2010-05-18 19:58:33 +04:00
|
|
|
/*
|
2010-09-09 16:16:39 +04:00
|
|
|
* Copyright (c) 2010 The WebM project authors. All Rights Reserved.
|
2010-05-18 19:58:33 +04:00
|
|
|
*
|
2010-06-18 20:39:21 +04:00
|
|
|
* Use of this source code is governed by a BSD-style license
|
2010-06-05 00:19:40 +04:00
|
|
|
* 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 20:39:21 +04:00
|
|
|
* in the file PATENTS. All contributing project authors may
|
2010-06-05 00:19:40 +04:00
|
|
|
* 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>
|
2012-11-28 22:41:40 +04:00
|
|
|
#include "vp9/encoder/vp9_onyx_int.h"
|
|
|
|
#include "vp9/encoder/vp9_tokenize.h"
|
2010-05-18 19:58:33 +04:00
|
|
|
#include "vpx_mem/vpx_mem.h"
|
|
|
|
|
2012-11-28 01:59:17 +04:00
|
|
|
#include "vp9/common/vp9_pred_common.h"
|
|
|
|
#include "vp9/common/vp9_seg_common.h"
|
|
|
|
#include "vp9/common/vp9_entropy.h"
|
2011-10-05 14:26:00 +04:00
|
|
|
|
2010-05-18 19:58:33 +04:00
|
|
|
/* Global event counters used for accumulating statistics across several
|
2012-11-28 01:59:17 +04:00
|
|
|
compressions, then generating vp9_context.c = initial stats. */
|
2010-05-18 19:58:33 +04:00
|
|
|
|
|
|
|
#ifdef ENTROPY_STATS
|
2013-07-27 04:15:37 +04:00
|
|
|
vp9_coeff_accum context_counters[TX_SIZES][BLOCK_TYPES];
|
|
|
|
extern vp9_coeff_stats tree_update_hist[TX_SIZES][BLOCK_TYPES];
|
2012-10-10 22:20:33 +04:00
|
|
|
#endif /* ENTROPY_STATS */
|
|
|
|
|
2013-05-23 14:08:00 +04:00
|
|
|
DECLARE_ALIGNED(16, extern const uint8_t,
|
|
|
|
vp9_pt_energy_class[MAX_ENTROPY_TOKENS]);
|
|
|
|
|
2012-07-14 02:21:29 +04:00
|
|
|
static TOKENVALUE dct_value_tokens[DCT_MAX_VALUE * 2];
|
2012-10-31 04:12:12 +04:00
|
|
|
const TOKENVALUE *vp9_dct_value_tokens_ptr;
|
2012-07-14 02:21:29 +04:00
|
|
|
static int dct_value_cost[DCT_MAX_VALUE * 2];
|
2012-10-31 04:12:12 +04:00
|
|
|
const int *vp9_dct_value_cost_ptr;
|
2011-02-15 01:18:18 +03:00
|
|
|
|
2012-07-14 02:21:29 +04:00
|
|
|
static void fill_value_tokens() {
|
2010-05-18 19:58:33 +04:00
|
|
|
|
2012-07-14 02:21:29 +04:00
|
|
|
TOKENVALUE *const t = dct_value_tokens + DCT_MAX_VALUE;
|
2013-08-01 05:51:18 +04:00
|
|
|
const vp9_extra_bit *const e = vp9_extra_bits;
|
2010-05-18 19:58:33 +04:00
|
|
|
|
2012-07-14 02:21:29 +04:00
|
|
|
int i = -DCT_MAX_VALUE;
|
|
|
|
int sign = 1;
|
2010-05-18 19:58:33 +04:00
|
|
|
|
2012-07-14 02:21:29 +04:00
|
|
|
do {
|
|
|
|
if (!i)
|
|
|
|
sign = 0;
|
2010-05-18 19:58:33 +04:00
|
|
|
|
2012-07-14 02:21:29 +04:00
|
|
|
{
|
|
|
|
const int a = sign ? -i : i;
|
|
|
|
int eb = sign;
|
2010-05-18 19:58:33 +04:00
|
|
|
|
2012-07-14 02:21:29 +04:00
|
|
|
if (a > 4) {
|
|
|
|
int j = 4;
|
2010-05-18 19:58:33 +04:00
|
|
|
|
2012-07-14 02:21:29 +04:00
|
|
|
while (++j < 11 && e[j].base_val <= a) {}
|
2010-05-18 19:58:33 +04:00
|
|
|
|
2013-04-19 03:18:08 +04:00
|
|
|
t[i].token = --j;
|
2012-07-14 02:21:29 +04:00
|
|
|
eb |= (a - e[j].base_val) << 1;
|
|
|
|
} else
|
2013-04-19 03:18:08 +04:00
|
|
|
t[i].token = a;
|
2010-05-18 19:58:33 +04:00
|
|
|
|
2013-04-19 03:18:08 +04:00
|
|
|
t[i].extra = eb;
|
2012-07-14 02:21:29 +04:00
|
|
|
}
|
2010-05-18 19:58:33 +04:00
|
|
|
|
2012-07-14 02:21:29 +04:00
|
|
|
// initialize the cost for extra bits for all possible coefficient value.
