Newer
Older
* 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 "vp9/encoder/vp9_encodeframe.h"
#include "vp9/encoder/vp9_encodemb.h"
#include "vp9/encoder/vp9_encodemv.h"
#include "vp9/encoder/vp9_onyx_int.h"
#include "vp9/common/vp9_entropymode.h"
#include "vp9/common/vp9_quant_common.h"
#include "vp9/encoder/vp9_segmentation.h"
#include "vp9/encoder/vp9_encodeintra.h"
#include "vp9/common/vp9_findnearmv.h"
#include "vp9/common/vp9_reconintra.h"
#include "vp9/common/vp9_seg_common.h"
#include "vp9/common/vp9_tile_common.h"
#include <limits.h>
#include "vpx_ports/vpx_timer.h"
#include "vp9/common/vp9_pred_common.h"
#include "vp9/common/vp9_mvref_common.h"
static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, int output_enabled,
int mi_row, int mi_col, BLOCK_SIZE_TYPE bsize);
static void adjust_act_zbin(VP9_COMP *cpi, MACROBLOCK *x);
/* activity_avg must be positive, or flat regions could get a zero weight
* (infinite lambda), which confounds analysis.
* This also avoids the need for divide by zero checks in
* vp9_activity_masking().
/* This is used as a reference when computing the source variance for the
* purposes of activity masking.
* Eventually this should be replaced by custom no-reference routines,
* which will be faster.
*/
static const uint8_t VP9_VAR_OFFS[16] = {128, 128, 128, 128, 128, 128, 128, 128,
128, 128, 128, 128, 128, 128, 128, 128};
// Original activity measure from Tim T's code.
static unsigned int tt_activity_measure(VP9_COMP *cpi, MACROBLOCK *x) {
unsigned int act;
unsigned int sse;
/* TODO: This could also be done over smaller areas (8x8), but that would
* require extensive changes elsewhere, as lambda is assumed to be fixed
* over an entire MB in most of the code.
* Another option is to compute four 8x8 variances, and pick a single
* lambda using a non-linear combination (e.g., the smallest, or second
* smallest, etc.).
*/
act = vp9_variance16x16(x->plane[0].src.buf, x->plane[0].src.stride,
VP9_VAR_OFFS, 0, &sse);
/* If the region is flat, lower the activity some more. */
if (act < 8 << 12)
act = act < 5 << 12 ? act : 5 << 12;
return act;
// Stub for alternative experimental activity measures.
static unsigned int alt_activity_measure(VP9_COMP *cpi, MACROBLOCK *x,
int use_dc_pred) {
return vp9_encode_intra(cpi, x, use_dc_pred);
DECLARE_ALIGNED(16, static const uint8_t, vp9_64x64_zeros[64*64]) = {0};
// Measure the activity of the current macroblock
// What we measure here is TBD so abstracted to this function
static unsigned int mb_activity_measure(VP9_COMP *cpi, MACROBLOCK *x,
if (ALT_ACT_MEASURE) {
int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
// Or use and alternative.
mb_activity = alt_activity_measure(cpi, x, use_dc_pred);
} else {
// Original activity measure from Tim T's code.
mb_activity = tt_activity_measure(cpi, x);
}
if (mb_activity < VP9_ACTIVITY_AVG_MIN)
mb_activity = VP9_ACTIVITY_AVG_MIN;
}
// Calculate an "average" mb activity value for the frame
static void calc_av_activity(VP9_COMP *cpi, int64_t activity_sum) {
// Find median: Simple n^2 algorithm for experimentation
{
unsigned int median;
unsigned int i, j;
unsigned int *sortlist;
unsigned int tmp;
// Create a list to sort to
CHECK_MEM_ERROR(&cpi->common, sortlist, vpx_calloc(sizeof(unsigned int),
cpi->common.MBs));
// Copy map to sort list
vpx_memcpy(sortlist, cpi->mb_activity_map,
// Ripple each value down to its correct position
for (i = 1; i < cpi->common.MBs; i ++) {
for (j = i; j > 0; j --) {
if (sortlist[j] < sortlist[j - 1]) {
// Swap values
tmp = sortlist[j - 1];
sortlist[j - 1] = sortlist[j];
sortlist[j] = tmp;
} else
// Even number MBs so estimate median as mean of two either side.
median = (1 + sortlist[cpi->common.MBs >> 1] +
cpi->activity_avg = (unsigned int) (activity_sum / cpi->common.MBs);
if (cpi->activity_avg < VP9_ACTIVITY_AVG_MIN)
cpi->activity_avg = VP9_ACTIVITY_AVG_MIN;
// Experimental code: return fixed value normalized for several clips
if (ALT_ACT_MEASURE)
cpi->activity_avg = 100000;
// Calculate an activity index for each mb
static void calc_activity_index(VP9_COMP *cpi, MACROBLOCK *x) {
VP9_COMMON *const cm = &cpi->common;
FILE *f = fopen("norm_act.stt", "a");
fprintf(f, "\n%12d\n", cpi->activity_avg);
// Reset pointers to start of activity map
x->mb_activity_ptr = cpi->mb_activity_map;
// Calculate normalized mb activity number.
