly because of improved edge use in the intra4x4 prediction. With 2 tiles, we lose about ~1% on derf, ~0.35% on HD and ~0.55% on STD/HD. With 4 tiles, we lose another ~1.5% on derf ~0.77% on HD and ~0.85% on STD/HD. Most of this loss is concentrated in the low-bitrate end of clips, and most of it is because of the loss of edges at tile boundaries and the resulting loss of intra predictors. TODO: - more tiles (perhaps allow row-based tiling also, and max. 8 tiles)? - maybe optionally (for EC purposes), motion vectors themselves should not cross tile edges, or we should emulate such borders as if they were off-frame, to limit error propagation to within one tile only. This doesn't have to be the default behaviour but could be an optional bitstream flag. Change-Id: I5951c3a0742a767b20bc9fb5af685d9892c2c96f
2013-02-01 18:35:28 +01:00
leftmv.as_int = left_block_mv(xd, m, k);
abovemv.as_int = above_block_mv(m, k, mis);
mv_contz = vp9_mv_cont(&leftmv, &abovemv);
write_sub_mv_ref(bc, blockmode,
cpi->common.fc.sub_mv_ref_prob[mv_contz]);
cpi->sub_mv_ref_count[mv_contz][blockmode - LEFT4X4]++;
if (blockmode == NEW4X4) {
#ifdef ENTROPY_STATS
active_section = 11;
#endif
write_nmv(cpi, bc, &blockmv.as_mv, &mi->best_mv,
(const nmv_context*) nmvc,
xd->allow_high_precision_mv);
if (mi->second_ref_frame > 0) {
write_nmv(cpi, bc,
&cpi->mb.partition_info->bmi[j].second_mv.as_mv,
&mi->best_second_mv,
(const nmv_context*) nmvc,
xd->allow_high_precision_mv);
}
}
} while (++j < cpi->mb.partition_info->count);
break;
}
default:
break;
}
}
2010-05-18 17:58:33 +02:00
if (((rf == INTRA_FRAME && mode <= I8X8_PRED) ||
(rf != INTRA_FRAME && !(mode == SPLITMV &&
mi->partitioning == PARTITIONING_4X4))) &&
pc->txfm_mode == TX_MODE_SELECT &&
!(skip_coeff || vp9_segfeature_active(xd, segment_id,
SEG_LVL_SKIP))) {
TX_SIZE sz = mi->txfm_size;
// FIXME(rbultje) code ternary symbol once all experiments are merged
vp9_write(bc, sz != TX_4X4, pc->prob_tx[0]);
if (sz != TX_4X4 && mode != I8X8_PRED && mode != SPLITMV) {
vp9_write(bc, sz != TX_8X8, pc->prob_tx[1]);
if (mi->sb_type >= BLOCK_SIZE_SB32X32 && sz != TX_8X8)
vp9_write(bc, sz != TX_16X16, pc->prob_tx[2]);
2010-05-18 17:58:33 +02:00
}
}
2010-05-18 17:58:33 +02:00
}
static void write_mb_modes_kf(const VP9_COMP *cpi,
MODE_INFO *m,
vp9_writer *bc, int mi_row, int mi_col) {
const VP9_COMMON *const c = &cpi->common;
const MACROBLOCKD *const xd = &cpi->mb.e_mbd;
const int mis = c->mode_info_stride;
const int ym = m->mbmi.mode;
const int segment_id = m->mbmi.segment_id;
int skip_coeff;
if (xd->update_mb_segmentation_map) {
write_mb_segid(bc, &m->mbmi, xd);
}
if (vp9_segfeature_active(xd, segment_id, SEG_LVL_SKIP)) {
skip_coeff = 1;
} else {
skip_coeff = m->mbmi.mb_skip_coeff;
vp9_write(bc, skip_coeff,
vp9_get_pred_prob(c, xd, PRED_MBSKIP));
}
if (m->mbmi.sb_type > BLOCK_SIZE_MB16X16) {
sb_kfwrite_ymode(bc, ym,
c->sb_kf_ymode_prob[c->kf_ymode_probs_index]);
} else {
kfwrite_ymode(bc, ym,
c->kf_ymode_prob[c->kf_ymode_probs_index]);
}
if (ym == I4X4_PRED) {
int i = 0;
do {
const B_PREDICTION_MODE A = above_block_mode(m, i, mis);
[WIP] Add column-based tiling. This patch adds column-based tiling. The idea is to make each tile independently decodable (after reading the common frame header) and also independendly encodable (minus within-frame cost adjustments in the RD loop) to speed-up hardware & software en/decoders if they used multi-threading. Column-based tiling has the added advantage (over other tiling methods) that it minimizes realtime use-case latency, since all threads can start encoding data as soon as the first SB-row worth of data is available to the encoder. There is some test code that does random tile ordering in the decoder, to confirm that each tile is indeed independently decodable from other tiles in the same frame. At tile edges, all contexts assume default values (i.e. 0, 0 motion vector, no coefficients, DC intra4x4 mode), and motion vector search and ordering do not cross tiles in the same frame. t log Tile independence is not maintained between frames ATM, i.e. tile 0 of frame 1 is free to use motion vectors that point into any tile of frame 0. We support 1 (i.e. no tiling), 2 or 4 column-tiles. The loopfilter crosses tile boundaries. I discussed this briefly with Aki and he says that's OK. An in-loop loopfilter would need to do some sync between tile threads, but that shouldn't be a big issue. Resuls: with tiling disabled, we go up slightly because of improved edge use in the intra4x4 prediction. With 2 tiles, we lose about ~1% on derf, ~0.35% on HD and ~0.55% on STD/HD. With 4 tiles, we lose another ~1.5% on derf ~0.77% on HD and ~0.85% on STD/HD. Most of this loss is concentrated in the low-bitrate end of clips, and most of it is because of the loss of edges at tile boundaries and the resulting loss of intra predictors. TODO: - more tiles (perhaps allow row-based tiling also, and max. 8 tiles)? - maybe optionally (for EC purposes), motion vectors themselves should not cross tile edges, or we should emulate such borders as if they were off-frame, to limit error propagation to within one tile only. This doesn't have to be the default behaviour but could be an optional bitstream flag. Change-Id: I5951c3a0742a767b20bc9fb5af685d9892c2c96f
2013-02-01 18:35:28 +01:00
const B_PREDICTION_MODE L = (xd->left_available || (i & 3)) ?
left_block_mode(m, i) : B_DC_PRED;
const int bm = m->bmi[i].as_mode.first;
#ifdef ENTROPY_STATS
++intra_mode_stats [A] [L] [bm];
#endif
write_kf_bmode(bc, bm, c->kf_bmode_prob[A][L]);
} while (++i < 16);
}
if (ym == I8X8_PRED) {
write_i8x8_mode(bc, m->bmi[0].as_mode.first,
c->fc.i8x8_mode_prob);
// printf(" mode: %d\n", m->bmi[0].as_mode.first); fflush(stdout);
write_i8x8_mode(bc, m->bmi[2].as_mode.first,
c->fc.i8x8_mode_prob);
// printf(" mode: %d\n", m->bmi[2].as_mode.first); fflush(stdout);
write_i8x8_mode(bc, m->bmi[8].as_mode.first,
c->fc.i8x8_mode_prob);
// printf(" mode: %d\n", m->bmi[8].as_mode.first); fflush(stdout);
write_i8x8_mode(bc, m->bmi[10].as_mode.first,
c->fc.i8x8_mode_prob);
// printf(" mode: %d\n", m->bmi[10].as_mode.first); fflush(stdout);
} else
write_uv_mode(bc, m->mbmi.uv_mode, c->kf_uv_mode_prob[ym]);
if (ym <= I8X8_PRED && c->txfm_mode == TX_MODE_SELECT &&
!(skip_coeff || vp9_segfeature_active(xd, segment_id, SEG_LVL_SKIP))) {
TX_SIZE sz = m->mbmi.txfm_size;
// FIXME(rbultje) code ternary symbol once all experiments are merged
vp9_write(bc, sz != TX_4X4, c->prob_tx[0]);
32x32 transform for superblocks. This adds Debargha's DCT/DWT hybrid and a regular 32x32 DCT, and adds code all over the place to wrap that in the bitstream/encoder/decoder/RD. Some implementation notes (these probably need careful review): - token range is extended by 1 bit, since the value range out of this transform is [-16384,16383]. - the coefficients coming out of the FDCT are manually scaled back by 1 bit, or else they won't fit in int16_t (they are 17 bits). Because of this, the RD error scoring does not right-shift the MSE score by two (unlike for 4x4/8x8/16x16). - to compensate for this loss in precision, the quantizer is halved also. This is currently a little hacky. - FDCT and IDCT is double-only right now. Needs a fixed-point impl. - There are no default probabilities for the 32x32 transform yet; I'm simply using the 16x16 luma ones. A future commit will add newly generated probabilities for all transforms. - No ADST version. I don't think we'll add one for this level; if an ADST is desired, transform-size selection can scale back to 16x16 or lower, and use an ADST at that level. Additional notes specific to Debargha's DWT/DCT hybrid: - coefficient scale is different for the top/left 16x16 (DCT-over-DWT) block than for the rest (DWT pixel differences) of the block. Therefore, RD error scoring isn't easily scalable between coefficient and pixel domain. Thus, unfortunately, we need to compute the RD distortion in the pixel domain until we figure out how to scale these appropriately. Change-Id: I00386f20f35d7fabb19aba94c8162f8aee64ef2b
2012-12-07 23:45:05 +01:00
if (sz != TX_4X4 && ym <= TM_PRED) {
vp9_write(bc, sz != TX_8X8, c->prob_tx[1]);
if (m->mbmi.sb_type >= BLOCK_SIZE_SB32X32 && sz != TX_8X8)
32x32 transform for superblocks. This adds Debargha's DCT/DWT hybrid and a regular 32x32 DCT, and adds code all over the place to wrap that in the bitstream/encoder/decoder/RD. Some implementation notes (these probably need careful review): - token range is extended by 1 bit, since the value range out of this transform is [-16384,16383]. - the coefficients coming out of the FDCT are manually scaled back by 1 bit, or else they won't fit in int16_t (they are 17 bits). Because of this, the RD error scoring does not right-shift the MSE score by two (unlike for 4x4/8x8/16x16). - to compensate for this loss in precision, the quantizer is halved also. This is currently a little hacky. - FDCT and IDCT is double-only right now. Needs a fixed-point impl. - There are no default probabilities for the 32x32 transform yet; I'm simply using the 16x16 luma ones. A future commit will add newly generated probabilities for all transforms. - No ADST version. I don't think we'll add one for this level; if an ADST is desired, transform-size selection can scale back to 16x16 or lower, and use an ADST at that level. Additional notes specific to Debargha's DWT/DCT hybrid: - coefficient scale is different for the top/left 16x16 (DCT-over-DWT) block than for the rest (DWT pixel differences) of the block. Therefore, RD error scoring isn't easily scalable between coefficient and pixel domain. Thus, unfortunately, we need to compute the RD distortion in the pixel domain until we figure out how to scale these appropriately. Change-Id: I00386f20f35d7fabb19aba94c8162f8aee64ef2b
2012-12-07 23:45:05 +01:00
vp9_write(bc, sz != TX_16X16, c->prob_tx[2]);
}
}
}
#if CONFIG_CODE_ZEROGROUP
#ifdef ZPC_STATS
void init_zpcstats() {
vp9_zero(zpc_stats_4x4);
vp9_zero(zpc_stats_8x8);
vp9_zero(zpc_stats_16x16);
vp9_zero(zpc_stats_32x32);
}
void update_zpcstats(VP9_COMMON *const cm) {
int r, b, p, n;
for (r = 0; r < REF_TYPES; ++r) {
for (b = 0; b < ZPC_BANDS; ++b) {
for (p = 0; p < ZPC_PTOKS; ++p) {
for (n = 0; n < ZPC_NODES; ++n) {
zpc_stats_4x4[r][b][p][n][0] += cm->fc.zpc_counts_4x4[r][b][p][n][0];
zpc_stats_4x4[r][b][p][n][1] += cm->fc.zpc_counts_4x4[r][b][p][n][1];
zpc_stats_8x8[r][b][p][n][0] += cm->fc.zpc_counts_8x8[r][b][p][n][0];
zpc_stats_8x8[r][b][p][n][1] += cm->fc.zpc_counts_8x8[r][b][p][n][1];
zpc_stats_16x16[r][b][p][n][0] +=
cm->fc.zpc_counts_16x16[r][b][p][n][0];
zpc_stats_16x16[r][b][p][n][1] +=
cm->fc.zpc_counts_16x16[r][b][p][n][1];
zpc_stats_32x32[r][b][p][n][0] +=
cm->fc.zpc_counts_32x32[r][b][p][n][0];
zpc_stats_32x32[r][b][p][n][1] +=
cm->fc.zpc_counts_32x32[r][b][p][n][1];
}
}
}
}
}
void print_zpcstats() {
int r, b, p, n;
FILE *f;
printf(
"static const unsigned int default_zpc_probs_4x4[REF_TYPES]\n"
" [ZPC_BANDS]\n"
" [ZPC_PTOKS]\n"
" [ZPC_NODES] = {\n");
for (r = 0; r < REF_TYPES; ++r) {
printf(" {\n");
for (b = 0; b < ZPC_BANDS; ++b) {
printf(" {\n");
for (p = 0; p < ZPC_PTOKS; ++p) {
printf(" {");
for (n = 0; n < ZPC_NODES; ++n) {
vp9_prob prob = get_binary_prob(zpc_stats_4x4[r][b][p][n][0],
zpc_stats_4x4[r][b][p][n][1]);
printf(" %-3d [%d/%d],", prob, zpc_stats_4x4[r][b][p][n][0],
zpc_stats_4x4[r][b][p][n][1]);
}
printf(" },\n");
}
printf(" },\n");
}
printf(" },\n");
}
printf("};\n");
printf(
"static const unsigned int default_zpc_probs_8x8[REF_TYPES]\n"
" [ZPC_BANDS]\n"
" [ZPC_PTOKS]\n"
" [ZPC_NODES] = {\n");
for (r = 0; r < REF_TYPES; ++r) {
printf(" {\n");
for (b = 0; b < ZPC_BANDS; ++b) {
printf(" {\n");
for (p = 0; p < ZPC_PTOKS; ++p) {
printf(" {");
for (n = 0; n < ZPC_NODES; ++n) {
vp9_prob prob = get_binary_prob(zpc_stats_8x8[r][b][p][n][0],
zpc_stats_8x8[r][b][p][n][1]);
printf(" %-3d [%d/%d],", prob, zpc_stats_8x8[r][b][p][n][0],
zpc_stats_8x8[r][b][p][n][1]);
}
printf(" },\n");
}
printf(" },\n");
}
printf(" },\n");
}
printf("};\n");
printf(
"static const unsigned int default_zpc_probs_16x16[REF_TYPES]\n"
" [ZPC_BANDS]\n"
" [ZPC_PTOKS]\n"
" [ZPC_NODES] = {\n");
for (r = 0; r < REF_TYPES; ++r) {
printf(" {\n");
for (b = 0; b < ZPC_BANDS; ++b) {
printf(" {\n");
for (p = 0; p < ZPC_PTOKS; ++p) {
printf(" {");
for (n = 0; n < ZPC_NODES; ++n) {
vp9_prob prob = get_binary_prob(zpc_stats_16x16[r][b][p][n][0],
zpc_stats_16x16[r][b][p][n][1]);
printf(" %-3d [%d/%d],", prob, zpc_stats_16x16[r][b][p][n][0],
zpc_stats_16x16[r][b][p][n][1]);
}
printf(" },\n");
}
printf(" },\n");
}
printf(" },\n");
}
printf("};\n");
printf(
"static const unsigned int default_zpc_probs_32x32[REF_TYPES]\n"
" [ZPC_BANDS]\n"
" [ZPC_PTOKS]\n"
" [ZPC_NODES] = {\n");
for (r = 0; r < REF_TYPES; ++r) {
printf(" {\n");
for (b = 0; b < ZPC_BANDS; ++b) {
printf(" {\n");
for (p = 0; p < ZPC_PTOKS; ++p) {
printf(" {");
