blob: 9574c097aaf0b2de441152d8112b72117e11712c [file] [log] [blame]
/*
* Copyright (c) 2010 The WebM project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include <limits.h>
#include <math.h>
#include <stdio.h>
#include "./vp9_rtcd.h"
#include "./vpx_dsp_rtcd.h"
#include "./vpx_config.h"
#include "vpx_dsp/vpx_dsp_common.h"
#include "vpx_ports/mem.h"
#include "vpx_ports/vpx_timer.h"
#include "vpx_ports/system_state.h"
#include "vp9/common/vp9_common.h"
#include "vp9/common/vp9_entropy.h"
#include "vp9/common/vp9_entropymode.h"
#include "vp9/common/vp9_idct.h"
#include "vp9/common/vp9_mvref_common.h"
#include "vp9/common/vp9_pred_common.h"
#include "vp9/common/vp9_quant_common.h"
#include "vp9/common/vp9_reconintra.h"
#include "vp9/common/vp9_reconinter.h"
#include "vp9/common/vp9_seg_common.h"
#include "vp9/common/vp9_tile_common.h"
#include "vp9/encoder/vp9_aq_360.h"
#include "vp9/encoder/vp9_aq_complexity.h"
#include "vp9/encoder/vp9_aq_cyclicrefresh.h"
#include "vp9/encoder/vp9_aq_variance.h"
#include "vp9/encoder/vp9_encodeframe.h"
#include "vp9/encoder/vp9_encodemb.h"
#include "vp9/encoder/vp9_encodemv.h"
#include "vp9/encoder/vp9_ethread.h"
#include "vp9/encoder/vp9_extend.h"
#include "vp9/encoder/vp9_pickmode.h"
#include "vp9/encoder/vp9_rd.h"
#include "vp9/encoder/vp9_rdopt.h"
#include "vp9/encoder/vp9_segmentation.h"
#include "vp9/encoder/vp9_tokenize.h"
static void encode_superblock(VP9_COMP *cpi, ThreadData *td, TOKENEXTRA **t,
int output_enabled, int mi_row, int mi_col,
BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx);
// Machine learning-based early termination parameters.
static const double train_mean[24] = {
303501.697372, 3042630.372158, 24.694696, 1.392182,
689.413511, 162.027012, 1.478213, 0.0,
135382.260230, 912738.513263, 28.845217, 1.515230,
544.158492, 131.807995, 1.436863, 0.0,
43682.377587, 208131.711766, 28.084737, 1.356677,
138.254122, 119.522553, 1.252322, 0.0
};
static const double train_stdm[24] = {
673689.212982, 5996652.516628, 0.024449, 1.989792,
985.880847, 0.014638, 2.001898, 0.0,
208798.775332, 1812548.443284, 0.018693, 1.838009,
396.986910, 0.015657, 1.332541, 0.0,
55888.847031, 448587.962714, 0.017900, 1.904776,
98.652832, 0.016598, 1.320992, 0.0
};
// Error tolerance: 0.01%-0.0.05%-0.1%
static const double classifiers[24] = {
0.111736, 0.289977, 0.042219, 0.204765, 0.120410, -0.143863,
0.282376, 0.847811, 0.637161, 0.131570, 0.018636, 0.202134,
0.112797, 0.028162, 0.182450, 1.124367, 0.386133, 0.083700,
0.050028, 0.150873, 0.061119, 0.109318, 0.127255, 0.625211
};
// This is used as a reference when computing the source variance for the
// purpose of activity masking.
// Eventually this should be replaced by custom no-reference routines,
// which will be faster.
static const uint8_t VP9_VAR_OFFS[64] = {
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128
};
#if CONFIG_VP9_HIGHBITDEPTH
static const uint16_t VP9_HIGH_VAR_OFFS_8[64] = {
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128
};
static const uint16_t VP9_HIGH_VAR_OFFS_10[64] = {
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4
};
static const uint16_t VP9_HIGH_VAR_OFFS_12[64] = {
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
128 * 16
};
#endif // CONFIG_VP9_HIGHBITDEPTH
unsigned int vp9_get_sby_perpixel_variance(VP9_COMP *cpi,
const struct buf_2d *ref,
BLOCK_SIZE bs) {
unsigned int sse;
const unsigned int var =
cpi->fn_ptr[bs].vf(ref->buf, ref->stride, VP9_VAR_OFFS, 0, &sse);
return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
}
#if CONFIG_VP9_HIGHBITDEPTH
unsigned int vp9_high_get_sby_perpixel_variance(VP9_COMP *cpi,
const struct buf_2d *ref,
BLOCK_SIZE bs, int bd) {
unsigned int var, sse;
switch (bd) {
case 10:
var =
cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10), 0, &sse);
break;
case 12:
var =
cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12), 0, &sse);
break;
case 8:
default:
var =
cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8), 0, &sse);
break;
}
return (unsigned int)ROUND64_POWER_OF_TWO((int64_t)var,
num_pels_log2_lookup[bs]);
}
#endif // CONFIG_VP9_HIGHBITDEPTH
static unsigned int get_sby_perpixel_diff_variance(VP9_COMP *cpi,
const struct buf_2d *ref,
int mi_row, int mi_col,
BLOCK_SIZE bs) {
unsigned int sse, var;
uint8_t *last_y;
const YV12_BUFFER_CONFIG *last = get_ref_frame_buffer(cpi, LAST_FRAME);
assert(last != NULL);
last_y =
&last->y_buffer[mi_row * MI_SIZE * last->y_stride + mi_col * MI_SIZE];
var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride, last_y, last->y_stride, &sse);
return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
}
static BLOCK_SIZE get_rd_var_based_fixed_partition(VP9_COMP *cpi, MACROBLOCK *x,
int mi_row, int mi_col) {
unsigned int var = get_sby_perpixel_diff_variance(
cpi, &x->plane[0].src, mi_row, mi_col, BLOCK_64X64);
if (var < 8)
return BLOCK_64X64;
else if (var < 128)
return BLOCK_32X32;
else if (var < 2048)
return BLOCK_16X16;
else
return BLOCK_8X8;
}
// Lighter version of set_offsets that only sets the mode info
// pointers.
static INLINE void set_mode_info_offsets(VP9_COMMON *const cm,
MACROBLOCK *const x,
MACROBLOCKD *const xd, int mi_row,
int mi_col) {
const int idx_str = xd->mi_stride * mi_row + mi_col;
xd->mi = cm->mi_grid_visible + idx_str;
xd->mi[0] = cm->mi + idx_str;
x->mbmi_ext = x->mbmi_ext_base + (mi_row * cm->mi_cols + mi_col);
}
static void set_offsets(VP9_COMP *cpi, const TileInfo *const tile,
MACROBLOCK *const x, int mi_row, int mi_col,
BLOCK_SIZE bsize) {
VP9_COMMON *const cm = &cpi->common;
MACROBLOCKD *const xd = &x->e_mbd;
MODE_INFO *mi;
const int mi_width = num_8x8_blocks_wide_lookup[bsize];
const int mi_height = num_8x8_blocks_high_lookup[bsize];
const struct segmentation *const seg = &cm->seg;
MvLimits *const mv_limits = &x->mv_limits;
set_skip_context(xd, mi_row, mi_col);
set_mode_info_offsets(cm, x, xd, mi_row, mi_col);
mi = xd->mi[0];
// Set up destination pointers.
vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
// Set up limit values for MV components.
// Mv beyond the range do not produce new/different prediction block.
mv_limits->row_min = -(((mi_row + mi_height) * MI_SIZE) + VP9_INTERP_EXTEND);
mv_limits->col_min = -(((mi_col + mi_width) * MI_SIZE) + VP9_INTERP_EXTEND);
mv_limits->row_max = (cm->mi_rows - mi_row) * MI_SIZE + VP9_INTERP_EXTEND;
mv_limits->col_max = (cm->mi_cols - mi_col) * MI_SIZE + VP9_INTERP_EXTEND;
// Set up distance of MB to edge of frame in 1/8th pel units.
assert(!(mi_col & (mi_width - 1)) && !(mi_row & (mi_height - 1)));
set_mi_row_col(xd, tile, mi_row, mi_height, mi_col, mi_width, cm->mi_rows,
cm->mi_cols);
// Set up source buffers.
vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
// R/D setup.
x->rddiv = cpi->rd.RDDIV;
x->rdmult = cpi->rd.RDMULT;
// Setup segment ID.
if (seg->enabled) {
if (cpi->oxcf.aq_mode != VARIANCE_AQ && cpi->oxcf.aq_mode != LOOKAHEAD_AQ &&
cpi->oxcf.aq_mode != EQUATOR360_AQ) {
const uint8_t *const map =
seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
}
vp9_init_plane_quantizers(cpi, x);
x->encode_breakout = cpi->segment_encode_breakout[mi->segment_id];
} else {
mi->segment_id = 0;
x->encode_breakout = cpi->encode_breakout;
}
// required by vp9_append_sub8x8_mvs_for_idx() and vp9_find_best_ref_mvs()
xd->tile = *tile;
}
static void duplicate_mode_info_in_sb(VP9_COMMON *cm, MACROBLOCKD *xd,
int mi_row, int mi_col,
BLOCK_SIZE bsize) {
const int block_width =
VPXMIN(num_8x8_blocks_wide_lookup[bsize], cm->mi_cols - mi_col);
const int block_height =
VPXMIN(num_8x8_blocks_high_lookup[bsize], cm->mi_rows - mi_row);
const int mi_stride = xd->mi_stride;
MODE_INFO *const src_mi = xd->mi[0];
int i, j;
for (j = 0; j < block_height; ++j)
for (i = 0; i < block_width; ++i) xd->mi[j * mi_stride + i] = src_mi;
}
static void set_block_size(VP9_COMP *const cpi, MACROBLOCK *const x,
MACROBLOCKD *const xd, int mi_row, int mi_col,
BLOCK_SIZE bsize) {
if (cpi->common.mi_cols > mi_col && cpi->common.mi_rows > mi_row) {
set_mode_info_offsets(&cpi->common, x, xd, mi_row, mi_col);
xd->mi[0]->sb_type = bsize;
}
}
typedef struct {
// This struct is used for computing variance in choose_partitioning(), where
// the max number of samples within a superblock is 16x16 (with 4x4 avg). Even
// in high bitdepth, uint32_t is enough for sum_square_error (2^12 * 2^12 * 16
// * 16 = 2^32).
