vpx/vp9/decoder/vp9_detokenize.c

304 lines
9.7 KiB
C
Raw Normal View History

2010-05-18 17:58:33 +02:00
/*
* Copyright (c) 2010 The WebM project authors. All Rights Reserved.
2010-05-18 17:58:33 +02:00
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
2010-05-18 17:58:33 +02:00
*/
#include "vpx_mem/vpx_mem.h"
#include "vpx_ports/mem.h"
2010-05-18 17:58:33 +02:00
#include "vp9/common/vp9_blockd.h"
#include "vp9/common/vp9_common.h"
#include "vp9/common/vp9_seg_common.h"
2011-10-05 12:26:00 +02:00
#include "vp9/decoder/vp9_detokenize.h"
#include "vp9/decoder/vp9_onyxd_int.h"
2010-05-18 17:58:33 +02:00
#define EOB_CONTEXT_NODE 0
#define ZERO_CONTEXT_NODE 1
#define ONE_CONTEXT_NODE 2
#define LOW_VAL_CONTEXT_NODE 3
#define TWO_CONTEXT_NODE 4
#define THREE_CONTEXT_NODE 5
#define HIGH_LOW_CONTEXT_NODE 6
#define CAT_ONE_CONTEXT_NODE 7
#define CAT_THREEFOUR_CONTEXT_NODE 8
#define CAT_THREE_CONTEXT_NODE 9
#define CAT_FIVE_CONTEXT_NODE 10
#define CAT1_MIN_VAL 5
#define CAT2_MIN_VAL 7
#define CAT3_MIN_VAL 11
#define CAT4_MIN_VAL 19
#define CAT5_MIN_VAL 35
#define CAT6_MIN_VAL 67
#define CAT1_PROB0 159
#define CAT2_PROB0 145
#define CAT2_PROB1 165
#define CAT3_PROB0 140
#define CAT3_PROB1 148
#define CAT3_PROB2 173
#define CAT4_PROB0 135
#define CAT4_PROB1 140
#define CAT4_PROB2 155
#define CAT4_PROB3 176
#define CAT5_PROB0 130
#define CAT5_PROB1 134
#define CAT5_PROB2 141
#define CAT5_PROB3 157
#define CAT5_PROB4 180
static const vp9_prob cat6_prob[15] = {
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
254, 254, 254, 252, 249, 243, 230, 196, 177, 153, 140, 133, 130, 129, 0
};
2010-05-18 17:58:33 +02:00
DECLARE_ALIGNED(16, extern const uint8_t,
vp9_pt_energy_class[MAX_ENTROPY_TOKENS]);
#define INCREMENT_COUNT(token) \
do { \
coef_counts[type][ref][band][pt] \
[token >= TWO_TOKEN ? \
(token == DCT_EOB_TOKEN ? DCT_EOB_MODEL_TOKEN : TWO_TOKEN) : \
token]++; \
token_cache[scan[c]] = vp9_pt_energy_class[token]; \
} while (0)
#define WRITE_COEF_CONTINUE(val, token) \
{ \
qcoeff_ptr[scan[c]] = vp9_read_and_apply_sign(r, val) * \
dq[c > 0] / (1 + (txfm_size == TX_32X32)); \
INCREMENT_COUNT(token); \
c++; \
continue; \
}
2010-05-18 17:58:33 +02:00
#define ADJUST_COEF(prob, bits_count) \
do { \
if (vp9_read(r, prob)) \
val += 1 << bits_count; \
} while (0);
2010-05-18 17:58:33 +02:00
static int decode_coefs(VP9_COMMON *cm, const MACROBLOCKD *xd,
vp9_reader *r, int block_idx,
PLANE_TYPE type, int seg_eob, int16_t *qcoeff_ptr,
TX_SIZE txfm_size, const int16_t *dq,
ENTROPY_CONTEXT *A, ENTROPY_CONTEXT *L) {
FRAME_CONTEXT *const fc = &cm->fc;
FRAME_COUNTS *const counts = &cm->counts;
ENTROPY_CONTEXT above_ec, left_ec;
int pt, c = 0;
int band;
vp9_prob (*coef_probs)[PREV_COEF_CONTEXTS][UNCONSTRAINED_NODES];
vp9_prob coef_probs_full[COEF_BANDS][PREV_COEF_CONTEXTS][ENTROPY_NODES];
