aom/vp8/encoder/tokenize.c

558 строки
14 KiB
C
Исходник Обычный вид История

2010-05-18 19:58:33 +04:00
/*
* Copyright (c) 2010 The VP8 project authors. All Rights Reserved.
*
* 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 19:58:33 +04:00
*/
#include <math.h>
#include <stdio.h>
#include <string.h>
#include <assert.h>
#include "onyx_int.h"
#include "tokenize.h"
#include "vpx_mem/vpx_mem.h"
/* Global event counters used for accumulating statistics across several
compressions, then generating context.c = initial stats. */
#ifdef ENTROPY_STATS
_int64 context_counters[BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [vp8_coef_tokens];
#endif
void vp8_stuff_mb(VP8_COMP *cpi, MACROBLOCKD *x, TOKENEXTRA **t) ;
void vp8_fix_contexts(MACROBLOCKD *x);
2010-05-18 19:58:33 +04:00
TOKENEXTRA vp8_dct_value_tokens[DCT_MAX_VALUE*2];
Add trellis quantization. Replace the exponential search for optimal rounding during quantization with a linear Viterbi trellis and enable it by default when using --best. Right now this operates on top of the output of the adaptive zero-bin quantizer in vp8_regular_quantize_b() and gives a small gain. It can be tested as a replacement for that quantizer by enabling the call to vp8_strict_quantize_b(), which uses normal rounding and no zero bin offset. Ultimately, the quantizer will have to become a function of lambda in order to take advantage of activity masking, since there is limited ability to change the quantization factor itself. However, currently vp8_strict_quantize_b() plus the trellis quantizer (which is lambda-dependent) loses to vp8_regular_quantize_b() alone (which is not) on my test clip. Patch Set 3: Fix an issue related to the cost evaluation of successor states when a coefficient is reduced to zero. With this issue fixed, now the trellis search almost exactly matches the exponential search. Patch Set 2: Overall, the goal of this patch set is to make "trellis" search to produce encodings that match the exponential search version. There are three main differences between Patch Set 2 and 1: a. Patch set 1 did not properly account for the scale of 2nd order error, so patch set 2 disable it all together for 2nd blocks. b. Patch set 1 was not consistent on when to enable the the quantization optimization. Patch set 2 restore the condition to be consistent. c. Patch set 1 checks quantized level L-1, and L for any input coefficient was quantized to L. Patch set 2 limits the candidate coefficient to those that were rounded up to L. It is worth noting here that a strategy to check L and L+1 for coefficients that were truncated down to L might work. (a and b get trellis quant to basically match the exponential search on all mid/low rate encodings on cif set, without a, b, trellis quant can hurt the psnr by 0.2 to .3db at 200kbps for some cif clips) (c gets trellis quant to match the exponential search to match at Q0 encoding, without c, trellis quant can be 1.5 to 2db lower for encodings with fixed Q at 0 on most derf cif clips) Change-Id: Ib1a043b665d75fbf00cb0257b7c18e90eebab95e
2010-07-03 01:35:53 +04:00
const TOKENEXTRA *vp8_dct_value_tokens_ptr;
2010-05-18 19:58:33 +04:00
int vp8_dct_value_cost[DCT_MAX_VALUE*2];