uint32_t sum_square_error;
int32_t sum_error;
int log2_count;
int variance;
} var;
typedef struct {
var none;
var horz[2];
var vert[2];
} partition_variance;
typedef struct {
partition_variance part_variances;
var split[4];
} v4x4;
typedef struct {
partition_variance part_variances;
v4x4 split[4];
} v8x8;
typedef struct {
partition_variance part_variances;
v8x8 split[4];
} v16x16;
typedef struct {
partition_variance part_variances;
v16x16 split[4];
} v32x32;
typedef struct {
partition_variance part_variances;
v32x32 split[4];
} v64x64;
typedef struct {
partition_variance *part_variances;
var *split[4];
} variance_node;
typedef enum {
V16X16,
V32X32,
V64X64,
} TREE_LEVEL;
static void tree_to_node(void *data, BLOCK_SIZE bsize, variance_node *node) {
int i;
node->part_variances = NULL;
switch (bsize) {
case BLOCK_64X64: {
v64x64 *vt = (v64x64 *)data;
node->part_variances = &vt->part_variances;
for (i = 0; i < 4; i++)
node->split[i] = &vt->split[i].part_variances.none;
break;
}
case BLOCK_32X32: {
v32x32 *vt = (v32x32 *)data;
node->part_variances = &vt->part_variances;
for (i = 0; i < 4; i++)
node->split[i] = &vt->split[i].part_variances.none;
break;
}
case BLOCK_16X16: {
v16x16 *vt = (v16x16 *)data;
node->part_variances = &vt->part_variances;
for (i = 0; i < 4; i++)
node->split[i] = &vt->split[i].part_variances.none;
break;
}
case BLOCK_8X8: {
v8x8 *vt = (v8x8 *)data;
node->part_variances = &vt->part_variances;
for (i = 0; i < 4; i++)
node->split[i] = &vt->split[i].part_variances.none;
break;
}
case BLOCK_4X4: {
v4x4 *vt = (v4x4 *)data;
node->part_variances = &vt->part_variances;
for (i = 0; i < 4; i++) node->split[i] = &vt->split[i];
break;
}
default: {
assert(0);
break;
}
}
}
// Set variance values given sum square error, sum error, count.
static void fill_variance(uint32_t s2, int32_t s, int c, var *v) {
v->sum_square_error = s2;
v->sum_error = s;
v->log2_count = c;
}
static void get_variance(var *v) {
v->variance =
(int)(256 * (v->sum_square_error -
((v->sum_error * v->sum_error) >> v->log2_count)) >>
v->log2_count);
}
static void sum_2_variances(const var *a, const var *b, var *r) {
assert(a->log2_count == b->log2_count);
fill_variance(a->sum_square_error + b->sum_square_error,
a->sum_error + b->sum_error, a->log2_count + 1, r);
}
static void fill_variance_tree(void *data, BLOCK_SIZE bsize) {
variance_node node;
memset(&node, 0, sizeof(node));
tree_to_node(data, bsize, &node);
sum_2_variances(node.split[0], node.split[1], &node.part_variances->horz[0]);
sum_2_variances(node.split[2], node.split[3], &node.part_variances->horz[1]);
sum_2_variances(node.split[0], node.split[2], &node.part_variances->vert[0]);
sum_2_variances(node.split[1], node.split[3], &node.part_variances->vert[1]);
sum_2_variances(&node.part_variances->vert[0], &node.part_variances->vert[1],
&node.part_variances->none);
}
static int set_vt_partitioning(VP9_COMP *cpi, MACROBLOCK *const x,
MACROBLOCKD *const xd, void *data,
BLOCK_SIZE bsize, int mi_row, int mi_col,
int64_t threshold, BLOCK_SIZE bsize_min,
int force_split) {
VP9_COMMON *const cm = &cpi->common;
variance_node vt;
const int block_width = num_8x8_blocks_wide_lookup[bsize];
const int block_height = num_8x8_blocks_high_lookup[bsize];
assert(block_height == block_width);
tree_to_node(data, bsize, &vt);
if (force_split == 1) return 0;
// For bsize=bsize_min (16x16/8x8 for 8x8/4x4 downsampling), select if
// variance is below threshold, otherwise split will be selected.
// No check for vert/horiz split as too few samples for variance.
if (bsize == bsize_min) {
// Variance already computed to set the force_split.
if (cm->frame_type == KEY_FRAME) get_variance(&vt.part_variances->none);
if (mi_col + block_width / 2 < cm->mi_cols &&
mi_row + block_height / 2 < cm->mi_rows &&
vt.part_variances->none.variance < threshold) {
set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
return 1;
}
return 0;
} else if (bsize > bsize_min) {
// Variance already computed to set the force_split.
if (cm->frame_type == KEY_FRAME) get_variance(&vt.part_variances->none);
// For key frame: take split for bsize above 32X32 or very high variance.
if (cm->frame_type == KEY_FRAME &&
(bsize > BLOCK_32X32 ||
vt.part_variances->none.variance > (threshold << 4))) {
return 0;
}
// If variance is low, take the bsize (no split).
if (mi_col + block_width / 2 < cm->mi_cols &&
mi_row + block_height / 2 < cm->mi_rows &&
vt.part_variances->none.variance < threshold) {
set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
return 1;
}
// Check vertical split.
if (mi_row + block_height / 2 < cm->mi_rows) {
BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_VERT);
get_variance(&vt.part_variances->vert[0]);
get_variance(&vt.part_variances->vert[1]);
if (vt.part_variances->vert[0].variance < threshold &&
vt.part_variances->vert[1].variance < threshold &&
get_plane_block_size(subsize, &xd->plane[1]) < BLOCK_INVALID) {
set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
set_block_size(cpi, x, xd, mi_row, mi_col + block_width / 2, subsize);
return 1;
}
}
// Check horizontal split.
if (mi_col + block_width / 2 < cm->mi_cols) {
BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_HORZ);
get_variance(&vt.part_variances->horz[0]);
get_variance(&vt.part_variances->horz[1]);
if (vt.part_variances->horz[0].variance < threshold &&
vt.part_variances->horz[1].variance < threshold &&
get_plane_block_size(subsize, &xd->plane[1]) < BLOCK_INVALID) {
set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
set_block_size(cpi, x, xd, mi_row + block_height / 2, mi_col, subsize);
return 1;
}
}
return 0;
}
return 0;
}
static int64_t scale_part_thresh_sumdiff(int64_t threshold_base, int speed,
int width, int height,
int content_state) {
if (speed >= 8) {
if (width <= 640 && height <= 480)
return (5 * threshold_base) >> 2;
else if ((content_state == kLowSadLowSumdiff) ||
(content_state == kHighSadLowSumdiff) ||
(content_state == kLowVarHighSumdiff))
return (5 * threshold_base) >> 2;
} else if (speed == 7) {
if ((content_state == kLowSadLowSumdiff) ||
(content_state == kHighSadLowSumdiff) ||
(content_state == kLowVarHighSumdiff)) {
return (5 * threshold_base) >> 2;
}
}
return threshold_base;
}
// Set the variance split thresholds for following the block sizes:
// 0 - threshold_64x64, 1 - threshold_32x32, 2 - threshold_16x16,
// 3 - vbp_threshold_8x8. vbp_threshold_8x8 (to split to 4x4 partition) is
// currently only used on key frame.
static void set_vbp_thresholds(VP9_COMP *cpi, int64_t thresholds[], int q,
int content_state) {
VP9_COMMON *const cm = &cpi->common;
const int is_key_frame = (cm->frame_type == KEY_FRAME);
const int threshold_multiplier = is_key_frame ? 20 : 1;
int64_t threshold_base =
(int64_t)(threshold_multiplier * cpi->y_dequant[q][1]);
if (is_key_frame) {
thresholds[0] = threshold_base;
thresholds[1] = threshold_base >> 2;
thresholds[2] = threshold_base >> 2;
thresholds[3] = threshold_base << 2;
} else {
// Increase base variance threshold based on estimated noise level.
if (cpi->noise_estimate.enabled && cm->width >= 640 && cm->height >= 480) {
NOISE_LEVEL noise_level =
vp9_noise_estimate_extract_level(&cpi->noise_estimate);
if (noise_level == kHigh)
threshold_base = 3 * threshold_base;
else if (noise_level == kMedium)
threshold_base = threshold_base << 1;
else if (noise_level < kLow)
threshold_base = (7 * threshold_base) >> 3;
}
#if CONFIG_VP9_TEMPORAL_DENOISING
if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) &&
cpi->oxcf.speed > 5 && cpi->denoiser.denoising_level >= kDenLow)
threshold_base =
vp9_scale_part_thresh(threshold_base, cpi->denoiser.denoising_level,
content_state, cpi->svc.temporal_layer_id);
else
threshold_base =
scale_part_thresh_sumdiff(threshold_base, cpi->oxcf.speed, cm->width,
cm->height, content_state);
#else
// Increase base variance threshold based on content_state/sum_diff level.
threshold_base = scale_part_thresh_sumdiff(
threshold_base, cpi->oxcf.speed, cm->width, cm->height, content_state);
#endif
thresholds[0] = threshold_base;
thresholds[2] = threshold_base << cpi->oxcf.speed;
if (cm->width >= 1280 && cm->height >= 720 && cpi->oxcf.speed < 7)
thresholds[2] = thresholds[2] << 1;
if (cm->width <= 352 && cm->height <= 288) {
thresholds[0] = threshold_base >> 3;
thresholds[1] = threshold_base >> 1;
thresholds[2] = threshold_base << 3;
} else if (cm->width < 1280 && cm->height < 720) {
thresholds[1] = (5 * threshold_base) >> 2;
} else if (cm->width < 1920 && cm->height < 1080) {
thresholds[1] = threshold_base << 1;
} else {
thresholds[1] = (5 * threshold_base) >> 1;
}
}
}
void vp9_set_variance_partition_thresholds(VP9_COMP *cpi, int q,
int content_state) {
VP9_COMMON *const cm = &cpi->common;
SPEED_FEATURES *const sf = &cpi->sf;
const int is_key_frame = (cm->frame_type == KEY_FRAME);
if (sf->partition_search_type != VAR_BASED_PARTITION &&
sf->partition_search_type != REFERENCE_PARTITION) {
return;
} else {
set_vbp_thresholds(cpi, cpi->vbp_thresholds, q, content_state);
// The thresholds below are not changed locally.
if (is_key_frame) {
cpi->vbp_threshold_sad = 0;
cpi->vbp_threshold_copy = 0;
cpi->vbp_bsize_min = BLOCK_8X8;
} else {
if (cm->width <= 352 && cm->height <= 288)
cpi->vbp_threshold_sad = 10;
else
cpi->vbp_threshold_sad = (cpi->y_dequant[q][1] << 1) > 1000
? (cpi->y_dequant[q][1] << 1)
: 1000;
cpi->vbp_bsize_min = BLOCK_16X16;
if (cm->width <= 352 && cm->height <= 288)
cpi->vbp_threshold_copy = 4000;
else if (cm->width <= 640 && cm->height <= 360)
cpi->vbp_threshold_copy = 8000;
else
cpi->vbp_threshold_copy = (cpi->y_dequant[q][1] << 3) > 8000
? (cpi->y_dequant[q][1] << 3)
: 8000;
}
cpi->vbp_threshold_minmax = 15 + (q >> 3);
}
}
// Compute the minmax over the 8x8 subblocks.