for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
// for each macroblock col in image
for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
// Read activity from the map
act = *(x->mb_activity_ptr);
// Calculate a normalized activity number
a = act + 4 * cpi->activity_avg;
*(x->activity_ptr) = (int)((b + (a >> 1)) / a) - 1;
*(x->activity_ptr) = 1 - (int)((a + (b >> 1)) / b);
// Increment activity map pointers
x->mb_activity_ptr++;
}
// Loop through all MBs. Note activity of each, average activity and
// calculate a normalized activity for each
static void build_activity_map(VP9_COMP *cpi) {
YV12_BUFFER_CONFIG *new_yv12 = &cm->yv12_fb[cm->new_fb_idx];
int recon_yoffset;
int recon_y_stride = new_yv12->y_stride;
int mb_row, mb_col;
unsigned int mb_activity;
int64_t activity_sum = 0;
x->mb_activity_ptr = cpi->mb_activity_map;
// for each macroblock row in image
for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
// reset above block coeffs
xd->up_available = (mb_row != 0);
recon_yoffset = (mb_row * recon_y_stride * 16);
// for each macroblock col in image
for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
xd->plane[0].dst.buf = new_yv12->y_buffer + recon_yoffset;
xd->left_available = (mb_col != 0);
recon_yoffset += 16;
// measure activity
mb_activity = mb_activity_measure(cpi, x, mb_row, mb_col);
// Store MB level activity details.
*x->mb_activity_ptr = mb_activity;
// Increment activity map pointer
x->mb_activity_ptr++;
x->plane[0].src.buf += 16 * x->plane[0].src.stride - 16 * cm->mb_cols;
// Calculate an "average" MB activity
calc_av_activity(cpi, activity_sum);
// Calculate an activity index number of each mb
calc_activity_index(cpi, x);
void vp9_activity_masking(VP9_COMP *cpi, MACROBLOCK *x) {
x->rdmult += *(x->mb_activity_ptr) * (x->rdmult >> 2);
x->errorperbit = x->rdmult * 100 / (110 * x->rddiv);
x->errorperbit += (x->errorperbit == 0);
int64_t a;
int64_t b;
int64_t act = *(x->mb_activity_ptr);
// Apply the masking to the RD multiplier.
a = act + (2 * cpi->activity_avg);
b = (2 * act) + cpi->activity_avg;
x->rdmult = (unsigned int) (((int64_t) x->rdmult * b + (a >> 1)) / a);
x->errorperbit = x->rdmult * 100 / (110 * x->rddiv);
x->errorperbit += (x->errorperbit == 0);
// Activity based Zbin adjustment
adjust_act_zbin(cpi, x);
static void update_state(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx,
BLOCK_SIZE_TYPE bsize, int output_enabled) {
MACROBLOCK * const x = &cpi->mb;
MACROBLOCKD * const xd = &x->e_mbd;
MB_MODE_INFO * const mbmi = &xd->mode_info_context->mbmi;
const int mis = cpi->common.mode_info_stride;
const int mi_height = num_8x8_blocks_high_lookup[bsize];
const int mi_width = num_8x8_blocks_wide_lookup[bsize];
assert(mi->mbmi.mode < MB_MODE_COUNT);
assert(mi->mbmi.ref_frame[0] < MAX_REF_FRAMES);
assert(mi->mbmi.ref_frame[1] < MAX_REF_FRAMES);
assert(mi->mbmi.sb_type == bsize);
// 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 + LOG2_MI_SIZE)) + mi_width > x_idx
&& (xd->mb_to_bottom_edge >> (3 + LOG2_MI_SIZE)) + mi_height > y) {
MODE_INFO *mi_addr = xd->mode_info_context + x_idx + y * mis;
// FIXME(rbultje) I'm pretty sure this should go to the end of this block
// (i.e. after the output_enabled)
if (bsize < BLOCK_SIZE_MB16X16)
ctx->txfm_rd_diff[ALLOW_16X16] = ctx->txfm_rd_diff[ALLOW_8X8];
ctx->txfm_rd_diff[ALLOW_32X32] = ctx->txfm_rd_diff[ALLOW_16X16];
if (mbmi->ref_frame[0] != INTRA_FRAME && mbmi->sb_type < BLOCK_SIZE_SB8X8) {
*x->partition_info = ctx->partition_info;
mbmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
mbmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
x->skip = ctx->skip;
if (!vp9_segfeature_active(&xd->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