for (n = 0; n < ZPC_NODES; ++n) {
vp9_prob prob = get_binary_prob(zpc_stats_32x32[r][b][p][n][0],
zpc_stats_32x32[r][b][p][n][1]);
printf(" %-3d [%d/%d],", prob, zpc_stats_32x32[r][b][p][n][0],
zpc_stats_32x32[r][b][p][n][1]);
}
printf(" },\n");
}
printf(" },\n");
}
printf(" },\n");
}
printf("};\n");
f = fopen("zpcstats.bin", "wb");
fwrite(zpc_stats_4x4, sizeof(zpc_stats_4x4), 1, f);
fwrite(zpc_stats_8x8, sizeof(zpc_stats_8x8), 1, f);
fwrite(zpc_stats_16x16, sizeof(zpc_stats_16x16), 1, f);
fwrite(zpc_stats_32x32, sizeof(zpc_stats_32x32), 1, f);
fclose(f);
}
#endif
#endif // CONFIG_CODE_ZEROGROUP
static void write_modes_b(VP9_COMP *cpi, MODE_INFO *m, vp9_writer *bc,
TOKENEXTRA **tok, TOKENEXTRA *tok_end,
int mi_row, int mi_col) {
VP9_COMMON *const cm = &cpi->common;
MACROBLOCKD *const xd = &cpi->mb.e_mbd;
xd->mode_info_context = m;
set_mi_row_col(&cpi->common, xd, mi_row,
1 << mi_height_log2(m->mbmi.sb_type),
mi_col, 1 << mi_width_log2(m->mbmi.sb_type));
if (cm->frame_type == KEY_FRAME) {
write_mb_modes_kf(cpi, m, bc, mi_row, mi_col);
#ifdef ENTROPY_STATS
active_section = 8;
#endif
} else {
pack_inter_mode_mvs(cpi, m, bc, mi_row, mi_col);
#ifdef ENTROPY_STATS
active_section = 1;
#endif
}
assert(*tok < tok_end);
pack_mb_tokens(bc, tok, tok_end);
}
static void write_modes_sb(VP9_COMP *cpi, MODE_INFO *m, vp9_writer *bc,
TOKENEXTRA **tok, TOKENEXTRA *tok_end,
int mi_row, int mi_col,
BLOCK_SIZE_TYPE bsize) {
VP9_COMMON *const cm = &cpi->common;
MACROBLOCKD *xd = &cpi->mb.e_mbd;
const int mis = cm->mode_info_stride;
int bwl, bhl;
int bw, bh;
int bsl = mi_width_log2(bsize), bs = (1 << bsl) / 2;
int n;
PARTITION_TYPE partition;
BLOCK_SIZE_TYPE subsize;
if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
return;
bwl = mi_width_log2(m->mbmi.sb_type);
bhl = mi_height_log2(m->mbmi.sb_type);
bw = 1 << bwl;
bh = 1 << bhl;
// parse the partition type
if ((bwl == bsl) && (bhl == bsl))
partition = PARTITION_NONE;
else if ((bwl == bsl) && (bhl < bsl))
partition = PARTITION_HORZ;
else if ((bwl < bsl) && (bhl == bsl))
partition = PARTITION_VERT;
else if ((bwl < bsl) && (bhl < bsl))
partition = PARTITION_SPLIT;
else
assert(0);
if (bsize > BLOCK_SIZE_MB16X16) {
int pl;
xd->left_seg_context =
cm->left_seg_context + ((mi_row >> CONFIG_SB8X8) & 3);
xd->above_seg_context = cm->above_seg_context + (mi_col >> CONFIG_SB8X8);
pl = partition_plane_context(xd, bsize);
// encode the partition information
write_token(bc, vp9_partition_tree, cm->fc.partition_prob[pl],
vp9_partition_encodings + partition);
}
switch (partition) {
case PARTITION_NONE:
subsize = bsize;
write_modes_b(cpi, m, bc, tok, tok_end, mi_row, mi_col);
break;
case PARTITION_HORZ:
subsize = (bsize == BLOCK_SIZE_SB64X64) ? BLOCK_SIZE_SB64X32 :
BLOCK_SIZE_SB32X16;
write_modes_b(cpi, m, bc, tok, tok_end, mi_row, mi_col);
if ((mi_row + bh) < cm->mi_rows)
write_modes_b(cpi, m + bh * mis, bc, tok, tok_end, mi_row + bh, mi_col);
break;
case PARTITION_VERT:
subsize = (bsize == BLOCK_SIZE_SB64X64) ? BLOCK_SIZE_SB32X64 :
BLOCK_SIZE_SB16X32;
write_modes_b(cpi, m, bc, tok, tok_end, mi_row, mi_col);
if ((mi_col + bw) < cm->mi_cols)
write_modes_b(cpi, m + bw, bc, tok, tok_end, mi_row, mi_col + bw);
break;
case PARTITION_SPLIT:
// TODO(jingning): support recursive partitioning down to 16x16 as for
// now. need to merge in 16x8, 8x16, 8x8, and smaller partitions.
if (bsize == BLOCK_SIZE_SB64X64)
subsize = BLOCK_SIZE_SB32X32;
else if (bsize == BLOCK_SIZE_SB32X32)
subsize = BLOCK_SIZE_MB16X16;
else
assert(0);
for (n = 0; n < 4; n++) {
int j = n >> 1, i = n & 0x01;
write_modes_sb(cpi, m + j * bs * mis + i * bs, bc, tok, tok_end,
mi_row + j * bs, mi_col + i * bs, subsize);
}
break;
default:
assert(0);
}
// update partition context
if ((partition == PARTITION_SPLIT) && (bsize > BLOCK_SIZE_SB32X32))
return;
xd->left_seg_context = cm->left_seg_context + ((mi_row >> CONFIG_SB8X8) & 3);
xd->above_seg_context = cm->above_seg_context + (mi_col >> CONFIG_SB8X8);
update_partition_context(xd, subsize, bsize);
}
[WIP] Add column-based tiling. This patch adds column-based tiling. The idea is to make each tile independently decodable (after reading the common frame header) and also independendly encodable (minus within-frame cost adjustments in the RD loop) to speed-up hardware & software en/decoders if they used multi-threading. Column-based tiling has the added advantage (over other tiling methods) that it minimizes realtime use-case latency, since all threads can start encoding data as soon as the first SB-row worth of data is available to the encoder. There is some test code that does random tile ordering in the decoder, to confirm that each tile is indeed independently decodable from other tiles in the same frame. At tile edges, all contexts assume default values (i.e. 0, 0 motion vector, no coefficients, DC intra4x4 mode), and motion vector search and ordering do not cross tiles in the same frame. t log Tile independence is not maintained between frames ATM, i.e. tile 0 of frame 1 is free to use motion vectors that point into any tile of frame 0. We support 1 (i.e. no tiling), 2 or 4 column-tiles. The loopfilter crosses tile boundaries. I discussed this briefly with Aki and he says that's OK. An in-loop loopfilter would need to do some sync between tile threads, but that shouldn't be a big issue. Resuls: with tiling disabled, we go up slightly because of improved edge use in the intra4x4 prediction. With 2 tiles, we lose about ~1% on derf, ~0.35% on HD and ~0.55% on STD/HD. With 4 tiles, we lose another ~1.5% on derf ~0.77% on HD and ~0.85% on STD/HD. Most of this loss is concentrated in the low-bitrate end of clips, and most of it is because of the loss of edges at tile boundaries and the resulting loss of intra predictors. TODO: - more tiles (perhaps allow row-based tiling also, and max. 8 tiles)? - maybe optionally (for EC purposes), motion vectors themselves should not cross tile edges, or we should emulate such borders as if they were off-frame, to limit error propagation to within one tile only. This doesn't have to be the default behaviour but could be an optional bitstream flag. Change-Id: I5951c3a0742a767b20bc9fb5af685d9892c2c96f
2013-02-01 18:35:28 +01:00
static void write_modes(VP9_COMP *cpi, vp9_writer* const bc,
TOKENEXTRA **tok, TOKENEXTRA *tok_end) {
VP9_COMMON *const c = &cpi->common;
const int mis = c->mode_info_stride;
MODE_INFO *m, *m_ptr = c->mi;
int mi_row, mi_col;
2010-05-18 17:58:33 +02:00
m_ptr += c->cur_tile_mi_col_start + c->cur_tile_mi_row_start * mis;
vpx_memset(c->above_seg_context, 0, sizeof(PARTITION_CONTEXT) *
mb_cols_aligned_to_sb(c));
for (mi_row = c->cur_tile_mi_row_start;
mi_row < c->cur_tile_mi_row_end;
mi_row += (4 << CONFIG_SB8X8), m_ptr += (4 << CONFIG_SB8X8) * mis) {
m = m_ptr;
vpx_memset(c->left_seg_context, 0, sizeof(c->left_seg_context));
for (mi_col = c->cur_tile_mi_col_start;
mi_col < c->cur_tile_mi_col_end;
mi_col += (4 << CONFIG_SB8X8), m += (4 << CONFIG_SB8X8))
write_modes_sb(cpi, m, bc, tok, tok_end, mi_row, mi_col,
BLOCK_SIZE_SB64X64);
}
2010-05-18 17:58:33 +02:00
}
/* This function is used for debugging probability trees. */
static void print_prob_tree(vp9_coeff_probs *coef_probs, int block_types) {
/* print coef probability tree */
int i, j, k, l, m;
FILE *f = fopen("enc_tree_probs.txt", "a");
fprintf(f, "{\n");
for (i = 0; i < block_types; i++) {
fprintf(f, " {\n");
for (j = 0; j < REF_TYPES; ++j) {
fprintf(f, " {\n");
for (k = 0; k < COEF_BANDS; k++) {
fprintf(f, " {\n");
for (l = 0; l < PREV_COEF_CONTEXTS; l++) {
fprintf(f, " {");
for (m = 0; m < ENTROPY_NODES; m++) {
fprintf(f, "%3u, ",
(unsigned int)(coef_probs[i][j][k][l][m]));
}
}
fprintf(f, " }\n");
}
fprintf(f, " }\n");
}
fprintf(f, " }\n");
}
fprintf(f, "}\n");
fclose(f);
}
static void build_tree_distribution(vp9_coeff_probs *coef_probs,
vp9_coeff_count *coef_counts,
unsigned int (*eob_branch_ct)[REF_TYPES]
[COEF_BANDS]
[PREV_COEF_CONTEXTS],
#ifdef ENTROPY_STATS
VP9_COMP *cpi,
vp9_coeff_accum *context_counters,
#endif
vp9_coeff_stats *coef_branch_ct,
int block_types) {
int i, j, k, l;
#ifdef ENTROPY_STATS
int t = 0;
#endif
for (i = 0; i < block_types; ++i) {
for (j = 0; j < REF_TYPES; ++j) {
for (k = 0; k < COEF_BANDS; ++k) {
for (l = 0; l < PREV_COEF_CONTEXTS; ++l) {
if (l >= 3 && k == 0)
continue;
vp9_tree_probs_from_distribution(vp9_coef_tree,
coef_probs[i][j][k][l],
coef_branch_ct[i][j][k][l],
coef_counts[i][j][k][l], 0);
coef_branch_ct[i][j][k][l][0][1] = eob_branch_ct[i][j][k][l] -
coef_branch_ct[i][j][k][l][0][0];
coef_probs[i][j][k][l][0] =
get_binary_prob(coef_branch_ct[i][j][k][l][0][0],
coef_branch_ct[i][j][k][l][0][1]);
#ifdef ENTROPY_STATS
if (!cpi->dummy_packing) {
for (t = 0; t < MAX_ENTROPY_TOKENS; ++t)
context_counters[i][j][k][l][t] += coef_counts[i][j][k][l][t];
context_counters[i][j][k][l][MAX_ENTROPY_TOKENS] +=
eob_branch_ct[i][j][k][l];
}
#endif
}
}
}
}
}
static void build_coeff_contexts(VP9_COMP *cpi) {
build_tree_distribution(cpi->frame_coef_probs_4x4,
cpi->coef_counts_4x4,
cpi->common.fc.eob_branch_counts[TX_4X4],
#ifdef ENTROPY_STATS
cpi, context_counters_4x4,
#endif
cpi->frame_branch_ct_4x4, BLOCK_TYPES);
build_tree_distribution(cpi->frame_coef_probs_8x8,
cpi->coef_counts_8x8,
cpi->common.fc.eob_branch_counts[TX_8X8],
#ifdef ENTROPY_STATS
cpi, context_counters_8x8,
#endif
cpi->frame_branch_ct_8x8, BLOCK_TYPES);
build_tree_distribution(cpi->frame_coef_probs_16x16,
cpi->coef_counts_16x16,
cpi->common.fc.eob_branch_counts[TX_16X16],
32x32 transform for superblocks. This adds Debargha's DCT/DWT hybrid and a regular 32x32 DCT, and adds code all over the place to wrap that in the bitstream/encoder/decoder/RD. Some implementation notes (these probably need careful review): - token range is extended by 1 bit, since the value range out of this transform is [-16384,16383]. - the coefficients coming out of the FDCT are manually scaled back by 1 bit, or else they won't fit in int16_t (they are 17 bits). Because of this, the RD error scoring does not right-shift the MSE score by two (unlike for 4x4/8x8/16x16). - to compensate for this loss in precision, the quantizer is halved also. This is currently a little hacky. - FDCT and IDCT is double-only right now. Needs a fixed-point impl. - There are no default probabilities for the 32x32 transform yet; I'm simply using the 16x16 luma ones. A future commit will add newly generated probabilities for all transforms. - No ADST version. I don't think we'll add one for this level; if an ADST is desired, transform-size selection can scale back to 16x16 or lower, and use an ADST at that level. Additional notes specific to Debargha's DWT/DCT hybrid: - coefficient scale is different for the top/left 16x16 (DCT-over-DWT) block than for the rest (DWT pixel differences) of the block. Therefore, RD error scoring isn't easily scalable between coefficient and pixel domain. Thus, unfortunately, we need to compute the RD distortion in the pixel domain until we figure out how to scale these appropriately. Change-Id: I00386f20f35d7fabb19aba94c8162f8aee64ef2b
2012-12-07 23:45:05 +01:00
#ifdef ENTROPY_STATS
cpi, context_counters_16x16,
#endif
cpi->frame_branch_ct_16x16, BLOCK_TYPES);
build_tree_distribution(cpi->frame_coef_probs_32x32,
cpi->coef_counts_32x32,
cpi->common.fc.eob_branch_counts[TX_32X32],
#ifdef ENTROPY_STATS
cpi, context_counters_32x32,
#endif
cpi->frame_branch_ct_32x32, BLOCK_TYPES);
2010-05-18 17:58:33 +02:00
}
#if CONFIG_CODE_ZEROGROUP
static void update_zpc_probs_common(VP9_COMP* cpi,
vp9_writer* const bc,
TX_SIZE tx_size) {
int r, b, p, n;
VP9_COMMON *const cm = &cpi->common;
int update[2] = {0, 0};
int savings = 0;
vp9_zpc_probs newprobs;
vp9_zpc_probs *zpc_probs;
vp9_zpc_count *zpc_counts;
vp9_prob upd = ZPC_UPDATE_PROB;
if (!get_zpc_used(tx_size)) return;
if (tx_size == TX_32X32) {
zpc_probs = &cm->fc.zpc_probs_32x32;
zpc_counts = &cm->fc.zpc_counts_32x32;
} else if (tx_size == TX_16X16) {
zpc_probs = &cm->fc.zpc_probs_16x16;
zpc_counts = &cm->fc.zpc_counts_16x16;
} else if (tx_size == TX_8X8) {
zpc_probs = &cm->fc.zpc_probs_8x8;
zpc_counts = &cm->fc.zpc_counts_8x8;
} else {
zpc_probs = &cm->fc.zpc_probs_4x4;
zpc_counts = &cm->fc.zpc_counts_4x4;
}
for (r = 0; r < REF_TYPES; ++r) {
for (b = 0; b < ZPC_BANDS; ++b) {
for (p = 0; p < ZPC_PTOKS; ++p) {
for (n = 0; n < ZPC_NODES; ++n) {
newprobs[r][b][p][n] = get_binary_prob((*zpc_counts)[r][b][p][n][0],
(*zpc_counts)[r][b][p][n][1]);
}
}
}
}
for (r = 0; r < REF_TYPES; ++r) {
for (b = 0; b < ZPC_BANDS; ++b) {
for (p = 0; p < ZPC_PTOKS; ++p) {
for (n = 0; n < ZPC_NODES; ++n) {
vp9_prob newp = newprobs[r][b][p][n];
vp9_prob oldp = (*zpc_probs)[r][b][p][n];
int s, u = 0;
#if USE_ZPC_EXTRA == 0
if (n == 1) continue;
#endif
#if defined(SEARCH_NEWP)
s = prob_diff_update_savings_search((*zpc_counts)[r][b][p][n],
oldp, &newp, upd);
if (s > 0 && newp != oldp)
u = 1;
if (u)
savings += s - (int)(vp9_cost_zero(upd));