|
|
|
|
{
|
|
|
|
int cost = 0;
|
2013-08-01 05:51:18 +04:00
|
|
|
const vp9_extra_bit *p = vp9_extra_bits + t[i].token;
|
2010-05-18 19:58:33 +04:00
|
|
|
|
2012-07-14 02:21:29 +04:00
|
|
|
if (p->base_val) {
|
2013-04-19 03:18:08 +04:00
|
|
|
const int extra = t[i].extra;
|
2013-04-19 22:14:33 +04:00
|
|
|
const int length = p->len;
|
2010-05-18 19:58:33 +04:00
|
|
|
|
2013-04-19 22:14:33 +04:00
|
|
|
if (length)
|
|
|
|
cost += treed_cost(p->tree, p->prob, extra >> 1, length);
|
2010-05-18 19:58:33 +04:00
|
|
|
|
2012-11-01 01:40:53 +04:00
|
|
|
cost += vp9_cost_bit(vp9_prob_half, extra & 1); /* sign */
|
2012-07-14 02:21:29 +04:00
|
|
|
dct_value_cost[i + DCT_MAX_VALUE] = cost;
|
|
|
|
}
|
2010-05-18 19:58:33 +04:00
|
|
|
|
|
|
|
}
|
|
|
|
|
2012-07-14 02:21:29 +04:00
|
|
|
} while (++i < DCT_MAX_VALUE);
|
|
|
|
|
2012-10-31 04:12:12 +04:00
|
|
|
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 19:58:33 +04:00
|
|
|
}
|
2012-11-28 03:51:06 +04:00
|
|
|
|
2013-04-29 22:57:48 +04:00
|
|
|
struct tokenize_b_args {
|
|
|
|
VP9_COMP *cpi;
|
|
|
|
MACROBLOCKD *xd;
|
|
|
|
TOKENEXTRA **tp;
|
|
|
|
TX_SIZE tx_size;
|
|
|
|
};
|
2013-06-06 17:07:09 +04:00
|
|
|
|
2013-07-30 01:54:31 +04:00
|
|
|
static void set_entropy_context_b(int plane, int block, BLOCK_SIZE_TYPE bsize,
|
|
|
|
int ss_txfrm_size, void *arg) {
|
|
|
|
struct tokenize_b_args* const args = arg;
|
|
|
|
TX_SIZE tx_size = ss_txfrm_size >> 1;
|
|
|
|
MACROBLOCKD *xd = args->xd;
|
|
|
|
const int bwl = b_width_log2(bsize);
|
|
|
|
const int off = block >> (2 * tx_size);
|
|
|
|
const int mod = bwl - tx_size - xd->plane[plane].subsampling_x;
|
|
|
|
const int aoff = (off & ((1 << mod) - 1)) << tx_size;
|
|
|
|
const int loff = (off >> mod) << tx_size;
|
|
|
|
ENTROPY_CONTEXT *A = xd->plane[plane].above_context + aoff;
|
|
|
|
ENTROPY_CONTEXT *L = xd->plane[plane].left_context + loff;
|
|
|
|
const int eob = xd->plane[plane].eobs[block];
|
2013-08-03 02:52:26 +04:00
|
|
|
const int tx_size_in_blocks = 1 << tx_size;
|
2013-07-30 01:54:31 +04:00
|
|
|
|
|
|
|
if (xd->mb_to_right_edge < 0 || xd->mb_to_bottom_edge < 0) {
|
2013-08-03 02:52:26 +04:00
|
|
|
set_contexts_on_border(xd, bsize, plane, tx_size_in_blocks, eob, aoff, loff,
|
|
|
|
A, L);
|
2013-07-30 01:54:31 +04:00
|
|
|
} else {
|
2013-08-03 02:52:26 +04:00
|
|
|
vpx_memset(A, eob > 0, sizeof(ENTROPY_CONTEXT) * tx_size_in_blocks);
|
|
|
|
vpx_memset(L, eob > 0, sizeof(ENTROPY_CONTEXT) * tx_size_in_blocks);
|
2013-07-30 01:54:31 +04:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-04-29 22:57:48 +04:00
|
|
|
static void tokenize_b(int plane, int block, BLOCK_SIZE_TYPE bsize,
|
|
|
|
int ss_txfrm_size, void *arg) {
|
|
|
|
struct tokenize_b_args* const args = arg;
|
|
|
|
VP9_COMP *cpi = args->cpi;
|
|
|
|
MACROBLOCKD *xd = args->xd;
|
|
|
|
TOKENEXTRA **tp = args->tp;
|
2013-08-03 02:52:26 +04:00
|
|
|
const TX_SIZE tx_size = ss_txfrm_size >> 1;
|
|
|
|