Add trellis quantization. Replace the exponential search for optimal rounding during quantization with a linear Viterbi trellis and enable it by default when using --best. Right now this operates on top of the output of the adaptive zero-bin quantizer in vp8_regular_quantize_b() and gives a small gain. It can be tested as a replacement for that quantizer by enabling the call to vp8_strict_quantize_b(), which uses normal rounding and no zero bin offset. Ultimately, the quantizer will have to become a function of lambda in order to take advantage of activity masking, since there is limited ability to change the quantization factor itself. However, currently vp8_strict_quantize_b() plus the trellis quantizer (which is lambda-dependent) loses to vp8_regular_quantize_b() alone (which is not) on my test clip. Patch Set 3: Fix an issue related to the cost evaluation of successor states when a coefficient is reduced to zero. With this issue fixed, now the trellis search almost exactly matches the exponential search. Patch Set 2: Overall, the goal of this patch set is to make "trellis" search to produce encodings that match the exponential search version. There are three main differences between Patch Set 2 and 1: a. Patch set 1 did not properly account for the scale of 2nd order error, so patch set 2 disable it all together for 2nd blocks. b. Patch set 1 was not consistent on when to enable the the quantization optimization. Patch set 2 restore the condition to be consistent. c. Patch set 1 checks quantized level L-1, and L for any input coefficient was quantized to L. Patch set 2 limits the candidate coefficient to those that were rounded up to L. It is worth noting here that a strategy to check L and L+1 for coefficients that were truncated down to L might work. (a and b get trellis quant to basically match the exponential search on all mid/low rate encodings on cif set, without a, b, trellis quant can hurt the psnr by 0.2 to .3db at 200kbps for some cif clips) (c gets trellis quant to match the exponential search to match at Q0 encoding, without c, trellis quant can be 1.5 to 2db lower for encodings with fixed Q at 0 on most derf cif clips) Change-Id: Ib1a043b665d75fbf00cb0257b7c18e90eebab95e
2010-07-03 01:35:53 +04:00
const int *vp8_dct_value_cost_ptr;
2010-05-18 19:58:33 +04:00
#if 0
int skip_true_count = 0;
int skip_false_count = 0;
#endif
static void fill_value_tokens()
{
TOKENEXTRA *const t = vp8_dct_value_tokens + DCT_MAX_VALUE;
vp8_extra_bit_struct *const e = vp8_extra_bits;
int i = -DCT_MAX_VALUE;
int sign = 1;
do
{
if (!i)
sign = 0;
{
const int a = sign ? -i : i;
int eb = sign;
if (a > 4)
{
int j = 4;
while (++j < 11 && e[j].base_val <= a) {}
t[i].Token = --j;
eb |= (a - e[j].base_val) << 1;
}
else
t[i].Token = a;
t[i].Extra = eb;
}
// initialize the cost for extra bits for all possible coefficient value.
{
int cost = 0;
vp8_extra_bit_struct *p = vp8_extra_bits + t[i].Token;
if (p->base_val)
{
const int extra = t[i].Extra;
const int Length = p->Len;
if (Length)
cost += vp8_treed_cost(p->tree, p->prob, extra >> 1, Length);
cost += vp8_cost_bit(vp8_prob_half, extra & 1); /* sign */
vp8_dct_value_cost[i + DCT_MAX_VALUE] = cost;
}
}
}
while (++i < DCT_MAX_VALUE);
vp8_dct_value_tokens_ptr = vp8_dct_value_tokens + DCT_MAX_VALUE;
vp8_dct_value_cost_ptr = vp8_dct_value_cost + DCT_MAX_VALUE;