else
savings -= (int)(vp9_cost_zero(upd));
#else
s = prob_update_savings((*zpc_counts)[r][b][p][n],
oldp, newp, upd);
if (s > 0)
u = 1;
if (u)
savings += s;
#endif
update[u]++;
}
}
}
}
if (update[1] == 0 || savings < 0) {
vp9_write_bit(bc, 0);
return;
}
vp9_write_bit(bc, 1);
for (r = 0; r < REF_TYPES; ++r) {
for (b = 0; b < ZPC_BANDS; ++b) {
for (p = 0; p < ZPC_PTOKS; ++p) {
for (n = 0; n < ZPC_NODES; ++n) {
vp9_prob newp = newprobs[r][b][p][n];
vp9_prob *oldp = &(*zpc_probs)[r][b][p][n];
int s, u = 0;
#if USE_ZPC_EXTRA == 0
if (n == 1) continue;
#endif
#if defined(SEARCH_NEWP)
s = prob_diff_update_savings_search((*zpc_counts)[r][b][p][n],
*oldp, &newp, upd);
if (s > 0 && newp != *oldp)
u = 1;
#else
s = prob_update_savings((*zpc_counts)[r][b][p][n],
*oldp, newp, upd);
if (s > 0)
u = 1;
#endif
vp9_write(bc, u, upd);
if (u) {
/* send/use new probability */
write_prob_diff_update(bc, newp, *oldp);
*oldp = newp;
}
}
}
}
}
}
static void update_zpc_probs(VP9_COMP* cpi,
vp9_writer* const bc) {
update_zpc_probs_common(cpi, bc, TX_4X4);
if (cpi->common.txfm_mode != ONLY_4X4)
update_zpc_probs_common(cpi, bc, TX_8X8);
if (cpi->common.txfm_mode > ALLOW_8X8)
update_zpc_probs_common(cpi, bc, TX_16X16);
if (cpi->common.txfm_mode > ALLOW_16X16)
update_zpc_probs_common(cpi, bc, TX_32X32);
#ifdef ZPC_STATS
if (!cpi->dummy_packing)
update_zpcstats(&cpi->common);
#endif
}
#endif // CONFIG_CODE_ZEROGROUP
static void update_coef_probs_common(vp9_writer* const bc,
VP9_COMP *cpi,
#ifdef ENTROPY_STATS
vp9_coeff_stats *tree_update_hist,
#endif
vp9_coeff_probs *new_frame_coef_probs,
vp9_coeff_probs *old_frame_coef_probs,
vp9_coeff_stats *frame_branch_ct,
TX_SIZE tx_size) {
int i, j, k, l, t;
int update[2] = {0, 0};
int savings;
Modeling default coef probs with distribution Replaces the default tables for single coefficient magnitudes with those obtained from an appropriate distribution. The EOB node is left unchanged. The model is represeted as a 256-size codebook where the index corresponds to the probability of the Zero or the One node. Two variations are implemented corresponding to whether the Zero node or the One-node is used as the peg. The main advantage is that the default prob tables will become considerably smaller and manageable. Besides there is substantially less risk of over-fitting for a training set. Various distributions are tried and the one that gives the best results is the family of Generalized Gaussian distributions with shape parameter 0.75. The results are within about 0.2% of fully trained tables for the Zero peg variant, and within 0.1% of the One peg variant. The forward updates are optionally (controlled by a macro) model-based, i.e. restricted to only convey probabilities from the codebook. Backward updates can also be optionally (controlled by another macro) model-based, but is turned off by default. Currently model-based forward updates work about the same as unconstrained updates, but there is a drop in performance with backward-updates being model based. The model based approach also allows the probabilities for the key frames to be adjusted from the defaults based on the base_qindex of the frame. Currently the adjustment function is a placeholder that adjusts the prob of EOB and Zero node from the nominal one at higher quality (lower qindex) or lower quality (higher qindex) ends of the range. The rest of the probabilities are then derived based on the model from the adjusted prob of zero. Change-Id: Iae050f3cbcc6d8b3f204e8dc395ae47b3b2192c9
2013-03-13 19:03:17 +01:00
#if CONFIG_MODELCOEFPROB && MODEL_BASED_UPDATE
const int entropy_nodes_update = UNCONSTRAINED_UPDATE_NODES;
#else
const int entropy_nodes_update = ENTROPY_NODES;
#endif
// vp9_prob bestupd = find_coef_update_prob(cpi);
const int tstart = 0;
/* dry run to see if there is any udpate at all needed */
savings = 0;
for (i = 0; i < BLOCK_TYPES; ++i) {
for (j = 0; j < REF_TYPES; ++j) {
for (k = 0; k < COEF_BANDS; ++k) {
Modeling default coef probs with distribution Replaces the default tables for single coefficient magnitudes with those obtained from an appropriate distribution. The EOB node is left unchanged. The model is represeted as a 256-size codebook where the index corresponds to the probability of the Zero or the One node. Two variations are implemented corresponding to whether the Zero node or the One-node is used as the peg. The main advantage is that the default prob tables will become considerably smaller and manageable. Besides there is substantially less risk of over-fitting for a training set. Various distributions are tried and the one that gives the best results is the family of Generalized Gaussian distributions with shape parameter 0.75. The results are within about 0.2% of fully trained tables for the Zero peg variant, and within 0.1% of the One peg variant. The forward updates are optionally (controlled by a macro) model-based, i.e. restricted to only convey probabilities from the codebook. Backward updates can also be optionally (controlled by another macro) model-based, but is turned off by default. Currently model-based forward updates work about the same as unconstrained updates, but there is a drop in performance with backward-updates being model based. The model based approach also allows the probabilities for the key frames to be adjusted from the defaults based on the base_qindex of the frame. Currently the adjustment function is a placeholder that adjusts the prob of EOB and Zero node from the nominal one at higher quality (lower qindex) or lower quality (higher qindex) ends of the range. The rest of the probabilities are then derived based on the model from the adjusted prob of zero. Change-Id: Iae050f3cbcc6d8b3f204e8dc395ae47b3b2192c9
2013-03-13 19:03:17 +01:00
// int prev_coef_savings[ENTROPY_NODES] = {0};
for (l = 0; l < PREV_COEF_CONTEXTS; ++l) {
for (t = tstart; t < entropy_nodes_update; ++t) {
vp9_prob newp = new_frame_coef_probs[i][j][k][l][t];
const vp9_prob oldp = old_frame_coef_probs[i][j][k][l][t];
Modeling default coef probs with distribution Replaces the default tables for single coefficient magnitudes with those obtained from an appropriate distribution. The EOB node is left unchanged. The model is represeted as a 256-size codebook where the index corresponds to the probability of the Zero or the One node. Two variations are implemented corresponding to whether the Zero node or the One-node is used as the peg. The main advantage is that the default prob tables will become considerably smaller and manageable. Besides there is substantially less risk of over-fitting for a training set. Various distributions are tried and the one that gives the best results is the family of Generalized Gaussian distributions with shape parameter 0.75. The results are within about 0.2% of fully trained tables for the Zero peg variant, and within 0.1% of the One peg variant. The forward updates are optionally (controlled by a macro) model-based, i.e. restricted to only convey probabilities from the codebook. Backward updates can also be optionally (controlled by another macro) model-based, but is turned off by default. Currently model-based forward updates work about the same as unconstrained updates, but there is a drop in performance with backward-updates being model based. The model based approach also allows the probabilities for the key frames to be adjusted from the defaults based on the base_qindex of the frame. Currently the adjustment function is a placeholder that adjusts the prob of EOB and Zero node from the nominal one at higher quality (lower qindex) or lower quality (higher qindex) ends of the range. The rest of the probabilities are then derived based on the model from the adjusted prob of zero. Change-Id: Iae050f3cbcc6d8b3f204e8dc395ae47b3b2192c9
2013-03-13 19:03:17 +01:00
const vp9_prob upd = vp9_coef_update_prob[t];
int s; // = prev_coef_savings[t];
int u = 0;
if (l >= 3 && k == 0)
continue;
#if defined(SEARCH_NEWP)
Modeling default coef probs with distribution Replaces the default tables for single coefficient magnitudes with those obtained from an appropriate distribution. The EOB node is left unchanged. The model is represeted as a 256-size codebook where the index corresponds to the probability of the Zero or the One node. Two variations are implemented corresponding to whether the Zero node or the One-node is used as the peg. The main advantage is that the default prob tables will become considerably smaller and manageable. Besides there is substantially less risk of over-fitting for a training set. Various distributions are tried and the one that gives the best results is the family of Generalized Gaussian distributions with shape parameter 0.75. The results are within about 0.2% of fully trained tables for the Zero peg variant, and within 0.1% of the One peg variant. The forward updates are optionally (controlled by a macro) model-based, i.e. restricted to only convey probabilities from the codebook. Backward updates can also be optionally (controlled by another macro) model-based, but is turned off by default. Currently model-based forward updates work about the same as unconstrained updates, but there is a drop in performance with backward-updates being model based. The model based approach also allows the probabilities for the key frames to be adjusted from the defaults based on the base_qindex of the frame. Currently the adjustment function is a placeholder that adjusts the prob of EOB and Zero node from the nominal one at higher quality (lower qindex) or lower quality (higher qindex) ends of the range. The rest of the probabilities are then derived based on the model from the adjusted prob of zero. Change-Id: Iae050f3cbcc6d8b3f204e8dc395ae47b3b2192c9
2013-03-13 19:03:17 +01:00
#if CONFIG_MODELCOEFPROB && MODEL_BASED_UPDATE
if (t == UNCONSTRAINED_NODES - 1)
s = prob_diff_update_savings_search_model(
frame_branch_ct[i][j][k][l][0],
old_frame_coef_probs[i][j][k][l], &newp, upd, i, j,
cpi->common.base_qindex);
Modeling default coef probs with distribution Replaces the default tables for single coefficient magnitudes with those obtained from an appropriate distribution. The EOB node is left unchanged. The model is represeted as a 256-size codebook where the index corresponds to the probability of the Zero or the One node. Two variations are implemented corresponding to whether the Zero node or the One-node is used as the peg. The main advantage is that the default prob tables will become considerably smaller and manageable. Besides there is substantially less risk of over-fitting for a training set. Various distributions are tried and the one that gives the best results is the family of Generalized Gaussian distributions with shape parameter 0.75. The results are within about 0.2% of fully trained tables for the Zero peg variant, and within 0.1% of the One peg variant. The forward updates are optionally (controlled by a macro) model-based, i.e. restricted to only convey probabilities from the codebook. Backward updates can also be optionally (controlled by another macro) model-based, but is turned off by default. Currently model-based forward updates work about the same as unconstrained updates, but there is a drop in performance with backward-updates being model based. The model based approach also allows the probabilities for the key frames to be adjusted from the defaults based on the base_qindex of the frame. Currently the adjustment function is a placeholder that adjusts the prob of EOB and Zero node from the nominal one at higher quality (lower qindex) or lower quality (higher qindex) ends of the range. The rest of the probabilities are then derived based on the model from the adjusted prob of zero. Change-Id: Iae050f3cbcc6d8b3f204e8dc395ae47b3b2192c9
2013-03-13 19:03:17 +01:00
else
#endif
s = prob_diff_update_savings_search(
frame_branch_ct[i][j][k][l][t], oldp, &newp, upd);
if (s > 0 && newp != oldp)
u = 1;
if (u)
savings += s - (int)(vp9_cost_zero(upd));
else
savings -= (int)(vp9_cost_zero(upd));
#else
s = prob_update_savings(frame_branch_ct[i][j][k][l][t],
oldp, newp, upd);
if (s > 0)
u = 1;
if (u)
savings += s;
#endif
update[u]++;
}
}
}
}
}
// printf("Update %d %d, savings %d\n", update[0], update[1], savings);
/* Is coef updated at all */
if (update[1] == 0 || savings < 0) {
vp9_write_bit(bc, 0);