const int tx_size_in_blocks = 1 << tx_size;
|
2013-03-05 02:12:17 +04:00
|
|
|
MB_MODE_INFO *mbmi = &xd->mode_info_context->mbmi;
|
2012-08-03 04:03:14 +04:00
|
|
|
int pt; /* near block/prev token context index */
|
2013-03-28 21:42:23 +04:00
|
|
|
int c = 0, rc = 0;
|
2012-08-03 04:03:14 +04:00
|
|
|
TOKENEXTRA *t = *tp; /* store tokens starting here */
|
2013-04-29 22:57:48 +04:00
|
|
|
const int eob = xd->plane[plane].eobs[block];
|
2013-07-24 23:55:45 +04:00
|
|
|
const PLANE_TYPE type = xd->plane[plane].plane_type;
|
2013-04-29 22:57:48 +04:00
|
|
|
const int16_t *qcoeff_ptr = BLOCK_OFFSET(xd->plane[plane].qcoeff, block, 16);
|
2013-07-30 03:03:02 +04:00
|
|
|
const int bwl = b_width_log2(bsize);
|
2013-04-29 22:57:48 +04:00
|
|
|
const int off = block >> (2 * tx_size);
|
|
|
|
const int mod = bwl - tx_size - xd->plane[plane].subsampling_x;
|
2013-04-29 21:37:25 +04:00
|
|
|
const int aoff = (off & ((1 << mod) - 1)) << tx_size;
|
|
|
|
const int loff = (off >> mod) << tx_size;
|
2013-04-29 22:57:48 +04:00
|
|
|
ENTROPY_CONTEXT *A = xd->plane[plane].above_context + aoff;
|
|
|
|
ENTROPY_CONTEXT *L = xd->plane[plane].left_context + loff;
|
2013-07-01 21:40:00 +04:00
|
|
|
int seg_eob;
|
2013-03-05 02:12:17 +04:00
|
|
|
const int segment_id = mbmi->segment_id;
|
2013-07-01 22:36:07 +04:00
|
|
|
const int16_t *scan, *nb;
|
2012-12-08 04:09:59 +04:00
|
|
|
vp9_coeff_count *counts;
|
2013-05-17 17:40:25 +04:00
|
|
|
vp9_coeff_probs_model *coef_probs;
|
2013-06-07 00:44:34 +04:00
|
|
|
const int ref = mbmi->ref_frame[0] != INTRA_FRAME;
|
2013-04-29 21:37:25 +04:00
|
|
|
ENTROPY_CONTEXT above_ec, left_ec;
|
2013-03-27 03:46:09 +04:00
|
|
|
uint8_t token_cache[1024];
|
2013-06-25 04:56:06 +04:00
|
|
|
const uint8_t *band_translate;
|
2013-04-29 22:57:48 +04:00
|
|
|
assert((!type && !plane) || (type && plane));
|
2012-12-18 23:25:24 +04:00
|
|
|
|
2013-05-31 20:18:59 +04:00
|
|
|
counts = cpi->coef_counts[tx_size];
|
|
|
|
coef_probs = cpi->common.fc.coef_probs[tx_size];
|
2012-10-25 20:14:21 +04:00
|
|
|
switch (tx_size) {
|
|
|
|
default:
|
2013-07-24 23:55:45 +04:00
|
|
|
case TX_4X4:
|
2013-04-29 21:37:25 +04:00
|
|
|
above_ec = A[0] != 0;
|
|
|
|
left_ec = L[0] != 0;
|
2012-10-25 20:14:21 +04:00
|
|
|
seg_eob = 16;
|
2013-07-24 23:55:45 +04:00
|
|
|
scan = get_scan_4x4(get_tx_type_4x4(type, xd, block));
|
2013-05-09 21:47:58 +04:00
|
|
|
band_translate = vp9_coefband_trans_4x4;
|
2012-10-25 20:14:21 +04:00
|
|
|
break;
|
2013-07-24 23:55:45 +04:00
|
|
|
case TX_8X8:
|
2013-07-30 03:03:02 +04:00
|
|
|
above_ec = !!*(uint16_t *)A;
|
|
|
|
left_ec = !!*(uint16_t *)L;
|
2013-02-15 22:15:42 +04:00
|
|
|
seg_eob = 64;
|
2013-07-24 23:55:45 +04:00
|
|
|
scan = get_scan_8x8(get_tx_type_8x8(type, xd));
|
2013-05-09 21:47:58 +04:00
|
|
|
band_translate = vp9_coefband_trans_8x8plus;
|
2012-10-25 20:14:21 +04:00
|
|
|
break;
|
2013-07-24 23:55:45 +04:00
|
|
|
case TX_16X16:
|
