2010-05-18 17:58:33 +02:00
|
|
|
/*
|
2010-09-09 14:16:39 +02:00
|
|
|
* Copyright (c) 2010 The WebM project authors. All Rights Reserved.
|
2010-05-18 17:58:33 +02:00
|
|
|
*
|
2010-06-18 18:39:21 +02:00
|
|
|
* Use of this source code is governed by a BSD-style license
|
2010-06-04 22:19:40 +02:00
|
|
|
* that can be found in the LICENSE file in the root of the source
|
|
|
|
* tree. An additional intellectual property rights grant can be found
|
2010-06-18 18:39:21 +02:00
|
|
|
* in the file PATENTS. All contributing project authors may
|
2010-06-04 22:19:40 +02:00
|
|
|
* 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>
|
2012-11-28 19:41:40 +01:00
|
|
|
#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"
|
|
|
|
|
2012-11-27 22:59:17 +01:00
|
|
|
#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
|
|
|
|
2010-05-18 17:58:33 +02:00
|
|
|
/* Global event counters used for accumulating statistics across several
|
2012-11-27 22:59:17 +01:00
|
|
|
compressions, then generating vp9_context.c = initial stats. */
|
2010-05-18 17:58:33 +02:00
|
|
|
|
|
|
|
#ifdef ENTROPY_STATS
|
2012-12-08 01:09:59 +01:00
|
|
|
vp9_coeff_accum context_counters_4x4[BLOCK_TYPES_4X4];
|
|
|
|
vp9_coeff_accum hybrid_context_counters_4x4[BLOCK_TYPES_4X4];
|
|
|
|
vp9_coeff_accum context_counters_8x8[BLOCK_TYPES_8X8];
|
|
|
|
vp9_coeff_accum hybrid_context_counters_8x8[BLOCK_TYPES_8X8];
|
|
|
|
vp9_coeff_accum context_counters_16x16[BLOCK_TYPES_16X16];
|
|
|
|
vp9_coeff_accum hybrid_context_counters_16x16[BLOCK_TYPES_16X16];
|
|
|
|
vp9_coeff_accum context_counters_32x32[BLOCK_TYPES_32X32];
|
|
|
|
|
|
|
|
extern vp9_coeff_stats tree_update_hist_4x4[BLOCK_TYPES_4X4];
|
|
|
|
extern vp9_coeff_stats hybrid_tree_update_hist_4x4[BLOCK_TYPES_4X4];
|
|
|
|
extern vp9_coeff_stats tree_update_hist_8x8[BLOCK_TYPES_8X8];
|
|
|
|
extern vp9_coeff_stats hybrid_tree_update_hist_8x8[BLOCK_TYPES_8X8];
|
|
|
|
extern vp9_coeff_stats tree_update_hist_16x16[BLOCK_TYPES_16X16];
|
|
|
|
extern vp9_coeff_stats hybrid_tree_update_hist_16x16[BLOCK_TYPES_16X16];
|
|
|
|
extern vp9_coeff_stats tree_update_hist_32x32[BLOCK_TYPES_32X32];
|
2012-10-10 20:20:33 +02:00
|
|
|
#endif /* ENTROPY_STATS */
|
|
|
|
|
2012-07-14 00:21:29 +02:00
|
|
|
static TOKENVALUE dct_value_tokens[DCT_MAX_VALUE * 2];
|
2012-10-31 01:12:12 +01:00
|
|
|
const TOKENVALUE *vp9_dct_value_tokens_ptr;
|
2012-07-14 00:21:29 +02:00
|
|
|
static int dct_value_cost[DCT_MAX_VALUE * 2];
|
2012-10-31 01:12:12 +01:00
|
|
|
const int *vp9_dct_value_cost_ptr;
|
2011-02-14 23:18:18 +01:00
|
|
|
|
2012-07-14 00:21:29 +02:00
|
|
|
static void fill_value_tokens() {
|
2010-05-18 17:58:33 +02:00
|
|
|
|
2012-07-14 00:21:29 +02:00
|
|
|
TOKENVALUE *const t = dct_value_tokens + DCT_MAX_VALUE;
|
2012-10-31 22:40:53 +01:00
|
|
|
vp9_extra_bit_struct *const e = vp9_extra_bits;
|
2010-05-18 17:58:33 +02:00
|
|
|
|
2012-07-14 00:21:29 +02:00
|
|
|
int i = -DCT_MAX_VALUE;
|
|
|
|
int sign = 1;
|
2010-05-18 17:58:33 +02:00
|
|
|
|
2012-07-14 00:21:29 +02:00
|
|
|
do {
|
|
|
|
if (!i)
|
|
|
|
sign = 0;
|
2010-05-18 17:58:33 +02:00
|
|
|
|
2012-07-14 00:21:29 +02:00
|
|
|
{
|
|
|
|
const int a = sign ? -i : i;
|
|
|
|
int eb = sign;
|
2010-05-18 17:58:33 +02:00
|
|
|
|
2012-07-14 00:21:29 +02:00
|
|
|
if (a > 4) {
|
|
|
|
int j = 4;
|
2010-05-18 17:58:33 +02:00
|
|
|
|
2012-07-14 00:21:29 +02:00
|
|
|
while (++j < 11 && e[j].base_val <= a) {}
|
2010-05-18 17:58:33 +02:00
|
|
|
|
2012-07-14 00:21:29 +02:00
|
|
|
t[i].Token = --j;
|
|
|
|
eb |= (a - e[j].base_val) << 1;
|
|
|
|
} else
|
|
|
|
t[i].Token = a;
|
2010-05-18 17:58:33 +02:00
|
|
|
|
2012-07-14 00:21:29 +02:00
|
|
|
t[i].Extra = eb;
|
|
|
|
}
|
2010-05-18 17:58:33 +02:00
|
|
|
|
2012-07-14 00:21:29 +02:00
|
|
|
// initialize the cost for extra bits for all possible coefficient value.
|
|
|
|
{
|
|
|
|
int cost = 0;
|
2012-10-31 22:40:53 +01:00
|
|
|
vp9_extra_bit_struct *p = vp9_extra_bits + t[i].Token;
|
2010-05-18 17:58:33 +02:00
|
|
|
|
2012-07-14 00:21:29 +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
|
|
|
|
2012-07-14 00:21:29 +02:00
|
|
|
if (Length)
|
2012-10-31 22:40:53 +01:00
|
|
|
cost += treed_cost(p->tree, p->prob, extra >> 1, Length);
|
2010-05-18 17:58:33 +02:00
|
|
|
|
2012-10-31 22:40:53 +01:00
|
|
|
cost += vp9_cost_bit(vp9_prob_half, extra & 1); /* sign */
|
2012-07-14 00:21:29 +02:00
|
|
|
dct_value_cost[i + DCT_MAX_VALUE] = cost;
|
|
|
|
}
|
2010-05-18 17:58:33 +02:00
|
|
|
|
|
|
|
}
|
|
|
|
|
2012-07-14 00:21:29 +02:00
|
|
|
} while (++i < DCT_MAX_VALUE);
|
|
|
|
|
2012-10-31 01:12:12 +01:00
|
|
|
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
|
|
|
}
|
2012-11-28 00:51:06 +01:00
|
|
|
|
|
|
|
#if CONFIG_NEWCOEFCONTEXT
|
|
|
|
#define PT pn
|
|
|
|
#else
|
|
|
|
#define PT pt
|
|
|
|
#endif
|
|
|
|
|
2012-10-31 01:53:32 +01:00
|
|
|
static void tokenize_b(VP9_COMP *cpi,
|
2012-10-25 18:14:21 +02:00
|
|
|
MACROBLOCKD *xd,
|
2012-12-06 21:40:57 +01:00
|
|
|
const int ib,
|
2012-10-25 18:14:21 +02:00
|
|
|
TOKENEXTRA **tp,
|
|
|
|
PLANE_TYPE type,
|
|
|
|
TX_SIZE tx_size,
|
|
|
|
int dry_run) {
|
2012-08-03 02:03:14 +02:00
|
|
|
int pt; /* near block/prev token context index */
|
2012-10-15 00:29:56 +02:00
|
|
|
int c = (type == PLANE_TYPE_Y_NO_DC) ? 1 : 0;
|
2012-12-06 21:40:57 +01:00
|
|
|
const BLOCKD * const b = xd->block + ib;
|
2013-01-03 18:00:30 +01:00
|
|
|
const int eob = b->eob; /* one beyond last nonzero coeff */
|
2012-08-03 02:03:14 +02:00
|
|
|
TOKENEXTRA *t = *tp; /* store tokens starting here */
|
2013-01-03 18:00:30 +01:00
|
|
|
int16_t *qcoeff_ptr = b->qcoeff;
|
2012-10-25 18:14:21 +02:00
|
|
|
int seg_eob;
|
2012-12-06 21:40:57 +01:00
|
|
|
const int segment_id = xd->mode_info_context->mbmi.segment_id;
|
2012-10-25 18:14:21 +02:00
|
|
|
const int *bands, *scan;
|
2012-12-08 01:09:59 +01:00
|
|
|
vp9_coeff_count *counts;
|
|
|
|
vp9_coeff_probs *probs;
|
2012-10-25 18:14:21 +02:00
|
|
|
const TX_TYPE tx_type = (type == PLANE_TYPE_Y_WITH_DC) ?