Modeling default coef probs with distribution Replaces the default tables for single coefficient magnitudes with those obtained from an appropriate distribution. The EOB node is left unchanged. The model is represeted as a 256-size codebook where the index corresponds to the probability of the Zero or the One node. Two variations are implemented corresponding to whether the Zero node or the One-node is used as the peg. The main advantage is that the default prob tables will become considerably smaller and manageable. Besides there is substantially less risk of over-fitting for a training set. Various distributions are tried and the one that gives the best results is the family of Generalized Gaussian distributions with shape parameter 0.75. The results are within about 0.2% of fully trained tables for the Zero peg variant, and within 0.1% of the One peg variant. The forward updates are optionally (controlled by a macro) model-based, i.e. restricted to only convey probabilities from the codebook. Backward updates can also be optionally (controlled by another macro) model-based, but is turned off by default. Currently model-based forward updates work about the same as unconstrained updates, but there is a drop in performance with backward-updates being model based. The model based approach also allows the probabilities for the key frames to be adjusted from the defaults based on the base_qindex of the frame. Currently the adjustment function is a placeholder that adjusts the prob of EOB and Zero node from the nominal one at higher quality (lower qindex) or lower quality (higher qindex) ends of the range. The rest of the probabilities are then derived based on the model from the adjusted prob of zero. Change-Id: Iae050f3cbcc6d8b3f204e8dc395ae47b3b2192c9
2013-03-13 19:03:17 +01:00
return;
}
vp9_write_bit(bc, 1);
for (i = 0; i < BLOCK_TYPES; ++i) {
Modeling default coef probs with distribution Replaces the default tables for single coefficient magnitudes with those obtained from an appropriate distribution. The EOB node is left unchanged. The model is represeted as a 256-size codebook where the index corresponds to the probability of the Zero or the One node. Two variations are implemented corresponding to whether the Zero node or the One-node is used as the peg. The main advantage is that the default prob tables will become considerably smaller and manageable. Besides there is substantially less risk of over-fitting for a training set. Various distributions are tried and the one that gives the best results is the family of Generalized Gaussian distributions with shape parameter 0.75. The results are within about 0.2% of fully trained tables for the Zero peg variant, and within 0.1% of the One peg variant. The forward updates are optionally (controlled by a macro) model-based, i.e. restricted to only convey probabilities from the codebook. Backward updates can also be optionally (controlled by another macro) model-based, but is turned off by default. Currently model-based forward updates work about the same as unconstrained updates, but there is a drop in performance with backward-updates being model based. The model based approach also allows the probabilities for the key frames to be adjusted from the defaults based on the base_qindex of the frame. Currently the adjustment function is a placeholder that adjusts the prob of EOB and Zero node from the nominal one at higher quality (lower qindex) or lower quality (higher qindex) ends of the range. The rest of the probabilities are then derived based on the model from the adjusted prob of zero. Change-Id: Iae050f3cbcc6d8b3f204e8dc395ae47b3b2192c9
2013-03-13 19:03:17 +01:00
for (j = 0; j < REF_TYPES; ++j) {
for (k = 0; k < COEF_BANDS; ++k) {
// int prev_coef_savings[ENTROPY_NODES] = {0};
for (l = 0; l < PREV_COEF_CONTEXTS; ++l) {
// calc probs and branch cts for this frame only
for (t = tstart; t < entropy_nodes_update; ++t) {
Modeling default coef probs with distribution Replaces the default tables for single coefficient magnitudes with those obtained from an appropriate distribution. The EOB node is left unchanged. The model is represeted as a 256-size codebook where the index corresponds to the probability of the Zero or the One node. Two variations are implemented corresponding to whether the Zero node or the One-node is used as the peg. The main advantage is that the default prob tables will become considerably smaller and manageable. Besides there is substantially less risk of over-fitting for a training set. Various distributions are tried and the one that gives the best results is the family of Generalized Gaussian distributions with shape parameter 0.75. The results are within about 0.2% of fully trained tables for the Zero peg variant, and within 0.1% of the One peg variant. The forward updates are optionally (controlled by a macro) model-based, i.e. restricted to only convey probabilities from the codebook. Backward updates can also be optionally (controlled by another macro) model-based, but is turned off by default. Currently model-based forward updates work about the same as unconstrained updates, but there is a drop in performance with backward-updates being model based. The model based approach also allows the probabilities for the key frames to be adjusted from the defaults based on the base_qindex of the frame. Currently the adjustment function is a placeholder that adjusts the prob of EOB and Zero node from the nominal one at higher quality (lower qindex) or lower quality (higher qindex) ends of the range. The rest of the probabilities are then derived based on the model from the adjusted prob of zero. Change-Id: Iae050f3cbcc6d8b3f204e8dc395ae47b3b2192c9
2013-03-13 19:03:17 +01:00
vp9_prob newp = new_frame_coef_probs[i][j][k][l][t];
vp9_prob *oldp = old_frame_coef_probs[i][j][k][l] + t;
const vp9_prob upd = vp9_coef_update_prob[t];
int s; // = prev_coef_savings[t];
int u = 0;
if (l >= 3 && k == 0)
continue;
#if defined(SEARCH_NEWP)
Modeling default coef probs with distribution Replaces the default tables for single coefficient magnitudes with those obtained from an appropriate distribution. The EOB node is left unchanged. The model is represeted as a 256-size codebook where the index corresponds to the probability of the Zero or the One node. Two variations are implemented corresponding to whether the Zero node or the One-node is used as the peg. The main advantage is that the default prob tables will become considerably smaller and manageable. Besides there is substantially less risk of over-fitting for a training set. Various distributions are tried and the one that gives the best results is the family of Generalized Gaussian distributions with shape parameter 0.75. The results are within about 0.2% of fully trained tables for the Zero peg variant, and within 0.1% of the One peg variant. The forward updates are optionally (controlled by a macro) model-based, i.e. restricted to only convey probabilities from the codebook. Backward updates can also be optionally (controlled by another macro) model-based, but is turned off by default. Currently model-based forward updates work about the same as unconstrained updates, but there is a drop in performance with backward-updates being model based. The model based approach also allows the probabilities for the key frames to be adjusted from the defaults based on the base_qindex of the frame. Currently the adjustment function is a placeholder that adjusts the prob of EOB and Zero node from the nominal one at higher quality (lower qindex) or lower quality (higher qindex) ends of the range. The rest of the probabilities are then derived based on the model from the adjusted prob of zero. Change-Id: Iae050f3cbcc6d8b3f204e8dc395ae47b3b2192c9
2013-03-13 19:03:17 +01:00
#if CONFIG_MODELCOEFPROB && MODEL_BASED_UPDATE
if (t == UNCONSTRAINED_NODES - 1)
s = prob_diff_update_savings_search_model(
frame_branch_ct[i][j][k][l][0],
old_frame_coef_probs[i][j][k][l], &newp, upd, i, j,
cpi->common.base_qindex);
Modeling default coef probs with distribution Replaces the default tables for single coefficient magnitudes with those obtained from an appropriate distribution. The EOB node is left unchanged. The model is represeted as a 256-size codebook where the index corresponds to the probability of the Zero or the One node. Two variations are implemented corresponding to whether the Zero node or the One-node is used as the peg. The main advantage is that the default prob tables will become considerably smaller and manageable. Besides there is substantially less risk of over-fitting for a training set. Various distributions are tried and the one that gives the best results is the family of Generalized Gaussian distributions with shape parameter 0.75. The results are within about 0.2% of fully trained tables for the Zero peg variant, and within 0.1% of the One peg variant. The forward updates are optionally (controlled by a macro) model-based, i.e. restricted to only convey probabilities from the codebook. Backward updates can also be optionally (controlled by another macro) model-based, but is turned off by default. Currently model-based forward updates work about the same as unconstrained updates, but there is a drop in performance with backward-updates being model based. The model based approach also allows the probabilities for the key frames to be adjusted from the defaults based on the base_qindex of the frame. Currently the adjustment function is a placeholder that adjusts the prob of EOB and Zero node from the nominal one at higher quality (lower qindex) or lower quality (higher qindex) ends of the range. The rest of the probabilities are then derived based on the model from the adjusted prob of zero. Change-Id: Iae050f3cbcc6d8b3f204e8dc395ae47b3b2192c9
2013-03-13 19:03:17 +01:00
else
#endif
s = prob_diff_update_savings_search(
Modeling default coef probs with distribution Replaces the default tables for single coefficient magnitudes with those obtained from an appropriate distribution. The EOB node is left unchanged. The model is represeted as a 256-size codebook where the index corresponds to the probability of the Zero or the One node. Two variations are implemented corresponding to whether the Zero node or the One-node is used as the peg. The main advantage is that the default prob tables will become considerably smaller and manageable. Besides there is substantially less risk of over-fitting for a training set. Various distributions are tried and the one that gives the best results is the family of Generalized Gaussian distributions with shape parameter 0.75. The results are within about 0.2% of fully trained tables for the Zero peg variant, and within 0.1% of the One peg variant. The forward updates are optionally (controlled by a macro) model-based, i.e. restricted to only convey probabilities from the codebook. Backward updates can also be optionally (controlled by another macro) model-based, but is turned off by default. Currently model-based forward updates work about the same as unconstrained updates, but there is a drop in performance with backward-updates being model based. The model based approach also allows the probabilities for the key frames to be adjusted from the defaults based on the base_qindex of the frame. Currently the adjustment function is a placeholder that adjusts the prob of EOB and Zero node from the nominal one at higher quality (lower qindex) or lower quality (higher qindex) ends of the range. The rest of the probabilities are then derived based on the model from the adjusted prob of zero. Change-Id: Iae050f3cbcc6d8b3f204e8dc395ae47b3b2192c9
2013-03-13 19:03:17 +01:00
frame_branch_ct[i][j][k][l][t],
*oldp, &newp, upd);
if (s > 0 && newp != *oldp)
u = 1;
#else
Modeling default coef probs with distribution Replaces the default tables for single coefficient magnitudes with those obtained from an appropriate distribution. The EOB node is left unchanged. The model is represeted as a 256-size codebook where the index corresponds to the probability of the Zero or the One node. Two variations are implemented corresponding to whether the Zero node or the One-node is used as the peg. The main advantage is that the default prob tables will become considerably smaller and manageable. Besides there is substantially less risk of over-fitting for a training set. Various distributions are tried and the one that gives the best results is the family of Generalized Gaussian distributions with shape parameter 0.75. The results are within about 0.2% of fully trained tables for the Zero peg variant, and within 0.1% of the One peg variant. The forward updates are optionally (controlled by a macro) model-based, i.e. restricted to only convey probabilities from the codebook. Backward updates can also be optionally (controlled by another macro) model-based, but is turned off by default. Currently model-based forward updates work about the same as unconstrained updates, but there is a drop in performance with backward-updates being model based. The model based approach also allows the probabilities for the key frames to be adjusted from the defaults based on the base_qindex of the frame. Currently the adjustment function is a placeholder that adjusts the prob of EOB and Zero node from the nominal one at higher quality (lower qindex) or lower quality (higher qindex) ends of the range. The rest of the probabilities are then derived based on the model from the adjusted prob of zero. Change-Id: Iae050f3cbcc6d8b3f204e8dc395ae47b3b2192c9