2013-07-30 03:03:02 +04:00
|
|
|
above_ec = !!*(uint32_t *)A;
|
|
|
|
left_ec = !!*(uint32_t *)L;
|
2012-10-25 20:14:21 +04:00
|
|
|
seg_eob = 256;
|
2013-07-24 23:55:45 +04:00
|
|
|
scan = get_scan_16x16(get_tx_type_16x16(type, xd));
|
2013-05-09 21:47:58 +04:00
|
|
|
band_translate = vp9_coefband_trans_8x8plus;
|
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
2012-12-08 02:45:05 +04:00
|
|
|
break;
|
|
|
|
case TX_32X32:
|
2013-07-30 03:03:02 +04:00
|
|
|
above_ec = !!*(uint64_t *)A;
|
|
|
|
left_ec = !!*(uint64_t *)L;
|
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
2012-12-08 02:45:05 +04:00
|
|
|
seg_eob = 1024;
|
2013-05-22 17:09:55 +04:00
|
|
|
scan = vp9_default_scan_32x32;
|
2013-05-09 21:47:58 +04:00
|
|
|
band_translate = vp9_coefband_trans_8x8plus;
|
2012-10-25 20:14:21 +04:00
|
|
|
break;
|
2012-07-14 02:21:29 +04:00
|
|
|
}
|
|
|
|
|
2013-04-29 21:37:25 +04:00
|
|
|
pt = combine_entropy_contexts(above_ec, left_ec);
|
2013-07-01 21:40:00 +04:00
|
|
|
nb = vp9_get_coef_neighbors_handle(scan);
|
2012-12-07 00:40:57 +04:00
|
|
|
|
2013-07-10 23:29:43 +04:00
|
|
|
if (vp9_segfeature_active(&xd->seg, segment_id, SEG_LVL_SKIP))
|
2013-01-28 19:22:53 +04:00
|
|
|
seg_eob = 0;
|
2012-07-14 02:21:29 +04:00
|
|
|
|
2013-03-28 21:42:23 +04:00
|
|
|
c = 0;
|
2012-10-13 22:46:21 +04:00
|
|
|
do {
|
2013-05-09 21:47:58 +04:00
|
|
|
const int band = get_coef_band(band_translate, c);
|
2012-10-13 22:46:21 +04:00
|
|
|
int token;
|
2013-02-20 22:16:24 +04:00
|
|
|
int v = 0;
|
2013-03-28 21:42:23 +04:00
|
|
|
rc = scan[c];
|
|
|
|
if (c)
|
2013-07-01 21:40:00 +04:00
|
|
|
pt = get_coef_context(nb, token_cache, c);
|
2012-10-13 22:46:21 +04:00
|
|
|
if (c < eob) {
|
2013-02-20 22:16:24 +04:00
|
|
|
v = qcoeff_ptr[rc];
|
2012-10-25 20:14:21 +04:00
|
|
|
assert(-DCT_MAX_VALUE <= v && v < DCT_MAX_VALUE);
|
|
|
|
|
2013-04-19 03:18:08 +04:00
|
|
|
t->extra = vp9_dct_value_tokens_ptr[v].extra;
|
|
|
|
token = vp9_dct_value_tokens_ptr[v].token;
|
2012-10-25 20:14:21 +04:00
|
|
|
} else {
|
2013-04-22 21:58:49 +04:00
|
|
|
token = DCT_EOB_TOKEN;
|
2012-10-25 20:14:21 +04:00
|
|
|
}
|
2012-07-14 02:21:29 +04:00
|
|
|
|
2013-04-19 03:18:08 +04:00
|
|
|
t->token = token;
|
2013-03-28 21:42:23 +04:00
|
|
|
t->context_tree = coef_probs[type][ref][band][pt];
|
2013-05-17 17:40:25 +04:00
|
|
|
t->skip_eob_node = (c > 0) && (token_cache[scan[c - 1]] == 0);
|
2013-05-08 21:04:14 +04:00
|
|
|
|
2013-04-19 03:18:08 +04:00
|
|
|
assert(vp9_coef_encodings[t->token].len - t->skip_eob_node > 0);
|
2013-05-07 20:24:21 +04:00
|
|
|
|
2013-07-30 04:06:18 +04:00
|
|
|
++counts[type][ref][band][pt][token];
|
|
|
|
if (!t->skip_eob_node)
|
|
|
|
++cpi->common.counts.eob_branch[tx_size][type][ref][band][pt];
|
|
|
|
|
2013-07-30 03:03:02 +04:00
|
|
|
token_cache[rc] = vp9_pt_energy_class[token];
|
2012-07-24 01:58:07 +04:00
|
|
|
++t;
|
2013-04-22 21:58:49 +04:00
|
|
|
} while (c < eob && ++c < seg_eob);