uint8_t load_map[COEF_BANDS][PREV_COEF_CONTEXTS] = {
{0, 0, 0, 0, 0, 0},
{0, 0, 0, 0, 0, 0},
{0, 0, 0, 0, 0, 0},
{0, 0, 0, 0, 0, 0},
{0, 0, 0, 0, 0, 0},
{0, 0, 0, 0, 0, 0},
};
vp9_prob *prob;
vp9_coeff_count_model *coef_counts;
const int ref = xd->mode_info_context->mbmi.ref_frame[0] != INTRA_FRAME;
const int16_t *scan, *nb;
uint8_t token_cache[1024];
const uint8_t * band_translate;
coef_probs = fc->coef_probs[txfm_size][type][ref];
coef_counts = counts->coef[txfm_size];
switch (txfm_size) {
default:
case TX_4X4: {
scan = get_scan_4x4(get_tx_type_4x4(type, xd, block_idx));
above_ec = A[0] != 0;
left_ec = L[0] != 0;
band_translate = vp9_coefband_trans_4x4;
break;
}
case TX_8X8: {
scan = get_scan_8x8(get_tx_type_8x8(type, xd));
above_ec = (A[0] + A[1]) != 0;
left_ec = (L[0] + L[1]) != 0;
band_translate = vp9_coefband_trans_8x8plus;
break;
}
case TX_16X16: {
scan = get_scan_16x16(get_tx_type_16x16(type, xd));
above_ec = (A[0] + A[1] + A[2] + A[3]) != 0;
left_ec = (L[0] + L[1] + L[2] + L[3]) != 0;
band_translate = vp9_coefband_trans_8x8plus;
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:
scan = vp9_default_scan_32x32;
above_ec = (A[0] + A[1] + A[2] + A[3] + A[4] + A[5] + A[6] + A[7]) != 0;
left_ec = (L[0] + L[1] + L[2] + L[3] + L[4] + L[5] + L[6] + L[7]) != 0;
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
2012-12-07 23:45:05 +01:00
break;
}
pt = combine_entropy_contexts(above_ec, left_ec);
nb = vp9_get_coef_neighbors_handle(scan);
while (1) {
int val;
const uint8_t *cat6 = cat6_prob;
if (c >= seg_eob)
break;
if (c)
pt = get_coef_context(nb, token_cache, c);
band = get_coef_band(band_translate, c);
prob = coef_probs[band][pt];
counts->eob_branch[txfm_size][type][ref][band][pt]++;
if (!vp9_read(r, prob[EOB_CONTEXT_NODE]))
break;
SKIP_START:
if (c >= seg_eob)
break;
if (c)
pt = get_coef_context(nb, token_cache, c);
band = get_coef_band(band_translate, c);
prob = coef_probs[band][pt];
if (!vp9_read(r, prob[ZERO_CONTEXT_NODE])) {
INCREMENT_COUNT(ZERO_TOKEN);
++c;
goto SKIP_START;
}
// ONE_CONTEXT_NODE_0_
if (!vp9_read(r, prob[ONE_CONTEXT_NODE])) {
WRITE_COEF_CONTINUE(1, ONE_TOKEN);
}
// Load full probabilities if not already loaded
if (!load_map[band][pt]) {
vp9_model_to_full_probs(coef_probs[band][pt],
coef_probs_full[band][pt]);
load_map[band][pt] = 1;
}
prob = coef_probs_full[band][pt];
// LOW_VAL_CONTEXT_NODE_0_
if (!vp9_read(r, prob[LOW_VAL_CONTEXT_NODE])) {
if (!vp9_read(r, prob[TWO_CONTEXT_NODE])) {
WRITE_COEF_CONTINUE(2, TWO_TOKEN);
}
if (!vp9_read(r, prob[THREE_CONTEXT_NODE])) {
WRITE_COEF_CONTINUE(3, THREE_TOKEN);
}
WRITE_COEF_CONTINUE(4, FOUR_TOKEN);
}
// HIGH_LOW_CONTEXT_NODE_0_
if (!vp9_read(r, prob[HIGH_LOW_CONTEXT_NODE])) {
if (!vp9_read(r, prob[CAT_ONE_CONTEXT_NODE])) {
val = CAT1_MIN_VAL;
ADJUST_COEF(CAT1_PROB0, 0);
WRITE_COEF_CONTINUE(val, DCT_VAL_CATEGORY1);
}