2013-03-13 19:03:17 +01:00
s = prob_update_savings(frame_branch_ct[i][j][k][l][t],
*oldp, newp, upd);
if (s > 0)
u = 1;
#endif
Modeling default coef probs with distribution Replaces the default tables for single coefficient magnitudes with those obtained from an appropriate distribution. The EOB node is left unchanged. The model is represeted as a 256-size codebook where the index corresponds to the probability of the Zero or the One node. Two variations are implemented corresponding to whether the Zero node or the One-node is used as the peg. The main advantage is that the default prob tables will become considerably smaller and manageable. Besides there is substantially less risk of over-fitting for a training set. Various distributions are tried and the one that gives the best results is the family of Generalized Gaussian distributions with shape parameter 0.75. The results are within about 0.2% of fully trained tables for the Zero peg variant, and within 0.1% of the One peg variant. The forward updates are optionally (controlled by a macro) model-based, i.e. restricted to only convey probabilities from the codebook. Backward updates can also be optionally (controlled by another macro) model-based, but is turned off by default. Currently model-based forward updates work about the same as unconstrained updates, but there is a drop in performance with backward-updates being model based. The model based approach also allows the probabilities for the key frames to be adjusted from the defaults based on the base_qindex of the frame. Currently the adjustment function is a placeholder that adjusts the prob of EOB and Zero node from the nominal one at higher quality (lower qindex) or lower quality (higher qindex) ends of the range. The rest of the probabilities are then derived based on the model from the adjusted prob of zero. Change-Id: Iae050f3cbcc6d8b3f204e8dc395ae47b3b2192c9
2013-03-13 19:03:17 +01:00
vp9_write(bc, u, upd);
#ifdef ENTROPY_STATS
Modeling default coef probs with distribution Replaces the default tables for single coefficient magnitudes with those obtained from an appropriate distribution. The EOB node is left unchanged. The model is represeted as a 256-size codebook where the index corresponds to the probability of the Zero or the One node. Two variations are implemented corresponding to whether the Zero node or the One-node is used as the peg. The main advantage is that the default prob tables will become considerably smaller and manageable. Besides there is substantially less risk of over-fitting for a training set. Various distributions are tried and the one that gives the best results is the family of Generalized Gaussian distributions with shape parameter 0.75. The results are within about 0.2% of fully trained tables for the Zero peg variant, and within 0.1% of the One peg variant. The forward updates are optionally (controlled by a macro) model-based, i.e. restricted to only convey probabilities from the codebook. Backward updates can also be optionally (controlled by another macro) model-based, but is turned off by default. Currently model-based forward updates work about the same as unconstrained updates, but there is a drop in performance with backward-updates being model based. The model based approach also allows the probabilities for the key frames to be adjusted from the defaults based on the base_qindex of the frame. Currently the adjustment function is a placeholder that adjusts the prob of EOB and Zero node from the nominal one at higher quality (lower qindex) or lower quality (higher qindex) ends of the range. The rest of the probabilities are then derived based on the model from the adjusted prob of zero. Change-Id: Iae050f3cbcc6d8b3f204e8dc395ae47b3b2192c9
2013-03-13 19:03:17 +01:00
if (!cpi->dummy_packing)
++tree_update_hist[i][j][k][l][t][u];
#endif
if (u) {
/* send/use new probability */
write_prob_diff_update(bc, newp, *oldp);
*oldp = newp;
#if CONFIG_MODELCOEFPROB && MODEL_BASED_UPDATE
if (t == UNCONSTRAINED_NODES - 1)
vp9_get_model_distribution(
newp, old_frame_coef_probs[i][j][k][l], i, j);
2010-05-18 17:58:33 +02:00
#endif
}
}
}
}
}
}
}
2010-05-18 17:58:33 +02:00
static void update_coef_probs(VP9_COMP* const cpi, vp9_writer* const bc) {
vp9_clear_system_state();
// Build the cofficient contexts based on counts collected in encode loop
build_coeff_contexts(cpi);
update_coef_probs_common(bc,
cpi,
#ifdef ENTROPY_STATS
tree_update_hist_4x4,
#endif
cpi->frame_coef_probs_4x4,
cpi->common.fc.coef_probs_4x4,
cpi->frame_branch_ct_4x4,
TX_4X4);
/* do not do this if not even allowed */
if (cpi->common.txfm_mode != ONLY_4X4) {
update_coef_probs_common(bc,
cpi,
#ifdef ENTROPY_STATS
tree_update_hist_8x8,
#endif
cpi->frame_coef_probs_8x8,
cpi->common.fc.coef_probs_8x8,
cpi->frame_branch_ct_8x8,
TX_8X8);
}
if (cpi->common.txfm_mode > ALLOW_8X8) {
update_coef_probs_common(bc,
cpi,
#ifdef ENTROPY_STATS
tree_update_hist_16x16,
#endif
cpi->frame_coef_probs_16x16,
cpi->common.fc.coef_probs_16x16,
cpi->frame_branch_ct_16x16,
TX_16X16);
}
32x32 transform for superblocks. This adds Debargha's DCT/DWT hybrid and a regular 32x32 DCT, and adds code all over the place to wrap that in the bitstream/encoder/decoder/RD. Some implementation notes (these probably need careful review): - token range is extended by 1 bit, since the value range out of this transform is [-16384,16383]. - the coefficients coming out of the FDCT are manually scaled back by 1 bit, or else they won't fit in int16_t (they are 17 bits). Because of this, the RD error scoring does not right-shift the MSE score by two (unlike for 4x4/8x8/16x16). - to compensate for this loss in precision, the quantizer is halved also. This is currently a little hacky. - FDCT and IDCT is double-only right now. Needs a fixed-point impl. - There are no default probabilities for the 32x32 transform yet; I'm simply using the 16x16 luma ones. A future commit will add newly generated probabilities for all transforms. - No ADST version. I don't think we'll add one for this level; if an ADST is desired, transform-size selection can scale back to 16x16 or lower, and use an ADST at that level. Additional notes specific to Debargha's DWT/DCT hybrid: - coefficient scale is different for the top/left 16x16 (DCT-over-DWT) block than for the rest (DWT pixel differences) of the block. Therefore, RD error scoring isn't easily scalable between coefficient and pixel domain. Thus, unfortunately, we need to compute the RD distortion in the pixel domain until we figure out how to scale these appropriately. Change-Id: I00386f20f35d7fabb19aba94c8162f8aee64ef2b
2012-12-07 23:45:05 +01:00
if (cpi->common.txfm_mode > ALLOW_16X16) {
update_coef_probs_common(bc,
cpi,
#ifdef ENTROPY_STATS
tree_update_hist_32x32,
#endif
32x32 transform for superblocks. This adds Debargha's DCT/DWT hybrid and a regular 32x32 DCT, and adds code all over the place to wrap that in the bitstream/encoder/decoder/RD. Some implementation notes (these probably need careful review): - token range is extended by 1 bit, since the value range out of this transform is [-16384,16383]. - the coefficients coming out of the FDCT are manually scaled back by 1 bit, or else they won't fit in int16_t (they are 17 bits). Because of this, the RD error scoring does not right-shift the MSE score by two (unlike for 4x4/8x8/16x16). - to compensate for this loss in precision, the quantizer is halved also. This is currently a little hacky. - FDCT and IDCT is double-only right now. Needs a fixed-point impl. - There are no default probabilities for the 32x32 transform yet; I'm simply using the 16x16 luma ones. A future commit will add newly generated probabilities for all transforms. - No ADST version. I don't think we'll add one for this level; if an ADST is desired, transform-size selection can scale back to 16x16 or lower, and use an ADST at that level. Additional notes specific to Debargha's DWT/DCT hybrid: - coefficient scale is different for the top/left 16x16 (DCT-over-DWT) block than for the rest (DWT pixel differences) of the block. Therefore, RD error scoring isn't easily scalable between coefficient and pixel domain. Thus, unfortunately, we need to compute the RD distortion in the pixel domain until we figure out how to scale these appropriately. Change-Id: I00386f20f35d7fabb19aba94c8162f8aee64ef2b
2012-12-07 23:45:05 +01:00
cpi->frame_coef_probs_32x32,
cpi->common.fc.coef_probs_32x32,
cpi->frame_branch_ct_32x32,
TX_32X32);
32x32 transform for superblocks. This adds Debargha's DCT/DWT hybrid and a regular 32x32 DCT, and adds code all over the place to wrap that in the bitstream/encoder/decoder/RD. Some implementation notes (these probably need careful review): - token range is extended by 1 bit, since the value range out of this transform is [-16384,16383]. - the coefficients coming out of the FDCT are manually scaled back by 1 bit, or else they won't fit in int16_t (they are 17 bits). Because of this, the RD error scoring does not right-shift the MSE score by two (unlike for 4x4/8x8/16x16). - to compensate for this loss in precision, the quantizer is halved also. This is currently a little hacky. - FDCT and IDCT is double-only right now. Needs a fixed-point impl. - There are no default probabilities for the 32x32 transform yet; I'm simply using the 16x16 luma ones. A future commit will add newly generated probabilities for all transforms. - No ADST version. I don't think we'll add one for this level; if an ADST is desired, transform-size selection can scale back to 16x16 or lower, and use an ADST at that level. Additional notes specific to Debargha's DWT/DCT hybrid: - coefficient scale is different for the top/left 16x16 (DCT-over-DWT) block than for the rest (DWT pixel differences) of the block. Therefore, RD error scoring isn't easily scalable between coefficient and pixel domain. Thus, unfortunately, we need to compute the RD distortion in the pixel domain until we figure out how to scale these appropriately. Change-Id: I00386f20f35d7fabb19aba94c8162f8aee64ef2b
2012-12-07 23:45:05 +01:00
}
2010-05-18 17:58:33 +02:00
}
2010-05-18 17:58:33 +02:00
#ifdef PACKET_TESTING
FILE *vpxlogc = 0;
#endif
static void put_delta_q(vp9_writer *bc, int delta_q) {
if (delta_q != 0) {
vp9_write_bit(bc, 1);
vp9_write_literal(bc, abs(delta_q), 4);
vp9_write_bit(bc, delta_q < 0);
} else {
vp9_write_bit(bc, 0);
}
2010-05-18 17:58:33 +02:00
}
static void decide_kf_ymode_entropy(VP9_COMP *cpi) {
int mode_cost[MB_MODE_COUNT];
int cost;
int bestcost = INT_MAX;
int bestindex = 0;
int i, j;
for (i = 0; i < 8; i++) {
vp9_cost_tokens(mode_cost, cpi->common.kf_ymode_prob[i], vp9_kf_ymode_tree);
cost = 0;
for (j = 0; j < VP9_YMODES; j++) {
cost += mode_cost[j] * cpi->ymode_count[j];
}
vp9_cost_tokens(mode_cost, cpi->common.sb_kf_ymode_prob[i],
vp9_sb_ymode_tree);
for (j = 0; j < VP9_I32X32_MODES; j++) {
cost += mode_cost[j] * cpi->sb_ymode_count[j];
}
if (cost < bestcost) {
bestindex = i;
bestcost = cost;
}
}
cpi->common.kf_ymode_probs_index = bestindex;
}
static void segment_reference_frames(VP9_COMP *cpi) {
VP9_COMMON *oci = &cpi->common;
MODE_INFO *mi = oci->mi;
int ref[MAX_MB_SEGMENTS] = {0};
int i, j;
int mb_index = 0;
MACROBLOCKD *const xd = &cpi->mb.e_mbd;
for (i = 0; i < oci->mb_rows; i++) {
for (j = 0; j < oci->mb_cols; j++, mb_index++)
ref[mi[mb_index].mbmi.segment_id] |= (1 << mi[mb_index].mbmi.ref_frame);
mb_index++;
}
for (i = 0; i < MAX_MB_SEGMENTS; i++) {
vp9_enable_segfeature(xd, i, SEG_LVL_REF_FRAME);
vp9_set_segdata(xd, i, SEG_LVL_REF_FRAME, ref[i]);
}
}
void vp9_pack_bitstream(VP9_COMP *cpi, unsigned char *dest,
unsigned long *size) {
int i, j;
VP9_HEADER oh;
VP9_COMMON *const pc = &cpi->common;
vp9_writer header_bc, residual_bc;
MACROBLOCKD *const xd = &cpi->mb.e_mbd;
int extra_bytes_packed = 0;
2010-05-18 17:58:33 +02:00
unsigned char *cx_data = dest;
2010-05-18 17:58:33 +02:00
oh.show_frame = (int) pc->show_frame;
oh.type = (int)pc->frame_type;
oh.version = pc->version;
oh.first_partition_length_in_bytes = 0;
2010-05-18 17:58:33 +02:00
cx_data += 3;
2010-05-18 17:58:33 +02:00
#if defined(SECTIONBITS_OUTPUT)
Sectionbits[active_section = 1] += sizeof(VP9_HEADER) * 8 * 256;
2010-05-18 17:58:33 +02:00
#endif
compute_update_table();
/* vp9_kf_default_bmode_probs() is called in vp9_setup_key_frame() once
* for each K frame before encode frame. pc->kf_bmode_prob doesn't get
* changed anywhere else. No need to call it again here. --yw
* vp9_kf_default_bmode_probs( pc->kf_bmode_prob);
*/
2010-05-18 17:58:33 +02:00
/* every keyframe send startcode, width, height, scale factor, clamp
* and color type.