}
static void tokenize2nd_order_b
(
const BLOCKD *const b,
TOKENEXTRA **tp,
const int type, /* which plane: 0=Y no DC, 1=Y2, 2=UV, 3=Y with DC */
const FRAME_TYPE frametype,
ENTROPY_CONTEXT *a,
ENTROPY_CONTEXT *l,
VP8_COMP *cpi
)
{
int pt; /* near block/prev token context index */
int c = 0; /* start at DC */
const int eob = b->eob; /* one beyond last nonzero coeff */
TOKENEXTRA *t = *tp; /* store tokens starting here */
int x;
const short *qcoeff_ptr = b->qcoeff;
VP8_COMBINEENTROPYCONTEXTS(pt, *a, *l);
do
{
const int band = vp8_coef_bands[c];
if (c < eob)
{
int rc = vp8_default_zig_zag1d[c];
const int v = qcoeff_ptr[rc];
assert(-DCT_MAX_VALUE <= v && v < (DCT_MAX_VALUE));
t->Extra = vp8_dct_value_tokens_ptr[v].Extra;
x = vp8_dct_value_tokens_ptr[v].Token;
}
else
x = DCT_EOB_TOKEN;
t->Token = x;
t->context_tree = cpi->common.fc.coef_probs [type] [band] [pt];
t->section = frametype * BLOCK_TYPES * 2 + 2 * type + (c == 0);
t->skip_eob_node = pt == 0 && ((band > 0 && type > 0) || (band > 1 && type == 0));
++cpi->coef_counts [type] [band] [pt] [x];
}
while (pt = vp8_prev_token_class[x], ++t, c < eob && ++c < 16);
*tp = t;
pt = (c != !type); /* 0 <-> all coeff data is zero */
*a = *l = pt;
}
static void tokenize1st_order_b
(
const BLOCKD *const b,
TOKENEXTRA **tp,
const int type, /* which plane: 0=Y no DC, 1=Y2, 2=UV, 3=Y with DC */
const FRAME_TYPE frametype,
ENTROPY_CONTEXT *a,
ENTROPY_CONTEXT *l,
VP8_COMP *cpi
)
{
int pt; /* near block/prev token context index */
int c = type ? 0 : 1; /* start at DC unless type 0 */
const int eob = b->eob; /* one beyond last nonzero coeff */
TOKENEXTRA *t = *tp; /* store tokens starting here */
int x;
const short *qcoeff_ptr = b->qcoeff;
VP8_COMBINEENTROPYCONTEXTS(pt, *a, *l);
do
{
const int band = vp8_coef_bands[c];
x = DCT_EOB_TOKEN;
if (c < eob)
{
int rc = vp8_default_zig_zag1d[c];
const int v = qcoeff_ptr[rc];
assert(-DCT_MAX_VALUE <= v && v < (DCT_MAX_VALUE));
t->Extra = vp8_dct_value_tokens_ptr[v].Extra;
x = vp8_dct_value_tokens_ptr[v].Token;
}
t->Token = x;
t->context_tree = cpi->common.fc.coef_probs [type] [band] [pt];
t->section = frametype * BLOCK_TYPES * 2 + 2 * type + (c == 0);
t->skip_eob_node = pt == 0 && ((band > 0 && type > 0) || (band > 1 && type == 0));
++cpi->coef_counts [type] [band] [pt] [x];
}
while (pt = vp8_prev_token_class[x], ++t, c < eob && ++c < 16);
*tp = t;
pt = (c != !type); /* 0 <-> all coeff data is zero */
*a = *l = pt;
}
void vp8_tokenize_mb(VP8_COMP *cpi, MACROBLOCKD *x, TOKENEXTRA **t)
{
ENTROPY_CONTEXT * A = (ENTROPY_CONTEXT *)x->above_context;
ENTROPY_CONTEXT * L = (ENTROPY_CONTEXT *)x->left_context;
2010-05-18 19:58:33 +04:00
int plane_type;
int b;
TOKENEXTRA *start = *t;
TOKENEXTRA *tp = *t;
x->mode_info_context->mbmi.dc_diff = 1;
2010-05-18 19:58:33 +04:00
#if 0
if (x->mbmi.force_no_skip)
{
x->mbmi.mb_skip_coeff = 1;
//reset for next_mb.
x->mbmi.force_no_skip = 0;
}
#endif
#if 1
if (x->mode_info_context->mbmi.mb_skip_coeff)
2010-05-18 19:58:33 +04:00
{
cpi->skip_true_count++;
if (!cpi->common.mb_no_coeff_skip)
vp8_stuff_mb(cpi, x, t) ;
else
{
vp8_fix_contexts(x);
2010-05-18 19:58:33 +04:00
}
if (x->mode_info_context->mbmi.mode != B_PRED && x->mode_info_context->mbmi.mode != SPLITMV)
x->mode_info_context->mbmi.dc_diff = 0;