|
|
|
|
get_tx_type(xd, b) : DCT_DCT;
|
2012-11-28 00:51:06 +01:00
|
|
|
#if CONFIG_NEWCOEFCONTEXT
|
|
|
|
const int *neighbors;
|
|
|
|
int pn;
|
|
|
|
#endif
|
2012-08-03 02:03:14 +02:00
|
|
|
|
2012-12-06 21:40:57 +01:00
|
|
|
ENTROPY_CONTEXT *const a = (ENTROPY_CONTEXT *)xd->above_context +
|
|
|
|
vp9_block2above[tx_size][ib];
|
|
|
|
ENTROPY_CONTEXT *const l = (ENTROPY_CONTEXT *)xd->left_context +
|
|
|
|
vp9_block2left[tx_size][ib];
|
|
|
|
ENTROPY_CONTEXT a_ec = *a, l_ec = *l;
|
|
|
|
|
2012-12-18 20:25:24 +01:00
|
|
|
ENTROPY_CONTEXT *const a1 = (ENTROPY_CONTEXT *)(&xd->above_context[1]) +
|
|
|
|
vp9_block2above[tx_size][ib];
|
|
|
|
ENTROPY_CONTEXT *const l1 = (ENTROPY_CONTEXT *)(&xd->left_context[1]) +
|
|
|
|
vp9_block2left[tx_size][ib];
|
|
|
|
|
|
|
|
|
2012-10-25 18:14:21 +02:00
|
|
|
switch (tx_size) {
|
|
|
|
default:
|
|
|
|
case TX_4X4:
|
|
|
|
seg_eob = 16;
|
2012-12-18 19:49:10 +01:00
|
|
|
bands = vp9_coef_bands_4x4;
|
|
|
|
scan = vp9_default_zig_zag1d_4x4;
|
2012-10-25 18:14:21 +02:00
|
|
|
if (tx_type != DCT_DCT) {
|
2012-12-08 01:09:59 +01:00
|
|
|
counts = cpi->hybrid_coef_counts_4x4;
|
|
|
|
probs = cpi->common.fc.hybrid_coef_probs_4x4;
|
2012-10-25 18:14:21 +02:00
|
|
|
if (tx_type == ADST_DCT) {
|
2012-12-18 19:49:10 +01:00
|
|
|
scan = vp9_row_scan_4x4;
|
2012-10-25 18:14:21 +02:00
|
|
|
} else if (tx_type == DCT_ADST) {
|
2012-12-18 19:49:10 +01:00
|
|
|
scan = vp9_col_scan_4x4;
|
2012-10-25 18:14:21 +02:00
|
|
|
}
|
|
|
|
} else {
|
2012-12-08 01:09:59 +01:00
|
|
|
counts = cpi->coef_counts_4x4;
|
|
|
|
probs = cpi->common.fc.coef_probs_4x4;
|
2012-10-25 18:14:21 +02:00
|
|
|
}
|
|
|
|
break;
|
|
|
|
case TX_8X8:
|
|
|
|
if (type == PLANE_TYPE_Y2) {
|
|
|
|
seg_eob = 4;
|
2012-12-18 19:49:10 +01:00
|
|
|
bands = vp9_coef_bands_4x4;
|
|
|
|
scan = vp9_default_zig_zag1d_4x4;
|
2012-10-25 18:14:21 +02:00
|
|
|
} else {
|
2012-12-06 21:40:57 +01:00
|
|
|
#if CONFIG_CNVCONTEXT
|
|
|
|
a_ec = (a[0] + a[1]) != 0;
|
|
|
|
l_ec = (l[0] + l[1]) != 0;
|
|
|
|
#endif
|
2012-10-25 18:14:21 +02:00
|
|
|
seg_eob = 64;
|
2012-10-31 01:12:12 +01:00
|
|
|
bands = vp9_coef_bands_8x8;
|
|
|
|
scan = vp9_default_zig_zag1d_8x8;
|
2012-10-25 18:14:21 +02:00
|
|
|
}
|
|
|
|
if (tx_type != DCT_DCT) {
|
|
|
|
counts = cpi->hybrid_coef_counts_8x8;
|
|
|
|
probs = cpi->common.fc.hybrid_coef_probs_8x8;
|
|
|
|
} else {
|
|
|
|
counts = cpi->coef_counts_8x8;
|
|
|
|
probs = cpi->common.fc.coef_probs_8x8;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case TX_16X16:
|
2012-12-06 21:40:57 +01:00
|
|
|
#if CONFIG_CNVCONTEXT
|
2012-12-13 23:58:30 +01:00
|
|
|
if (type != PLANE_TYPE_UV) {
|
|
|
|
a_ec = (a[0] + a[1] + a[2] + a[3]) != 0;
|
|
|
|
l_ec = (l[0] + l[1] + l[2] + l[3]) != 0;
|
2012-12-18 20:25:24 +01:00
|
|
|
} else {
|
|
|
|
a_ec = (a[0] + a[1] + a1[0] + a1[1]) != 0;
|
|
|
|
l_ec = (l[0] + l[1] + l1[0] + l1[1]) != 0;
|
2012-12-13 23:58:30 +01:00
|
|
|
}
|
2012-12-06 21:40:57 +01:00
|
|
|
#endif
|
2012-10-25 18:14:21 +02:00
|
|
|
seg_eob = 256;
|
2012-10-31 01:12:12 +01:00
|
|
|
bands = vp9_coef_bands_16x16;
|
|
|
|
scan = vp9_default_zig_zag1d_16x16;
|
2012-10-25 18:14:21 +02:00
|
|
|
if (tx_type != DCT_DCT) {
|
|
|
|
counts = cpi->hybrid_coef_counts_16x16;
|
|
|
|
probs = cpi->common.fc.hybrid_coef_probs_16x16;
|
|
|
|
} else {
|
|
|
|
counts = cpi->coef_counts_16x16;
|
|
|
|
probs = 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
|
|
|
if (type == PLANE_TYPE_UV) {
|
2013-01-03 18:00:30 +01:00
|
|
|
int uv_idx = (ib - 16) >> 2;
|
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
|
|
|
qcoeff_ptr = xd->sb_coeff_data.qcoeff + 1024 + 256 * uv_idx;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case TX_32X32:
|
2012-12-18 20:25:24 +01:00
|
|
|
#if CONFIG_CNVCONTEXT
|
|
|
|
a_ec = a[0] + a[1] + a[2] + a[3] +
|
|
|
|
a1[0] + a1[1] + a1[2] + a1[3];
|
|
|
|
l_ec = l[0] + l[1] + l[2] + l[3] +
|
|
|
|
l1[0] + l1[1] + l1[2] + l1[3];
|
|
|
|
a_ec = a_ec != 0;
|
|
|
|
l_ec = l_ec != 0;
|
|
|
|
#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
|
|
|
seg_eob = 1024;
|
|
|
|
bands = vp9_coef_bands_32x32;
|
|
|
|
scan = vp9_default_zig_zag1d_32x32;
|
|
|
|
counts = cpi->coef_counts_32x32;
|
|
|
|
probs = cpi->common.fc.coef_probs_32x32;
|
|
|
|
qcoeff_ptr = xd->sb_coeff_data.qcoeff;
|
2012-10-25 18:14:21 +02:00
|
|
|
break;
|
2012-07-14 00:21:29 +02:00
|
|
|
}
|
|
|
|
|
2012-12-06 21:40:57 +01:00
|
|
|
VP9_COMBINEENTROPYCONTEXTS(pt, a_ec, l_ec);
|
2012-11-28 00:51:06 +01:00
|
|
|
#if CONFIG_NEWCOEFCONTEXT
|
|
|
|
neighbors = vp9_get_coef_neighbors_handle(scan);
|
|
|
|
pn = pt;
|
|
|
|
#endif
|
2012-12-06 21:40:57 +01:00
|
|
|
|
2013-01-28 16:22:53 +01:00
|
|
|
if (vp9_segfeature_active(xd, segment_id, SEG_LVL_SKIP))
|
|
|
|
seg_eob = 0;
|
2012-07-14 00:21:29 +02:00
|
|
|
|
2012-10-13 20:46:21 +02:00
|
|
|
do {
|
2012-10-25 18:14:21 +02:00
|
|
|
const int band = bands[c];
|
2012-10-13 20:46:21 +02:00
|
|
|
int token;
|
|
|
|
|
|
|
|
if (c < eob) {
|
2012-10-25 18:14:21 +02:00
|
|
|
const int rc = scan[c];
|
2012-10-13 20:46:21 +02:00
|
|
|
const int v = qcoeff_ptr[rc];
|
2012-10-25 18:14:21 +02:00
|
|
|
assert(-DCT_MAX_VALUE <= v && v < DCT_MAX_VALUE);
|
|
|
|
|
2012-10-31 01:12:12 +01:00
|
|
|
t->Extra = vp9_dct_value_tokens_ptr[v].Extra;
|
|
|
|
token = vp9_dct_value_tokens_ptr[v].Token;
|
2012-10-25 18:14:21 +02:00
|
|
|
} else {
|
|
|
|
token = DCT_EOB_TOKEN;
|
|
|
|
}
|
2012-07-14 00:21:29 +02:00
|
|
|
|
|
|
|
t->Token = token;
|
2012-11-28 00:51:06 +01:00
|
|
|
t->context_tree = probs[type][band][PT];
|
2012-10-25 18:14:21 +02:00
|
|
|
t->skip_eob_node = (pt == 0) && ((band > 0 && type != PLANE_TYPE_Y_NO_DC) ||
|
|
|
|
(band > 1 && type == PLANE_TYPE_Y_NO_DC));
|
2012-10-31 01:12:12 +01:00
|
|
|
assert(vp9_coef_encodings[t->Token].Len - t->skip_eob_node > 0);
|
2012-09-10 07:42:35 +02:00
|
|
|
if (!dry_run) {
|
2012-11-28 00:51:06 +01:00
|
|
|
++counts[type][band][PT][token];
|
2012-09-10 07:42:35 +02:00
|
|
|
}
|
2012-10-31 01:12:12 +01:00
|
|
|
pt = vp9_prev_token_class[token];
|
2012-11-28 00:51:06 +01:00
|
|
|
#if CONFIG_NEWCOEFCONTEXT
|
|
|
|
if (c < seg_eob - 1 && NEWCOEFCONTEXT_BAND_COND(bands[c + 1]))
|
|
|
|
