vpx/vp9/encoder/vp9_tokenize.c

338 lines
10 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 <math.h>
#include <stdio.h>
#include <string.h>
#include <assert.h>
#include "vp9/encoder/vp9_onyx_int.h"
#include "vp9/encoder/vp9_tokenize.h"
2010-05-18 17:58:33 +02:00
#include "vpx_mem/vpx_mem.h"
#include "vp9/common/vp9_pred_common.h"
#include "vp9/common/vp9_seg_common.h"
#include "vp9/common/vp9_entropy.h"
2011-10-05 12:26:00 +02:00
static TOKENVALUE dct_value_tokens[DCT_MAX_VALUE * 2];
const TOKENVALUE *vp9_dct_value_tokens_ptr;
static int dct_value_cost[DCT_MAX_VALUE * 2];
const int *vp9_dct_value_cost_ptr;
// Array indices are identical to previously-existing CONTEXT_NODE indices
const vp9_tree_index vp9_coef_tree[TREE_SIZE(ENTROPY_TOKENS)] = {
-EOB_TOKEN, 2, // 0 = EOB
-ZERO_TOKEN, 4, // 1 = ZERO
-ONE_TOKEN, 6, // 2 = ONE
8, 12, // 3 = LOW_VAL
-TWO_TOKEN, 10, // 4 = TWO
-THREE_TOKEN, -FOUR_TOKEN, // 5 = THREE
14, 16, // 6 = HIGH_LOW
-CATEGORY1_TOKEN, -CATEGORY2_TOKEN, // 7 = CAT_ONE
18, 20, // 8 = CAT_THREEFOUR
-CATEGORY3_TOKEN, -CATEGORY4_TOKEN, // 9 = CAT_THREE
-CATEGORY5_TOKEN, -CATEGORY6_TOKEN // 10 = CAT_FIVE
};
// Unconstrained Node Tree
const vp9_tree_index vp9_coef_con_tree[TREE_SIZE(ENTROPY_TOKENS)] = {
2, 6, // 0 = LOW_VAL
-TWO_TOKEN, 4, // 1 = TWO
-THREE_TOKEN, -FOUR_TOKEN, // 2 = THREE
8, 10, // 3 = HIGH_LOW
-CATEGORY1_TOKEN, -CATEGORY2_TOKEN, // 4 = CAT_ONE
12, 14, // 5 = CAT_THREEFOUR
-CATEGORY3_TOKEN, -CATEGORY4_TOKEN, // 6 = CAT_THREE
-CATEGORY5_TOKEN, -CATEGORY6_TOKEN // 7 = CAT_FIVE
};
static const vp9_prob Pcat1[] = { 159};
static const vp9_prob Pcat2[] = { 165, 145};
static const vp9_prob Pcat3[] = { 173, 148, 140};
static const vp9_prob Pcat4[] = { 176, 155, 140, 135};
static const vp9_prob Pcat5[] = { 180, 157, 141, 134, 130};
static const vp9_prob Pcat6[] = {
254, 254, 254, 252, 249, 243, 230, 196, 177, 153, 140, 133, 130, 129
};
static vp9_tree_index cat1[2], cat2[4], cat3[6], cat4[8], cat5[10], cat6[28];
static void init_bit_tree(vp9_tree_index *p, int n) {
int i = 0;
while (++i < n) {
p[0] = p[1] = i << 1;
p += 2;
}
p[0] = p[1] = 0;
}
static void init_bit_trees() {
init_bit_tree(cat1, 1);
init_bit_tree(cat2, 2);
init_bit_tree(cat3, 3);
init_bit_tree(cat4, 4);
init_bit_tree(cat5, 5);
init_bit_tree(cat6, 14);
}
const vp9_extra_bit vp9_extra_bits[ENTROPY_TOKENS] = {
{0, 0, 0, 0}, // ZERO_TOKEN
{0, 0, 0, 1}, // ONE_TOKEN
{0, 0, 0, 2}, // TWO_TOKEN
{0, 0, 0, 3}, // THREE_TOKEN
{0, 0, 0, 4}, // FOUR_TOKEN
{cat1, Pcat1, 1, 5}, // CATEGORY1_TOKEN
{cat2, Pcat2, 2, 7}, // CATEGORY2_TOKEN
{cat3, Pcat3, 3, 11}, // CATEGORY3_TOKEN
{cat4, Pcat4, 4, 19}, // CATEGORY4_TOKEN
{cat5, Pcat5, 5, 35}, // CATEGORY5_TOKEN
{cat6, Pcat6, 14, 67}, // CATEGORY6_TOKEN
{0, 0, 0, 0} // EOB_TOKEN
};
struct vp9_token vp9_coef_encodings[ENTROPY_TOKENS];
void vp9_coef_tree_initialize() {
init_bit_trees();
vp9_tokens_from_tree(vp9_coef_encodings, vp9_coef_tree);
}
static void fill_value_tokens() {
TOKENVALUE *const t = dct_value_tokens + DCT_MAX_VALUE;
const vp9_extra_bit *const e = vp9_extra_bits;
2010-05-18 17:58:33 +02:00
int i = -DCT_MAX_VALUE;
int sign = 1;
2010-05-18 17:58:33 +02:00
do {
if (!i)
sign = 0;
2010-05-18 17:58:33 +02:00
{
const int a = sign ? -i : i;
int eb = sign;
2010-05-18 17:58:33 +02:00
if (a > 4) {
int j = 4;
2010-05-18 17:58:33 +02:00
while (++j < 11 && e[j].base_val <= a) {}
2010-05-18 17:58:33 +02:00
t[i].token = --j;
eb |= (a - e[j].base_val) << 1;
} else {
t[i].token = a;
}
t[i].extra = eb;
}
2010-05-18 17:58:33 +02:00
// initialize the cost for extra bits for all possible coefficient value.