2010-05-18 19:58:33 +04:00
else
x->mode_info_context->mbmi.dc_diff = 1;
2010-05-18 19:58:33 +04:00
return;
}
cpi->skip_false_count++;
#endif
#if 0
if (x->mbmi.mode == B_PRED || x->mbmi.mode == SPLITMV)
{
int i, skip = 1;
for (i = 0; i < 24; i++)
skip &= (!x->block[i].eob);
if (skip != x->mbmi.mb_skip_coeff)
skip += 0;
x->mbmi.mb_skip_coeff = skip;
}
else
{
int i, skip = 1;
for (i = 0; i < 16; i++)
skip &= (x->block[i].eob < 2);
for (i = 16; i < 25; i++)
skip &= (!x->block[i].eob);
if (skip != x->mbmi.mb_skip_coeff)
skip += 0;
x->mbmi.mb_skip_coeff = skip;
}
vpx_memcpy(cpi->coef_counts_backup, cpi->coef_counts, sizeof(cpi->coef_counts));
#endif
if (x->mode_info_context->mbmi.mode == B_PRED || x->mode_info_context->mbmi.mode == SPLITMV)
2010-05-18 19:58:33 +04:00
{
plane_type = 3;
}
else
{
tokenize2nd_order_b(x->block + 24, t, 1, x->frame_type,
A + vp8_block2above[24], L + vp8_block2left[24], cpi);
2010-05-18 19:58:33 +04:00
plane_type = 0;
}
for (b = 0; b < 16; b++)
tokenize1st_order_b(x->block + b, t, plane_type, x->frame_type,
A + vp8_block2above[b],
L + vp8_block2left[b], cpi);
2010-05-18 19:58:33 +04:00
for (b = 16; b < 24; b++)
tokenize1st_order_b(x->block + b, t, 2, x->frame_type,
A + vp8_block2above[b],
L + vp8_block2left[b], cpi);
2010-05-18 19:58:33 +04:00
#if 0
if (cpi->common.mb_no_coeff_skip)
{
int skip = 1;
while ((tp != *t) && skip)
{
skip = (skip && (tp->Token == DCT_EOB_TOKEN));
tp ++;
}
if (skip != x->mbmi.mb_skip_coeff)
skip += 0;
x->mbmi.mb_skip_coeff = skip;
if (x->mbmi.mb_skip_coeff == 1)
{
x->mbmi.dc_diff = 0;
//redo the coutnts
vpx_memcpy(cpi->coef_counts, cpi->coef_counts_backup, sizeof(cpi->coef_counts));
*t = start;
cpi->skip_true_count++;
//skip_true_count++;
}
else
{
cpi->skip_false_count++;
//skip_false_count++;
}
}
#endif
}
2010-05-18 19:58:33 +04:00
#ifdef ENTROPY_STATS
void init_context_counters(void)
{
vpx_memset(context_counters, 0, sizeof(context_counters));
}
void print_context_counters()
{
int type, band, pt, t;
FILE *const f = fopen("context.c", "w");
fprintf(f, "#include \"entropy.h\"\n");
fprintf(f, "\n/* *** GENERATED FILE: DO NOT EDIT *** */\n\n");
fprintf(f, "int Contexts[BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [vp8_coef_tokens];\n\n");
fprintf(f, "const int default_contexts[BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [vp8_coef_tokens] = {");
# define Comma( X) (X? ",":"")
type = 0;
do
{
fprintf(f, "%s\n { /* block Type %d */", Comma(type), type);
band = 0;
do
{
fprintf(f, "%s\n { /* Coeff Band %d */", Comma(band), band);
pt = 0;
do
{
fprintf(f, "%s\n {", Comma(pt));
t = 0;
do
{
const _int64 x = context_counters [type] [band] [pt] [t];
const int y = (int) x;
assert(x == (_int64) y); /* no overflow handling yet */
fprintf(f, "%s %d", Comma(t), y);
}
while (++t < vp8_coef_tokens);
fprintf(f, "}");
}
while (++pt < PREV_COEF_CONTEXTS);
fprintf(f, "\n }");
}
while (++band < COEF_BANDS);
fprintf(f, "\n }");
}
while (++type < BLOCK_TYPES);
fprintf(f, "\n};\n");
fclose(f);
}
#endif
void vp8_tokenize_initialize()
{
fill_value_tokens();
}
static __inline void stuff2nd_order_b
(
const BLOCKD *const b,
TOKENEXTRA **tp,
const int type, /* which plane: 0=Y no DC, 1=Y2, 2=UV, 3=Y with DC */
const FRAME_TYPE frametype,
ENTROPY_CONTEXT *a,