pn = vp9_get_coef_neighbor_context(
|
|
|
|
qcoeff_ptr, (type == PLANE_TYPE_Y_NO_DC), neighbors, scan[c + 1]);
|
|
|
|
else
|
|
|
|
pn = pt;
|
|
|
|
#endif
|
2012-07-23 23:58:07 +02:00
|
|
|
++t;
|
2012-10-13 20:46:21 +02:00
|
|
|
} while (c < eob && ++c < seg_eob);
|
|
|
|
|
|
|
|
*tp = t;
|
2012-12-06 21:40:57 +01:00
|
|
|
a_ec = l_ec = (c > !type); /* 0 <-> all coeff data is zero */
|
|
|
|
a[0] = a_ec;
|
|
|
|
l[0] = l_ec;
|
|
|
|
|
|
|
|
if (tx_size == TX_8X8 && type != PLANE_TYPE_Y2) {
|
|
|
|
a[1] = a_ec;
|
|
|
|
l[1] = l_ec;
|
|
|
|
} else if (tx_size == TX_16X16) {
|
2012-12-13 23:58:30 +01:00
|
|
|
if (type != PLANE_TYPE_UV) {
|
|
|
|
a[1] = a[2] = a[3] = a_ec;
|
|
|
|
l[1] = l[2] = l[3] = l_ec;
|
|
|
|
} else {
|
2012-12-18 20:25:24 +01:00
|
|
|
a1[0] = a1[1] = a[1] = a_ec;
|
|
|
|
l1[0] = l1[1] = l[1] = l_ec;
|
2012-12-13 23:58:30 +01:00
|
|
|
}
|
2012-12-18 20:25:24 +01:00
|
|
|
} else if (tx_size == TX_32X32) {
|
|
|
|
a[1] = a[2] = a[3] = a_ec;
|
|
|
|
l[1] = l[2] = l[3] = l_ec;
|
|
|
|
a1[0] = a1[1] = a1[2] = a1[3] = a_ec;
|
|
|
|
l1[0] = l1[1] = l1[2] = l1[3] = l_ec;
|
2012-12-06 21:40:57 +01:00
|
|
|
}
|
2010-05-18 17:58:33 +02:00
|
|
|
}
|
|
|
|
|
2012-11-16 00:14:38 +01:00
|
|
|
int vp9_mby_is_skippable_4x4(MACROBLOCKD *xd, int has_2nd_order) {
|
2012-07-14 00:21:29 +02:00
|
|
|
int skip = 1;
|
|
|
|
int i = 0;
|
2010-10-07 07:39:16 +02:00
|
|
|
|
2012-11-16 00:14:38 +01:00
|
|
|
if (has_2nd_order) {
|
2012-07-14 00:21:29 +02:00
|
|
|
for (i = 0; i < 16; i++)
|
2012-08-15 12:00:53 +02:00
|
|
|
skip &= (xd->block[i].eob < 2);
|
|
|
|
skip &= (!xd->block[24].eob);
|
2012-07-14 00:21:29 +02:00
|
|
|
} else {
|
|
|
|
for (i = 0; i < 16; i++)
|
2012-08-15 12:00:53 +02:00
|
|
|
skip &= (!xd->block[i].eob);
|
2012-07-14 00:21:29 +02:00
|
|
|
}
|
|
|
|
return skip;
|
2012-02-22 03:10:18 +01:00
|
|
|
}
|
2010-10-07 07:39:16 +02:00
|
|
|
|
2012-10-30 05:07:40 +01:00
|
|
|
int vp9_mbuv_is_skippable_4x4(MACROBLOCKD *xd) {
|
2012-07-14 00:21:29 +02:00
|
|
|
int skip = 1;
|
|
|
|
int i;
|
2010-10-07 07:39:16 +02:00
|
|
|
|
2012-07-14 00:21:29 +02:00
|
|
|
for (i = 16; i < 24; i++)
|
2012-08-15 12:00:53 +02:00
|
|
|
skip &= (!xd->block[i].eob);
|
2012-07-14 00:21:29 +02:00
|
|
|
return skip;
|
2010-10-07 07:39:16 +02:00
|
|
|
}
|
|
|
|
|
2012-11-16 00:14:38 +01:00
|
|
|
static int mb_is_skippable_4x4(MACROBLOCKD *xd, int has_2nd_order) {
|
|
|
|
return (vp9_mby_is_skippable_4x4(xd, has_2nd_order) &
|
2012-10-30 05:07:40 +01:00
|
|
|
vp9_mbuv_is_skippable_4x4(xd));
|
2012-02-22 03:10:18 +01:00
|
|
|
}
|
2012-02-29 02:11:12 +01:00
|
|
|
|
2012-11-16 00:14:38 +01:00
|
|
|
int vp9_mby_is_skippable_8x8(MACROBLOCKD *xd, int has_2nd_order) {
|
2012-07-14 00:21:29 +02:00
|
|
|
int skip = 1;
|
|
|
|
int i = 0;
|
2011-11-18 03:31:00 +01:00
|
|
|
|
2012-11-16 00:14:38 +01:00
|
|
|
if (has_2nd_order) {
|
2012-09-12 04:36:28 +02:00
|
|
|
for (i = 0; i < 16; i += 4)
|
|
|
|
skip &= (xd->block[i].eob < 2);
|
|
|
|
skip &= (!xd->block[24].eob);
|
|
|
|
} else {
|
|
|
|
for (i = 0; i < 16; i += 4)
|
|
|
|
skip &= (!xd->block[i].eob);
|
|
|
|
}
|
2012-07-14 00:21:29 +02:00
|
|
|
return skip;
|
2012-02-22 03:10:18 +01:00
|
|
|
}
|
2011-02-14 23:18:18 +01:00
|
|
|
|
2012-10-30 05:07:40 +01:00
|
|
|
int vp9_mbuv_is_skippable_8x8(MACROBLOCKD *xd) {
|
2012-08-15 12:00:53 +02:00
|
|
|
return (!xd->block[16].eob) & (!xd->block[20].eob);
|
2012-02-22 03:10:18 +01:00
|
|
|
}
|
2011-02-14 23:18:18 +01:00
|
|
|
|
2012-11-16 00:14:38 +01:00
|
|
|
static int mb_is_skippable_8x8(MACROBLOCKD *xd, int has_2nd_order) {
|
|
|
|
return (vp9_mby_is_skippable_8x8(xd, has_2nd_order) &
|
2012-10-30 05:07:40 +01:00
|
|
|
vp9_mbuv_is_skippable_8x8(xd));
|
2012-09-12 04:36:28 +02:00
|
|
|
}
|
|
|
|
|
2012-11-16 00:14:38 +01:00
|
|
|
static int mb_is_skippable_8x8_4x4uv(MACROBLOCKD *xd, int has_2nd_order) {
|
|
|
|
return (vp9_mby_is_skippable_8x8(xd, has_2nd_order) &
|
2012-10-30 05:07:40 +01:00
|
|
|
vp9_mbuv_is_skippable_4x4(xd));
|
2011-02-14 23:18:18 +01:00
|
|
|
}
|
2012-02-29 02:11:12 +01:00
|
|
|
|
2012-10-30 05:07:40 +01:00
|
|
|
int vp9_mby_is_skippable_16x16(MACROBLOCKD *xd) {
|
2012-08-03 02:03:14 +02:00
|
|
|
int skip = 1;
|
2012-08-15 12:00:53 +02:00
|
|
|
skip &= !xd->block[0].eob;
|
2012-08-03 02:03:14 +02:00
|
|
|
return skip;
|
|
|
|
}
|
|
|
|
|
2012-10-30 05:07:40 +01:00
|
|
|
static int mb_is_skippable_16x16(MACROBLOCKD *xd) {
|
|
|
|
return (vp9_mby_is_skippable_16x16(xd) & vp9_mbuv_is_skippable_8x8(xd));
|
2012-08-03 02:03:14 +02:00
|
|
|
}
|
2010-10-07 07:39:16 +02:00
|
|
|
|
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
|
|
|
int vp9_sby_is_skippable_32x32(MACROBLOCKD *xd) {
|
|
|
|
int skip = 1;
|
|
|
|
skip &= !xd->block[0].eob;
|
|
|
|
return skip;
|
|
|
|
}
|
|
|
|
|
|
|
|
int vp9_sbuv_is_skippable_16x16(MACROBLOCKD *xd) {
|
|
|
|
return (!xd->block[16].eob) & (!xd->block[20].eob);
|
|
|
|
}
|
|
|
|
|
|
|
|
static int sb_is_skippable_32x32(MACROBLOCKD *xd) {
|
|
|
|
return vp9_sby_is_skippable_32x32(xd) &&
|
|
|
|
vp9_sbuv_is_skippable_16x16(xd);
|
|
|
|
}
|
|
|
|
|
|
|
|
void vp9_tokenize_sb(VP9_COMP *cpi,
|
|
|
|
MACROBLOCKD *xd,
|
|
|
|
TOKENEXTRA **t,
|
|
|
|
int dry_run) {
|
|
|
|
VP9_COMMON * const cm = &cpi->common;
|
|
|
|
MB_MODE_INFO * const mbmi = &xd->mode_info_context->mbmi;
|
|
|
|
TOKENEXTRA *t_backup = *t;
|
|
|
|
ENTROPY_CONTEXT *A[2] = { (ENTROPY_CONTEXT *) (xd->above_context + 0),
|
|
|
|
(ENTROPY_CONTEXT *) (xd->above_context + 1), };
|
|
|
|
ENTROPY_CONTEXT *L[2] = { (ENTROPY_CONTEXT *) (xd->left_context + 0),
|
|
|
|
(ENTROPY_CONTEXT *) (xd->left_context + 1), };
|
|
|
|
const int mb_skip_context = vp9_get_pred_context(cm, xd, PRED_MBSKIP);
|
|
|
|
const int segment_id = mbmi->segment_id;
|
2013-01-28 16:22:53 +01:00
|
|
|
const int skip_inc = !vp9_segfeature_active(xd, segment_id, SEG_LVL_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
|
|
|
int b;
|
|
|
|
|
|
|
|
mbmi->mb_skip_coeff = sb_is_skippable_32x32(xd);
|
|
|
|
|
|
|
|
if (mbmi->mb_skip_coeff) {
|
|
|
|
if (!dry_run)
|
|
|
|
cpi->skip_true_count[mb_skip_context] += skip_inc;
|
|
|
|
if (!cm->mb_no_coeff_skip) {
|
|
|
|
vp9_stuff_sb(cpi, xd, t, dry_run);
|
|
|
|
} else {
|
|
|
|
vp9_fix_contexts_sb(xd);