|
2012-10-13 22:46:21 +04:00
|
|
|
|
|
|
|
*tp = t;
|
2013-06-06 17:07:09 +04:00
|
|
|
if (xd->mb_to_right_edge < 0 || xd->mb_to_bottom_edge < 0) {
|
2013-08-03 02:52:26 +04:00
|
|
|
set_contexts_on_border(xd, bsize, plane, tx_size_in_blocks, c, aoff, loff,
|
|
|
|
A, L);
|
2013-06-06 17:07:09 +04:00
|
|
|
} else {
|
2013-08-03 02:52:26 +04:00
|
|
|
vpx_memset(A, c > 0, sizeof(ENTROPY_CONTEXT) * tx_size_in_blocks);
|
|
|
|
vpx_memset(L, c > 0, sizeof(ENTROPY_CONTEXT) * tx_size_in_blocks);
|
2012-12-07 00:40:57 +04:00
|
|
|
}
|
2010-05-18 19:58:33 +04:00
|
|
|
}
|
|
|
|
|
2013-04-11 22:14:31 +04:00
|
|
|
struct is_skippable_args {
|
|
|
|
MACROBLOCKD *xd;
|
|
|
|
int *skippable;
|
|
|
|
};
|
|
|
|
static void is_skippable(int plane, int block,
|
2013-04-13 01:12:05 +04:00
|
|
|
BLOCK_SIZE_TYPE bsize, int ss_txfrm_size, void *argv) {
|
2013-04-11 22:14:31 +04:00
|
|
|
struct is_skippable_args *args = argv;
|
|
|
|
args->skippable[0] &= (!args->xd->plane[plane].eobs[block]);
|
2012-09-12 06:36:28 +04:00
|
|
|
}
|
|
|
|
|
2013-04-11 22:14:31 +04:00
|
|
|
int vp9_sb_is_skippable(MACROBLOCKD *xd, BLOCK_SIZE_TYPE bsize) {
|
|
|
|
int result = 1;
|
|
|
|
struct is_skippable_args args = {xd, &result};
|
2013-04-13 01:12:05 +04:00
|
|
|
foreach_transformed_block(xd, bsize, is_skippable, &args);
|
2013-04-11 22:14:31 +04:00
|
|
|
return result;
|
2013-03-05 02:12:17 +04:00
|
|
|
}
|
|
|
|
|
2013-04-11 22:14:31 +04:00
|
|
|
int vp9_sby_is_skippable(MACROBLOCKD *xd, BLOCK_SIZE_TYPE bsize) {
|
|
|
|
int result = 1;
|
|
|
|
struct is_skippable_args args = {xd, &result};
|
2013-06-25 04:56:06 +04:00
|
|
|
foreach_transformed_block_in_plane(xd, bsize, 0, is_skippable, &args);
|
2013-04-11 22:14:31 +04:00
|
|
|
return result;
|
2013-03-05 02:12:17 +04:00
|
|
|
}
|
|
|
|
|
2013-04-11 22:14:31 +04:00
|
|
|
int vp9_sbuv_is_skippable(MACROBLOCKD *xd, BLOCK_SIZE_TYPE bsize) {
|
|
|
|
int result = 1;
|
|
|
|
struct is_skippable_args args = {xd, &result};
|
2013-04-13 01:12:05 +04:00
|
|
|
foreach_transformed_block_uv(xd, bsize, is_skippable, &args);
|
2013-04-11 22:14:31 +04:00
|
|
|
return result;
|
2013-03-05 02:12:17 +04:00
|
|
|
}
|
|
|
|
|
2013-07-12 05:39:10 +04:00
|
|
|
void vp9_tokenize_sb(VP9_COMP *cpi, TOKENEXTRA **t, int dry_run,
|
|
|
|
BLOCK_SIZE_TYPE bsize) {
|
|
|
|
VP9_COMMON *const cm = &cpi->common;
|
|
|
|
MACROBLOCKD *const xd = &cpi->mb.e_mbd;
|
|
|
|
MB_MODE_INFO *const mbmi = &xd->mode_info_context->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
2012-12-08 02:45:05 +04:00
|
|
|
TOKENEXTRA *t_backup = *t;
|
2013-07-12 05:39:10 +04:00
|
|
|
const int mb_skip_context = vp9_get_pred_context_mbskip(xd);
|
2013-07-10 23:29:43 +04:00
|
|
|
const int skip_inc = !vp9_segfeature_active(&xd->seg, mbmi->segment_id,
|
|
|
|
SEG_LVL_SKIP);
|
2013-07-30 04:06:18 +04:00
|
|
|
struct tokenize_b_args arg = {cpi, xd, t, mbmi->txfm_size};
|
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