*/
if (oh.type == KEY_FRAME) {
// Start / synch code
cx_data[0] = 0x9D;
cx_data[1] = 0x01;
cx_data[2] = 0x2a;
extra_bytes_packed = 3;
cx_data += extra_bytes_packed;
}
2010-05-18 17:58:33 +02:00
if (pc->width != pc->display_width || pc->height != pc->display_height) {
write_le16(cx_data, pc->display_width);
write_le16(cx_data + 2, pc->display_height);
cx_data += 4;
extra_bytes_packed += 4;
}
2010-05-18 17:58:33 +02:00
write_le16(cx_data, pc->width);
write_le16(cx_data + 2, pc->height);
extra_bytes_packed += 4;
cx_data += 4;
vp9_start_encode(&header_bc, cx_data);
2010-05-18 17:58:33 +02:00
// TODO(jkoleszar): remove these two unused bits?
vp9_write_bit(&header_bc, pc->clr_type);
vp9_write_bit(&header_bc, pc->clamp_type);
// error resilient mode
vp9_write_bit(&header_bc, pc->error_resilient_mode);
// lossless mode: note this needs to be before loopfilter
vp9_write_bit(&header_bc, cpi->mb.e_mbd.lossless);
// Encode the loop filter level and type
vp9_write_bit(&header_bc, pc->filter_type);
vp9_write_literal(&header_bc, pc->filter_level, 6);
vp9_write_literal(&header_bc, pc->sharpness_level, 3);
#if CONFIG_LOOP_DERING
if (pc->dering_enabled) {
vp9_write_bit(&header_bc, 1);
vp9_write_literal(&header_bc, pc->dering_enabled - 1, 4);
} else {
vp9_write_bit(&header_bc, 0);
}
#endif
// Write out loop filter deltas applied at the MB level based on mode or ref frame (if they are enabled).
vp9_write_bit(&header_bc, (xd->mode_ref_lf_delta_enabled) ? 1 : 0);
if (xd->mode_ref_lf_delta_enabled) {
// Do the deltas need to be updated
vp9_write_bit(&header_bc, xd->mode_ref_lf_delta_update);
if (xd->mode_ref_lf_delta_update) {
// Send update
for (i = 0; i < MAX_REF_LF_DELTAS; i++) {
const int delta = xd->ref_lf_deltas[i];
// Frame level data
if (delta != xd->last_ref_lf_deltas[i]) {
xd->last_ref_lf_deltas[i] = delta;
vp9_write_bit(&header_bc, 1);
if (delta > 0) {
vp9_write_literal(&header_bc, delta & 0x3F, 6);
vp9_write_bit(&header_bc, 0); // sign
} else {
assert(delta < 0);
vp9_write_literal(&header_bc, (-delta) & 0x3F, 6);
vp9_write_bit(&header_bc, 1); // sign
}
} else {
vp9_write_bit(&header_bc, 0);
}
}
// Send update
for (i = 0; i < MAX_MODE_LF_DELTAS; i++) {
const int delta = xd->mode_lf_deltas[i];
if (delta != xd->last_mode_lf_deltas[i]) {
xd->last_mode_lf_deltas[i] = delta;
vp9_write_bit(&header_bc, 1);
if (delta > 0) {
vp9_write_literal(&header_bc, delta & 0x3F, 6);
vp9_write_bit(&header_bc, 0); // sign
} else {
assert(delta < 0);
vp9_write_literal(&header_bc, (-delta) & 0x3F, 6);
vp9_write_bit(&header_bc, 1); // sign
}
} else {
vp9_write_bit(&header_bc, 0);
}
}
}
}
2010-05-18 17:58:33 +02:00
// TODO(jkoleszar): remove these unused bits
vp9_write_literal(&header_bc, 0, 2);
2010-05-18 17:58:33 +02:00
// Frame Q baseline quantizer index
vp9_write_literal(&header_bc, pc->base_qindex, QINDEX_BITS);
2010-05-18 17:58:33 +02:00
// Transmit Dc, Second order and Uv quantizer delta information
put_delta_q(&header_bc, pc->y_dc_delta_q);
put_delta_q(&header_bc, pc->uv_dc_delta_q);
put_delta_q(&header_bc, pc->uv_ac_delta_q);
2010-05-18 17:58:33 +02:00
// When there is a key frame all reference buffers are updated using the new key frame
if (pc->frame_type != KEY_FRAME) {
int refresh_mask;
// Should the GF or ARF be updated using the transmitted frame or buffer
#if CONFIG_MULTIPLE_ARF
if (!cpi->multi_arf_enabled && cpi->refresh_golden_frame &&
!cpi->refresh_alt_ref_frame) {
#else
if (cpi->refresh_golden_frame && !cpi->refresh_alt_ref_frame) {
#endif
/* Preserve the previously existing golden frame and update the frame in
* the alt ref slot instead. This is highly specific to the use of
* alt-ref as a forward reference, and this needs to be generalized as
* other uses are implemented (like RTC/temporal scaling)
*
* gld_fb_idx and alt_fb_idx need to be swapped for future frames, but
* that happens in vp9_onyx_if.c:update_reference_frames() so that it can
* be done outside of the recode loop.
*/
refresh_mask = (cpi->refresh_last_frame << cpi->lst_fb_idx) |
(cpi->refresh_golden_frame << cpi->alt_fb_idx);
} else {
int arf_idx = cpi->alt_fb_idx;
#if CONFIG_MULTIPLE_ARF
// Determine which ARF buffer to use to encode this ARF frame.
if (cpi->multi_arf_enabled) {
int sn = cpi->sequence_number;
arf_idx = (cpi->frame_coding_order[sn] < 0) ?
cpi->arf_buffer_idx[sn + 1] :
cpi->arf_buffer_idx[sn];
}
#endif
refresh_mask = (cpi->refresh_last_frame << cpi->lst_fb_idx) |
(cpi->refresh_golden_frame << cpi->gld_fb_idx) |
(cpi->refresh_alt_ref_frame << arf_idx);
}
vp9_write_literal(&header_bc, refresh_mask, NUM_REF_FRAMES);
vp9_write_literal(&header_bc, cpi->lst_fb_idx, NUM_REF_FRAMES_LG2);
vp9_write_literal(&header_bc, cpi->gld_fb_idx, NUM_REF_FRAMES_LG2);
vp9_write_literal(&header_bc, cpi->alt_fb_idx, NUM_REF_FRAMES_LG2);
2010-05-18 17:58:33 +02:00
// Indicate reference frame sign bias for Golden and ARF frames (always 0 for last frame buffer)
vp9_write_bit(&header_bc, pc->ref_frame_sign_bias[GOLDEN_FRAME]);
vp9_write_bit(&header_bc, pc->ref_frame_sign_bias[ALTREF_FRAME]);
Supporting high precision 1/8-pel motion vectors This is the initial patch for supporting 1/8th pel motion. Currently if we configure with enable-high-precision-mv, all motion vectors would default to 1/8 pel. Encode and decode syncs fine with the current code. In the next phase the code will be refactored so that we can choose the 1/8 pel mode adaptively at a frame/segment/mb level. Derf results: http://www.corp.google.com/~debargha/vp8_results/enhinterp_hpmv.html (about 0.83% better than 8-tap interpoaltion) Patch 3: Rebased. Also adding 1/16th pel interpolation for U and V Patch 4: HD results. http://www.corp.google.com/~debargha/vp8_results/enhinterp_hd_hpmv.html Seems impressive (unless I am doing something wrong). Patch 5: Added mmx/sse for bilateral filtering, as well as enforced use of c-versions of subpel filters with 8-taps and 1/16th pel; Also redesigned the 8-tap filters to reduce the cut-off in order to introduce a denoising effect. There is a new configure option sixteenth-subpel-uv which will use 1/16 th pel interpolation for uv, if the motion vectors have 1/8 pel accuracy. With the fixes the results are promising on the derf set. The enhanced interpolation option with 8-taps alone gives 3% improvement over thei derf set: http://www.corp.google.com/~debargha/vp8_results/enhinterpn.html Results on high precision mv and on the hd set are to follow. Patch 6: Adding a missing condition for CONFIG_SIXTEENTH_SUBPEL_UV in vp8/common/x86/x86_systemdependent.c Patch 7: Cleaning up various debug messages. Patch 8: Merge conflict Change-Id: I5b1d844457aefd7414a9e4e0e06c6ed38fd8cc04
2012-02-16 18:29:54 +01:00
// Signal whether to allow high MV precision
vp9_write_bit(&header_bc, (xd->allow_high_precision_mv) ? 1 : 0);
if (pc->mcomp_filter_type == SWITCHABLE) {
/* Check to see if only one of the filters is actually used */
int count[VP9_SWITCHABLE_FILTERS];
int i, j, c = 0;
for (i = 0; i < VP9_SWITCHABLE_FILTERS; ++i) {
count[i] = 0;
for (j = 0; j <= VP9_SWITCHABLE_FILTERS; ++j) {
count[i] += cpi->switchable_interp_count[j][i];
}
c += (count[i] > 0);
}
if (c == 1) {
/* Only one filter is used. So set the filter at frame level */
for (i = 0; i < VP9_SWITCHABLE_FILTERS; ++i) {
if (count[i]) {
pc->mcomp_filter_type = vp9_switchable_interp[i];
break;
}
}
}
}
// Signal the type of subpel filter to use
vp9_write_bit(&header_bc, (pc->mcomp_filter_type == SWITCHABLE));
if (pc->mcomp_filter_type != SWITCHABLE)
vp9_write_literal(&header_bc, (pc->mcomp_filter_type), 2);
#if CONFIG_COMP_INTERINTRA_PRED
// printf("Counts: %d %d\n", cpi->interintra_count[0],
// cpi->interintra_count[1]);
if (!cpi->dummy_packing && pc->use_interintra)
pc->use_interintra = (cpi->interintra_count[1] > 0);
vp9_write_bit(&header_bc, pc->use_interintra);
if (!pc->use_interintra)
vp9_zero(cpi->interintra_count);
#endif
}
2010-05-18 17:58:33 +02:00
if (!pc->error_resilient_mode) {
vp9_write_bit(&header_bc, pc->refresh_entropy_probs);
vp9_write_bit(&header_bc, pc->frame_parallel_decoding_mode);
}
vp9_write_literal(&header_bc, pc->frame_context_idx,
NUM_FRAME_CONTEXTS_LG2);
2010-05-18 17:58:33 +02:00
#ifdef ENTROPY_STATS
if (pc->frame_type == INTER_FRAME)
active_section = 0;
else
active_section = 7;
2010-05-18 17:58:33 +02:00
#endif
// Signal whether or not Segmentation is enabled
vp9_write_bit(&header_bc, (xd->segmentation_enabled) ? 1 : 0);
// Indicate which features are enabled
if (xd->segmentation_enabled) {
// Indicate whether or not the segmentation map is being updated.
vp9_write_bit(&header_bc, (xd->update_mb_segmentation_map) ? 1 : 0);
#if CONFIG_IMPLICIT_SEGMENTATION
vp9_write_bit(&header_bc, (xd->allow_implicit_segment_update) ? 1 : 0);
#endif
// If it is, then indicate the method that will be used.
if (xd->update_mb_segmentation_map) {
// Select the coding strategy (temporal or spatial)
vp9_choose_segmap_coding_method(cpi);
// Send the tree probabilities used to decode unpredicted
// macro-block segments
for (i = 0; i < MB_SEG_TREE_PROBS; i++) {
const int prob = xd->mb_segment_tree_probs[i];
if (prob != 255) {
vp9_write_bit(&header_bc, 1);
vp9_write_prob(&header_bc, prob);
} else {
vp9_write_bit(&header_bc, 0);
}
}
// Write out the chosen coding method.
vp9_write_bit(&header_bc, (pc->temporal_update) ? 1 : 0);
if (pc->temporal_update) {
for (i = 0; i < PREDICTION_PROBS; i++) {
const int prob = pc->segment_pred_probs[i];
if (prob != 255) {
vp9_write_bit(&header_bc, 1);
vp9_write_prob(&header_bc, prob);
} else {
vp9_write_bit(&header_bc, 0);
}
}
}
}
vp9_write_bit(&header_bc, (xd->update_mb_segmentation_data) ? 1 : 0);
// segment_reference_frames(cpi);
if (xd->update_mb_segmentation_data) {
vp9_write_bit(&header_bc, (xd->mb_segment_abs_delta) ? 1 : 0);
// For each segments id...
for (i = 0; i < MAX_MB_SEGMENTS; i++) {
// For each segmentation codable feature...
for (j = 0; j < SEG_LVL_MAX; j++) {
const int8_t data = vp9_get_segdata(xd, i, j);
const int data_max = vp9_seg_feature_data_max(j);
// If the feature is enabled...
if (vp9_segfeature_active(xd, i, j)) {
vp9_write_bit(&header_bc, 1);
// Is the segment data signed..