for (i = 0; i < NB_TXFM_MODES; i++) {
cpi->rd_tx_select_diff[i] += ctx->txfm_rd_diff[i];
if (cpi->common.frame_type == KEY_FRAME) {
// Restore the coding modes to that held in the coding context
// for (i = 0; i < 16; i++)
// {
// xd->block[i].bmi.as_mode =
// xd->mode_info_context->bmi[i].as_mode;
// assert(xd->mode_info_context->bmi[i].as_mode < MB_MODE_COUNT);
// }
#if CONFIG_INTERNAL_STATS
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_D27_PRED /*D27_PRED*/,
THR_D63_PRED /*D63_PRED*/,
THR_TM /*TM_PRED*/,
cpi->mode_chosen_counts[kf_mode_index[mi->mbmi.mode]]++;
} else {
// Note how often each mode chosen as best
cpi->mode_chosen_counts[mb_mode_index]++;
if (mbmi->ref_frame[0] != INTRA_FRAME
&& (mbmi->sb_type < BLOCK_SIZE_SB8X8 || mbmi->mode == NEWMV)) {
const MV_REFERENCE_FRAME rf1 = mbmi->ref_frame[0];
const MV_REFERENCE_FRAME rf2 = mbmi->ref_frame[1];
best_mv.as_int = ctx->best_ref_mv.as_int;
best_second_mv.as_int = ctx->second_best_ref_mv.as_int;
if (mbmi->mode == NEWMV) {
best_mv.as_int = mbmi->ref_mvs[rf1][0].as_int;
best_second_mv.as_int = mbmi->ref_mvs[rf2][0].as_int;
}
mbmi->best_mv.as_int = best_mv.as_int;
mbmi->best_second_mv.as_int = best_second_mv.as_int;
vp9_update_nmv_count(cpi, x, &best_mv, &best_second_mv);
if (bsize > BLOCK_SIZE_SB8X8 && mbmi->mode == NEWMV) {
int i, j;
for (j = 0; j < mi_height; ++j)
for (i = 0; i < mi_width; ++i)
if ((xd->mb_to_right_edge >> (3 + LOG2_MI_SIZE)) + mi_width > i
&& (xd->mb_to_bottom_edge >> (3 + LOG2_MI_SIZE)) + mi_height > j)
xd->mode_info_context[mis * j + i].mbmi = *mbmi;
if (cpi->common.mcomp_filter_type == SWITCHABLE
&& is_inter_mode(mbmi->mode)) {
++cpi->common.counts.switchable_interp[
vp9_get_pred_context_switchable_interp(xd)]
[vp9_switchable_interp_map[mbmi->interp_filter]];
cpi->rd_comp_pred_diff[SINGLE_PREDICTION_ONLY] += ctx->single_pred_diff;
cpi->rd_comp_pred_diff[COMP_PREDICTION_ONLY] += ctx->comp_pred_diff;
cpi->rd_comp_pred_diff[HYBRID_PREDICTION] += ctx->hybrid_pred_diff;
for (i = 0; i <= VP9_SWITCHABLE_FILTERS; i++) {
cpi->rd_filter_diff[i] += ctx->best_filter_diff[i];
}
void vp9_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src,
uint8_t *buffers[4] = {src->y_buffer, src->u_buffer, src->v_buffer, src
->alpha_buffer};
int strides[4] = {src->y_stride, src->uv_stride, src->uv_stride, src
->alpha_stride};
int i;
for (i = 0; i < MAX_MB_PLANE; i++) {
setup_pred_plane(&x->plane[i].src, buffers[i], strides[i], mb_row, mb_col,
NULL, x->e_mbd.plane[i].subsampling_x,
x->e_mbd.plane[i].subsampling_y);
}
static void set_offsets(VP9_COMP *cpi, int mi_row, int mi_col,
BLOCK_SIZE_TYPE bsize) {
MACROBLOCK * const x = &cpi->mb;
VP9_COMMON * const cm = &cpi->common;
MACROBLOCKD * const xd = &x->e_mbd;
MB_MODE_INFO *mbmi;
const int dst_fb_idx = cm->new_fb_idx;
const int idx_str = xd->mode_info_stride * mi_row + mi_col;
const int mi_width = num_8x8_blocks_wide_lookup[bsize];
const int mi_height = num_8x8_blocks_high_lookup[bsize];
const int mb_row = mi_row >> 1;
const int mb_col = mi_col >> 1;
const int idx_map = mb_row * cm->mb_cols + mb_col;
for (i = 0; i < MAX_MB_PLANE; i++) {
xd->plane[i].above_context = cm->above_context[i]
+ (mi_col * 2 >> xd->plane[i].subsampling_x);
xd->plane[i].left_context = cm->left_context[i]
+ (((mi_row * 2) & 15) >> xd->plane[i].subsampling_y);
set_partition_seg_context(cm, xd, mi_row, mi_col);
// Activity map pointer
x->mb_activity_ptr = &cpi->mb_activity_map[idx_map];
x->active_ptr = cpi->active_map + idx_map;
/* pointers to mode info contexts */
x->partition_info = x->pi + idx_str;
xd->mode_info_context = cm->mi + idx_str;
// Special case: if prev_mi is NULL, the previous mode info context
// cannot be used.