static int compute_minmax_8x8(const uint8_t *s, int sp, const uint8_t *d,
int dp, int x16_idx, int y16_idx,
#if CONFIG_VP9_HIGHBITDEPTH
int highbd_flag,
#endif
int pixels_wide, int pixels_high) {
int k;
int minmax_max = 0;
int minmax_min = 255;
// Loop over the 4 8x8 subblocks.
for (k = 0; k < 4; k++) {
int x8_idx = x16_idx + ((k & 1) << 3);
int y8_idx = y16_idx + ((k >> 1) << 3);
int min = 0;
int max = 0;
if (x8_idx < pixels_wide && y8_idx < pixels_high) {
#if CONFIG_VP9_HIGHBITDEPTH
if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
vpx_highbd_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
d + y8_idx * dp + x8_idx, dp, &min, &max);
} else {
vpx_minmax_8x8(s + y8_idx * sp + x8_idx, sp, d + y8_idx * dp + x8_idx,
dp, &min, &max);
}
#else
vpx_minmax_8x8(s + y8_idx * sp + x8_idx, sp, d + y8_idx * dp + x8_idx, dp,
&min, &max);
#endif
if ((max - min) > minmax_max) minmax_max = (max - min);
if ((max - min) < minmax_min) minmax_min = (max - min);
}
}
return (minmax_max - minmax_min);
}
static void fill_variance_4x4avg(const uint8_t *s, int sp, const uint8_t *d,
int dp, int x8_idx, int y8_idx, v8x8 *vst,
#if CONFIG_VP9_HIGHBITDEPTH
int highbd_flag,
#endif
int pixels_wide, int pixels_high,
int is_key_frame) {
int k;
for (k = 0; k < 4; k++) {
int x4_idx = x8_idx + ((k & 1) << 2);
int y4_idx = y8_idx + ((k >> 1) << 2);
unsigned int sse = 0;
int sum = 0;
if (x4_idx < pixels_wide && y4_idx < pixels_high) {
int s_avg;
int d_avg = 128;
#if CONFIG_VP9_HIGHBITDEPTH
if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
s_avg = vpx_highbd_avg_4x4(s + y4_idx * sp + x4_idx, sp);
if (!is_key_frame)
d_avg = vpx_highbd_avg_4x4(d + y4_idx * dp + x4_idx, dp);
} else {
s_avg = vpx_avg_4x4(s + y4_idx * sp + x4_idx, sp);
if (!is_key_frame) d_avg = vpx_avg_4x4(d + y4_idx * dp + x4_idx, dp);
}
#else
s_avg = vpx_avg_4x4(s + y4_idx * sp + x4_idx, sp);
if (!is_key_frame) d_avg = vpx_avg_4x4(d + y4_idx * dp + x4_idx, dp);
#endif
sum = s_avg - d_avg;
sse = sum * sum;
}
fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
}
}
static void fill_variance_8x8avg(const uint8_t *s, int sp, const uint8_t *d,
int dp, int x16_idx, int y16_idx, v16x16 *vst,
#if CONFIG_VP9_HIGHBITDEPTH
int highbd_flag,
#endif
int pixels_wide, int pixels_high,
int is_key_frame) {
int k;
for (k = 0; k < 4; k++) {
int x8_idx = x16_idx + ((k & 1) << 3);
int y8_idx = y16_idx + ((k >> 1) << 3);
unsigned int sse = 0;
int sum = 0;
if (x8_idx < pixels_wide && y8_idx < pixels_high) {
int s_avg;
int d_avg = 128;
#if CONFIG_VP9_HIGHBITDEPTH
if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
s_avg = vpx_highbd_avg_8x8(s + y8_idx * sp + x8_idx, sp);
if (!is_key_frame)
d_avg = vpx_highbd_avg_8x8(d + y8_idx * dp + x8_idx, dp);
} else {
s_avg = vpx_avg_8x8(s + y8_idx * sp + x8_idx, sp);
if (!is_key_frame) d_avg = vpx_avg_8x8(d + y8_idx * dp + x8_idx, dp);
}
#else
s_avg = vpx_avg_8x8(s + y8_idx * sp + x8_idx, sp);
if (!is_key_frame) d_avg = vpx_avg_8x8(d + y8_idx * dp + x8_idx, dp);
#endif
sum = s_avg - d_avg;
sse = sum * sum;
}
fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
}
}
// Check if most of the superblock is skin content, and if so, force split to
// 32x32, and set x->sb_is_skin for use in mode selection.
static int skin_sb_split(VP9_COMP *cpi, MACROBLOCK *x, const int low_res,
int mi_row, int mi_col, int *force_split) {
VP9_COMMON *const cm = &cpi->common;
#if CONFIG_VP9_HIGHBITDEPTH
if (cm->use_highbitdepth) return 0;
#endif
// Avoid checking superblocks on/near boundary and avoid low resolutions.
// Note superblock may still pick 64X64 if y_sad is very small
// (i.e., y_sad < cpi->vbp_threshold_sad) below. For now leave this as is.
if (!low_res && (mi_col >= 8 && mi_col + 8 < cm->mi_cols && mi_row >= 8 &&
mi_row + 8 < cm->mi_rows)) {
int num_16x16_skin = 0;
int num_16x16_nonskin = 0;
uint8_t *ysignal = x->plane[0].src.buf;
uint8_t *usignal = x->plane[1].src.buf;
uint8_t *vsignal = x->plane[2].src.buf;
int sp = x->plane[0].src.stride;
int spuv = x->plane[1].src.stride;
const int block_index = mi_row * cm->mi_cols + mi_col;
const int bw = num_8x8_blocks_wide_lookup[BLOCK_64X64];
const int bh = num_8x8_blocks_high_lookup[BLOCK_64X64];
const int xmis = VPXMIN(cm->mi_cols - mi_col, bw);
const int ymis = VPXMIN(cm->mi_rows - mi_row, bh);
// Loop through the 16x16 sub-blocks.
int i, j;
for (i = 0; i < ymis; i += 2) {
for (j = 0; j < xmis; j += 2) {
int bl_index = block_index + i * cm->mi_cols + j;
int is_skin = cpi->skin_map[bl_index];
num_16x16_skin += is_skin;
num_16x16_nonskin += (1 - is_skin);
if (num_16x16_nonskin > 3) {
// Exit loop if at least 4 of the 16x16 blocks are not skin.
i = ymis;
break;
}
ysignal += 16;
usignal += 8;
vsignal += 8;
}
ysignal += (sp << 4) - 64;
usignal += (spuv << 3) - 32;
vsignal += (spuv << 3) - 32;
}
if (num_16x16_skin > 12) {
*force_split = 1;
return 1;
}
}
return 0;
}
static void set_low_temp_var_flag(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
v64x64 *vt, int64_t thresholds[],
MV_REFERENCE_FRAME ref_frame_partition,
int mi_col, int mi_row) {
int i, j;
VP9_COMMON *const cm = &cpi->common;
const int mv_thr = cm->width > 640 ? 8 : 4;
// Check temporal variance for bsize >= 16x16, if LAST_FRAME was selected and
// int_pro mv is small. If the temporal variance is small set the flag
// variance_low for the block. The variance threshold can be adjusted, the
// higher the more aggressive.
if (ref_frame_partition == LAST_FRAME &&
(cpi->sf.short_circuit_low_temp_var == 1 ||
(xd->mi[0]->mv[0].as_mv.col < mv_thr &&
xd->mi[0]->mv[0].as_mv.col > -mv_thr &&
xd->mi[0]->mv[0].as_mv.row < mv_thr &&
xd->mi[0]->mv[0].as_mv.row > -mv_thr))) {
if (xd->mi[0]->sb_type == BLOCK_64X64) {
if ((vt->part_variances).none.variance < (thresholds[0] >> 1))
x->variance_low[0] = 1;
} else if (xd->mi[0]->sb_type == BLOCK_64X32) {
for (i = 0; i < 2; i++) {
if (vt->part_variances.horz[i].variance < (thresholds[0] >> 2))
x->variance_low[i + 1] = 1;
}
} else if (xd->mi[0]->sb_type == BLOCK_32X64) {
for (i = 0; i < 2; i++) {
if (vt->part_variances.vert[i].variance < (thresholds[0] >> 2))
x->variance_low[i + 3] = 1;
}
} else {
for (i = 0; i < 4; i++) {
const int idx[4][2] = { { 0, 0 }, { 0, 4 }, { 4, 0 }, { 4, 4 } };
const int idx_str =
cm->mi_stride * (mi_row + idx[i][0]) + mi_col + idx[i][1];
MODE_INFO **this_mi = cm->mi_grid_visible + idx_str;
if (cm->mi_cols <= mi_col + idx[i][1] ||
cm->mi_rows <= mi_row + idx[i][0])
continue;
if ((*this_mi)->sb_type == BLOCK_32X32) {
int64_t threshold_32x32 = (cpi->sf.short_circuit_low_temp_var == 1 ||
cpi->sf.short_circuit_low_temp_var == 3)
? ((5 * thresholds[1]) >> 3)
: (thresholds[1] >> 1);
if (vt->split[i].part_variances.none.variance < threshold_32x32)
x->variance_low[i + 5] = 1;
} else if (cpi->sf.short_circuit_low_temp_var >= 2) {
// For 32x16 and 16x32 blocks, the flag is set on each 16x16 block
// inside.
if ((*this_mi)->sb_type == BLOCK_16X16 ||
(*this_mi)->sb_type == BLOCK_32X16 ||
(*this_mi)->sb_type == BLOCK_16X32) {
for (j = 0; j < 4; j++) {
if (vt->split[i].split[j].part_variances.none.variance <
(thresholds[2] >> 8))
x->variance_low[(i << 2) + j + 9] = 1;
}
}
}
}
}
}
}
static void copy_partitioning_helper(VP9_COMP *cpi, MACROBLOCK *x,
MACROBLOCKD *xd, BLOCK_SIZE bsize,
int mi_row, int mi_col) {
VP9_COMMON *const cm = &cpi->common;
BLOCK_SIZE *prev_part = cpi->prev_partition;
int start_pos = mi_row * cm->mi_stride + mi_col;
const int bsl = b_width_log2_lookup[bsize];
const int bs = (1 << bsl) >> 2;
BLOCK_SIZE subsize;
PARTITION_TYPE partition;
if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
partition = partition_lookup[bsl][prev_part[start_pos]];
subsize = get_subsize(bsize, partition);
if (subsize < BLOCK_8X8) {
set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
} else {
switch (partition) {
case PARTITION_NONE:
set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
break;
case PARTITION_HORZ:
set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
set_block_size(cpi, x, xd, mi_row + bs, mi_col, subsize);
break;
case PARTITION_VERT:
set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
set_block_size(cpi, x, xd, mi_row, mi_col + bs, subsize);
break;
case PARTITION_SPLIT:
copy_partitioning_helper(cpi, x, xd, subsize, mi_row, mi_col);
copy_partitioning_helper(cpi, x, xd, subsize, mi_row + bs, mi_col);
copy_partitioning_helper(cpi, x, xd, subsize, mi_row, mi_col + bs);
copy_partitioning_helper(cpi, x, xd, subsize, mi_row + bs, mi_col + bs);
break;
default: assert(0);
}
}
}
static int copy_partitioning(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
int mi_row, int mi_col, int segment_id,
int sb_offset) {
int svc_copy_allowed = 1;
int frames_since_key_thresh = 1;
if (cpi->use_svc) {
// For SVC, don't allow copy if base spatial layer is key frame, or if
// frame is not a temporal enhancement layer frame.