val = CAT2_MIN_VAL;
ADJUST_COEF(CAT2_PROB1, 1);
ADJUST_COEF(CAT2_PROB0, 0);
WRITE_COEF_CONTINUE(val, DCT_VAL_CATEGORY2);
}
// CAT_THREEFOUR_CONTEXT_NODE_0_
if (!vp9_read(r, prob[CAT_THREEFOUR_CONTEXT_NODE])) {
if (!vp9_read(r, prob[CAT_THREE_CONTEXT_NODE])) {
val = CAT3_MIN_VAL;
ADJUST_COEF(CAT3_PROB2, 2);
ADJUST_COEF(CAT3_PROB1, 1);
ADJUST_COEF(CAT3_PROB0, 0);
WRITE_COEF_CONTINUE(val, DCT_VAL_CATEGORY3);
}
val = CAT4_MIN_VAL;
ADJUST_COEF(CAT4_PROB3, 3);
ADJUST_COEF(CAT4_PROB2, 2);
ADJUST_COEF(CAT4_PROB1, 1);
ADJUST_COEF(CAT4_PROB0, 0);
WRITE_COEF_CONTINUE(val, DCT_VAL_CATEGORY4);
}
// CAT_FIVE_CONTEXT_NODE_0_:
if (!vp9_read(r, prob[CAT_FIVE_CONTEXT_NODE])) {
val = CAT5_MIN_VAL;
ADJUST_COEF(CAT5_PROB4, 4);
ADJUST_COEF(CAT5_PROB3, 3);
ADJUST_COEF(CAT5_PROB2, 2);
ADJUST_COEF(CAT5_PROB1, 1);
ADJUST_COEF(CAT5_PROB0, 0);
WRITE_COEF_CONTINUE(val, DCT_VAL_CATEGORY5);
}
val = 0;
while (*cat6) {
val = (val << 1) | vp9_read(r, *cat6++);
}
val += CAT6_MIN_VAL;
WRITE_COEF_CONTINUE(val, DCT_VAL_CATEGORY6);
}
if (c < seg_eob)
coef_counts[type][ref][band][pt][DCT_EOB_MODEL_TOKEN]++;
return c;
}
2010-05-18 17:58:33 +02:00
static int get_eob(struct segmentation *seg, int segment_id, int eob_max) {
return vp9_segfeature_active(seg, segment_id, SEG_LVL_SKIP) ? 0 : eob_max;
}
struct decode_block_args {
VP9D_COMP *pbi;
vp9_reader *r;
int *eobtotal;
};
static void decode_block(int plane, int block,
BLOCK_SIZE_TYPE bsize,
int ss_txfrm_size,
void *argv) {
const struct decode_block_args* const arg = argv;
const int bw = b_width_log2(bsize);
// find the maximum eob for this transform size, adjusted by segment
MACROBLOCKD *xd = &arg->pbi->mb;
struct macroblockd_plane* pd = &xd->plane[plane];
const int segment_id = xd->mode_info_context->mbmi.segment_id;
const TX_SIZE ss_tx_size = ss_txfrm_size / 2;
const int seg_eob = get_eob(&xd->seg, segment_id, 16 << ss_txfrm_size);
const int off = block >> ss_txfrm_size;
const int mod = bw - ss_tx_size - pd->subsampling_x;
const int aoff = (off & ((1 << mod) - 1)) << ss_tx_size;
const int loff = (off >> mod) << ss_tx_size;
ENTROPY_CONTEXT *A = pd->above_context + aoff;
ENTROPY_CONTEXT *L = pd->left_context + loff;
const int eob = decode_coefs(&arg->pbi->common, xd, arg->r, block,
pd->plane_type, seg_eob,
BLOCK_OFFSET(pd->qcoeff, block, 16),
ss_tx_size, pd->dequant, A, L);
if (xd->mb_to_right_edge < 0 || xd->mb_to_bottom_edge < 0) {
set_contexts_on_border(xd, bsize, plane, ss_tx_size, eob, aoff, loff, A, L);
} else {
int pt;
for (pt = 0; pt < (1 << ss_tx_size); pt++)
A[pt] = L[pt] = eob > 0;
}
pd->eobs[block] = eob;
*arg->eobtotal += eob;
}
int vp9_decode_tokens(VP9D_COMP *pbi, vp9_reader *r, BLOCK_SIZE_TYPE bsize) {
int eobtotal = 0;
struct decode_block_args args = {pbi, r, &eobtotal};
foreach_transformed_block(&pbi->mb, bsize, decode_block, &args);
return eobtotal;
}