xd->prev_mode_info_context = cm->prev_mi ? cm->prev_mi + idx_str : NULL;
setup_dst_planes(xd, &cm->yv12_fb[dst_fb_idx], mi_row, mi_col);
/* Set up limit values for MV components to prevent them from
* extending beyond the UMV borders assuming 16x16 block size */
x->mv_row_min = -((mi_row * MI_SIZE)+ VP9BORDERINPIXELS - VP9_INTERP_EXTEND);
x->mv_col_min = -((mi_col * MI_SIZE)+ VP9BORDERINPIXELS - VP9_INTERP_EXTEND);
x->mv_row_max = ((cm->mi_rows - mi_row) * MI_SIZE
+ (VP9BORDERINPIXELS - MI_SIZE * mi_height - VP9_INTERP_EXTEND));
x->mv_col_max = ((cm->mi_cols - mi_col) * MI_SIZE
+ (VP9BORDERINPIXELS - MI_SIZE * mi_width - 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(cm, xd, mi_row, mi_height, mi_col, mi_width);
vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
/* R/D setup */
x->rddiv = cpi->RDDIV;
x->rdmult = cpi->RDMULT;
/* segment ID */
if (xd->seg.enabled) {
uint8_t *map = xd->seg.update_map ? cpi->segmentation_map
: cm->last_frame_seg_map;
mbmi->segment_id = vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
if (xd->seg.enabled && cpi->seg0_cnt > 0
&& !vp9_segfeature_active(&xd->seg, 0, SEG_LVL_REF_FRAME)
&& vp9_segfeature_active(&xd->seg, 1, SEG_LVL_REF_FRAME)) {
cpi->seg0_progress = (cpi->seg0_idx << 16) / cpi->seg0_cnt;
} else {
const int y = mb_row & ~3;
const int x = mb_col & ~3;
const int p16 = ((mb_row & 1) << 1) + (mb_col & 1);
const int p32 = ((mb_row & 2) << 2) + ((mb_col & 2) << 1);
const int tile_progress = cm->cur_tile_mi_col_start * cm->mb_rows >> 1;
const int mb_cols = (cm->cur_tile_mi_col_end - cm->cur_tile_mi_col_start)
>> 1;
cpi->seg0_progress = ((y * mb_cols + x * 4 + p32 + p16 + tile_progress)
<< 16) / cm->MBs;
x->encode_breakout = cpi->segment_encode_breakout[mbmi->segment_id];
x->encode_breakout = cpi->oxcf.encode_breakout;
static void pick_sb_modes(VP9_COMP *cpi, int mi_row, int mi_col,
int *totalrate, int64_t *totaldist,
BLOCK_SIZE_TYPE bsize, PICK_MODE_CONTEXT *ctx,
int64_t best_rd) {
VP9_COMMON *const cm = &cpi->common;
MACROBLOCK *const x = &cpi->mb;
MACROBLOCKD *const xd = &x->e_mbd;
if (bsize < BLOCK_SIZE_SB8X8)
if (xd->ab_index != 0)
return;
set_offsets(cpi, mi_row, mi_col, bsize);
xd->mode_info_context->mbmi.sb_type = bsize;
vp9_activity_masking(cpi, x);
// Find best coding mode & reconstruct the MB so it is available
// as a predictor for MBs that follow in the SB
if (cm->frame_type == KEY_FRAME)
vp9_rd_pick_intra_mode_sb(cpi, x, totalrate, totaldist, bsize, ctx,
best_rd);
vp9_rd_pick_inter_mode_sb(cpi, x, mi_row, mi_col, totalrate, totaldist,
static void update_stats(VP9_COMP *cpi, int mi_row, int mi_col) {
VP9_COMMON * const cm = &cpi->common;
MACROBLOCK * const x = &cpi->mb;
MACROBLOCKD * const xd = &x->e_mbd;
const int seg_ref_active = vp9_segfeature_active(&xd->seg, mbmi->segment_id,
SEG_LVL_REF_FRAME);
if (!seg_ref_active)
cpi->intra_inter_count[vp9_get_pred_context_intra_inter(xd)][mbmi
// 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 ((mbmi->ref_frame[0] > INTRA_FRAME) && !seg_ref_active) {
if (cm->comp_pred_mode == HYBRID_PREDICTION)
cpi->comp_inter_count[vp9_get_pred_context_comp_inter_inter(cm, xd)]
[mbmi->ref_frame[1] > INTRA_FRAME]++;
cpi->comp_ref_count[vp9_get_pred_context_comp_ref_p(cm, xd)][mbmi
cpi->single_ref_count[vp9_get_pred_context_single_ref_p1(xd)]
[0][mbmi->ref_frame[0] != LAST_FRAME]++;
cpi->single_ref_count[vp9_get_pred_context_single_ref_p2(xd)][1]
[mbmi->ref_frame[0] != GOLDEN_FRAME]++;
if ((mbmi->mode == ZEROMV) && (mbmi->ref_frame[0] == LAST_FRAME))
// TODO(jingning): the variables used here are little complicated. need further
// refactoring on organizing the the temporary buffers, when recursive
// partition down to 4x4 block size is enabled.