int layer = LAYER_IDS_TO_IDX(0, cpi->svc.temporal_layer_id,
cpi->svc.number_temporal_layers);
const LAYER_CONTEXT *lc = &cpi->svc.layer_context[layer];
if (lc->is_key_frame ||
(cpi->svc.temporal_layer_id != cpi->svc.number_temporal_layers - 1 &&
cpi->svc.number_temporal_layers > 1))
svc_copy_allowed = 0;
frames_since_key_thresh = cpi->svc.number_spatial_layers << 1;
}
if (cpi->rc.frames_since_key > frames_since_key_thresh && svc_copy_allowed &&
!cpi->resize_pending && segment_id == CR_SEGMENT_ID_BASE &&
cpi->prev_segment_id[sb_offset] == CR_SEGMENT_ID_BASE &&
cpi->copied_frame_cnt[sb_offset] < cpi->max_copied_frame) {
if (cpi->prev_partition != NULL) {
copy_partitioning_helper(cpi, x, xd, BLOCK_64X64, mi_row, mi_col);
cpi->copied_frame_cnt[sb_offset] += 1;
memcpy(x->variance_low, &(cpi->prev_variance_low[sb_offset * 25]),
sizeof(x->variance_low));
return 1;
}
}
return 0;
}
static void copy_partitioning_svc(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
BLOCK_SIZE bsize, int mi_rowref,
int mi_colref, int mi_row, int mi_col,
int mi_row_high, int mi_col_high) {
VP9_COMMON *const cm = &cpi->common;
SVC *const svc = &cpi->svc;
BLOCK_SIZE *prev_part = svc->prev_partition_svc;
// Variables with _high are for higher resolution.
int bsize_high = 0;
int subsize_high = 0;
int bsl_high = 0;
int bs_high = 0;
int shift_row = 0;
int shift_col = 0;
// Find corresponding (mi_col/mi_row) block down-scaled by 2x2.
int start_pos = mi_row * (svc->mi_stride[svc->spatial_layer_id - 1]) + mi_col;
const int bsl = b_width_log2_lookup[bsize];
const int bs = (1 << bsl) >> 2;
BLOCK_SIZE subsize;
PARTITION_TYPE partition;
const int bw = b_width_log2_lookup[bsize];
const int bh = b_height_log2_lookup[bsize];
// For block size >= 32x32 shift_row/col can stay 0.
if (bw == 1)
shift_col = 3;
else if (bw == 2)
shift_col = 2;
if (bh == 1)
shift_row = 3;
else if (bh == 2)
shift_row = 2;
if (mi_row_high >= cm->mi_rows || mi_col_high >= cm->mi_cols) return;
if (mi_row >= (cm->mi_rows >> 1) || mi_col >= (cm->mi_cols >> 1)) return;
if ((mi_row - mi_rowref > shift_row) || (mi_col - mi_colref > shift_col))
return;
partition = partition_lookup[bsl][prev_part[start_pos]];
subsize = get_subsize(bsize, partition);
// Scale up block size by 2x2. Force 64x64 for size larger than 32x32.
if (bsize < BLOCK_32X32) {
bsize_high = bsize + 3;
subsize_high = subsize + 3;
} else if (bsize >= BLOCK_32X32) {
bsize_high = BLOCK_64X64;
subsize_high = BLOCK_64X64;
}
bsl_high = b_width_log2_lookup[bsize_high];
bs_high = (1 << bsl_high) / 4;
if (subsize < BLOCK_8X8) {
set_block_size(cpi, x, xd, mi_row_high, mi_col_high, bsize_high);
} else {
switch (partition) {
case PARTITION_NONE:
set_block_size(cpi, x, xd, mi_row_high, mi_col_high, bsize_high);
break;
case PARTITION_HORZ:
set_block_size(cpi, x, xd, mi_row_high, mi_col_high, subsize_high);
if (subsize_high < BLOCK_64X64)
set_block_size(cpi, x, xd, mi_row_high + bs_high, mi_col,
subsize_high);
break;
case PARTITION_VERT:
set_block_size(cpi, x, xd, mi_row_high, mi_col_high, subsize_high);
if (subsize_high < BLOCK_64X64)
set_block_size(cpi, x, xd, mi_row_high, mi_col_high + bs_high,
subsize_high);
break;
case PARTITION_SPLIT:
copy_partitioning_svc(cpi, x, xd, subsize, mi_rowref, mi_colref, mi_row,
mi_col, mi_row_high, mi_col_high);
copy_partitioning_svc(cpi, x, xd, subsize, mi_rowref, mi_colref,
mi_row + bs, mi_col, mi_row_high + bs_high,
mi_col_high);
copy_partitioning_svc(cpi, x, xd, subsize, mi_rowref, mi_colref, mi_row,
mi_col + bs, mi_row_high, mi_col_high + bs_high);
copy_partitioning_svc(cpi, x, xd, subsize, mi_rowref, mi_colref,
mi_row + bs, mi_col + bs, mi_row_high + bs_high,
mi_col_high + bs_high);
break;
default: assert(0);
}
}
}
static void update_partition_svc(VP9_COMP *cpi, BLOCK_SIZE bsize, int mi_row,
int mi_col) {
VP9_COMMON *const cm = &cpi->common;
BLOCK_SIZE *prev_part = cpi->svc.prev_partition_svc;
int start_pos = mi_row * cm->mi_stride + mi_col;
const int bsl = b_width_log2_lookup[bsize];
const int bs = (1 << bsl) >> 2;
BLOCK_SIZE subsize;
PARTITION_TYPE partition;
const MODE_INFO *mi = NULL;
int xx, yy;
if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
mi = cm->mi_grid_visible[start_pos];
partition = partition_lookup[bsl][mi->sb_type];
subsize = get_subsize(bsize, partition);
if (subsize < BLOCK_8X8) {
prev_part[start_pos] = bsize;
} else {
switch (partition) {
case PARTITION_NONE:
prev_part[start_pos] = bsize;
if (bsize == BLOCK_64X64) {
for (xx = 0; xx < 8; xx += 4)
for (yy = 0; yy < 8; yy += 4) {
if ((mi_row + xx < cm->mi_rows) && (mi_col + yy < cm->mi_cols))
prev_part[start_pos + xx * cm->mi_stride + yy] = bsize;
}
}
break;
case PARTITION_HORZ:
prev_part[start_pos] = subsize;
if (mi_row + bs < cm->mi_rows)
prev_part[start_pos + bs * cm->mi_stride] = subsize;
break;
case PARTITION_VERT:
prev_part[start_pos] = subsize;
if (mi_col + bs < cm->mi_cols) prev_part[start_pos + bs] = subsize;
break;
case PARTITION_SPLIT:
update_partition_svc(cpi, subsize, mi_row, mi_col);
update_partition_svc(cpi, subsize, mi_row + bs, mi_col);
update_partition_svc(cpi, subsize, mi_row, mi_col + bs);
update_partition_svc(cpi, subsize, mi_row + bs, mi_col + bs);
break;
default: assert(0);
}
}
}
static void update_prev_partition(VP9_COMP *cpi, BLOCK_SIZE bsize, int mi_row,
int mi_col) {
VP9_COMMON *const cm = &cpi->common;
BLOCK_SIZE *prev_part = cpi->prev_partition;
int start_pos = mi_row * cm->mi_stride + mi_col;
const int bsl = b_width_log2_lookup[bsize];
const int bs = (1 << bsl) >> 2;
BLOCK_SIZE subsize;
PARTITION_TYPE partition;
const MODE_INFO *mi = NULL;
if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
mi = cm->mi_grid_visible[start_pos];
partition = partition_lookup[bsl][mi->sb_type];
subsize = get_subsize(bsize, partition);
if (subsize < BLOCK_8X8) {
prev_part[start_pos] = bsize;
} else {
switch (partition) {
case PARTITION_NONE: prev_part[start_pos] = bsize; break;
case PARTITION_HORZ:
prev_part[start_pos] = subsize;
if (mi_row + bs < cm->mi_rows)
prev_part[start_pos + bs * cm->mi_stride] = subsize;
break;
case PARTITION_VERT:
prev_part[start_pos] = subsize;
if (mi_col + bs < cm->mi_cols) prev_part[start_pos + bs] = subsize;
break;
case PARTITION_SPLIT:
update_prev_partition(cpi, subsize, mi_row, mi_col);
update_prev_partition(cpi, subsize, mi_row + bs, mi_col);
update_prev_partition(cpi, subsize, mi_row, mi_col + bs);
update_prev_partition(cpi, subsize, mi_row + bs, mi_col + bs);
break;
default: assert(0);
}
}
}
static void chroma_check(VP9_COMP *cpi, MACROBLOCK *x, int bsize,
unsigned int y_sad, int is_key_frame) {
int i;
MACROBLOCKD *xd = &x->e_mbd;
if (is_key_frame) return;
// For speed >= 8, avoid the chroma check if y_sad is above threshold.
if (cpi->oxcf.speed >= 8) {
if (y_sad > cpi->vbp_thresholds[1] &&
(!cpi->noise_estimate.enabled ||
vp9_noise_estimate_extract_level(&cpi->noise_estimate) < kMedium))
return;
}
for (i = 1; i <= 2; ++i) {
unsigned int uv_sad = UINT_MAX;
struct macroblock_plane *p = &x->plane[i];
struct macroblockd_plane *pd = &xd->plane[i];
const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
if (bs != BLOCK_INVALID)
uv_sad = cpi->fn_ptr[bs].sdf(p->src.buf, p->src.stride, pd->dst.buf,
pd->dst.stride);
// TODO(marpan): Investigate if we should lower this threshold if
// superblock is detected as skin.