{
int cost = 0;
const vp9_extra_bit *p = &vp9_extra_bits[t[i].token];
2010-05-18 17:58:33 +02:00
if (p->base_val) {
const int extra = t[i].extra;
const int length = p->len;
2010-05-18 17:58:33 +02:00
if (length)
cost += treed_cost(p->tree, p->prob, extra >> 1, length);
2010-05-18 17:58:33 +02:00
cost += vp9_cost_bit(vp9_prob_half, extra & 1); /* sign */
dct_value_cost[i + DCT_MAX_VALUE] = cost;
}
2010-05-18 17:58:33 +02:00
}
} while (++i < DCT_MAX_VALUE);
vp9_dct_value_tokens_ptr = dct_value_tokens + DCT_MAX_VALUE;
vp9_dct_value_cost_ptr = dct_value_cost + DCT_MAX_VALUE;
2010-05-18 17:58:33 +02:00
}
struct tokenize_b_args {
VP9_COMP *cpi;
MACROBLOCKD *xd;
TOKENEXTRA **tp;
uint8_t *token_cache;
};
static void set_entropy_context_b(int plane, int block, BLOCK_SIZE plane_bsize,
TX_SIZE tx_size, void *arg) {
struct tokenize_b_args* const args = arg;
MACROBLOCKD *const xd = args->xd;
struct macroblock_plane *p = &args->cpi->mb.plane[plane];
struct macroblockd_plane *pd = &xd->plane[plane];
int aoff, loff;
txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &aoff, &loff);
vp9_set_contexts(xd, pd, plane_bsize, tx_size, p->eobs[block] > 0,
aoff, loff);
}
static INLINE void add_token(TOKENEXTRA **t, const vp9_prob *context_tree,
int16_t extra, uint8_t token,
uint8_t skip_eob_node,
unsigned int *counts) {
(*t)->token = token;
(*t)->extra = extra;
(*t)->context_tree = context_tree;
(*t)->skip_eob_node = skip_eob_node;
(*t)++;
++counts[token];
}
static INLINE void add_token_no_extra(TOKENEXTRA **t,
const vp9_prob *context_tree,
uint8_t token,
uint8_t skip_eob_node,
unsigned int *counts) {
(*t)->token = token;
(*t)->context_tree = context_tree;
(*t)->skip_eob_node = skip_eob_node;
(*t)++;
++counts[token];
}
static void tokenize_b(int plane, int block, BLOCK_SIZE plane_bsize,
TX_SIZE tx_size, void *arg) {
struct tokenize_b_args* const args = arg;
VP9_COMP *cpi = args->cpi;
MACROBLOCKD *xd = args->xd;
TOKENEXTRA **tp = args->tp;
uint8_t *token_cache = args->token_cache;
struct macroblock_plane *p = &cpi->mb.plane[plane];
struct macroblockd_plane *pd = &xd->plane[plane];
MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
int pt; /* near block/prev token context index */
int c;
TOKENEXTRA *t = *tp; /* store tokens starting here */
int eob = p->eobs[block];
const PLANE_TYPE type = pd->plane_type;
const int16_t *qcoeff_ptr = BLOCK_OFFSET(p->qcoeff, block);
const int segment_id = mbmi->segment_id;
const int16_t *scan, *nb;
const scan_order *so;
const int ref = is_inter_block(mbmi);
unsigned int (*const counts)[COEFF_CONTEXTS][ENTROPY_TOKENS] =
cpi->coef_counts[tx_size][type][ref];
vp9_prob (*const coef_probs)[COEFF_CONTEXTS][UNCONSTRAINED_NODES] =
cpi->common.fc.coef_probs[tx_size][type][ref];
unsigned int (*const eob_branch)[COEFF_CONTEXTS] =
cpi->common.counts.eob_branch[tx_size][type][ref];
const uint8_t *const band = get_band_translate(tx_size);
const int seg_eob = get_tx_eob(&cpi->common.seg, segment_id, tx_size);
int aoff, loff;
txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &aoff, &loff);
pt = get_entropy_context(tx_size, pd->above_context + aoff,
pd->left_context + loff);
so = get_scan(xd, tx_size, type, block);
scan = so->scan;
nb = so->neighbors;
c = 0;
while (c < eob) {
int v = 0;
int skip_eob = 0;
v = qcoeff_ptr[scan[c]];
while (!v) {
add_token_no_extra(&t, coef_probs[band[c]][pt], ZERO_TOKEN, skip_eob,
counts[band[c]][pt]);
eob_branch[band[c]][pt] += !skip_eob;
skip_eob = 1;
token_cache[scan[c]] = 0;
++c;
pt = get_coef_context(nb, token_cache, c);
v = qcoeff_ptr[scan[c]];
}
add_token(&t, coef_probs[band[c]][pt],
vp9_dct_value_tokens_ptr[v].extra,
vp9_dct_value_tokens_ptr[v].token, skip_eob,
counts[band[c]][pt]);
eob_branch[band[c]][pt] += !skip_eob;
token_cache[scan[c]] =
vp9_pt_energy_class[vp9_dct_value_tokens_ptr[v].token];
++c;
pt = get_coef_context(nb, token_cache, c);
}
if (c < seg_eob) {
add_token_no_extra(&t, coef_probs[band[c]][pt], EOB_TOKEN, 0,
counts[band[c]][pt]);
++eob_branch[band[c]][pt];
}
*tp = t;
vp9_set_contexts(xd, pd, plane_bsize, tx_size, c > 0, aoff, loff);
2010-05-18 17:58:33 +02:00
}
struct is_skippable_args {
MACROBLOCK *x;
int *skippable;
};
static void is_skippable(int plane, int block,
BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
void *argv) {
struct is_skippable_args *args = argv;
args->skippable[0] &= (!args->x->plane[plane].eobs[block]);
}
static int sb_is_skippable(MACROBLOCK *x, BLOCK_SIZE bsize) {
int result = 1;
struct is_skippable_args args = {x, &result};
vp9_foreach_transformed_block(&x->e_mbd, bsize, is_skippable, &args);
return result;
}
int vp9_is_skippable_in_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
int result = 1;
struct is_skippable_args args = {x, &result};
vp9_foreach_transformed_block_in_plane(&x->e_mbd, bsize, plane, is_skippable,
&args);
return result;
}
void vp9_tokenize_sb(VP9_COMP *cpi, TOKENEXTRA **t, int dry_run,
BLOCK_SIZE bsize) {
VP9_COMMON *const cm = &cpi->common;
MACROBLOCKD *const xd = &cpi->mb.e_mbd;
MB_MODE_INFO *const mbmi = &xd->mi_8x8[0]->mbmi;
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
TOKENEXTRA *t_backup = *t;
const int ctx = vp9_get_skip_context(xd);
const int skip_inc = !vp9_segfeature_active(&cm->seg, mbmi->segment_id,
SEG_LVL_SKIP);
struct tokenize_b_args arg = {cpi, xd, t, cpi->mb.token_cache};
if (mbmi->skip) {
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 (!dry_run)
cm->counts.skip[ctx][1] += skip_inc;
reset_skip_context(xd, bsize);
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 (dry_run)
*t = t_backup;
return;
}
if (!dry_run) {
cm->counts.skip[ctx][0] += skip_inc;
vp9_foreach_transformed_block(xd, bsize, tokenize_b, &arg);
} else {
vp9_foreach_transformed_block(xd, bsize, set_entropy_context_b, &arg);
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
*t = t_backup;
}
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
}
void vp9_tokenize_initialize() {
fill_value_tokens();
2010-05-18 17:58:33 +02:00
}