|
|
|
|
}
|
|
|
|
if (dry_run)
|
|
|
|
*t = t_backup;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!dry_run)
|
|
|
|
cpi->skip_false_count[mb_skip_context] += skip_inc;
|
|
|
|
|
2012-12-13 21:18:38 +01:00
|
|
|
tokenize_b(cpi, xd, 0, t, PLANE_TYPE_Y_WITH_DC,
|
|
|
|
TX_32X32, dry_run);
|
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
|
|
|
|
|
|
|
for (b = 16; b < 24; b += 4) {
|
2012-12-13 21:18:38 +01:00
|
|
|
tokenize_b(cpi, xd, b, t, PLANE_TYPE_UV,
|
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
|
|
|
TX_16X16, dry_run);
|
|
|
|
}
|
2012-12-18 20:25:24 +01:00
|
|
|
A[0][8] = L[0][8] = A[1][8] = L[1][8] = 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 (dry_run)
|
|
|
|
*t = t_backup;
|
|
|
|
}
|
|
|
|
|
2012-10-31 01:53:32 +01:00
|
|
|
void vp9_tokenize_mb(VP9_COMP *cpi,
|
2012-08-15 12:00:53 +02:00
|
|
|
MACROBLOCKD *xd,
|
2012-08-06 19:51:20 +02:00
|
|
|
TOKENEXTRA **t,
|
|
|
|
int dry_run) {
|
2012-10-15 00:29:56 +02:00
|
|
|
PLANE_TYPE plane_type;
|
2012-11-16 00:14:38 +01:00
|
|
|
int has_2nd_order;
|
2012-07-14 00:21:29 +02:00
|
|
|
int b;
|
2012-09-10 07:42:35 +02:00
|
|
|
int tx_size = xd->mode_info_context->mbmi.txfm_size;
|
2012-10-29 14:44:18 +01:00
|
|
|
int mb_skip_context = vp9_get_pred_context(&cpi->common, xd, PRED_MBSKIP);
|
2012-08-06 19:51:20 +02:00
|
|
|
TOKENEXTRA *t_backup = *t;
|
2011-11-04 19:29:51 +01:00
|
|
|
|
2012-07-14 00:21:29 +02:00
|
|
|
// If the MB is going to be skipped because of a segment level flag
|
|
|
|
// exclude this from the skip count stats used to calculate the
|
|
|
|
// transmitted skip probability;
|
|
|
|
int skip_inc;
|
2012-08-15 12:00:53 +02:00
|
|
|
int segment_id = xd->mode_info_context->mbmi.segment_id;
|
2011-10-28 16:27:23 +02:00
|
|
|
|
2013-01-28 16:22:53 +01:00
|
|
|
if (!vp9_segfeature_active(xd, segment_id, SEG_LVL_SKIP)) {
|
2012-07-14 00:21:29 +02:00
|
|
|
skip_inc = 1;
|
|
|
|
} else
|
|
|
|
skip_inc = 0;
|
2011-10-28 16:27:23 +02:00
|
|
|
|
2012-11-16 00:14:38 +01:00
|
|
|
has_2nd_order = get_2nd_order_usage(xd);
|
2010-05-18 17:58:33 +02:00
|
|
|
|
2012-09-10 07:42:35 +02:00
|
|
|
switch (tx_size) {
|
2012-08-03 02:03:14 +02:00
|
|
|
case TX_16X16:
|
2012-12-18 20:25:24 +01:00
|
|
|
|
2012-08-15 18:43:36 +02:00
|
|
|
xd->mode_info_context->mbmi.mb_skip_coeff = mb_is_skippable_16x16(xd);
|
2012-08-03 02:03:14 +02:00
|
|
|
break;
|
|
|
|
case TX_8X8:
|
2012-10-22 20:49:00 +02:00
|
|
|
if (xd->mode_info_context->mbmi.mode == I8X8_PRED ||
|
|
|
|
xd->mode_info_context->mbmi.mode == SPLITMV)
|
2012-11-16 00:14:38 +01:00
|
|
|
xd->mode_info_context->mbmi.mb_skip_coeff =
|
|
|
|
mb_is_skippable_8x8_4x4uv(xd, 0);
|
2012-09-12 04:36:28 +02:00
|
|
|
else
|
2012-11-16 00:14:38 +01:00
|
|
|
xd->mode_info_context->mbmi.mb_skip_coeff =
|
|
|
|
mb_is_skippable_8x8(xd, has_2nd_order);
|
2012-08-03 02:03:14 +02:00
|
|
|
break;
|
2012-09-12 04:36:28 +02:00
|
|
|
|
2012-08-03 02:03:14 +02:00
|
|
|
default:
|
2012-11-16 00:14:38 +01:00
|
|
|
xd->mode_info_context->mbmi.mb_skip_coeff =
|
|
|
|
mb_is_skippable_4x4(xd, has_2nd_order);
|
2012-08-03 02:03:14 +02:00
|
|
|
break;
|
|
|
|
}
|
2011-02-14 23:18:18 +01:00
|
|
|
|
2012-08-15 12:00:53 +02:00
|
|
|
if (xd->mode_info_context->mbmi.mb_skip_coeff) {
|
2012-08-06 19:51:20 +02:00
|
|
|
if (!dry_run)
|
|
|
|
cpi->skip_true_count[mb_skip_context] += skip_inc;
|
2012-07-14 00:21:29 +02:00
|
|
|
if (!cpi->common.mb_no_coeff_skip) {
|
2012-10-30 20:58:42 +01:00
|
|
|
vp9_stuff_mb(cpi, xd, t, dry_run);
|
2012-07-14 00:21:29 +02:00
|
|
|
} else {
|
2012-11-29 00:15:51 +01:00
|
|
|
vp9_reset_mb_tokens_context(xd);
|
2010-05-18 17:58:33 +02:00
|
|
|
}
|
2012-12-06 21:40:57 +01:00
|
|
|
|
2012-08-06 19:51:20 +02:00
|
|
|
if (dry_run)
|
|
|
|
*t = t_backup;
|
2012-07-14 00:21:29 +02:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2012-08-06 19:51:20 +02:00
|
|
|
if (!dry_run)
|
|
|
|
cpi->skip_false_count[mb_skip_context] += skip_inc;
|
2011-04-21 20:38:36 +02:00
|
|
|
|
2012-11-16 00:14:38 +01:00
|
|
|
if (has_2nd_order) {
|
2012-12-06 21:40:57 +01:00
|
|
|
tokenize_b(cpi, xd, 24, t, PLANE_TYPE_Y2, tx_size, dry_run);
|
2012-10-15 00:29:56 +02:00
|
|
|
plane_type = PLANE_TYPE_Y_NO_DC;
|
2012-11-28 01:02:35 +01:00
|
|
|
} else {
|
2012-11-29 00:15:51 +01:00
|
|
|
xd->above_context->y2 = 0;
|
|
|
|
xd->left_context->y2 = 0;
|
2012-10-15 00:29:56 +02:00
|
|
|
plane_type = PLANE_TYPE_Y_WITH_DC;
|
2012-11-28 01:02:35 +01:00
|
|
|
}
|
2012-07-14 00:21:29 +02:00
|
|
|
|
2012-09-10 07:42:35 +02:00
|
|
|
if (tx_size == TX_16X16) {
|
2012-12-06 21:40:57 +01:00
|
|
|
tokenize_b(cpi, xd, 0, t, PLANE_TYPE_Y_WITH_DC, TX_16X16, dry_run);
|
2012-08-03 02:03:14 +02:00
|
|
|
for (b = 16; b < 24; b += 4) {
|
2012-12-06 21:40:57 +01:00
|
|
|
tokenize_b(cpi, xd, b, t, PLANE_TYPE_UV, TX_8X8, dry_run);
|
2012-08-03 02:03:14 +02:00
|
|
|
}
|
2012-10-25 18:14:21 +02:00
|
|
|
} else if (tx_size == TX_8X8) {
|
2012-07-14 00:21:29 +02:00
|
|
|
for (b = 0; b < 16; b += 4) {
|
2012-12-06 21:40:57 +01:00
|
|
|
tokenize_b(cpi, xd, b, t, plane_type, TX_8X8, dry_run);
|
2012-07-14 00:21:29 +02:00
|
|
|
}
|
2012-10-22 20:49:00 +02:00
|
|
|
if (xd->mode_info_context->mbmi.mode == I8X8_PRED ||
|
|
|
|
xd->mode_info_context->mbmi.mode == SPLITMV) {
|
2012-10-25 18:14:21 +02:00
|
|
|
for (b = 16; b < 24; b++) {
|
2012-12-06 21:40:57 +01:00
|
|
|
tokenize_b(cpi, xd, b, t, PLANE_TYPE_UV, TX_4X4, dry_run);
|
2012-10-25 18:14:21 +02:00
|
|
|
}
|
2012-10-05 14:08:20 +02:00
|
|
|
} else {
|
2012-09-12 04:36:28 +02:00
|
|
|
for (b = 16; b < 24; b += 4) {
|
2012-12-06 21:40:57 +01:00
|
|
|
tokenize_b(cpi, xd, b, t, PLANE_TYPE_UV, TX_8X8, dry_run);
|
2012-09-12 04:36:28 +02:00
|
|
|
}
|
2011-02-14 23:18:18 +01:00
|
|
|
}
|
2012-06-25 21:26:09 +02:00
|
|
|
} else {
|
2012-12-06 21:40:57 +01:00
|
|
|
for (b = 0; b < 24; b++) {
|
|
|
|
if (b >= 16)
|
|
|
|
plane_type = PLANE_TYPE_UV;
|
|
|
|
tokenize_b(cpi, xd, b, t, plane_type, TX_4X4, dry_run);
|
2012-10-25 18:14:21 +02:00
|
|
|
}
|
2012-06-25 21:26:09 +02:00
|
|
|
}
|
2012-08-06 19:51:20 +02:00
|
|
|
if (dry_run)
|
|
|
|
*t = t_backup;
|
2010-05-18 17:58:33 +02:00
|
|
|
}
|
2010-08-31 16:49:57 +02:00
|
|
|
|
2010-05-18 17:58:33 +02:00