x->color_sensitivity[i - 1] = uv_sad > (y_sad >> 2);
}
}
static uint64_t avg_source_sad(VP9_COMP *cpi, MACROBLOCK *x, int shift,
int sb_offset) {
unsigned int tmp_sse;
uint64_t tmp_sad;
unsigned int tmp_variance;
const BLOCK_SIZE bsize = BLOCK_64X64;
uint8_t *src_y = cpi->Source->y_buffer;
int src_ystride = cpi->Source->y_stride;
uint8_t *last_src_y = cpi->Last_Source->y_buffer;
int last_src_ystride = cpi->Last_Source->y_stride;
uint64_t avg_source_sad_threshold = 10000;
uint64_t avg_source_sad_threshold2 = 12000;
#if CONFIG_VP9_HIGHBITDEPTH
if (cpi->common.use_highbitdepth) return 0;
#endif
src_y += shift;
last_src_y += shift;
tmp_sad =
cpi->fn_ptr[bsize].sdf(src_y, src_ystride, last_src_y, last_src_ystride);
tmp_variance = vpx_variance64x64(src_y, src_ystride, last_src_y,
last_src_ystride, &tmp_sse);
// Note: tmp_sse - tmp_variance = ((sum * sum) >> 12)
if (tmp_sad < avg_source_sad_threshold)
x->content_state_sb = ((tmp_sse - tmp_variance) < 25) ? kLowSadLowSumdiff
: kLowSadHighSumdiff;
else
x->content_state_sb = ((tmp_sse - tmp_variance) < 25) ? kHighSadLowSumdiff
: kHighSadHighSumdiff;
// Detect large lighting change.
if (cpi->oxcf.content != VP9E_CONTENT_SCREEN &&
cpi->oxcf.rc_mode == VPX_CBR && tmp_variance < (tmp_sse >> 3) &&
(tmp_sse - tmp_variance) > 10000)
x->content_state_sb = kLowVarHighSumdiff;
else if (tmp_sad > (avg_source_sad_threshold << 1))
x->content_state_sb = kVeryHighSad;
if (cpi->content_state_sb_fd != NULL) {
if (tmp_sad < avg_source_sad_threshold2) {
// Cap the increment to 255.
if (cpi->content_state_sb_fd[sb_offset] < 255)
cpi->content_state_sb_fd[sb_offset]++;
} else {
cpi->content_state_sb_fd[sb_offset] = 0;
}
}
return tmp_sad;
}
// This function chooses partitioning based on the variance between source and
// reconstructed last, where variance is computed for down-sampled inputs.
static int choose_partitioning(VP9_COMP *cpi, const TileInfo *const tile,
MACROBLOCK *x, int mi_row, int mi_col) {
VP9_COMMON *const cm = &cpi->common;
MACROBLOCKD *xd = &x->e_mbd;
int i, j, k, m;
v64x64 vt;
v16x16 *vt2 = NULL;
int force_split[21];
int avg_32x32;
int max_var_32x32 = 0;
int min_var_32x32 = INT_MAX;
int var_32x32;
int avg_16x16[4];
int maxvar_16x16[4];
int minvar_16x16[4];
int64_t threshold_4x4avg;
NOISE_LEVEL noise_level = kLow;
int content_state = 0;
uint8_t *s;
const uint8_t *d;
int sp;
int dp;
int compute_minmax_variance = 1;
unsigned int y_sad = UINT_MAX;
BLOCK_SIZE bsize = BLOCK_64X64;
// Ref frame used in partitioning.
MV_REFERENCE_FRAME ref_frame_partition = LAST_FRAME;
int pixels_wide = 64, pixels_high = 64;
int64_t thresholds[4] = { cpi->vbp_thresholds[0], cpi->vbp_thresholds[1],
cpi->vbp_thresholds[2], cpi->vbp_thresholds[3] };
// For the variance computation under SVC mode, we treat the frame as key if
// the reference (base layer frame) is key frame (i.e., is_key_frame == 1).
const int is_key_frame =
(cm->frame_type == KEY_FRAME ||
(is_one_pass_cbr_svc(cpi) &&
cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame));
// Always use 4x4 partition for key frame.
const int use_4x4_partition = cm->frame_type == KEY_FRAME;
const int low_res = (cm->width <= 352 && cm->height <= 288);
int variance4x4downsample[16];
int segment_id;
int sb_offset = (cm->mi_stride >> 3) * (mi_row >> 3) + (mi_col >> 3);
set_offsets(cpi, tile, x, mi_row, mi_col, BLOCK_64X64);
segment_id = xd->mi[0]->segment_id;
if (cpi->oxcf.speed >= 8 || (cpi->use_svc && cpi->svc.non_reference_frame))
compute_minmax_variance = 0;
if (cpi->sf.use_source_sad && !is_key_frame) {
int sb_offset2 = ((cm->mi_cols + 7) >> 3) * (mi_row >> 3) + (mi_col >> 3);
content_state = x->content_state_sb;
x->skip_low_source_sad = (content_state == kLowSadLowSumdiff ||
content_state == kLowSadHighSumdiff)
? 1
: 0;
x->lowvar_highsumdiff = (content_state == kLowVarHighSumdiff) ? 1 : 0;
if (cpi->content_state_sb_fd != NULL)
x->last_sb_high_content = cpi->content_state_sb_fd[sb_offset2];
// For SVC on top spatial layer and non_reference frame: copy partition
// from lower spatial resolution if svc_use_lowres_part is enabled.
// TODO(jianj): Fix to allow it to work on boundary.
if (cpi->sf.svc_use_lowres_part && cpi->svc.spatial_layer_id == 2 &&
cpi->svc.non_reference_frame && cpi->svc.prev_partition_svc != NULL &&
mi_row < cm->mi_rows - 8 && mi_col < cm->mi_cols - 8 &&
content_state != kVeryHighSad) {
copy_partitioning_svc(cpi, x, xd, BLOCK_64X64, mi_row >> 1, mi_col >> 1,
mi_row >> 1, mi_col >> 1, mi_row, mi_col);
return 0;
}
// If source_sad is low copy the partition without computing the y_sad.
if (x->skip_low_source_sad && cpi->sf.copy_partition_flag &&
copy_partitioning(cpi, x, xd, mi_row, mi_col, segment_id, sb_offset)) {
x->sb_use_mv_part = 1;
if (cpi->sf.svc_use_lowres_part && cpi->svc.spatial_layer_id == 1)
update_partition_svc(cpi, BLOCK_64X64, mi_row, mi_col);
return 0;
}
}
if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled &&
cyclic_refresh_segment_id_boosted(segment_id)) {
int q = vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex);
set_vbp_thresholds(cpi, thresholds, q, content_state);
} else {
set_vbp_thresholds(cpi, thresholds, cm->base_qindex, content_state);
}
// For non keyframes, disable 4x4 average for low resolution when speed = 8
threshold_4x4avg = (cpi->oxcf.speed < 8) ? thresholds[1] << 1 : INT64_MAX;
memset(x->variance_low, 0, sizeof(x->variance_low));
if (xd->mb_to_right_edge < 0) pixels_wide += (xd->mb_to_right_edge >> 3);
if (xd->mb_to_bottom_edge < 0) pixels_high += (xd->mb_to_bottom_edge >> 3);
s = x->plane[0].src.buf;
sp = x->plane[0].src.stride;
// Index for force_split: 0 for 64x64, 1-4 for 32x32 blocks,
// 5-20 for the 16x16 blocks.
force_split[0] = 0;
if (!is_key_frame) {
// In the case of spatial/temporal scalable coding, the assumption here is
// that the temporal reference frame will always be of type LAST_FRAME.
// TODO(marpan): If that assumption is broken, we need to revisit this code.
MODE_INFO *mi = xd->mi[0];
YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
const YV12_BUFFER_CONFIG *yv12_g = NULL;
unsigned int y_sad_g, y_sad_thr, y_sad_last;
bsize = BLOCK_32X32 + (mi_col + 4 < cm->mi_cols) * 2 +
(mi_row + 4 < cm->mi_rows);
assert(yv12 != NULL);
if (!(is_one_pass_cbr_svc(cpi) && cpi->svc.spatial_layer_id)) {
// For now, GOLDEN will not be used for non-zero spatial layers, since
// it may not be a temporal reference.
yv12_g = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
}
// Only compute y_sad_g (sad for golden reference) for speed < 8.
if (cpi->oxcf.speed < 8 && yv12_g && yv12_g != yv12 &&
(cpi->ref_frame_flags & VP9_GOLD_FLAG)) {
vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
&cm->frame_refs[GOLDEN_FRAME - 1].sf);
y_sad_g = cpi->fn_ptr[bsize].sdf(
x->plane[0].src.buf, x->plane[0].src.stride, xd->plane[0].pre[0].buf,
xd->plane[0].pre[0].stride);
} else {
y_sad_g = UINT_MAX;
}
if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR &&
cpi->rc.is_src_frame_alt_ref) {
yv12 = get_ref_frame_buffer(cpi, ALTREF_FRAME);
vp9_setup_pre_planes(xd, 0, yv12, mi_row, mi_col,
&cm->frame_refs[ALTREF_FRAME - 1].sf);
mi->ref_frame[0] = ALTREF_FRAME;
y_sad_g = UINT_MAX;
} else {
vp9_setup_pre_planes(xd, 0, yv12, mi_row, mi_col,
&cm->frame_refs[LAST_FRAME - 1].sf);
mi->ref_frame[0] = LAST_FRAME;
}
mi->ref_frame[1] = NONE;
mi->sb_type = BLOCK_64X64;
mi->mv[0].as_int = 0;
mi->interp_filter = BILINEAR;
if (cpi->oxcf.speed >= 8 && !low_res &&
x->content_state_sb != kVeryHighSad) {
y_sad = cpi->fn_ptr[bsize].sdf(
x->plane[0].src.buf, x->plane[0].src.stride, xd->plane[0].pre[0].buf,
xd->plane[0].pre[0].stride);
} else {
y_sad = vp9_int_pro_motion_estimation(cpi, x, bsize, mi_row, mi_col);
x->sb_use_mv_part = 1;
x->sb_mvcol_part = mi->mv[0].as_mv.col;
x->sb_mvrow_part = mi->mv[0].as_mv.row;
}
y_sad_last = y_sad;
// Pick ref frame for partitioning, bias last frame when y_sad_g and y_sad
// are close if short_circuit_low_temp_var is on.
y_sad_thr = cpi->sf.short_circuit_low_temp_var ? (y_sad * 7) >> 3 : y_sad;
if (y_sad_g < y_sad_thr) {
vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
&cm->frame_refs[GOLDEN_FRAME - 1].sf);
mi->ref_frame[0] = GOLDEN_FRAME;
mi->mv[0].as_int = 0;
y_sad = y_sad_g;
ref_frame_partition = GOLDEN_FRAME;
} else {
x->pred_mv[LAST_FRAME] = mi->mv[0].as_mv;
ref_frame_partition = LAST_FRAME;
}
set_ref_ptrs(cm, xd, mi->ref_frame[0], mi->ref_frame[1]);
vp9_build_inter_predictors_sb(xd, mi_row, mi_col, BLOCK_64X64);
if (cpi->use_skin_detection)
x->sb_is_skin =
skin_sb_split(cpi, x, low_res, mi_row, mi_col, force_split);
d = xd->plane[0].dst.buf;
dp = xd->plane[0].dst.stride;
// If the y_sad is very small, take 64x64 as partition and exit.