static PICK_MODE_CONTEXT *get_block_context(MACROBLOCK *x,
BLOCK_SIZE_TYPE bsize) {
switch (bsize) {
case BLOCK_SIZE_SB64X64:
return &x->sb64_context;
case BLOCK_SIZE_SB64X32:
return &x->sb64x32_context[xd->sb_index];
case BLOCK_SIZE_SB32X64:
return &x->sb32x64_context[xd->sb_index];
case BLOCK_SIZE_SB32X32:
return &x->sb32_context[xd->sb_index];
case BLOCK_SIZE_SB32X16:
return &x->sb32x16_context[xd->sb_index][xd->mb_index];
case BLOCK_SIZE_SB16X32:
return &x->sb16x32_context[xd->sb_index][xd->mb_index];
case BLOCK_SIZE_MB16X16:
return &x->mb_context[xd->sb_index][xd->mb_index];
case BLOCK_SIZE_SB16X8:
return &x->sb16x8_context[xd->sb_index][xd->mb_index][xd->b_index];
case BLOCK_SIZE_SB8X16:
return &x->sb8x16_context[xd->sb_index][xd->mb_index][xd->b_index];
case BLOCK_SIZE_SB8X8:
return &x->sb8x8_context[xd->sb_index][xd->mb_index][xd->b_index];
case BLOCK_SIZE_SB8X4:
return &x->sb8x4_context[xd->sb_index][xd->mb_index][xd->b_index];
case BLOCK_SIZE_SB4X8:
return &x->sb4x8_context[xd->sb_index][xd->mb_index][xd->b_index];
case BLOCK_SIZE_AB4X4:
return &x->ab4x4_context[xd->sb_index][xd->mb_index][xd->b_index];
default:
assert(0);
}
}
static BLOCK_SIZE_TYPE *get_sb_partitioning(MACROBLOCK *x,
BLOCK_SIZE_TYPE bsize) {
MACROBLOCKD *xd = &x->e_mbd;
switch (bsize) {
case BLOCK_SIZE_SB64X64:
return &x->sb64_partitioning;
case BLOCK_SIZE_SB32X32:
return &x->sb_partitioning[xd->sb_index];
case BLOCK_SIZE_MB16X16:
return &x->mb_partitioning[xd->sb_index][xd->mb_index];
case BLOCK_SIZE_SB8X8:
return &x->b_partitioning[xd->sb_index][xd->mb_index][xd->b_index];
}
}
static void restore_context(VP9_COMP *cpi, 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],
VP9_COMMON * const cm = &cpi->common;
MACROBLOCK * const x = &cpi->mb;
MACROBLOCKD * const xd = &x->e_mbd;
int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
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++) {
vpx_memcpy(
cm->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);
+ ((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);
}
vpx_memcpy(cm->above_seg_context + mi_col, sa,
vpx_memcpy(cm->left_seg_context + (mi_row & MI_MASK), sl,
static void save_context(VP9_COMP *cpi, 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_TYPE bsize) {
VP9_COMMON * const cm = &cpi->common;
MACROBLOCK * const x = &cpi->mb;
MACROBLOCKD * const xd = &x->e_mbd;
int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
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) {
cm->above_context[p] + (mi_col * 2 >> xd->plane[p].subsampling_x),
(sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
xd->plane[p].subsampling_x);
+ ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
(sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
xd->plane[p].subsampling_y);
}
vpx_memcpy(sl, cm->left_seg_context + (mi_row & MI_MASK),
static void encode_b(VP9_COMP *cpi, TOKENEXTRA **tp, int mi_row, int mi_col,
int output_enabled, BLOCK_SIZE_TYPE bsize, int sub_index) {
VP9_COMMON * const cm = &cpi->common;
MACROBLOCK * const x = &cpi->mb;
MACROBLOCKD * const xd = &x->e_mbd;
if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
return;
if (sub_index != -1)
*(get_sb_index(xd, bsize)) = sub_index;
if (bsize < BLOCK_SIZE_SB8X8)
if (xd->ab_index > 0)
return;
set_offsets(cpi, mi_row, mi_col, bsize);
update_state(cpi, get_block_context(x, bsize), bsize, output_enabled);
encode_superblock(cpi, tp, output_enabled, mi_row, mi_col, bsize);
if (output_enabled) {
update_stats(cpi, mi_row, mi_col);
(*tp)->token = EOSB_TOKEN;
(*tp)++;
static void encode_sb(VP9_COMP *cpi, TOKENEXTRA **tp, int mi_row, int mi_col,
int output_enabled, BLOCK_SIZE_TYPE bsize) {
VP9_COMMON * const cm = &cpi->common;
MACROBLOCK * const x = &cpi->mb;
MACROBLOCKD * const xd = &x->e_mbd;
BLOCK_SIZE_TYPE c1 = BLOCK_SIZE_SB8X8;
const int bsl = b_width_log2(bsize), bs = (1 << bsl) / 4;
PARTITION_TYPE partition;
BLOCK_SIZE_TYPE subsize;
int i;
if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
return;
if (bsize >= BLOCK_SIZE_SB8X8) {
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, bsize);