|
|
|
#ifdef ENTROPY_STATS
|
2012-07-14 00:21:29 +02:00
|
|
|
void init_context_counters(void) {
|
|
|
|
FILE *f = fopen("context.bin", "rb");
|
|
|
|
if (!f) {
|
2012-12-08 01:09:59 +01:00
|
|
|
vpx_memset(context_counters_4x4, 0, sizeof(context_counters_4x4));
|
|
|
|
vpx_memset(hybrid_context_counters_4x4, 0,
|
|
|
|
sizeof(hybrid_context_counters_4x4));
|
2012-07-14 00:21:29 +02:00
|
|
|
vpx_memset(context_counters_8x8, 0, sizeof(context_counters_8x8));
|
2012-12-08 01:09:59 +01:00
|
|
|
vpx_memset(hybrid_context_counters_8x8, 0,
|
|
|
|
sizeof(hybrid_context_counters_8x8));
|
2012-08-03 02:03:14 +02:00
|
|
|
vpx_memset(context_counters_16x16, 0, sizeof(context_counters_16x16));
|
2012-12-08 01:09:59 +01:00
|
|
|
vpx_memset(hybrid_context_counters_16x16, 0,
|
|
|
|
sizeof(hybrid_context_counters_16x16));
|
|
|
|
vpx_memset(context_counters_32x32, 0, sizeof(context_counters_32x32));
|
2012-07-14 00:21:29 +02:00
|
|
|
} else {
|
2012-12-08 01:09:59 +01:00
|
|
|
fread(context_counters_4x4, sizeof(context_counters_4x4), 1, f);
|
|
|
|
fread(hybrid_context_counters_4x4,
|
|
|
|
sizeof(hybrid_context_counters_4x4), 1, f);
|
2012-07-14 00:21:29 +02:00
|
|
|
fread(context_counters_8x8, sizeof(context_counters_8x8), 1, f);
|
2012-12-08 01:09:59 +01:00
|
|
|
fread(hybrid_context_counters_8x8,
|
|
|
|
sizeof(hybrid_context_counters_8x8), 1, f);
|
2012-08-03 02:03:14 +02:00
|
|
|
fread(context_counters_16x16, sizeof(context_counters_16x16), 1, f);
|
2012-12-08 01:09:59 +01:00
|
|
|
fread(hybrid_context_counters_16x16,
|
|
|
|
sizeof(hybrid_context_counters_16x16), 1, f);
|
|
|
|
fread(context_counters_32x32, sizeof(context_counters_32x32), 1, f);
|
2012-07-14 00:21:29 +02:00
|
|
|
fclose(f);
|
|
|
|
}
|
|
|
|
|
|
|
|
f = fopen("treeupdate.bin", "rb");
|
|
|
|
if (!f) {
|
2012-12-08 01:09:59 +01:00
|
|
|
vpx_memset(tree_update_hist_4x4, 0, sizeof(tree_update_hist_4x4));
|
|
|
|
vpx_memset(hybrid_tree_update_hist_4x4, 0,
|
|
|
|
sizeof(hybrid_tree_update_hist_4x4));
|
2012-07-14 00:21:29 +02:00
|
|
|
vpx_memset(tree_update_hist_8x8, 0, sizeof(tree_update_hist_8x8));
|
2012-12-08 01:09:59 +01:00
|
|
|
vpx_memset(hybrid_tree_update_hist_8x8, 0,
|
|
|
|
sizeof(hybrid_tree_update_hist_8x8));
|
2012-08-03 02:03:14 +02:00
|
|
|
vpx_memset(tree_update_hist_16x16, 0, sizeof(tree_update_hist_16x16));
|
2012-12-08 01:09:59 +01:00
|
|
|
vpx_memset(hybrid_tree_update_hist_16x16, 0,
|
|
|
|
sizeof(hybrid_tree_update_hist_16x16));
|
|
|
|
vpx_memset(tree_update_hist_32x32, 0, sizeof(tree_update_hist_32x32));
|
2012-07-14 00:21:29 +02:00
|
|
|
} else {
|
2012-12-08 01:09:59 +01:00
|
|
|
fread(tree_update_hist_4x4, sizeof(tree_update_hist_4x4), 1, f);
|
|
|
|
fread(hybrid_tree_update_hist_4x4,
|
|
|
|
sizeof(hybrid_tree_update_hist_4x4), 1, f);
|
2012-07-14 00:21:29 +02:00
|
|
|
fread(tree_update_hist_8x8, sizeof(tree_update_hist_8x8), 1, f);
|
2012-12-08 01:09:59 +01:00
|
|
|
fread(hybrid_tree_update_hist_8x8,
|
|
|
|
sizeof(hybrid_tree_update_hist_8x8), 1, f);
|
2012-08-03 02:03:14 +02:00
|
|
|
fread(tree_update_hist_16x16, sizeof(tree_update_hist_16x16), 1, f);
|
2012-12-08 01:09:59 +01:00
|
|
|
fread(hybrid_tree_update_hist_16x16,
|
|
|
|
sizeof(hybrid_tree_update_hist_16x16), 1, f);
|
|
|
|
fread(tree_update_hist_32x32, sizeof(tree_update_hist_32x32), 1, f);
|
2012-07-14 00:21:29 +02:00
|
|
|
fclose(f);
|
|
|
|
}
|
2010-05-18 17:58:33 +02:00
|
|
|
}
|
|
|
|
|
2012-12-08 01:09:59 +01:00
|
|
|
static void print_counter(FILE *f, vp9_coeff_accum *context_counters,
|
|
|
|
int block_types, const char *header) {
|
2012-07-14 00:21:29 +02:00
|
|
|
int type, band, pt, t;
|
2010-05-18 17:58:33 +02:00
|
|
|
|
2012-12-08 01:09:59 +01:00
|
|
|
fprintf(f, "static const vp9_coeff_count %s = {\n", header);
|
2010-05-18 17:58:33 +02:00
|
|
|
|
2012-12-12 19:25:58 +01:00
|
|
|
#define Comma(X) (X ? "," : "")
|
2012-07-14 00:21:29 +02:00
|
|
|
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 {
|
2012-12-14 21:35:33 +01:00
|
|
|
const int64_t x = context_counters[type][band][pt][t];
|
2012-07-14 00:21:29 +02:00
|
|
|
const int y = (int) x;
|
2011-02-14 23:18:18 +01:00
|
|
|
|
2012-12-14 21:35:33 +01:00
|
|
|
assert(x == (int64_t) y); /* no overflow handling yet */
|
2012-07-14 00:21:29 +02:00
|
|
|
fprintf(f, "%s %d", Comma(t), y);
|
|
|
|
} while (++t < MAX_ENTROPY_TOKENS);
|
|
|
|
fprintf(f, "}");
|
|
|
|
} while (++pt < PREV_COEF_CONTEXTS);
|
|
|
|
fprintf(f, "\n }");
|
|
|
|
} while (++band < COEF_BANDS);
|
|
|
|
fprintf(f, "\n }");
|
2012-12-08 01:09:59 +01:00
|
|
|
} while (++type < block_types);
|
2012-08-03 02:03:14 +02:00
|
|
|
fprintf(f, "\n};\n");
|
2012-12-08 01:09:59 +01:00
|
|
|
}
|
2012-03-21 23:22:21 +01:00
|
|
|
|
2012-12-08 01:09:59 +01:00
|
|
|
static void print_probs(FILE *f, vp9_coeff_accum *context_counters,
|
|
|
|
int block_types, const char *header) {
|
|
|
|
int type, band, pt, t;
|
2012-08-03 02:03:14 +02:00
|
|
|
|
2012-12-12 19:25:58 +01:00
|
|
|
fprintf(f, "static const vp9_coeff_probs %s = {", header);
|
2012-03-21 23:22:21 +01:00
|
|
|
|
2012-07-14 00:21:29 +02:00
|
|
|
type = 0;
|
2012-12-12 19:25:58 +01:00
|
|
|
#define Newline(x, spaces) (x ? " " : "\n" spaces)
|
2012-07-14 00:21:29 +02:00
|
|
|
do {
|
2012-12-12 19:25:58 +01:00
|
|
|
fprintf(f, "%s%s{ /* block Type %d */",
|
|
|
|
Comma(type), Newline(type, " "), type);
|
2012-07-14 00:21:29 +02:00
|
|
|
band = 0;
|
|
|
|
do {
|
2012-12-12 19:25:58 +01:00
|
|
|
fprintf(f, "%s%s{ /* Coeff Band %d */",
|
|
|
|
Comma(band), Newline(band, " "), band);
|
2012-07-14 00:21:29 +02:00
|
|
|
pt = 0;
|
|
|
|
do {
|
2012-12-12 19:25:58 +01:00
|
|
|
unsigned int branch_ct[ENTROPY_NODES][2];
|
2012-07-14 00:21:29 +02:00
|
|
|
unsigned int coef_counts[MAX_ENTROPY_TOKENS];
|
2012-10-31 22:40:53 +01:00
|
|
|
vp9_prob coef_probs[ENTROPY_NODES];
|
2012-12-08 01:09:59 +01:00
|
|
|
|
2012-07-14 00:21:29 +02:00
|
|
|
for (t = 0; t < MAX_ENTROPY_TOKENS; ++t)
|
2012-12-08 01:09:59 +01:00
|
|
|
coef_counts[t] = context_counters[type][band][pt][t];
|
|
|
|
vp9_tree_probs_from_distribution(MAX_ENTROPY_TOKENS,
|
|
|
|
vp9_coef_encodings, vp9_coef_tree,
|
Consistently use get_prob(), clip_prob() and newly added clip_pixel().