2012-12-08 02:45:05 +04:00
|
|
|
|
2013-04-11 22:14:31 +04:00
|
|
|
mbmi->mb_skip_coeff = vp9_sb_is_skippable(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
2012-12-08 02:45:05 +04:00
|
|
|
if (mbmi->mb_skip_coeff) {
|
|
|
|
if (!dry_run)
|
2013-07-24 04:02:08 +04:00
|
|
|
cm->counts.mbskip[mb_skip_context][1] += skip_inc;
|
2013-04-09 21:54:19 +04:00
|
|
|
vp9_reset_sb_tokens_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
2012-12-08 02:45:05 +04:00
|
|
|
if (dry_run)
|
|
|
|
*t = t_backup;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2013-07-30 01:54:31 +04:00
|
|
|
if (!dry_run) {
|
2013-07-24 04:02:08 +04:00
|
|
|
cm->counts.mbskip[mb_skip_context][0] += skip_inc;
|
2013-07-30 01:54:31 +04:00
|
|
|
foreach_transformed_block(xd, bsize, tokenize_b, &arg);
|
|
|
|
} else {
|
|
|
|
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
2012-12-08 02:45:05 +04:00
|
|
|
*t = t_backup;
|
2013-07-30 01:54:31 +04:00
|
|
|
}
|
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
2012-12-08 02:45:05 +04:00
|
|
|
}
|
|
|
|
|
2010-05-18 19:58:33 +04:00
|
|
|
#ifdef ENTROPY_STATS
|
2012-07-14 02:21:29 +04:00
|
|
|
void init_context_counters(void) {
|
|
|
|
FILE *f = fopen("context.bin", "rb");
|
|
|
|
if (!f) {
|
2013-05-31 20:18:59 +04:00
|
|
|
vp9_zero(context_counters);
|
2012-07-14 02:21:29 +04:00
|
|
|
} else {
|
2013-05-31 20:18:59 +04:00
|
|
|
fread(context_counters, sizeof(context_counters), 1, f);
|
2012-07-14 02:21:29 +04:00
|
|
|
fclose(f);
|
|
|
|
}
|
|
|
|
|
|
|
|
f = fopen("treeupdate.bin", "rb");
|
|
|
|
if (!f) {
|
2013-05-31 20:18:59 +04:00
|
|
|
vpx_memset(tree_update_hist, 0, sizeof(tree_update_hist));
|
2012-07-14 02:21:29 +04:00
|
|
|
} else {
|
2013-05-31 20:18:59 +04:00
|
|
|
fread(tree_update_hist, sizeof(tree_update_hist), 1, f);
|
2012-07-14 02:21:29 +04:00
|
|
|
fclose(f);
|
|
|
|
}
|
2010-05-18 19:58:33 +04:00
|
|
|
}
|
|
|
|
|
2012-12-08 04:09:59 +04:00
|
|
|
static void print_counter(FILE *f, vp9_coeff_accum *context_counters,
|
|
|
|
int block_types, const char *header) {
|
2013-02-20 01:36:38 +04:00
|
|
|
int type, ref, band, pt, t;
|
2010-05-18 19:58:33 +04:00
|
|
|
|
2012-12-08 04:09:59 +04:00
|
|
|
fprintf(f, "static const vp9_coeff_count %s = {\n", header);
|
2010-05-18 19:58:33 +04:00
|
|
|
|
2012-12-12 22:25:58 +04:00
|
|
|
#define Comma(X) (X ? "," : "")
|
2012-07-14 02:21:29 +04:00
|
|
|
type = 0;
|
|
|
|
do {
|
2013-02-20 01:36:38 +04:00
|
|
|
ref = 0;
|
2012-07-14 02:21:29 +04:00
|
|
|
fprintf(f, "%s\n { /* block Type %d */", Comma(type), type);
|
|
|
|
do {
|
2013-02-20 01:36:38 +04:00
|
|
|
fprintf(f, "%s\n { /* %s */", Comma(type), ref ? "Inter" : "Intra");
|
|
|
|
band = 0;
|
2012-07-14 02:21:29 +04:00
|
|
|
do {
|
2013-02-20 01:36:38 +04:00
|
|
|
fprintf(f, "%s\n { /* Coeff Band %d */", Comma(band), band);
|
|
|
|
pt = 0;
|
2012-07-14 02:21:29 +04:00
|
|
|
do {
|
2013-02-20 01:36:38 +04:00
|
|
|