c1 = *(get_sb_partitioning(x, bsize));
switch (partition) {
case PARTITION_NONE:
if (output_enabled && bsize >= BLOCK_SIZE_SB8X8)
cpi->partition_count[pl][PARTITION_NONE]++;
encode_b(cpi, tp, mi_row, mi_col, output_enabled, c1, -1);
break;
case PARTITION_VERT:
if (output_enabled)
cpi->partition_count[pl][PARTITION_VERT]++;
encode_b(cpi, tp, mi_row, mi_col, output_enabled, c1, 0);
encode_b(cpi, tp, mi_row, mi_col + bs, output_enabled, c1, 1);
break;
case PARTITION_HORZ:
if (output_enabled)
cpi->partition_count[pl][PARTITION_HORZ]++;
encode_b(cpi, tp, mi_row, mi_col, output_enabled, c1, 0);
encode_b(cpi, tp, mi_row + bs, mi_col, output_enabled, c1, 1);
break;
case PARTITION_SPLIT:
subsize = get_subsize(bsize, PARTITION_SPLIT);
if (output_enabled)
cpi->partition_count[pl][PARTITION_SPLIT]++;
for (i = 0; i < 4; i++) {
const int x_idx = i & 1, y_idx = i >> 1;
*(get_sb_index(xd, subsize)) = i;
encode_sb(cpi, tp, mi_row + y_idx * bs, mi_col + x_idx * bs,
output_enabled, subsize);
}
break;
default:
assert(0);
break;
if (partition != PARTITION_SPLIT || bsize == BLOCK_SIZE_SB8X8) {
set_partition_seg_context(cm, xd, mi_row, mi_col);
update_partition_context(xd, c1, bsize);
static void set_partitioning(VP9_COMP *cpi, MODE_INFO *m,
BLOCK_SIZE_TYPE bsize) {
VP9_COMMON *const cm = &cpi->common;
const int mis = cm->mode_info_stride;
int block_row, block_col;
for (block_row = 0; block_row < 8; ++block_row) {
for (block_col = 0; block_col < 8; ++block_col) {
m[block_row * mis + block_col].mbmi.sb_type = bsize;
static void copy_partitioning(VP9_COMP *cpi, MODE_INFO *m, MODE_INFO *p) {
VP9_COMMON *const cm = &cpi->common;
const int mis = cm->mode_info_stride;
int block_row, block_col;
for (block_row = 0; block_row < 8; ++block_row) {
for (block_col = 0; block_col < 8; ++block_col) {
m[block_row * mis + block_col].mbmi.sb_type =
p[block_row * mis + block_col].mbmi.sb_type;
}
}
}
static void set_block_size(VP9_COMMON * const cm, MODE_INFO *m,
BLOCK_SIZE_TYPE bsize, int mis, int mi_row,
int mi_col) {
int bwl = b_width_log2(bsize);
int bhl = b_height_log2(bsize);
int bsl = (bwl > bhl ? bwl : bhl);
int bs = (1 << bsl) / 2; // Block size in units of 8 pels.
MODE_INFO *m2 = m + mi_row * mis + mi_col;
for (row = 0; row < bs; row++) {
for (col = 0; col < bs; col++) {
if (mi_row + row >= cm->mi_rows || mi_col + col >= cm->mi_cols)
typedef struct {
int64_t sum_square_error;
int64_t sum_error;
int count;
int variance;
} var;
typedef struct {
var none;
var horz[2];
var vert[2];
} partition_variance;
#define VT(TYPE, BLOCKSIZE) \
typedef struct { \
BLOCKSIZE split[4]; } TYPE;
VT(v8x8, var)
VT(v16x16, v8x8)
VT(v32x32, v16x16)
VT(v64x64, v32x32)
typedef struct {
partition_variance *vt;
var *split[4];
} vt_node;
typedef enum {
V16X16,
V32X32,
V64X64,
} TREE_LEVEL;
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
static void tree_to_node(void *data, BLOCK_SIZE_TYPE block_size, vt_node *node) {
int i;
switch (block_size) {
case BLOCK_SIZE_SB64X64: {
v64x64 *vt = (v64x64 *) data;
node->vt = &vt->vt;
for (i = 0; i < 4; i++)
node->split[i] = &vt->split[i].vt.none;
break;
}
case BLOCK_SIZE_SB32X32: {
v32x32 *vt = (v32x32 *) data;
node->vt = &vt->vt;
for (i = 0; i < 4; i++)
node->split[i] = &vt->split[i].vt.none;
break;
}
case BLOCK_SIZE_MB16X16: {
v16x16 *vt = (v16x16 *) data;
node->vt = &vt->vt;
for (i = 0; i < 4; i++)
node->split[i] = &vt->split[i].vt.none;
break;
}
case BLOCK_SIZE_SB8X8: {
v8x8 *vt = (v8x8 *) data;
node->vt = &vt->vt;
for (i = 0; i < 4; i++)
node->split[i] = &vt->split[i];
break;
}
default:
node->vt = 0;
for (i = 0; i < 4; i++)
node->split[i] = 0;
assert(-1);
}
}
// Set variance values given sum square error, sum error, count.
static void fill_variance(var *v, int64_t s2, int64_t s, int c) {
v->sum_square_error = s2;
v->sum_error = s;
v->count = c;
if (c > 0)
v->variance = 256
* (v->sum_square_error - v->sum_error * v->sum_error / v->count)
/ v->count;
else
v->variance = 0;
}
// Combine 2 variance structures by summing the sum_error, sum_square_error,
// and counts and then calculating the new variance.