if (vp9_is_segfeature_signed(j)) {
// Encode the relevant feature data
if (data < 0) {
vp9_encode_unsigned_max(&header_bc, -data, data_max);
vp9_write_bit(&header_bc, 1);
} else {
vp9_encode_unsigned_max(&header_bc, data, data_max);
vp9_write_bit(&header_bc, 0);
}
} else {
// Unsigned data element so no sign bit needed
vp9_encode_unsigned_max(&header_bc, data, data_max);
}
} else {
vp9_write_bit(&header_bc, 0);
}
}
}
}
}
// Encode the common prediction model status flag probability updates for
// the reference frame
update_refpred_stats(cpi);
if (pc->frame_type != KEY_FRAME) {
for (i = 0; i < PREDICTION_PROBS; i++) {
if (cpi->ref_pred_probs_update[i]) {
vp9_write_bit(&header_bc, 1);
vp9_write_prob(&header_bc, pc->ref_pred_probs[i]);
} else {
vp9_write_bit(&header_bc, 0);
}
}
}
if (cpi->mb.e_mbd.lossless) {
pc->txfm_mode = ONLY_4X4;
} else {
if (pc->txfm_mode == TX_MODE_SELECT) {
pc->prob_tx[0] = get_prob(cpi->txfm_count_32x32p[TX_4X4] +
cpi->txfm_count_16x16p[TX_4X4] +
cpi->txfm_count_8x8p[TX_4X4],
cpi->txfm_count_32x32p[TX_4X4] +
cpi->txfm_count_32x32p[TX_8X8] +
cpi->txfm_count_32x32p[TX_16X16] +
cpi->txfm_count_32x32p[TX_32X32] +
cpi->txfm_count_16x16p[TX_4X4] +
cpi->txfm_count_16x16p[TX_8X8] +
cpi->txfm_count_16x16p[TX_16X16] +
cpi->txfm_count_8x8p[TX_4X4] +
cpi->txfm_count_8x8p[TX_8X8]);
pc->prob_tx[1] = get_prob(cpi->txfm_count_32x32p[TX_8X8] +
cpi->txfm_count_16x16p[TX_8X8],
cpi->txfm_count_32x32p[TX_8X8] +
cpi->txfm_count_32x32p[TX_16X16] +
cpi->txfm_count_32x32p[TX_32X32] +
cpi->txfm_count_16x16p[TX_8X8] +
cpi->txfm_count_16x16p[TX_16X16]);
pc->prob_tx[2] = get_prob(cpi->txfm_count_32x32p[TX_16X16],
cpi->txfm_count_32x32p[TX_16X16] +
cpi->txfm_count_32x32p[TX_32X32]);
} else {
pc->prob_tx[0] = 128;
pc->prob_tx[1] = 128;
pc->prob_tx[2] = 128;
}
vp9_write_literal(&header_bc, pc->txfm_mode <= 3 ? pc->txfm_mode : 3, 2);
if (pc->txfm_mode > ALLOW_16X16) {
vp9_write_bit(&header_bc, pc->txfm_mode == TX_MODE_SELECT);
}
if (pc->txfm_mode == TX_MODE_SELECT) {
vp9_write_prob(&header_bc, pc->prob_tx[0]);
vp9_write_prob(&header_bc, pc->prob_tx[1]);
vp9_write_prob(&header_bc, pc->prob_tx[2]);
}
}
// If appropriate update the inter mode probability context and code the
// changes in the bitstream.
if (pc->frame_type != KEY_FRAME) {
int i, j;
int new_context[INTER_MODE_CONTEXTS][4];
if (!cpi->dummy_packing) {
update_inter_mode_probs(pc, new_context);
} else {
// In dummy pack assume context unchanged.
vpx_memcpy(new_context, pc->fc.vp9_mode_contexts,
sizeof(pc->fc.vp9_mode_contexts));
}
for (i = 0; i < INTER_MODE_CONTEXTS; i++) {
for (j = 0; j < 4; j++) {
if (new_context[i][j] != pc->fc.vp9_mode_contexts[i][j]) {
vp9_write(&header_bc, 1, 252);
vp9_write_prob(&header_bc, new_context[i][j]);
// Only update the persistent copy if this is the "real pack"
if (!cpi->dummy_packing) {
pc->fc.vp9_mode_contexts[i][j] = new_context[i][j];
}
} else {
vp9_write(&header_bc, 0, 252);
}
}
}
}
vp9_clear_system_state(); // __asm emms;
vp9_copy(cpi->common.fc.pre_coef_probs_4x4,
cpi->common.fc.coef_probs_4x4);
vp9_copy(cpi->common.fc.pre_coef_probs_8x8,
cpi->common.fc.coef_probs_8x8);
vp9_copy(cpi->common.fc.pre_coef_probs_16x16,
cpi->common.fc.coef_probs_16x16);
32x32 transform for superblocks. This adds Debargha's DCT/DWT hybrid and a regular 32x32 DCT, and adds code all over the place to wrap that in the bitstream/encoder/decoder/RD. Some implementation notes (these probably need careful review): - token range is extended by 1 bit, since the value range out of this transform is [-16384,16383]. - the coefficients coming out of the FDCT are manually scaled back by 1 bit, or else they won't fit in int16_t (they are 17 bits). Because of this, the RD error scoring does not right-shift the MSE score by two (unlike for 4x4/8x8/16x16). - to compensate for this loss in precision, the quantizer is halved also. This is currently a little hacky. - FDCT and IDCT is double-only right now. Needs a fixed-point impl. - There are no default probabilities for the 32x32 transform yet; I'm simply using the 16x16 luma ones. A future commit will add newly generated probabilities for all transforms. - No ADST version. I don't think we'll add one for this level; if an ADST is desired, transform-size selection can scale back to 16x16 or lower, and use an ADST at that level. Additional notes specific to Debargha's DWT/DCT hybrid: - coefficient scale is different for the top/left 16x16 (DCT-over-DWT) block than for the rest (DWT pixel differences) of the block. Therefore, RD error scoring isn't easily scalable between coefficient and pixel domain. Thus, unfortunately, we need to compute the RD distortion in the pixel domain until we figure out how to scale these appropriately. Change-Id: I00386f20f35d7fabb19aba94c8162f8aee64ef2b
2012-12-07 23:45:05 +01:00
vp9_copy(cpi->common.fc.pre_coef_probs_32x32,
cpi->common.fc.coef_probs_32x32);
#if CONFIG_CODE_ZEROGROUP
vp9_copy(cpi->common.fc.pre_zpc_probs_4x4,
cpi->common.fc.zpc_probs_4x4);
vp9_copy(cpi->common.fc.pre_zpc_probs_8x8,
cpi->common.fc.zpc_probs_8x8);
vp9_copy(cpi->common.fc.pre_zpc_probs_16x16,
cpi->common.fc.zpc_probs_16x16);
vp9_copy(cpi->common.fc.pre_zpc_probs_32x32,
cpi->common.fc.zpc_probs_32x32);
#endif
vp9_copy(cpi->common.fc.pre_sb_ymode_prob, cpi->common.fc.sb_ymode_prob);
vp9_copy(cpi->common.fc.pre_ymode_prob, cpi->common.fc.ymode_prob);
vp9_copy(cpi->common.fc.pre_uv_mode_prob, cpi->common.fc.uv_mode_prob);
vp9_copy(cpi->common.fc.pre_bmode_prob, cpi->common.fc.bmode_prob);
vp9_copy(cpi->common.fc.pre_sub_mv_ref_prob, cpi->common.fc.sub_mv_ref_prob);
vp9_copy(cpi->common.fc.pre_mbsplit_prob, cpi->common.fc.mbsplit_prob);
vp9_copy(cpi->common.fc.pre_i8x8_mode_prob, cpi->common.fc.i8x8_mode_prob);
vp9_copy(cpi->common.fc.pre_partition_prob, cpi->common.fc.partition_prob);
cpi->common.fc.pre_nmvc = cpi->common.fc.nmvc;
#if CONFIG_COMP_INTERINTRA_PRED
cpi->common.fc.pre_interintra_prob = cpi->common.fc.interintra_prob;
#endif
vp9_zero(cpi->sub_mv_ref_count);
vp9_zero(cpi->mbsplit_count);
vp9_zero(cpi->common.fc.mv_ref_ct);
update_coef_probs(cpi, &header_bc);
#if CONFIG_CODE_ZEROGROUP
update_zpc_probs(cpi, &header_bc);
#endif
2010-05-18 17:58:33 +02:00
#ifdef ENTROPY_STATS
active_section = 2;
2010-05-18 17:58:33 +02:00
#endif
// TODO(jkoleszar): remove this unused bit
vp9_write_bit(&header_bc, 1);
vp9_update_skip_probs(cpi);
for (i = 0; i < MBSKIP_CONTEXTS; ++i) {
vp9_write_prob(&header_bc, pc->mbskip_pred_probs[i]);
}
if (pc->frame_type == KEY_FRAME) {
if (!pc->kf_ymode_probs_update) {
vp9_write_literal(&header_bc, pc->kf_ymode_probs_index, 3);
}
} else {
// Update the probabilities used to encode reference frame data
update_ref_probs(cpi);
#ifdef ENTROPY_STATS
active_section = 1;
#endif
if (pc->mcomp_filter_type == SWITCHABLE)
update_switchable_interp_probs(cpi, &header_bc);
#if CONFIG_COMP_INTERINTRA_PRED
if (pc->use_interintra) {
vp9_cond_prob_update(&header_bc,
&pc->fc.interintra_prob,
VP9_UPD_INTERINTRA_PROB,
cpi->interintra_count);
}
#endif
vp9_write_prob(&header_bc, pc->prob_intra_coded);
vp9_write_prob(&header_bc, pc->prob_last_coded);
vp9_write_prob(&header_bc, pc->prob_gf_coded);
{
const int comp_pred_mode = cpi->common.comp_pred_mode;
const int use_compound_pred = (comp_pred_mode != SINGLE_PREDICTION_ONLY);
const int use_hybrid_pred = (comp_pred_mode == HYBRID_PREDICTION);
vp9_write_bit(&header_bc, use_compound_pred);
if (use_compound_pred) {
vp9_write_bit(&header_bc, use_hybrid_pred);
if (use_hybrid_pred) {
for (i = 0; i < COMP_PRED_CONTEXTS; i++) {
pc->prob_comppred[i] = get_binary_prob(cpi->single_pred_count[i],
cpi->comp_pred_count[i]);
vp9_write_prob(&header_bc, pc->prob_comppred[i]);
}
}
}
}
update_mbintra_mode_probs(cpi, &header_bc);
vp9_write_nmv_probs(cpi, xd->allow_high_precision_mv, &header_bc);
}
2010-05-18 17:58:33 +02:00
[WIP] Add column-based tiling. This patch adds column-based tiling. The idea is to make each tile independently decodable (after reading the common frame header) and also independendly encodable (minus within-frame cost adjustments in the RD loop) to speed-up hardware & software en/decoders if they used multi-threading. Column-based tiling has the added advantage (over other tiling methods) that it minimizes realtime use-case latency, since all threads can start encoding data as soon as the first SB-row worth of data is available to the encoder. There is some test code that does random tile ordering in the decoder, to confirm that each tile is indeed independently decodable from other tiles in the same frame. At tile edges, all contexts assume default values (i.e. 0, 0 motion vector, no coefficients, DC intra4x4 mode), and motion vector search and ordering do not cross tiles in the same frame. t log Tile independence is not maintained between frames ATM, i.e. tile 0 of frame 1 is free to use motion vectors that point into any tile of frame 0. We support 1 (i.e. no tiling), 2 or 4 column-tiles. The loopfilter crosses tile boundaries. I discussed this briefly with Aki and he says that's OK. An in-loop loopfilter would need to do some sync between tile threads, but that shouldn't be a big issue. Resuls: with tiling disabled, we go up slightly because of improved edge use in the intra4x4 prediction. With 2 tiles, we lose about ~1% on derf, ~0.35% on HD and ~0.55% on STD/HD. With 4 tiles, we lose another ~1.5% on derf ~0.77% on HD and ~0.85% on STD/HD. Most of this loss is concentrated in the low-bitrate end of clips, and most of it is because of the loss of edges at tile boundaries and the resulting loss of intra predictors. TODO: - more tiles (perhaps allow row-based tiling also, and max. 8 tiles)? - maybe optionally (for EC purposes), motion vectors themselves should not cross tile edges, or we should emulate such borders as if they were off-frame, to limit error propagation to within one tile only. This doesn't have to be the default behaviour but could be an optional bitstream flag. Change-Id: I5951c3a0742a767b20bc9fb5af685d9892c2c96f
2013-02-01 18:35:28 +01:00
/* tiling */
{
int min_log2_tiles, delta_log2_tiles, n_tile_bits, n;
vp9_get_tile_n_bits(pc, &min_log2_tiles, &delta_log2_tiles);
n_tile_bits = pc->log2_tile_columns - min_log2_tiles;
for (n = 0; n < delta_log2_tiles; n++) {
if (n_tile_bits--) {
vp9_write_bit(&header_bc, 1);
} else {
vp9_write_bit(&header_bc, 0);
break;
}
}
vp9_write_bit(&header_bc, pc->log2_tile_rows != 0);
if (pc->log2_tile_rows != 0)
vp9_write_bit(&header_bc, pc->log2_tile_rows != 1);
[WIP] Add column-based tiling. This patch adds column-based tiling. The idea is to make each tile independently decodable (after reading the common frame header) and also independendly encodable (minus within-frame cost adjustments in the RD loop) to speed-up hardware & software en/decoders if they used multi-threading. Column-based tiling has the added advantage (over other tiling methods) that it minimizes realtime use-case latency, since all threads can start encoding data as soon as the first SB-row worth of data is available to the encoder. There is some test code that does random tile ordering in the decoder, to confirm that each tile is indeed independently decodable from other tiles in the same frame. At tile edges, all contexts assume default values (i.e. 0, 0 motion vector, no coefficients, DC intra4x4 mode), and motion vector search and ordering do not cross tiles in the same frame. t log Tile independence is not maintained between frames ATM, i.e. tile 0 of frame 1 is free to use motion vectors that point into any tile of frame 0. We support 1 (i.e. no tiling), 2 or 4 column-tiles. The loopfilter crosses tile boundaries. I discussed this briefly with Aki and he says that's OK. An in-loop loopfilter would need to do some sync between tile threads, but that shouldn't be a big issue. Resuls: with tiling disabled, we go up slightly because of improved edge use in the intra4x4 prediction. With 2 tiles, we lose about ~1% on derf, ~0.35% on HD and ~0.55% on STD/HD. With 4 tiles, we lose another ~1.5% on derf ~0.77% on HD and ~0.85% on STD/HD. Most of this loss is concentrated in the low-bitrate end of clips, and most of it is because of the loss of edges at tile boundaries and the resulting loss of intra predictors. TODO: - more tiles (perhaps allow row-based tiling also, and max. 8 tiles)? - maybe optionally (for EC purposes), motion vectors themselves should not cross tile edges, or we should emulate such borders as if they were off-frame, to limit error propagation to within one tile only. This doesn't have to be the default behaviour but could be an optional bitstream flag. Change-Id: I5951c3a0742a767b20bc9fb5af685d9892c2c96f