fprintf(f, "%s\n {", Comma(pt));
|
|
|
|
|
|
|
|
t = 0;
|
|
|
|
do {
|
|
|
|
const int64_t x = context_counters[type][ref][band][pt][t];
|
|
|
|
const int y = (int) x;
|
|
|
|
|
|
|
|
assert(x == (int64_t) y); /* no overflow handling yet */
|
|
|
|
fprintf(f, "%s %d", Comma(t), y);
|
2013-03-27 03:46:09 +04:00
|
|
|
} while (++t < 1 + MAX_ENTROPY_TOKENS);
|
2013-02-20 01:36:38 +04:00
|
|
|
fprintf(f, "}");
|
|
|
|
} while (++pt < PREV_COEF_CONTEXTS);
|
|
|
|
fprintf(f, "\n }");
|
|
|
|
} while (++band < COEF_BANDS);
|
2012-07-14 02:21:29 +04:00
|
|
|
fprintf(f, "\n }");
|
2013-02-20 01:36:38 +04:00
|
|
|
} while (++ref < REF_TYPES);
|
2012-07-14 02:21:29 +04:00
|
|
|
fprintf(f, "\n }");
|
2012-12-08 04:09:59 +04:00
|
|
|
} while (++type < block_types);
|
2012-08-03 04:03:14 +04:00
|
|
|
fprintf(f, "\n};\n");
|
2012-12-08 04:09:59 +04:00
|
|
|
}
|
2012-03-22 02:22:21 +04:00
|
|
|
|
2012-12-08 04:09:59 +04:00
|
|
|
static void print_probs(FILE *f, vp9_coeff_accum *context_counters,
|
|
|
|
int block_types, const char *header) {
|
2013-02-20 01:36:38 +04:00
|
|
|
int type, ref, band, pt, t;
|
2012-08-03 04:03:14 +04:00
|
|
|
|
2012-12-12 22:25:58 +04:00
|
|
|
fprintf(f, "static const vp9_coeff_probs %s = {", header);
|
2012-03-22 02:22:21 +04:00
|
|
|
|
2012-07-14 02:21:29 +04:00
|
|
|
type = 0;
|
2012-12-12 22:25:58 +04:00
|
|
|
#define Newline(x, spaces) (x ? " " : "\n" spaces)
|
2012-07-14 02:21:29 +04:00
|
|
|
do {
|
2012-12-12 22:25:58 +04:00
|
|
|
fprintf(f, "%s%s{ /* block Type %d */",
|
|
|
|
Comma(type), Newline(type, " "), type);
|
2013-02-20 01:36:38 +04:00
|
|
|
ref = 0;
|
2012-07-14 02:21:29 +04:00
|
|
|
do {
|
2013-02-20 01:36:38 +04:00
|
|
|
fprintf(f, "%s%s{ /* %s */",
|
|
|
|
Comma(band), Newline(band, " "), ref ? "Inter" : "Intra");
|
|
|
|
band = 0;
|
2012-07-14 02:21:29 +04:00
|
|
|
do {
|
2013-02-20 01:36:38 +04:00
|
|
|
fprintf(f, "%s%s{ /* Coeff Band %d */",
|
|
|
|
Comma(band), Newline(band, " "), band);
|
|
|
|
pt = 0;
|
2012-07-14 02:21:29 +04:00
|
|
|
do {
|
2013-02-20 01:36:38 +04:00
|
|
|
unsigned int branch_ct[ENTROPY_NODES][2];
|
2013-03-27 03:46:09 +04:00
|
|
|
unsigned int coef_counts[MAX_ENTROPY_TOKENS + 1];
|
2013-02-20 01:36:38 +04:00
|
|
|
vp9_prob coef_probs[ENTROPY_NODES];
|
|
|
|
|
2013-03-27 03:46:09 +04:00
|
|
|
if (pt >= 3 && band == 0)
|
|
|
|
break;
|
|
|
|
for (t = 0; t < MAX_ENTROPY_TOKENS + 1; ++t)
|
2013-02-20 01:36:38 +04:00
|
|
|
coef_counts[t] = context_counters[type][ref][band][pt][t];
|
2013-03-19 01:39:36 +04:00
|
|
|
vp9_tree_probs_from_distribution(vp9_coef_tree, coef_probs,
|
|
|
|
branch_ct, coef_counts, 0);
|
2013-03-27 03:46:09 +04:00
|
|
|
branch_ct[0][1] = coef_counts[MAX_ENTROPY_TOKENS] - branch_ct[0][0];
|
|
|
|
coef_probs[0] = get_binary_prob(branch_ct[0][0], branch_ct[0][1]);
|
2013-02-20 01:36:38 +04:00
|
|
|
fprintf(f, "%s\n {", Comma(pt));
|
|
|
|
|
|
|
|
t = 0;
|
|
|
|
do {
|
|
|
|
fprintf(f, "%s %3d", Comma(t), coef_probs[t]);