// Don't check on boosted segment for now, as 64x64 is suppressed there.
if (segment_id == CR_SEGMENT_ID_BASE && y_sad < cpi->vbp_threshold_sad) {
const int block_width = num_8x8_blocks_wide_lookup[BLOCK_64X64];
const int block_height = num_8x8_blocks_high_lookup[BLOCK_64X64];
if (mi_col + block_width / 2 < cm->mi_cols &&
mi_row + block_height / 2 < cm->mi_rows) {
set_block_size(cpi, x, xd, mi_row, mi_col, BLOCK_64X64);
x->variance_low[0] = 1;
chroma_check(cpi, x, bsize, y_sad, is_key_frame);
if (cpi->sf.svc_use_lowres_part && cpi->svc.spatial_layer_id == 1)
update_partition_svc(cpi, BLOCK_64X64, mi_row, mi_col);
return 0;
}
}
// If the y_sad is small enough, copy the partition of the superblock in the
// last frame to current frame only if the last frame is not a keyframe.
// Stop the copy every cpi->max_copied_frame to refresh the partition.
// TODO(jianj) : tune the threshold.
if (cpi->sf.copy_partition_flag && y_sad_last < cpi->vbp_threshold_copy &&
copy_partitioning(cpi, x, xd, mi_row, mi_col, segment_id, sb_offset)) {
chroma_check(cpi, x, bsize, y_sad, is_key_frame);
if (cpi->sf.svc_use_lowres_part && cpi->svc.spatial_layer_id == 1)
update_partition_svc(cpi, BLOCK_64X64, mi_row, mi_col);
return 0;
}
} else {
d = VP9_VAR_OFFS;
dp = 0;
#if CONFIG_VP9_HIGHBITDEPTH
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
switch (xd->bd) {
case 10: d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10); break;
case 12: d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12); break;
case 8:
default: d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8); break;
}
}
#endif // CONFIG_VP9_HIGHBITDEPTH
}
if (low_res && threshold_4x4avg < INT64_MAX)
CHECK_MEM_ERROR(cm, vt2, vpx_calloc(16, sizeof(*vt2)));
// Fill in the entire tree of 8x8 (or 4x4 under some conditions) variances
// for splits.
for (i = 0; i < 4; i++) {
const int x32_idx = ((i & 1) << 5);
const int y32_idx = ((i >> 1) << 5);
const int i2 = i << 2;
force_split[i + 1] = 0;
avg_16x16[i] = 0;
maxvar_16x16[i] = 0;
minvar_16x16[i] = INT_MAX;
for (j = 0; j < 4; j++) {
const int x16_idx = x32_idx + ((j & 1) << 4);
const int y16_idx = y32_idx + ((j >> 1) << 4);
const int split_index = 5 + i2 + j;
v16x16 *vst = &vt.split[i].split[j];
force_split[split_index] = 0;
variance4x4downsample[i2 + j] = 0;
if (!is_key_frame) {
fill_variance_8x8avg(s, sp, d, dp, x16_idx, y16_idx, vst,
#if CONFIG_VP9_HIGHBITDEPTH
xd->cur_buf->flags,
#endif
pixels_wide, pixels_high, is_key_frame);
fill_variance_tree(&vt.split[i].split[j], BLOCK_16X16);
get_variance(&vt.split[i].split[j].part_variances.none);
avg_16x16[i] += vt.split[i].split[j].part_variances.none.variance;
if (vt.split[i].split[j].part_variances.none.variance < minvar_16x16[i])
minvar_16x16[i] = vt.split[i].split[j].part_variances.none.variance;
if (vt.split[i].split[j].part_variances.none.variance > maxvar_16x16[i])
maxvar_16x16[i] = vt.split[i].split[j].part_variances.none.variance;
if (vt.split[i].split[j].part_variances.none.variance > thresholds[2]) {
// 16X16 variance is above threshold for split, so force split to 8x8
// for this 16x16 block (this also forces splits for upper levels).
force_split[split_index] = 1;
force_split[i + 1] = 1;
force_split[0] = 1;
} else if (compute_minmax_variance &&
vt.split[i].split[j].part_variances.none.variance >
thresholds[1] &&
!cyclic_refresh_segment_id_boosted(segment_id)) {
// We have some nominal amount of 16x16 variance (based on average),
// compute the minmax over the 8x8 sub-blocks, and if above threshold,
// force split to 8x8 block for this 16x16 block.
int minmax = compute_minmax_8x8(s, sp, d, dp, x16_idx, y16_idx,
#if CONFIG_VP9_HIGHBITDEPTH
xd->cur_buf->flags,
#endif
pixels_wide, pixels_high);
int thresh_minmax = (int)cpi->vbp_threshold_minmax;
if (x->content_state_sb == kVeryHighSad)
thresh_minmax = thresh_minmax << 1;
if (minmax > thresh_minmax) {
force_split[split_index] = 1;
force_split[i + 1] = 1;
force_split[0] = 1;
}
}
}
if (is_key_frame || (low_res &&
vt.split[i].split[j].part_variances.none.variance >
threshold_4x4avg)) {
force_split[split_index] = 0;
// Go down to 4x4 down-sampling for variance.
variance4x4downsample[i2 + j] = 1;
for (k = 0; k < 4; k++) {
int x8_idx = x16_idx + ((k & 1) << 3);
int y8_idx = y16_idx + ((k >> 1) << 3);
v8x8 *vst2 = is_key_frame ? &vst->split[k] : &vt2[i2 + j].split[k];
fill_variance_4x4avg(s, sp, d, dp, x8_idx, y8_idx, vst2,
#if CONFIG_VP9_HIGHBITDEPTH
xd->cur_buf->flags,
#endif
pixels_wide, pixels_high, is_key_frame);
}
}
}
}
if (cpi->noise_estimate.enabled)
noise_level = vp9_noise_estimate_extract_level(&cpi->noise_estimate);
// Fill the rest of the variance tree by summing split partition values.
avg_32x32 = 0;
for (i = 0; i < 4; i++) {
const int i2 = i << 2;
for (j = 0; j < 4; j++) {
if (variance4x4downsample[i2 + j] == 1) {
v16x16 *vtemp = (!is_key_frame) ? &vt2[i2 + j] : &vt.split[i].split[j];
for (m = 0; m < 4; m++) fill_variance_tree(&vtemp->split[m], BLOCK_8X8);
fill_variance_tree(vtemp, BLOCK_16X16);
// If variance of this 16x16 block is above the threshold, force block
// to split. This also forces a split on the upper levels.
get_variance(&vtemp->part_variances.none);
if (vtemp->part_variances.none.variance > thresholds[2]) {
force_split[5 + i2 + j] = 1;
force_split[i + 1] = 1;
force_split[0] = 1;
}
}
}
fill_variance_tree(&vt.split[i], BLOCK_32X32);
// If variance of this 32x32 block is above the threshold, or if its above
// (some threshold of) the average variance over the sub-16x16 blocks, then
// force this block to split. This also forces a split on the upper
// (64x64) level.
if (!force_split[i + 1]) {
get_variance(&vt.split[i].part_variances.none);
var_32x32 = vt.split[i].part_variances.none.variance;
max_var_32x32 = VPXMAX(var_32x32, max_var_32x32);
min_var_32x32 = VPXMIN(var_32x32, min_var_32x32);
if (vt.split[i].part_variances.none.variance > thresholds[1] ||
(!is_key_frame &&
vt.split[i].part_variances.none.variance > (thresholds[1] >> 1) &&
vt.split[i].part_variances.none.variance > (avg_16x16[i] >> 1))) {
force_split[i + 1] = 1;
force_split[0] = 1;
} else if (!is_key_frame && noise_level < kLow && cm->height <= 360 &&
(maxvar_16x16[i] - minvar_16x16[i]) > (thresholds[1] >> 1) &&
maxvar_16x16[i] > thresholds[1]) {
force_split[i + 1] = 1;
force_split[0] = 1;
}
avg_32x32 += var_32x32;
}
}
if (!force_split[0]) {
fill_variance_tree(&vt, BLOCK_64X64);
get_variance(&vt.part_variances.none);
// If variance of this 64x64 block is above (some threshold of) the average
// variance over the sub-32x32 blocks, then force this block to split.
// Only checking this for noise level >= medium for now.
if (!is_key_frame && noise_level >= kMedium &&
vt.part_variances.none.variance > (9 * avg_32x32) >> 5)
force_split[0] = 1;
// Else if the maximum 32x32 variance minus the miniumum 32x32 variance in
// a 64x64 block is greater than threshold and the maximum 32x32 variance is
// above a miniumum threshold, then force the split of a 64x64 block
// Only check this for low noise.
else if (!is_key_frame && noise_level < kMedium &&
(max_var_32x32 - min_var_32x32) > 3 * (thresholds[0] >> 3) &&
max_var_32x32 > thresholds[0] >> 1)
force_split[0] = 1;
}
// Now go through the entire structure, splitting every block size until
// we get to one that's got a variance lower than our threshold.
if (mi_col + 8 > cm->mi_cols || mi_row + 8 > cm->mi_rows ||
!set_vt_partitioning(cpi, x, xd, &vt, BLOCK_64X64, mi_row, mi_col,
thresholds[0], BLOCK_16X16, force_split[0])) {
for (i = 0; i < 4; ++i) {
const int x32_idx = ((i & 1) << 2);
const int y32_idx = ((i >> 1) << 2);
const int i2 = i << 2;
if (!set_vt_partitioning(cpi, x, xd, &vt.split[i], BLOCK_32X32,
(mi_row + y32_idx), (mi_col + x32_idx),
thresholds[1], BLOCK_16X16,
force_split[i + 1])) {
for (j = 0; j < 4; ++j) {
const int x16_idx = ((j & 1) << 1);
const int y16_idx = ((j >> 1) << 1);
// For inter frames: if variance4x4downsample[] == 1 for this 16x16
// block, then the variance is based on 4x4 down-sampling, so use vt2
// in set_vt_partioning(), otherwise use vt.