Add a function clip_pixel() to clip a pixel value to the [0,255] range
of allowed values, and use this where-ever appropriate (e.g. prediction,
reconstruction). Likewise, consistently use the recently added function
clip_prob(), which calculates a binary probability in the [1,255] range.
If possible, try to use get_prob() or its sister get_binary_prob() to
calculate binary probabilities, for consistency.
Since in some places, this means that binary probability calculations
are changed (we use {255,256}*count0/(total) in a range of places,
and all of these are now changed to use 256*count0+(total>>1)/total),
this changes the encoding result, so this patch warrants some extensive
testing.
Change-Id: Ibeeff8d886496839b8e0c0ace9ccc552351f7628
2012-12-10 21:09:07 +01:00
|
|
|
coef_probs, branch_ct, coef_counts);
|
2012-07-14 00:21:29 +02:00
|
|
|
fprintf(f, "%s\n {", Comma(pt));
|
2012-03-21 23:22:21 +01:00
|
|
|
|
2012-07-14 00:21:29 +02:00
|
|
|
t = 0;
|
|
|
|
do {
|
2012-12-12 19:25:58 +01:00
|
|
|
fprintf(f, "%s %3d", Comma(t), coef_probs[t]);
|
2012-07-14 00:21:29 +02:00
|
|
|
} while (++t < ENTROPY_NODES);
|
2012-03-21 23:22:21 +01:00
|
|
|
|
2012-12-12 19:25:58 +01:00
|
|
|
fprintf(f, " }");
|
2012-07-14 00:21:29 +02:00
|
|
|
} while (++pt < PREV_COEF_CONTEXTS);
|
|
|
|
fprintf(f, "\n }");
|
|
|
|
} while (++band < COEF_BANDS);
|
|
|
|
fprintf(f, "\n }");
|
2012-12-08 01:09:59 +01:00
|
|
|
} while (++type < block_types);
|
2012-07-14 00:21:29 +02:00
|
|
|
fprintf(f, "\n};\n");
|
2012-12-08 01:09:59 +01:00
|
|
|
}
|
2012-03-21 23:22:21 +01:00
|
|
|
|
2012-12-08 01:09:59 +01:00
|
|
|
void print_context_counters() {
|
|
|
|
FILE *f = fopen("vp9_context.c", "w");
|
2012-03-21 23:22:21 +01:00
|
|
|
|
2012-12-08 01:09:59 +01:00
|
|
|
fprintf(f, "#include \"vp9_entropy.h\"\n");
|
|
|
|
fprintf(f, "\n/* *** GENERATED FILE: DO NOT EDIT *** */\n\n");
|
2012-07-14 00:21:29 +02:00
|
|
|
|
2012-12-08 01:09:59 +01:00
|
|
|
/* print counts */
|
|
|
|
print_counter(f, context_counters_4x4, BLOCK_TYPES_4X4,
|
|
|
|
"vp9_default_coef_counts_4x4[BLOCK_TYPES_4X4]");
|
|
|
|
print_counter(f, hybrid_context_counters_4x4, BLOCK_TYPES_4X4,
|
|
|
|
"vp9_default_hybrid_coef_counts_4x4[BLOCK_TYPES_4X4]");
|
|
|
|
print_counter(f, context_counters_8x8, BLOCK_TYPES_8X8,
|
|
|
|
"vp9_default_coef_counts_8x8[BLOCK_TYPES_8X8]");
|
|
|
|
print_counter(f, hybrid_context_counters_8x8, BLOCK_TYPES_8X8,
|
|
|
|
"vp9_default_hybrid_coef_counts_8x8[BLOCK_TYPES_8X8]");
|
|
|
|
print_counter(f, context_counters_16x16, BLOCK_TYPES_16X16,
|
|
|
|
"vp9_default_coef_counts_16x16[BLOCK_TYPES_16X16]");
|
|
|
|
print_counter(f, hybrid_context_counters_16x16, BLOCK_TYPES_16X16,
|
|
|
|
"vp9_default_hybrid_coef_counts_16x16[BLOCK_TYPES_16X16]");
|
|
|
|
print_counter(f, context_counters_32x32, BLOCK_TYPES_32X32,
|
|
|
|
"vp9_default_coef_counts_32x32[BLOCK_TYPES_32X32]");
|
2012-08-03 02:03:14 +02:00
|
|
|
|
2012-12-08 01:09:59 +01:00
|
|
|
/* print coefficient probabilities */
|
|
|
|
print_probs(f, context_counters_4x4, BLOCK_TYPES_4X4,
|
2012-12-12 19:25:58 +01:00
|
|
|
"default_coef_probs_4x4[BLOCK_TYPES_4X4]");
|
2012-12-08 01:09:59 +01:00
|
|
|
print_probs(f, hybrid_context_counters_4x4, BLOCK_TYPES_4X4,
|
2012-12-12 19:25:58 +01:00
|
|
|
"default_hybrid_coef_probs_4x4[BLOCK_TYPES_4X4]");
|
2012-12-08 01:09:59 +01:00
|
|
|
print_probs(f, context_counters_8x8, BLOCK_TYPES_8X8,
|
2012-12-12 19:25:58 +01:00
|
|
|
"default_coef_probs_8x8[BLOCK_TYPES_8X8]");
|
2012-12-08 01:09:59 +01:00
|
|
|
print_probs(f, hybrid_context_counters_8x8, BLOCK_TYPES_8X8,
|
2012-12-12 19:25:58 +01:00
|
|
|
"default_hybrid_coef_probs_8x8[BLOCK_TYPES_8X8]");
|
2012-12-08 01:09:59 +01:00
|
|
|
print_probs(f, context_counters_16x16, BLOCK_TYPES_16X16,
|
2012-12-12 19:25:58 +01:00
|
|
|
"default_coef_probs_16x16[BLOCK_TYPES_16X16]");
|
2012-12-08 01:09:59 +01:00
|
|
|
print_probs(f, hybrid_context_counters_16x16, BLOCK_TYPES_16X16,
|
2012-12-12 19:25:58 +01:00
|
|
|
"default_hybrid_coef_probs_16x16[BLOCK_TYPES_16X16]");
|
2012-12-08 01:09:59 +01:00
|
|
|
print_probs(f, context_counters_32x32, BLOCK_TYPES_32X32,
|
2012-12-12 19:25:58 +01:00
|
|
|
"default_coef_probs_32x32[BLOCK_TYPES_32X32]");
|
2012-08-03 02:03:14 +02:00
|
|
|
|
2012-07-14 00:21:29 +02:00
|
|
|
fclose(f);
|
|
|
|
|
|
|
|
f = fopen("context.bin", "wb");
|
2012-12-08 01:09:59 +01:00
|
|
|
fwrite(context_counters_4x4, sizeof(context_counters_4x4), 1, f);
|
|
|
|
fwrite(hybrid_context_counters_4x4,
|
|
|
|
sizeof(hybrid_context_counters_4x4), 1, f);
|
2012-07-14 00:21:29 +02:00
|
|
|
fwrite(context_counters_8x8, sizeof(context_counters_8x8), 1, f);
|
2012-12-08 01:09:59 +01:00
|
|
|
fwrite(hybrid_context_counters_8x8,
|
|
|
|
sizeof(hybrid_context_counters_8x8), 1, f);
|
2012-08-03 02:03:14 +02:00
|
|
|
fwrite(context_counters_16x16, sizeof(context_counters_16x16), 1, f);
|
2012-12-08 01:09:59 +01:00
|
|
|
fwrite(hybrid_context_counters_16x16,
|
|
|
|
sizeof(hybrid_context_counters_16x16), 1, f);
|
|
|
|
fwrite(context_counters_32x32, sizeof(context_counters_32x32), 1, f);
|
2012-07-14 00:21:29 +02:00
|
|
|
fclose(f);
|
2010-05-18 17:58:33 +02:00
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2012-10-30 20:58:42 +01:00
|
|
|
void vp9_tokenize_initialize() {
|
2012-07-14 00:21:29 +02:00
|
|
|
fill_value_tokens();
|
2010-05-18 17:58:33 +02:00
|
|
|
}
|
|
|
|
|
2012-10-31 01:53:32 +01:00
|
|
|
static __inline void stuff_b(VP9_COMP *cpi,
|
2012-10-25 01:59:00 +02:00
|
|
|
MACROBLOCKD *xd,
|
2012-12-06 21:40:57 +01:00
|
|
|
const int ib,
|
2012-10-25 01:59:00 +02:00
|
|
|
TOKENEXTRA **tp,
|
|
|
|
PLANE_TYPE type,
|
|
|
|
TX_SIZE tx_size,
|
|
|
|
int dry_run) {
|
2012-12-06 21:40:57 +01:00
|
|
|
const BLOCKD * const b = xd->block + ib;
|
2012-10-25 01:59:00 +02:00
|
|
|
const int *bands;
|
2012-12-08 01:09:59 +01:00
|
|
|
vp9_coeff_count *counts;
|
|
|
|
vp9_coeff_probs *probs;
|
2012-10-25 01:59:00 +02:00
|
|
|
int pt, band;
|
|
|
|
TOKENEXTRA *t = *tp;
|
|
|
|
const TX_TYPE tx_type = (type == PLANE_TYPE_Y_WITH_DC) ?