|
|
|
|
} while (++t < ENTROPY_NODES);
|
|
|
|
|
|
|
|
fprintf(f, " }");
|
|
|
|
} while (++pt < PREV_COEF_CONTEXTS);
|
|
|
|
fprintf(f, "\n }");
|
|
|
|
} while (++band < COEF_BANDS);
|
2012-07-14 02:21:29 +04:00
|
|
|
fprintf(f, "\n }");
|
2013-02-20 01:36:38 +04:00
|
|
|
} while (++ref < REF_TYPES);
|
2012-07-14 02:21:29 +04:00
|
|
|
fprintf(f, "\n }");
|
2012-12-08 04:09:59 +04:00
|
|
|
} while (++type < block_types);
|
2012-07-14 02:21:29 +04:00
|
|
|
fprintf(f, "\n};\n");
|
2012-12-08 04:09:59 +04:00
|
|
|
}
|
2012-03-22 02:22:21 +04:00
|
|
|
|
2012-12-08 04:09:59 +04:00
|
|
|
void print_context_counters() {
|
|
|
|
FILE *f = fopen("vp9_context.c", "w");
|
2012-03-22 02:22:21 +04:00
|
|
|
|
2012-12-08 04:09:59 +04:00
|
|
|
fprintf(f, "#include \"vp9_entropy.h\"\n");
|
|
|
|
fprintf(f, "\n/* *** GENERATED FILE: DO NOT EDIT *** */\n\n");
|
2012-07-14 02:21:29 +04:00
|
|
|
|
2012-12-08 04:09:59 +04:00
|
|
|
/* print counts */
|
2013-05-31 20:18:59 +04:00
|
|
|
print_counter(f, context_counters[TX_4X4], BLOCK_TYPES,
|
2013-03-05 02:12:17 +04:00
|
|
|
"vp9_default_coef_counts_4x4[BLOCK_TYPES]");
|
2013-05-31 20:18:59 +04:00
|
|
|
print_counter(f, context_counters[TX_8X8], BLOCK_TYPES,
|
2013-03-05 02:12:17 +04:00
|
|
|
"vp9_default_coef_counts_8x8[BLOCK_TYPES]");
|
2013-05-31 20:18:59 +04:00
|
|
|
print_counter(f, context_counters[TX_16X16], BLOCK_TYPES,
|
2013-03-05 02:12:17 +04:00
|
|
|
"vp9_default_coef_counts_16x16[BLOCK_TYPES]");
|
2013-05-31 20:18:59 +04:00
|
|
|
print_counter(f, context_counters[TX_32X32], BLOCK_TYPES,
|
2013-03-05 02:12:17 +04:00
|
|
|
"vp9_default_coef_counts_32x32[BLOCK_TYPES]");
|
2012-08-03 04:03:14 +04:00
|
|
|
|
2012-12-08 04:09:59 +04:00
|
|
|
/* print coefficient probabilities */
|
2013-05-31 20:18:59 +04:00
|
|
|
print_probs(f, context_counters[TX_4X4], BLOCK_TYPES,
|
2013-03-05 02:12:17 +04:00
|
|
|
"default_coef_probs_4x4[BLOCK_TYPES]");
|
2013-05-31 20:18:59 +04:00
|
|
|
print_probs(f, context_counters[TX_8X8], BLOCK_TYPES,
|
2013-03-05 02:12:17 +04:00
|
|
|
"default_coef_probs_8x8[BLOCK_TYPES]");
|
2013-05-31 20:18:59 +04:00
|
|
|
print_probs(f, context_counters[TX_16X16], BLOCK_TYPES,
|
2013-03-05 02:12:17 +04:00
|
|
|
"default_coef_probs_16x16[BLOCK_TYPES]");
|
2013-05-31 20:18:59 +04:00
|
|
|
print_probs(f, context_counters[TX_32X32], BLOCK_TYPES,
|
2013-03-05 02:12:17 +04:00
|
|
|
"default_coef_probs_32x32[BLOCK_TYPES]");
|
2012-08-03 04:03:14 +04:00
|
|
|
|
2012-07-14 02:21:29 +04:00
|
|
|
fclose(f);
|
|
|
|
|
|
|
|
f = fopen("context.bin", "wb");
|
2013-05-31 20:18:59 +04:00
|
|
|
fwrite(context_counters, sizeof(context_counters), 1, f);
|
2012-07-14 02:21:29 +04:00
|
|
|
fclose(f);
|
2010-05-18 19:58:33 +04:00
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2012-10-30 23:58:42 +04:00
|
|
|
void vp9_tokenize_initialize() {
|
2012-07-14 02:21:29 +04:00
|
|
|
fill_value_tokens();
|
2010-05-18 19:58:33 +04:00
|
|
|
}
|