v16x16 *vtemp = (!is_key_frame && variance4x4downsample[i2 + j] == 1)
? &vt2[i2 + j]
: &vt.split[i].split[j];
if (!set_vt_partitioning(
cpi, x, xd, vtemp, BLOCK_16X16, mi_row + y32_idx + y16_idx,
mi_col + x32_idx + x16_idx, thresholds[2], cpi->vbp_bsize_min,
force_split[5 + i2 + j])) {
for (k = 0; k < 4; ++k) {
const int x8_idx = (k & 1);
const int y8_idx = (k >> 1);
if (use_4x4_partition) {
if (!set_vt_partitioning(cpi, x, xd, &vtemp->split[k],
BLOCK_8X8,
mi_row + y32_idx + y16_idx + y8_idx,
mi_col + x32_idx + x16_idx + x8_idx,
thresholds[3], BLOCK_8X8, 0)) {
set_block_size(
cpi, x, xd, (mi_row + y32_idx + y16_idx + y8_idx),
(mi_col + x32_idx + x16_idx + x8_idx), BLOCK_4X4);
}
} else {
set_block_size(
cpi, x, xd, (mi_row + y32_idx + y16_idx + y8_idx),
(mi_col + x32_idx + x16_idx + x8_idx), BLOCK_8X8);
}
}
}
}
}
}
}
if (cm->frame_type != KEY_FRAME && cpi->sf.copy_partition_flag) {
update_prev_partition(cpi, BLOCK_64X64, mi_row, mi_col);
cpi->prev_segment_id[sb_offset] = segment_id;
memcpy(&(cpi->prev_variance_low[sb_offset * 25]), x->variance_low,
sizeof(x->variance_low));
// Reset the counter for copy partitioning
if (cpi->copied_frame_cnt[sb_offset] == cpi->max_copied_frame)
cpi->copied_frame_cnt[sb_offset] = 0;
}
if (cm->frame_type != KEY_FRAME && cpi->sf.svc_use_lowres_part &&
cpi->svc.spatial_layer_id == 1)
update_partition_svc(cpi, BLOCK_64X64, mi_row, mi_col);
if (cpi->sf.short_circuit_low_temp_var) {
set_low_temp_var_flag(cpi, x, xd, &vt, thresholds, ref_frame_partition,
mi_col, mi_row);
}
chroma_check(cpi, x, bsize, y_sad, is_key_frame);
if (vt2) vpx_free(vt2);
return 0;
}
static void update_state(VP9_COMP *cpi, ThreadData *td, PICK_MODE_CONTEXT *ctx,
int mi_row, int mi_col, BLOCK_SIZE bsize,
int output_enabled) {
int i, x_idx, y;
VP9_COMMON *const cm = &cpi->common;
RD_COUNTS *const rdc = &td->rd_counts;
MACROBLOCK *const x = &td->mb;
MACROBLOCKD *const xd = &x->e_mbd;
struct macroblock_plane *const p = x->plane;
struct macroblockd_plane *const pd = xd->plane;
MODE_INFO *mi = &ctx->mic;
MODE_INFO *const xdmi = xd->mi[0];
MODE_INFO *mi_addr = xd->mi[0];
const struct segmentation *const seg = &cm->seg;
const int bw = num_8x8_blocks_wide_lookup[mi->sb_type];
const int bh = num_8x8_blocks_high_lookup[mi->sb_type];
const int x_mis = VPXMIN(bw, cm->mi_cols - mi_col);
const int y_mis = VPXMIN(bh, cm->mi_rows - mi_row);
MV_REF *const frame_mvs = cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
int w, h;
const int mis = cm->mi_stride;
const int mi_width = num_8x8_blocks_wide_lookup[bsize];
const int mi_height = num_8x8_blocks_high_lookup[bsize];
int max_plane;
assert(mi->sb_type == bsize);
*mi_addr = *mi;
*x->mbmi_ext = ctx->mbmi_ext;
// If segmentation in use
if (seg->enabled) {
// For in frame complexity AQ copy the segment id from the segment map.
if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
const uint8_t *const map =
seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
mi_addr->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
}
// Else for cyclic refresh mode update the segment map, set the segment id
// and then update the quantizer.
if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
vp9_cyclic_refresh_update_segment(cpi, xd->mi[0], mi_row, mi_col, bsize,
ctx->rate, ctx->dist, x->skip, p);
}
}
max_plane = is_inter_block(xdmi) ? MAX_MB_PLANE : 1;
for (i = 0; i < max_plane; ++i) {
p[i].coeff = ctx->coeff_pbuf[i][1];
p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
p[i].eobs = ctx->eobs_pbuf[i][1];
}
for (i = max_plane; i < MAX_MB_PLANE; ++i) {
p[i].coeff = ctx->coeff_pbuf[i][2];
p[i].qcoeff = ctx->qcoeff_pbuf[i][2];
pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][2];
p[i].eobs = ctx->eobs_pbuf[i][2];
}
// Restore the coding context of the MB to that that was in place
// when the mode was picked for it
for (y = 0; y < mi_height; y++)
for (x_idx = 0; x_idx < mi_width; x_idx++)
if ((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width > x_idx &&
(xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height > y) {
xd->mi[x_idx + y * mis] = mi_addr;
}
if (cpi->oxcf.aq_mode != NO_AQ) vp9_init_plane_quantizers(cpi, x);
if (is_inter_block(xdmi) && xdmi->sb_type < BLOCK_8X8) {
xdmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
xdmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
}
x->skip = ctx->skip;
memcpy(x->zcoeff_blk[xdmi->tx_size], ctx->zcoeff_blk,
sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
if (!output_enabled) return;
#if CONFIG_INTERNAL_STATS
if (frame_is_intra_only(cm)) {
static const int kf_mode_index[] = {
THR_DC /*DC_PRED*/, THR_V_PRED /*V_PRED*/,
THR_H_PRED /*H_PRED*/, THR_D45_PRED /*D45_PRED*/,
THR_D135_PRED /*D135_PRED*/, THR_D117_PRED /*D117_PRED*/,
THR_D153_PRED /*D153_PRED*/, THR_D207_PRED /*D207_PRED*/,
THR_D63_PRED /*D63_PRED*/, THR_TM /*TM_PRED*/,
};
++cpi->mode_chosen_counts[kf_mode_index[xdmi->mode]];
} else {
// Note how often each mode chosen as best
++cpi->mode_chosen_counts[ctx->best_mode_index];
}
#endif
if (!frame_is_intra_only(cm)) {
if (is_inter_block(xdmi)) {
vp9_update_mv_count(td);
if (cm->interp_filter == SWITCHABLE) {
const int ctx = get_pred_context_switchable_interp(xd);
++td->counts->switchable_interp[ctx][xdmi->interp_filter];
}
}
rdc->comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff;
rdc->comp_pred_diff[COMPOUND_REFERENCE] += ctx->comp_pred_diff;
rdc->comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff;
for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
rdc->filter_diff[i] += ctx->best_filter_diff[i];
}
for (h = 0; h < y_mis; ++h) {
MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
for (w = 0; w < x_mis; ++w) {
MV_REF *const mv = frame_mv + w;
mv->ref_frame[0] = mi->ref_frame[0];
mv->ref_frame[1] = mi->ref_frame[1];
mv->mv[0].as_int = mi->mv[0].as_int;
mv->mv[1].as_int = mi->mv[1].as_int;
}
}
}
void vp9_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src,
int mi_row, int mi_col) {
uint8_t *const buffers[3] = { src->y_buffer, src->u_buffer, src->v_buffer };
const int strides[3] = { src->y_stride, src->uv_stride, src->uv_stride };
int i;
// Set current frame pointer.
x->e_mbd.cur_buf = src;
for (i = 0; i < MAX_MB_PLANE; i++)
setup_pred_plane(&x->plane[i].src, buffers[i], strides[i], mi_row, mi_col,
NULL, x->e_mbd.plane[i].subsampling_x,
x->e_mbd.plane[i].subsampling_y);
}
static void set_mode_info_seg_skip(MACROBLOCK *x, TX_MODE tx_mode,
RD_COST *rd_cost, BLOCK_SIZE bsize) {
MACROBLOCKD *const xd = &x->e_mbd;
MODE_INFO *const mi = xd->mi[0];
INTERP_FILTER filter_ref;
filter_ref = get_pred_context_switchable_interp(xd);
if (filter_ref == SWITCHABLE_FILTERS) filter_ref = EIGHTTAP;
mi->sb_type = bsize;
mi->mode = ZEROMV;
mi->tx_size =
VPXMIN(max_txsize_lookup[bsize], tx_mode_to_biggest_tx_size[tx_mode]);
mi->skip = 1;
mi->uv_mode = DC_PRED;
mi->ref_frame[0] = LAST_FRAME;
mi->ref_frame[1] = NONE;
mi->mv[0].as_int = 0;
mi->interp_filter = filter_ref;
xd->mi[0]->bmi[0].as_mv[0].as_int = 0;
x->skip = 1;
vp9_rd_cost_init(rd_cost);
}
static int set_segment_rdmult(VP9_COMP *const cpi, MACROBLOCK *const x,
int8_t segment_id) {
int segment_qindex;
VP9_COMMON *const cm = &cpi->common;
vp9_init_plane_quantizers(cpi, x);
vpx_clear_system_state();
segment_qindex = vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex);
return vp9_compute_rd_mult(cpi, segment_qindex + cm->y_dc_delta_q);
}
static void rd_pick_sb_modes(VP9_COMP *cpi, TileDataEnc *tile_data,
MACROBLOCK *const x, int mi_row, int mi_col,
RD_COST *rd_cost, BLOCK_SIZE bsize,
PICK_MODE_CONTEXT *ctx, int64_t best_rd) {
VP9_COMMON *const cm = &cpi->common;
TileInfo *const tile_info = &tile_data->tile_info;
MACROBLOCKD *const xd = &x->e_mbd;
MODE_INFO *mi;
struct macroblock_plane *const p = x->plane;
struct macroblockd_plane *const pd = xd->plane;
const AQ_MODE aq_mode = cpi->oxcf.aq_mode;
int i, orig_rdmult;
vpx_clear_system_state();
// Use the lower precision, but faster, 32x32 fdct for mode selection.
x->use_lp32x32fdct = 1;
set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
mi = xd->mi[0];
mi->sb_type = bsize;
for (i = 0; i < MAX_MB_PLANE; ++i) {
p[i].coeff = ctx->coeff_pbuf[i][0];
p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
p[i].eobs = ctx->eobs_pbuf[i][0];
}
ctx->is_coded = 0;
ctx->skippable = 0;
ctx->pred_pixel_ready = 0;
x->skip_recode = 0;
// Set to zero to make sure we do not use the previous encoded frame stats
mi->skip = 0;
#if CONFIG_VP9_HIGHBITDEPTH
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
x->source_variance = vp9_high_get_sby_perpixel_variance(
cpi, &x->plane[0].src, bsize, xd->bd);
} else {
x->source_variance =
vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
}
#else
x->source_variance =
vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
#endif // CONFIG_VP9_HIGHBITDEPTH
// Save rdmult before it might be changed, so it can be restored later.