|
|
|
|
get_tx_type(xd, b) : DCT_DCT;
|
2012-12-06 21:40:57 +01:00
|
|
|
ENTROPY_CONTEXT *const a = (ENTROPY_CONTEXT *)xd->above_context +
|
|
|
|
vp9_block2above[tx_size][ib];
|
|
|
|
ENTROPY_CONTEXT *const l = (ENTROPY_CONTEXT *)xd->left_context +
|
|
|
|
vp9_block2left[tx_size][ib];
|
|
|
|
ENTROPY_CONTEXT a_ec = *a, l_ec = *l;
|
2012-12-18 20:25:24 +01:00
|
|
|
ENTROPY_CONTEXT *const a1 = (ENTROPY_CONTEXT *)(&xd->above_context[1]) +
|
|
|
|
vp9_block2above[tx_size][ib];
|
|
|
|
ENTROPY_CONTEXT *const l1 = (ENTROPY_CONTEXT *)(&xd->left_context[1]) +
|
|
|
|
vp9_block2left[tx_size][ib];
|
2012-07-14 00:21:29 +02:00
|
|
|
|
2012-10-25 01:59:00 +02:00
|
|
|
switch (tx_size) {
|
|
|
|
default:
|
|
|
|
case TX_4X4:
|
2012-12-18 19:49:10 +01:00
|
|
|
bands = vp9_coef_bands_4x4;
|
2012-10-25 01:59:00 +02:00
|
|
|
if (tx_type != DCT_DCT) {
|
2012-12-08 01:09:59 +01:00
|
|
|
counts = cpi->hybrid_coef_counts_4x4;
|
|
|
|
probs = cpi->common.fc.hybrid_coef_probs_4x4;
|
2012-10-25 01:59:00 +02:00
|
|
|
} else {
|
2012-12-08 01:09:59 +01:00
|
|
|
counts = cpi->coef_counts_4x4;
|
|
|
|
probs = cpi->common.fc.coef_probs_4x4;
|
2012-10-25 01:59:00 +02:00
|
|
|
}
|
|
|
|
break;
|
|
|
|
case TX_8X8:
|
2012-12-06 21:40:57 +01:00
|
|
|
#if CONFIG_CNVCONTEXT
|
|
|
|
if (type != PLANE_TYPE_Y2) {
|
|
|
|
a_ec = (a[0] + a[1]) != 0;
|
|
|
|
l_ec = (l[0] + l[1]) != 0;
|
|
|
|
}
|
|
|
|
#endif
|
2012-10-31 01:12:12 +01:00
|
|
|
bands = vp9_coef_bands_8x8;
|
2012-10-25 01:59:00 +02:00
|
|
|
if (tx_type != DCT_DCT) {
|
|
|
|
counts = cpi->hybrid_coef_counts_8x8;
|
|
|
|
probs = cpi->common.fc.hybrid_coef_probs_8x8;
|
|
|
|
} else {
|
|
|
|
counts = cpi->coef_counts_8x8;
|
|
|
|
probs = cpi->common.fc.coef_probs_8x8;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case TX_16X16:
|
2012-12-06 21:40:57 +01:00
|
|
|
#if CONFIG_CNVCONTEXT
|
2012-12-13 23:58:30 +01:00
|
|
|
if (type != PLANE_TYPE_UV) {
|
|
|
|
a_ec = (a[0] + a[1] + a[2] + a[3]) != 0;
|
|
|
|
l_ec = (l[0] + l[1] + l[2] + l[3]) != 0;
|
2012-12-18 20:25:24 +01:00
|
|
|
} else {
|
|
|
|
a_ec = (a[0] + a[1] + a1[0] + a1[1]) != 0;
|
|
|
|
l_ec = (l[0] + l[1] + l1[0] + l1[1]) != 0;
|
2012-12-13 23:58:30 +01:00
|
|
|
}
|
2012-12-06 21:40:57 +01:00
|
|
|
#endif
|
2012-10-31 01:12:12 +01:00
|
|
|
bands = vp9_coef_bands_16x16;
|
2012-10-25 01:59:00 +02:00
|
|
|
if (tx_type != DCT_DCT) {
|
|
|
|
counts = cpi->hybrid_coef_counts_16x16;
|
|
|
|
probs = cpi->common.fc.hybrid_coef_probs_16x16;
|
|
|
|
} else {
|
|
|
|
counts = cpi->coef_counts_16x16;
|
|
|
|
probs = cpi->common.fc.coef_probs_16x16;
|
|
|
|
}
|
|
|
|
break;
|
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
|
|
|
case TX_32X32:
|
2012-12-18 20:25:24 +01:00
|
|
|
#if CONFIG_CNVCONTEXT
|
|
|
|
a_ec = a[0] + a[1] + a[2] + a[3] +
|
|
|
|
a1[0] + a1[1] + a1[2] + a1[3];
|
|
|
|
l_ec = l[0] + l[1] + l[2] + l[3] +
|
|
|
|
l1[0] + l1[1] + l1[2] + l1[3];
|
|
|
|
a_ec = a_ec != 0;
|
|
|
|
l_ec = l_ec != 0;
|
|
|
|
#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
|
|
|
bands = vp9_coef_bands_32x32;
|
|
|
|
counts = cpi->coef_counts_32x32;
|
|
|
|
probs = cpi->common.fc.coef_probs_32x32;
|
|
|
|
break;
|
2012-10-25 01:59:00 +02:00
|
|
|
}
|
2012-12-06 21:40:57 +01:00
|
|
|
|
|
|
|
VP9_COMBINEENTROPYCONTEXTS(pt, a_ec, l_ec);
|
|
|
|
|
2012-10-25 01:59:00 +02:00
|
|
|
band = bands[(type == PLANE_TYPE_Y_NO_DC) ? 1 : 0];
|
2012-07-14 00:21:29 +02:00
|
|
|
t->Token = DCT_EOB_TOKEN;
|
2012-10-25 01:59:00 +02:00
|
|
|
t->context_tree = probs[type][band][pt];
|
2012-07-14 00:21:29 +02:00
|
|
|
t->skip_eob_node = 0;
|
|
|
|
++t;
|
|
|
|
*tp = t;
|
2012-10-25 01:59:00 +02:00
|
|
|
*a = *l = 0;
|
2012-12-06 21:40:57 +01:00
|
|
|
if (tx_size == TX_8X8 && type != PLANE_TYPE_Y2) {
|
|
|
|
a[1] = 0;
|
|
|
|
l[1] = 0;
|
|
|
|
} else if (tx_size == TX_16X16) {
|
2012-12-13 23:58:30 +01:00
|
|
|
if (type != PLANE_TYPE_UV) {
|
|
|
|
a[1] = a[2] = a[3] = 0;
|
|
|
|
l[1] = l[2] = l[3] = 0;
|
|
|
|
} else {
|
2012-12-18 20:25:24 +01:00
|
|
|
a1[0] = a1[1] = a[1] = a_ec;
|
|
|
|
l1[0] = l1[1] = l[1] = l_ec;
|
2012-12-13 23:58:30 +01:00
|
|
|
}
|
2012-12-18 20:25:24 +01:00
|
|
|
} else if (tx_size == TX_32X32) {
|
|
|
|
a[1] = a[2] = a[3] = a_ec;
|
|
|
|
l[1] = l[2] = l[3] = l_ec;
|
|
|
|
a1[0] = a1[1] = a1[2] = a1[3] = a_ec;
|
|
|
|
l1[0] = l1[1] = l1[2] = l1[3] = l_ec;
|
2012-12-06 21:40:57 +01:00
|
|
|
}
|
|
|
|
|
2012-09-10 07:42:35 +02:00
|
|
|
if (!dry_run) {
|
2012-10-25 01:59:00 +02:00
|
|
|
++counts[type][band][pt][DCT_EOB_TOKEN];
|
2012-09-10 07:42:35 +02:00
|
|
|
}
|
2011-02-14 23:18:18 +01:00
|
|
|
}
|
|
|
|
|
2012-10-31 22:40:53 +01:00
|
|
|
static void stuff_mb_8x8(VP9_COMP *cpi, MACROBLOCKD *xd,
|
|
|
|
TOKENEXTRA **t, int dry_run) {
|
2012-10-15 19:52:13 +02:00
|
|
|
PLANE_TYPE plane_type;
|
2012-07-14 00:21:29 +02:00
|
|
|
int b;
|
2012-11-16 00:14:38 +01:00
|
|
|
int has_2nd_order = get_2nd_order_usage(xd);
|
2012-07-14 00:21:29 +02:00
|
|
|
|
2012-11-16 00:14:38 +01:00
|
|
|
if (has_2nd_order) {
|
2012-12-06 21:40:57 +01:00
|
|
|
stuff_b(cpi, xd, 24, t, PLANE_TYPE_Y2, TX_8X8, dry_run);
|
2012-10-15 19:52:13 +02:00
|
|
|
plane_type = PLANE_TYPE_Y_NO_DC;
|
|
|
|
} else {
|
2012-11-29 00:15:51 +01:00
|
|
|
#if CONFIG_CNVCONTEXT
|
|
|
|
xd->above_context->y2 = 0;
|
|
|
|
xd->left_context->y2 = 0;
|
|
|
|
#endif
|
2012-10-15 19:52:13 +02:00