2013-02-01 18:35:28 +01:00
}
vp9_stop_encode(&header_bc);
2010-05-18 17:58:33 +02:00
oh.first_partition_length_in_bytes = header_bc.pos;
/* update frame tag */
{
int scaling = (pc->width != pc->display_width ||
pc->height != pc->display_height);
int v = (oh.first_partition_length_in_bytes << 8) |
(scaling << 5) |
(oh.show_frame << 4) |
(oh.version << 1) |
oh.type;
assert(oh.first_partition_length_in_bytes <= 0xffff);
dest[0] = v;
dest[1] = v >> 8;
dest[2] = v >> 16;
}
*size = VP9_HEADER_SIZE + extra_bytes_packed + header_bc.pos;
2010-05-18 17:58:33 +02:00
if (pc->frame_type == KEY_FRAME) {
decide_kf_ymode_entropy(cpi);
} else {
/* This is not required if the counts in cpi are consistent with the
* final packing pass */
// if (!cpi->dummy_packing) vp9_zero(cpi->NMVcount);
}
[WIP] Add column-based tiling. This patch adds column-based tiling. The idea is to make each tile independently decodable (after reading the common frame header) and also independendly encodable (minus within-frame cost adjustments in the RD loop) to speed-up hardware & software en/decoders if they used multi-threading. Column-based tiling has the added advantage (over other tiling methods) that it minimizes realtime use-case latency, since all threads can start encoding data as soon as the first SB-row worth of data is available to the encoder. There is some test code that does random tile ordering in the decoder, to confirm that each tile is indeed independently decodable from other tiles in the same frame. At tile edges, all contexts assume default values (i.e. 0, 0 motion vector, no coefficients, DC intra4x4 mode), and motion vector search and ordering do not cross tiles in the same frame. t log Tile independence is not maintained between frames ATM, i.e. tile 0 of frame 1 is free to use motion vectors that point into any tile of frame 0. We support 1 (i.e. no tiling), 2 or 4 column-tiles. The loopfilter crosses tile boundaries. I discussed this briefly with Aki and he says that's OK. An in-loop loopfilter would need to do some sync between tile threads, but that shouldn't be a big issue. Resuls: with tiling disabled, we go up slightly because of improved edge use in the intra4x4 prediction. With 2 tiles, we lose about ~1% on derf, ~0.35% on HD and ~0.55% on STD/HD. With 4 tiles, we lose another ~1.5% on derf ~0.77% on HD and ~0.85% on STD/HD. Most of this loss is concentrated in the low-bitrate end of clips, and most of it is because of the loss of edges at tile boundaries and the resulting loss of intra predictors. TODO: - more tiles (perhaps allow row-based tiling also, and max. 8 tiles)? - maybe optionally (for EC purposes), motion vectors themselves should not cross tile edges, or we should emulate such borders as if they were off-frame, to limit error propagation to within one tile only. This doesn't have to be the default behaviour but could be an optional bitstream flag. Change-Id: I5951c3a0742a767b20bc9fb5af685d9892c2c96f
2013-02-01 18:35:28 +01:00
{
int tile_row, tile_col, total_size = 0;
[WIP] Add column-based tiling. This patch adds column-based tiling. The idea is to make each tile independently decodable (after reading the common frame header) and also independendly encodable (minus within-frame cost adjustments in the RD loop) to speed-up hardware & software en/decoders if they used multi-threading. Column-based tiling has the added advantage (over other tiling methods) that it minimizes realtime use-case latency, since all threads can start encoding data as soon as the first SB-row worth of data is available to the encoder. There is some test code that does random tile ordering in the decoder, to confirm that each tile is indeed independently decodable from other tiles in the same frame. At tile edges, all contexts assume default values (i.e. 0, 0 motion vector, no coefficients, DC intra4x4 mode), and motion vector search and ordering do not cross tiles in the same frame. t log Tile independence is not maintained between frames ATM, i.e. tile 0 of frame 1 is free to use motion vectors that point into any tile of frame 0. We support 1 (i.e. no tiling), 2 or 4 column-tiles. The loopfilter crosses tile boundaries. I discussed this briefly with Aki and he says that's OK. An in-loop loopfilter would need to do some sync between tile threads, but that shouldn't be a big issue. Resuls: with tiling disabled, we go up slightly because of improved edge use in the intra4x4 prediction. With 2 tiles, we lose about ~1% on derf, ~0.35% on HD and ~0.55% on STD/HD. With 4 tiles, we lose another ~1.5% on derf ~0.77% on HD and ~0.85% on STD/HD. Most of this loss is concentrated in the low-bitrate end of clips, and most of it is because of the loss of edges at tile boundaries and the resulting loss of intra predictors. TODO: - more tiles (perhaps allow row-based tiling also, and max. 8 tiles)? - maybe optionally (for EC purposes), motion vectors themselves should not cross tile edges, or we should emulate such borders as if they were off-frame, to limit error propagation to within one tile only. This doesn't have to be the default behaviour but could be an optional bitstream flag. Change-Id: I5951c3a0742a767b20bc9fb5af685d9892c2c96f
2013-02-01 18:35:28 +01:00
unsigned char *data_ptr = cx_data + header_bc.pos;
TOKENEXTRA *tok[1 << 6], *tok_end;
tok[0] = cpi->tok;
for (tile_col = 1; tile_col < pc->tile_columns; tile_col++)
tok[tile_col] = tok[tile_col - 1] + cpi->tok_count[tile_col - 1];
for (tile_row = 0; tile_row < pc->tile_rows; tile_row++) {
vp9_get_tile_row_offsets(pc, tile_row);
tok_end = cpi->tok + cpi->tok_count[0];
for (tile_col = 0; tile_col < pc->tile_columns;
tile_col++, tok_end += cpi->tok_count[tile_col]) {
vp9_get_tile_col_offsets(pc, tile_col);
if (tile_col < pc->tile_columns - 1 || tile_row < pc->tile_rows - 1)
vp9_start_encode(&residual_bc, data_ptr + total_size + 4);
else
vp9_start_encode(&residual_bc, data_ptr + total_size);
write_modes(cpi, &residual_bc, &tok[tile_col], tok_end);
vp9_stop_encode(&residual_bc);
if (tile_col < pc->tile_columns - 1 || tile_row < pc->tile_rows - 1) {
// size of this tile
write_le32(data_ptr + total_size, residual_bc.pos);
total_size += 4;
}
2010-05-18 17:58:33 +02:00
total_size += residual_bc.pos;
}
[WIP] Add column-based tiling. This patch adds column-based tiling. The idea is to make each tile independently decodable (after reading the common frame header) and also independendly encodable (minus within-frame cost adjustments in the RD loop) to speed-up hardware & software en/decoders if they used multi-threading. Column-based tiling has the added advantage (over other tiling methods) that it minimizes realtime use-case latency, since all threads can start encoding data as soon as the first SB-row worth of data is available to the encoder. There is some test code that does random tile ordering in the decoder, to confirm that each tile is indeed independently decodable from other tiles in the same frame. At tile edges, all contexts assume default values (i.e. 0, 0 motion vector, no coefficients, DC intra4x4 mode), and motion vector search and ordering do not cross tiles in the same frame. t log Tile independence is not maintained between frames ATM, i.e. tile 0 of frame 1 is free to use motion vectors that point into any tile of frame 0. We support 1 (i.e. no tiling), 2 or 4 column-tiles. The loopfilter crosses tile boundaries. I discussed this briefly with Aki and he says that's OK. An in-loop loopfilter would need to do some sync between tile threads, but that shouldn't be a big issue. Resuls: with tiling disabled, we go up slightly because of improved edge use in the intra4x4 prediction. With 2 tiles, we lose about ~1% on derf, ~0.35% on HD and ~0.55% on STD/HD. With 4 tiles, we lose another ~1.5% on derf ~0.77% on HD and ~0.85% on STD/HD. Most of this loss is concentrated in the low-bitrate end of clips, and most of it is because of the loss of edges at tile boundaries and the resulting loss of intra predictors. TODO: - more tiles (perhaps allow row-based tiling also, and max. 8 tiles)? - maybe optionally (for EC purposes), motion vectors themselves should not cross tile edges, or we should emulate such borders as if they were off-frame, to limit error propagation to within one tile only. This doesn't have to be the default behaviour but could be an optional bitstream flag. Change-Id: I5951c3a0742a767b20bc9fb5af685d9892c2c96f
2013-02-01 18:35:28 +01:00
}
assert((unsigned int)(tok[0] - cpi->tok) == cpi->tok_count[0]);
for (tile_col = 1; tile_col < pc->tile_columns; tile_col++)
assert((unsigned int)(tok[tile_col] - tok[tile_col - 1]) ==
cpi->tok_count[tile_col]);
[WIP] Add column-based tiling. This patch adds column-based tiling. The idea is to make each tile independently decodable (after reading the common frame header) and also independendly encodable (minus within-frame cost adjustments in the RD loop) to speed-up hardware & software en/decoders if they used multi-threading. Column-based tiling has the added advantage (over other tiling methods) that it minimizes realtime use-case latency, since all threads can start encoding data as soon as the first SB-row worth of data is available to the encoder. There is some test code that does random tile ordering in the decoder, to confirm that each tile is indeed independently decodable from other tiles in the same frame. At tile edges, all contexts assume default values (i.e. 0, 0 motion vector, no coefficients, DC intra4x4 mode), and motion vector search and ordering do not cross tiles in the same frame. t log Tile independence is not maintained between frames ATM, i.e. tile 0 of frame 1 is free to use motion vectors that point into any tile of frame 0. We support 1 (i.e. no tiling), 2 or 4 column-tiles. The loopfilter crosses tile boundaries. I discussed this briefly with Aki and he says that's OK. An in-loop loopfilter would need to do some sync between tile threads, but that shouldn't be a big issue. Resuls: with tiling disabled, we go up slightly because of improved edge use in the intra4x4 prediction. With 2 tiles, we lose about ~1% on derf, ~0.35% on HD and ~0.55% on STD/HD. With 4 tiles, we lose another ~1.5% on derf ~0.77% on HD and ~0.85% on STD/HD. Most of this loss is concentrated in the low-bitrate end of clips, and most of it is because of the loss of edges at tile boundaries and the resulting loss of intra predictors. TODO: - more tiles (perhaps allow row-based tiling also, and max. 8 tiles)? - maybe optionally (for EC purposes), motion vectors themselves should not cross tile edges, or we should emulate such borders as if they were off-frame, to limit error propagation to within one tile only. This doesn't have to be the default behaviour but could be an optional bitstream flag. Change-Id: I5951c3a0742a767b20bc9fb5af685d9892c2c96f
2013-02-01 18:35:28 +01:00
*size += total_size;
}
2010-05-18 17:58:33 +02:00
}
#ifdef ENTROPY_STATS
static void print_tree_update_for_type(FILE *f,
vp9_coeff_stats *tree_update_hist,
int block_types, const char *header) {
int i, j, k, l, m;
fprintf(f, "const vp9_coeff_prob %s = {\n", header);
for (i = 0; i < block_types; i++) {
fprintf(f, " { \n");
for (j = 0; j < REF_TYPES; j++) {
fprintf(f, " { \n");
for (k = 0; k < COEF_BANDS; k++) {
fprintf(f, " {\n");
for (l = 0; l < PREV_COEF_CONTEXTS; l++) {
fprintf(f, " {");
for (m = 0; m < ENTROPY_NODES; m++) {
fprintf(f, "%3d, ",
get_binary_prob(tree_update_hist[i][j][k][l][m][0],
tree_update_hist[i][j][k][l][m][1]));
}
fprintf(f, "},\n");
2010-05-18 17:58:33 +02:00
}
fprintf(f, "},\n");
}
fprintf(f, " },\n");
2010-05-18 17:58:33 +02:00
}
fprintf(f, " },\n");
}
fprintf(f, "};\n");
}
void print_tree_update_probs() {
FILE *f = fopen("coefupdprob.h", "w");
fprintf(f, "\n/* Update probabilities for token entropy tree. */\n\n");
print_tree_update_for_type(f, tree_update_hist_4x4, BLOCK_TYPES,
"vp9_coef_update_probs_4x4[BLOCK_TYPES]");
print_tree_update_for_type(f, tree_update_hist_8x8, BLOCK_TYPES,
"vp9_coef_update_probs_8x8[BLOCK_TYPES]");
print_tree_update_for_type(f, tree_update_hist_16x16, BLOCK_TYPES,
"vp9_coef_update_probs_16x16[BLOCK_TYPES]");
print_tree_update_for_type(f, tree_update_hist_32x32, BLOCK_TYPES,
"vp9_coef_update_probs_32x32[BLOCK_TYPES]");
fclose(f);
f = fopen("treeupdate.bin", "wb");
fwrite(tree_update_hist_4x4, sizeof(tree_update_hist_4x4), 1, f);
fwrite(tree_update_hist_8x8, sizeof(tree_update_hist_8x8), 1, f);
fwrite(tree_update_hist_16x16, sizeof(tree_update_hist_16x16), 1, f);
fwrite(tree_update_hist_32x32, sizeof(tree_update_hist_32x32), 1, f);
fclose(f);
2010-05-18 17:58:33 +02:00
}
#endif