orig_rdmult = x->rdmult;
if ((cpi->sf.tx_domain_thresh > 0.0) || (cpi->sf.quant_opt_thresh > 0.0)) {
double logvar = vp9_log_block_var(cpi, x, bsize);
// Check block complexity as part of descision on using pixel or transform
// domain distortion in rd tests.
x->block_tx_domain = cpi->sf.allow_txfm_domain_distortion &&
(logvar >= cpi->sf.tx_domain_thresh);
// Check block complexity as part of descision on using quantized
// coefficient optimisation inside the rd loop.
x->block_qcoeff_opt =
cpi->sf.allow_quant_coeff_opt && (logvar <= cpi->sf.quant_opt_thresh);
} else {
x->block_tx_domain = cpi->sf.allow_txfm_domain_distortion;
x->block_qcoeff_opt = cpi->sf.allow_quant_coeff_opt;
}
if (aq_mode == VARIANCE_AQ) {
const int energy =
bsize <= BLOCK_16X16 ? x->mb_energy : vp9_block_energy(cpi, x, bsize);
if (cm->frame_type == KEY_FRAME || cpi->refresh_alt_ref_frame ||
cpi->force_update_segmentation ||
(cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
mi->segment_id = vp9_vaq_segment_id(energy);
} else {
const uint8_t *const map =
cm->seg.update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
}
x->rdmult = set_segment_rdmult(cpi, x, mi->segment_id);
} else if (aq_mode == LOOKAHEAD_AQ) {
const uint8_t *const map = cpi->segmentation_map;
// I do not change rdmult here consciously.
mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
} else if (aq_mode == EQUATOR360_AQ) {
if (cm->frame_type == KEY_FRAME || cpi->force_update_segmentation) {
mi->segment_id = vp9_360aq_segment_id(mi_row, cm->mi_rows);
} else {
const uint8_t *const map =
cm->seg.update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
}
x->rdmult = set_segment_rdmult(cpi, x, mi->segment_id);
} else if (aq_mode == COMPLEXITY_AQ) {
x->rdmult = set_segment_rdmult(cpi, x, mi->segment_id);
} else if (aq_mode == CYCLIC_REFRESH_AQ) {
const uint8_t *const map =
cm->seg.update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
// If segment is boosted, use rdmult for that segment.
if (cyclic_refresh_segment_id_boosted(
get_segment_id(cm, map, bsize, mi_row, mi_col)))
x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
}
// Find best coding mode & reconstruct the MB so it is available
// as a predictor for MBs that follow in the SB
if (frame_is_intra_only(cm)) {
vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, best_rd);
} else {
if (bsize >= BLOCK_8X8) {
if (segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_SKIP))
vp9_rd_pick_inter_mode_sb_seg_skip(cpi, tile_data, x, rd_cost, bsize,
ctx, best_rd);
else
vp9_rd_pick_inter_mode_sb(cpi, tile_data, x, mi_row, mi_col, rd_cost,
bsize, ctx, best_rd);
} else {
vp9_rd_pick_inter_mode_sub8x8(cpi, tile_data, x, mi_row, mi_col, rd_cost,
bsize, ctx, best_rd);
}
}
// Examine the resulting rate and for AQ mode 2 make a segment choice.
if ((rd_cost->rate != INT_MAX) && (aq_mode == COMPLEXITY_AQ) &&
(bsize >= BLOCK_16X16) &&
(cm->frame_type == KEY_FRAME || cpi->refresh_alt_ref_frame ||
(cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref))) {
vp9_caq_select_segment(cpi, x, bsize, mi_row, mi_col, rd_cost->rate);
}
x->rdmult = orig_rdmult;
// TODO(jingning) The rate-distortion optimization flow needs to be
// refactored to provide proper exit/return handle.
if (rd_cost->rate == INT_MAX) rd_cost->rdcost = INT64_MAX;
ctx->rate = rd_cost->rate;
ctx->dist = rd_cost->dist;
}
static void update_stats(VP9_COMMON *cm, ThreadData *td) {
const MACROBLOCK *x = &td->mb;
const MACROBLOCKD *const xd = &x->e_mbd;
const MODE_INFO *const mi = xd->mi[0];
const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
const BLOCK_SIZE bsize = mi->sb_type;
if (!frame_is_intra_only(cm)) {
FRAME_COUNTS *const counts = td->counts;
const int inter_block = is_inter_block(mi);
const int seg_ref_active =
segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_REF_FRAME);
if (!seg_ref_active) {
counts->intra_inter[get_intra_inter_context(xd)][inter_block]++;
// If the segment reference feature is enabled we have only a single
// reference frame allowed for the segment so exclude it from
// the reference frame counts used to work out probabilities.
if (inter_block) {
const MV_REFERENCE_FRAME ref0 = mi->ref_frame[0];
if (cm->reference_mode == REFERENCE_MODE_SELECT)
counts->comp_inter[vp9_get_reference_mode_context(cm, xd)]
[has_second_ref(mi)]++;
if (has_second_ref(mi)) {
counts->comp_ref[vp9_get_pred_context_comp_ref_p(cm, xd)]
[ref0 == GOLDEN_FRAME]++;
} else {
counts->single_ref[vp9_get_pred_context_single_ref_p1(xd)][0]
[ref0 != LAST_FRAME]++;
if (ref0 != LAST_FRAME)
counts->single_ref[vp9_get_pred_context_single_ref_p2(xd)][1]
[ref0 != GOLDEN_FRAME]++;
}
}
}
if (inter_block &&
!segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_SKIP)) {
const int mode_ctx = mbmi_ext->mode_context[mi->ref_frame[0]];
if (bsize >= BLOCK_8X8) {
const PREDICTION_MODE mode = mi->mode;
++counts->inter_mode[mode_ctx][INTER_OFFSET(mode)];
} else {
const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
int idx, idy;
for (idy = 0; idy < 2; idy += num_4x4_h) {
for (idx = 0; idx < 2; idx += num_4x4_w) {
const int j = idy * 2 + idx;
const PREDICTION_MODE b_mode = mi->bmi[j].as_mode;
++counts->inter_mode[mode_ctx][INTER_OFFSET(b_mode)];
}
}
}
}
}
}
static void restore_context(MACROBLOCK *const x, int mi_row, int mi_col,
ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
BLOCK_SIZE bsize) {
MACROBLOCKD *const xd = &x->e_mbd;
int p;
const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
int mi_width = num_8x8_blocks_wide_lookup[bsize];
int mi_height = num_8x8_blocks_high_lookup[bsize];
for (p = 0; p < MAX_MB_PLANE; p++) {
memcpy(xd->above_context[p] + ((mi_col * 2) >> xd->plane[p].subsampling_x),
a + num_4x4_blocks_wide * p,
(sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
xd->plane[p].subsampling_x);
memcpy(xd->left_context[p] +
((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
l + num_4x4_blocks_high * p,
(sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
xd->plane[p].subsampling_y);
}
memcpy(xd->above_seg_context + mi_col, sa,
sizeof(*xd->above_seg_context) * mi_width);
memcpy(xd->left_seg_context + (mi_row & MI_MASK), sl,
sizeof(xd->left_seg_context[0]) * mi_height);
}
static void save_context(MACROBLOCK *const x, int mi_row, int mi_col,
ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
BLOCK_SIZE bsize) {
const MACROBLOCKD *const xd = &x->e_mbd;
int p;
const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
int mi_width = num_8x8_blocks_wide_lookup[bsize];
int mi_height = num_8x8_blocks_high_lookup[bsize];
// buffer the above/left context information of the block in search.
for (p = 0; p < MAX_MB_PLANE; ++p) {
memcpy(a + num_4x4_blocks_wide * p,
xd->above_context[p] + (mi_col * 2 >> xd->plane[p].subsampling_x),
(sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
xd->plane[p].subsampling_x);
memcpy(l + num_4x4_blocks_high * p,
xd->left_context[p] +
((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
(sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
xd->plane[p].subsampling_y);
}
memcpy(sa, xd->above_seg_context + mi_col,
sizeof(*xd->above_seg_context) * mi_width);
memcpy(sl, xd->left_seg_context + (mi_row & MI_MASK),
sizeof(xd->left_seg_context[0]) * mi_height);
}
static void encode_b(VP9_COMP *cpi, const TileInfo *const tile, ThreadData *td,
TOKENEXTRA **tp, int mi_row, int mi_col,
int output_enabled, BLOCK_SIZE bsize,
PICK_MODE_CONTEXT *ctx) {
MACROBLOCK *const x = &td->mb;
set_offsets(cpi, tile, x, mi_row, mi_col, bsize);
update_state(cpi, td, ctx, mi_row, mi_col, bsize, output_enabled);
encode_superblock(cpi, td, tp, output_enabled, mi_row, mi_col, bsize, ctx);
if (output_enabled) {
update_stats(&cpi->common, td);
(*tp)->token = EOSB_TOKEN;
(*tp)++;
}
}
static void encode_sb(VP9_COMP *cpi, ThreadData *td, const TileInfo *const tile,
TOKENEXTRA **tp, int mi_row, int mi_col,
int output_enabled, BLOCK_SIZE bsize, PC_TREE *pc_tree) {
VP9_COMMON *const cm = &cpi->common;
MACROBLOCK *const x = &td->mb;
MACROBLOCKD *const xd = &x->e_mbd;
const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
int ctx;
PARTITION_TYPE partition;
BLOCK_SIZE subsize = bsize;
if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
if (bsize >= BLOCK_8X8) {
ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
subsize = get_subsize(bsize, pc_tree->partitioning);
} else {
ctx = 0;
subsize = BLOCK_4X4;
}
partition = partition_lookup[bsl][subsize];
if (output_enabled && bsize != BLOCK_4X4)
td->counts->partition[ctx][partition]++;
switch (partition) {
case PARTITION_NONE:
encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
&pc_tree->none);
break;
case PARTITION_VERT:
encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
&pc_tree->vertical[0]);
if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
encode_b(cpi, tile, td, tp, mi_row, mi_col + hbs, output_enabled,
subsize, &pc_tree->vertical[1]);
}
break;
case PARTITION_HORZ:
encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
&pc_tree->horizontal[0]);
if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
encode_b(cpi, tile, td, tp, mi_row + hbs, mi_col, output_enabled,
subsize, &pc_tree->horizontal[1]);
}
break;
case PARTITION_SPLIT:
if (bsize == BLOCK_8X8) {
encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
pc_tree->leaf_split[0]);
} else {
encode_sb(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
pc_tree->split[0]);
encode_sb(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
subsize, pc_tree->split[1]);
encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
subsize, pc_tree->split[2]);
encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col + hbs, output_enabled,
subsize, pc_tree->split[3]);