ENTROPY_CONTEXT *l,
VP8_COMP *cpi
)
{
int pt; /* near block/prev token context index */
TOKENEXTRA *t = *tp; /* store tokens starting here */
VP8_COMBINEENTROPYCONTEXTS(pt, *a, *l);
(void) frametype;
(void) type;
(void) b;
t->Token = DCT_EOB_TOKEN;
t->context_tree = cpi->common.fc.coef_probs [1] [0] [pt];
t->section = 11;
t->skip_eob_node = 0;
++cpi->coef_counts [1] [0] [pt] [DCT_EOB_TOKEN];
++t;
*tp = t;
pt = 0;
*a = *l = pt;
}
static __inline void stuff1st_order_b
(
const BLOCKD *const b,
TOKENEXTRA **tp,
const int type, /* which plane: 0=Y no DC, 1=Y2, 2=UV, 3=Y with DC */
const FRAME_TYPE frametype,
ENTROPY_CONTEXT *a,
ENTROPY_CONTEXT *l,
VP8_COMP *cpi
)
{
int pt; /* near block/prev token context index */
TOKENEXTRA *t = *tp; /* store tokens starting here */
VP8_COMBINEENTROPYCONTEXTS(pt, *a, *l);
(void) frametype;
(void) type;
(void) b;
t->Token = DCT_EOB_TOKEN;
t->context_tree = cpi->common.fc.coef_probs [0] [1] [pt];
t->section = 8;
t->skip_eob_node = 0;
++cpi->coef_counts [0] [1] [pt] [DCT_EOB_TOKEN];
++t;
*tp = t;
pt = 0; /* 0 <-> all coeff data is zero */
*a = *l = pt;
}
static __inline
void stuff1st_order_buv
(
const BLOCKD *const b,
TOKENEXTRA **tp,
const int type, /* which plane: 0=Y no DC, 1=Y2, 2=UV, 3=Y with DC */
const FRAME_TYPE frametype,
ENTROPY_CONTEXT *a,
ENTROPY_CONTEXT *l,
VP8_COMP *cpi
)
{
int pt; /* near block/prev token context index */
TOKENEXTRA *t = *tp; /* store tokens starting here */
VP8_COMBINEENTROPYCONTEXTS(pt, *a, *l);
(void) frametype;
(void) type;
(void) b;
t->Token = DCT_EOB_TOKEN;
t->context_tree = cpi->common.fc.coef_probs [2] [0] [pt];
t->section = 13;
t->skip_eob_node = 0;
++cpi->coef_counts[2] [0] [pt] [DCT_EOB_TOKEN];
++t;
*tp = t;
pt = 0; /* 0 <-> all coeff data is zero */
*a = *l = pt;
}
void vp8_stuff_mb(VP8_COMP *cpi, MACROBLOCKD *x, TOKENEXTRA **t)
{
ENTROPY_CONTEXT * A = (ENTROPY_CONTEXT *)x->above_context;
ENTROPY_CONTEXT * L = (ENTROPY_CONTEXT *)x->left_context;
2010-05-18 19:58:33 +04:00
int plane_type;
int b;
stuff2nd_order_b(x->block + 24, t, 1, x->frame_type,
A + vp8_block2above[24], L + vp8_block2left[24], cpi);
2010-05-18 19:58:33 +04:00
plane_type = 0;
if (x->mode_info_context->mbmi.mode != B_PRED && x->mode_info_context->mbmi.mode != SPLITMV)
x->mode_info_context->mbmi.dc_diff = 0;
2010-05-18 19:58:33 +04:00
else
x->mode_info_context->mbmi.dc_diff = 1;
2010-05-18 19:58:33 +04:00
for (b = 0; b < 16; b++)
stuff1st_order_b(x->block + b, t, plane_type, x->frame_type,
A + vp8_block2above[b],
L + vp8_block2left[b], cpi);
2010-05-18 19:58:33 +04:00
for (b = 16; b < 24; b++)
stuff1st_order_buv(x->block + b, t, 2, x->frame_type,
A + vp8_block2above[b],
L + vp8_block2left[b], cpi);
2010-05-18 19:58:33 +04:00
}
void vp8_fix_contexts(MACROBLOCKD *x)
2010-05-18 19:58:33 +04:00
{
/* Clear entropy contexts for Y2 blocks */
if (x->mode_info_context->mbmi.mode != B_PRED && x->mode_info_context->mbmi.mode != SPLITMV)
2010-05-18 19:58:33 +04:00
{
vpx_memset(x->above_context, 0, sizeof(ENTROPY_CONTEXT_PLANES));
vpx_memset(x->left_context, 0, sizeof(ENTROPY_CONTEXT_PLANES));
2010-05-18 19:58:33 +04:00
}
else
{
vpx_memset(x->above_context, 0, sizeof(ENTROPY_CONTEXT_PLANES)-1);
vpx_memset(x->left_context, 0, sizeof(ENTROPY_CONTEXT_PLANES)-1);
}
2010-05-18 19:58:33 +04:00
}