|
|
|
plane_type = PLANE_TYPE_Y_WITH_DC;
|
|
|
|
}
|
2012-07-14 00:21:29 +02:00
|
|
|
|
2012-12-06 21:40:57 +01:00
|
|
|
for (b = 0; b < 24; b += 4) {
|
|
|
|
if (b >= 16)
|
|
|
|
plane_type = PLANE_TYPE_UV;
|
|
|
|
stuff_b(cpi, xd, b, t, plane_type, TX_8X8, dry_run);
|
2012-07-14 00:21:29 +02:00
|
|
|
}
|
2012-08-03 02:03:14 +02:00
|
|
|
}
|
|
|
|
|
2012-10-31 22:40:53 +01:00
|
|
|
static void stuff_mb_16x16(VP9_COMP *cpi, MACROBLOCKD *xd,
|
|
|
|
TOKENEXTRA **t, int dry_run) {
|
2012-10-25 01:59:00 +02:00
|
|
|
int b;
|
2012-12-06 21:40:57 +01:00
|
|
|
stuff_b(cpi, xd, 0, t, PLANE_TYPE_Y_WITH_DC, TX_16X16, dry_run);
|
2012-08-03 02:03:14 +02:00
|
|
|
|
|
|
|
for (b = 16; b < 24; b += 4) {
|
2012-12-06 21:40:57 +01:00
|
|
|
stuff_b(cpi, xd, b, t, PLANE_TYPE_UV, TX_8X8, dry_run);
|
|
|
|
}
|
2012-11-29 00:15:51 +01:00
|
|
|
#if CONFIG_CNVCONTEXT
|
2012-12-06 21:40:57 +01:00
|
|
|
xd->above_context->y2 = 0;
|
|
|
|
xd->left_context->y2 = 0;
|
2012-11-29 00:15:51 +01:00
|
|
|
#endif
|
2010-05-18 17:58:33 +02:00
|
|
|
}
|
|
|
|
|
2012-10-31 22:40:53 +01:00
|
|
|
static void stuff_mb_4x4(VP9_COMP *cpi, MACROBLOCKD *xd,
|
|
|
|
TOKENEXTRA **t, int dry_run) {
|
2012-07-14 00:21:29 +02:00
|
|
|
int b;
|
2012-10-15 19:52:13 +02:00
|
|
|
PLANE_TYPE plane_type;
|
2012-11-29 00:15:51 +01:00
|
|
|
int has_2nd_order = get_2nd_order_usage(xd);
|
2010-05-18 17:58:33 +02:00
|
|
|
|
2012-11-16 00:14:38 +01:00
|
|
|
if (has_2nd_order) {
|
2012-12-06 21:40:57 +01:00
|
|
|
stuff_b(cpi, xd, 24, t, PLANE_TYPE_Y2, TX_4X4, dry_run);
|
2012-10-15 19:52:13 +02:00
|
|
|
plane_type = PLANE_TYPE_Y_NO_DC;
|
|
|
|
} else {
|
2012-11-29 00:15:51 +01:00
|
|
|
xd->above_context->y2 = 0;
|
|
|
|
xd->left_context->y2 = 0;
|
2012-10-15 19:52:13 +02:00
|
|
|
plane_type = PLANE_TYPE_Y_WITH_DC;
|
|
|
|
}
|
2010-05-18 17:58:33 +02:00
|
|
|
|
2012-12-06 21:40:57 +01:00
|
|
|
for (b = 0; b < 24; b++) {
|
|
|
|
if (b >= 16)
|
|
|
|
plane_type = PLANE_TYPE_UV;
|
|
|
|
stuff_b(cpi, xd, b, t, plane_type, TX_4X4, dry_run);
|
|
|
|
}
|
2010-05-18 17:58:33 +02:00
|
|
|
}
|
2012-09-12 04:36:28 +02:00
|
|
|
|
2012-10-31 22:40:53 +01:00
|
|
|
static void stuff_mb_8x8_4x4uv(VP9_COMP *cpi, MACROBLOCKD *xd,
|
|
|
|
TOKENEXTRA **t, int dry_run) {
|
2012-11-16 00:14:38 +01:00
|
|
|
PLANE_TYPE plane_type;
|
2012-09-12 04:36:28 +02:00
|
|
|
int b;
|
|
|
|
|
2012-11-16 00:14:38 +01:00
|
|
|
int has_2nd_order = get_2nd_order_usage(xd);
|
|
|
|
if (has_2nd_order) {
|
2012-12-06 21:40:57 +01:00
|
|
|
stuff_b(cpi, xd, 24, t, PLANE_TYPE_Y2, TX_8X8, dry_run);
|
2012-11-16 00:14:38 +01:00
|
|
|
plane_type = PLANE_TYPE_Y_NO_DC;
|
|
|
|
} else {
|
2012-11-29 00:15:51 +01:00
|
|
|
xd->above_context->y2 = 0;
|
|
|
|
xd->left_context->y2 = 0;
|
2012-11-16 00:14:38 +01:00
|
|
|
plane_type = PLANE_TYPE_Y_WITH_DC;
|
|
|
|
}
|
|
|
|
|
2012-09-12 04:36:28 +02:00
|
|
|
for (b = 0; b < 16; b += 4) {
|
2012-12-06 21:40:57 +01:00
|
|
|
stuff_b(cpi, xd, b, t, plane_type, TX_8X8, dry_run);
|
2012-09-12 04:36:28 +02:00
|
|
|
}
|
|
|
|
|
2012-12-06 21:40:57 +01:00
|
|
|
for (b = 16; b < 24; b++) {
|
|
|
|
stuff_b(cpi, xd, b, t, PLANE_TYPE_UV, TX_4X4, dry_run);
|
|
|
|
}
|
2012-09-12 04:36:28 +02:00
|
|
|
}
|
|
|
|
|
2012-10-31 01:53:32 +01:00
|
|
|
void vp9_stuff_mb(VP9_COMP *cpi, MACROBLOCKD *xd, TOKENEXTRA **t, int dry_run) {
|
2012-10-15 19:52:13 +02:00
|
|
|
TX_SIZE tx_size = xd->mode_info_context->mbmi.txfm_size;
|
2012-10-25 01:59:00 +02:00
|
|
|
TOKENEXTRA * const t_backup = *t;
|
2012-10-15 19:52:13 +02:00
|
|
|
|
|
|
|
if (tx_size == TX_16X16) {
|
2012-10-31 22:40:53 +01:00
|
|
|
stuff_mb_16x16(cpi, xd, t, dry_run);
|
2012-10-15 19:52:13 +02:00
|
|
|
} else if (tx_size == TX_8X8) {
|
2012-10-22 20:49:00 +02:00
|
|
|
if (xd->mode_info_context->mbmi.mode == I8X8_PRED ||
|
|
|
|
xd->mode_info_context->mbmi.mode == SPLITMV) {
|
2012-10-31 22:40:53 +01:00
|
|
|
stuff_mb_8x8_4x4uv(cpi, xd, t, dry_run);
|
2012-10-15 19:52:13 +02:00
|
|
|
} else {
|
2012-10-31 22:40:53 +01:00
|
|
|
stuff_mb_8x8(cpi, xd, t, dry_run);
|
2012-10-15 19:52:13 +02:00
|
|
|
}
|
|
|
|
} else {
|
2012-10-31 22:40:53 +01:00
|
|
|
stuff_mb_4x4(cpi, xd, t, dry_run);
|
2012-10-15 19:52:13 +02:00
|
|
|
}
|
2012-10-25 01:59:00 +02:00
|
|
|
|
|
|
|
if (dry_run) {
|
|
|
|
*t = t_backup;
|
|
|
|
}
|
2012-10-15 19:52:13 +02:00
|
|
|
}
|
|
|
|
|
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
|
|
|
static void stuff_sb_32x32(VP9_COMP *cpi, MACROBLOCKD *xd,
|
|
|
|
TOKENEXTRA **t, int dry_run) {
|
|
|
|
int b;
|
|
|
|
|
2012-12-13 21:18:38 +01:00
|
|
|
stuff_b(cpi, xd, 0, t, PLANE_TYPE_Y_WITH_DC, TX_32X32, dry_run);
|
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
|
|
|
for (b = 16; b < 24; b += 4) {
|
2012-12-13 21:18:38 +01:00
|
|
|
stuff_b(cpi, xd, b, t, PLANE_TYPE_UV, TX_16X16, dry_run);
|
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_stuff_sb(VP9_COMP *cpi, MACROBLOCKD *xd, TOKENEXTRA **t, int dry_run) {
|
|
|
|
TOKENEXTRA * const t_backup = *t;
|
|
|
|
|
|
|
|
stuff_sb_32x32(cpi, xd, t, dry_run);
|
|
|
|
|
|
|
|
if (dry_run) {
|
|
|
|
*t = t_backup;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void vp9_fix_contexts_sb(MACROBLOCKD *xd) {
|
|
|
|
vpx_memset(xd->above_context, 0, sizeof(ENTROPY_CONTEXT_PLANES) * 2);
|
|
|
|
vpx_memset(xd->left_context, 0, sizeof(ENTROPY_CONTEXT_PLANES) * 2);
|
|
|
|
}
|