void sum_2_variances(var *r, var *a, var*b) {
fill_variance(r, a->sum_square_error + b->sum_square_error,
a->sum_error + b->sum_error, a->count + b->count);
}
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
static void fill_variance_tree(void *data, BLOCK_SIZE_TYPE block_size) {
vt_node node;
tree_to_node(data, block_size, &node);
sum_2_variances(&node.vt->horz[0], node.split[0], node.split[1]);
sum_2_variances(&node.vt->horz[1], node.split[2], node.split[3]);
sum_2_variances(&node.vt->vert[0], node.split[0], node.split[2]);
sum_2_variances(&node.vt->vert[1], node.split[1], node.split[3]);
sum_2_variances(&node.vt->none, &node.vt->vert[0], &node.vt->vert[1]);
}
#if PERFORM_RANDOM_PARTITIONING
static int set_vt_partitioning(VP9_COMP *cpi, void *data, MODE_INFO *m,
BLOCK_SIZE_TYPE block_size, int mi_row,
int mi_col, int mi_size) {
VP9_COMMON * const cm = &cpi->common;
vt_node vt;
const int mis = cm->mode_info_stride;
int64_t threshold = 4 * cpi->common.base_qindex * cpi->common.base_qindex;
tree_to_node(data, block_size, &vt);
// split none is available only if we have more than half a block size
// in width and height inside the visible image
if (mi_col + mi_size < cm->mi_cols && mi_row + mi_size < cm->mi_rows &&
(rand() & 3) < 1) {
set_block_size(cm, m, block_size, mis, mi_row, mi_col);
return 1;
}
// vertical split is available on all but the bottom border
if (mi_row + mi_size < cm->mi_rows && vt.vt->vert[0].variance < threshold
&& (rand() & 3) < 1) {
set_block_size(cm, m, get_subsize(block_size, PARTITION_VERT), mis, mi_row,
mi_col);
return 1;
}
// horizontal split is available on all but the right border
if (mi_col + mi_size < cm->mi_cols && vt.vt->horz[0].variance < threshold
&& (rand() & 3) < 1) {
set_block_size(cm, m, get_subsize(block_size, PARTITION_HORZ), mis, mi_row,
mi_col);
return 1;
}
return 0;
}
#else
static int set_vt_partitioning(VP9_COMP *cpi, void *data, MODE_INFO *m,
BLOCK_SIZE_TYPE block_size, int mi_row,
int mi_col, int mi_size) {
VP9_COMMON * const cm = &cpi->common;
vt_node vt;
const int mis = cm->mode_info_stride;
int64_t threshold = 50 * cpi->common.base_qindex;
tree_to_node(data, block_size, &vt);
// split none is available only if we have more than half a block size
// in width and height inside the visible image
if (mi_col + mi_size < cm->mi_cols && mi_row + mi_size < cm->mi_rows
&& vt.vt->none.variance < threshold) {
set_block_size(cm, m, block_size, mis, mi_row, mi_col);
return 1;
}
// vertical split is available on all but the bottom border
if (mi_row + mi_size < cm->mi_rows && vt.vt->vert[0].variance < threshold
&& vt.vt->vert[1].variance < threshold) {
set_block_size(cm, m, get_subsize(block_size, PARTITION_VERT), mis, mi_row,
mi_col);
return 1;
// horizontal split is available on all but the right border
if (mi_col + mi_size < cm->mi_cols && vt.vt->horz[0].variance < threshold
&& vt.vt->horz[1].variance < threshold) {
set_block_size(cm, m, get_subsize(block_size, PARTITION_HORZ), mis, mi_row,
mi_col);
return 1;
}
return 0;
}
#endif
static void choose_partitioning(VP9_COMP *cpi, MODE_INFO *m, int mi_row,
int mi_col) {
VP9_COMMON * const cm = &cpi->common;
MACROBLOCK *x = &cpi->mb;
MACROBLOCKD *xd = &cpi->mb.e_mbd;
const int mis = cm->mode_info_stride;
// TODO(JBB): More experimentation or testing of this threshold;
int64_t threshold = 4;
int i, j, k;
v64x64 vt;
unsigned char * s;
int sp;
int pixels_wide = 64, pixels_high = 64;
set_offsets(cpi, mi_row, mi_col, BLOCK_SIZE_SB64X64);
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;
// TODO(JBB): Clearly the higher the quantizer the fewer partitions we want
// but this needs more experimentation.
threshold = threshold * cpi->common.base_qindex * cpi->common.base_qindex;
d = vp9_64x64_zeros;
dp = 64;
YV12_BUFFER_CONFIG *ref_fb = &cm->yv12_fb[0];
YV12_BUFFER_CONFIG *second_ref_fb = NULL;
setup_pre_planes(xd, 0, ref_fb, mi_row, mi_col,
setup_pre_planes(xd, 1, second_ref_fb, mi_row, mi_col,
xd->mode_info_context->mbmi.ref_frame[0] = LAST_FRAME;
xd->mode_info_context->mbmi.sb_type = BLOCK_SIZE_SB64X64;
vp9_find_best_ref_mvs(xd, m->mbmi.ref_mvs[m->mbmi.ref_frame[0]],
&nearest_mv, &near_mv);
xd->mode_info_context->mbmi.mv[0] = nearest_mv;
vp9_build_inter_predictors_sby(xd, mi_row, mi_col, BLOCK_SIZE_SB64X64);
d = xd->plane[0].dst.buf;
dp = xd->plane[0].dst.stride;
}
// Fill in the entire tree of 8x8 variances for splits.
for (i = 0; i < 4; i++) {
const int x32_idx = ((i & 1) << 5);
const int y32_idx = ((i >> 1) << 5);
for (j = 0; j < 4; j++) {
const int x16_idx = x32_idx + ((j & 1) << 4);
const int y16_idx = y32_idx + ((j >> 1) << 4);
v16x16 *vst = &vt.split[i].split[j];
for (k = 0; k < 4; k++) {
int x_idx = x16_idx + ((k & 1) << 3);
int y_idx = y16_idx + ((k >> 1) << 3);
unsigned int sse = 0;
int sum = 0;
if (x_idx < pixels_wide && y_idx < pixels_high)
vp9_get_sse_sum_8x8(s + y_idx * sp + x_idx, sp,
d + y_idx * dp + x_idx, dp, &sse, &sum);
fill_variance(&vst->split[k].vt.none, sse, sum, 64);
}
}
}
// Fill the rest of the variance tree by summing the split partition
// values.
for (i = 0; i < 4; i++) {
for (j = 0; j < 4; j++) {
fill_variance_tree(&vt.split[i].split[j], BLOCK_SIZE_MB16X16);
fill_variance_tree(&vt.split[i], BLOCK_SIZE_SB32X32);
fill_variance_tree(&vt, BLOCK_SIZE_SB64X64);
// Now go through the entire structure, splitting every block size until
// we get to one that's got a variance lower than our threshold, or we
// hit 8x8.
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
if (!set_vt_partitioning(cpi, &vt, m, BLOCK_SIZE_SB64X64, mi_row, mi_col,
4)) {
for (i = 0; i < 4; ++i) {
const int x32_idx = ((i & 1) << 2);
const int y32_idx = ((i >> 1) << 2);
if (!set_vt_partitioning(cpi, &vt.split[i], m, BLOCK_SIZE_SB32X32,
(mi_row + y32_idx), (mi_col + x32_idx), 2)) {
for (j = 0; j < 4; ++j) {
const int x16_idx = ((j & 1) << 1);
const int y16_idx = ((j >> 1) << 1);
if (!set_vt_partitioning(cpi, &vt.split[i].split[j], m,
BLOCK_SIZE_MB16X16,
(mi_row + y32_idx + y16_idx),
(mi_col + x32_idx + x16_idx), 1)) {
for (k = 0; k < 4; ++k) {
const int x8_idx = (k & 1);
const int y8_idx = (k >> 1);
set_block_size(cm, m, BLOCK_SIZE_SB8X8, mis,
(mi_row + y32_idx + y16_idx + y8_idx),
(mi_col + x32_idx + x16_idx + x8_idx));
}
}
}
}
}
}
}
static void rd_use_partition(VP9_COMP *cpi, MODE_INFO *m, TOKENEXTRA **tp,
int mi_row, int mi_col, BLOCK_SIZE_TYPE bsize,
int *rate, int64_t *dist, int do_recon) {
VP9_COMMON * const cm = &cpi->common;
MACROBLOCK * const x = &cpi->mb;
MACROBLOCKD *xd = &cpi->mb.e_mbd;
const int mis = cm->mode_info_stride;
int bsl = b_width_log2(bsize);
int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
int ms = num_4x4_blocks_wide / 2;
int mh = num_4x4_blocks_high / 2;
PARTITION_TYPE partition = PARTITION_NONE;
BLOCK_SIZE_TYPE subsize;
ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
PARTITION_CONTEXT sl[8], sa[8];
int last_part_rate = INT_MAX;
int64_t last_part_dist = INT_MAX;
int split_rate = INT_MAX;
int64_t split_dist = INT_MAX;
int none_rate = INT_MAX;
int64_t none_dist = INT_MAX;
int chosen_rate = INT_MAX;
int64_t chosen_dist = INT_MAX;
BLOCK_SIZE_TYPE sub_subsize = BLOCK_SIZE_AB4X4;
int splits_below = 0;
BLOCK_SIZE_TYPE bs_type = m->mbmi.sb_type;
if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
return;
subsize = get_subsize(bsize, partition);
if (bsize < BLOCK_SIZE_SB8X8) {
if (xd->ab_index != 0) {
*rate = 0;
*dist = 0;
return;
}
} else {
*(get_sb_partitioning(x, bsize)) = subsize;
}
save_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
x->fast_ms = 0;
x->pred_mv.as_int = 0;
x->subblock_ref = 0;
if (cpi->sf.adjust_partitioning_from_last_frame) {
// Check if any of the sub blocks are further split.
if (partition == PARTITION_SPLIT && subsize > BLOCK_SIZE_SB8X8) {
sub_subsize = get_subsize(subsize, PARTITION_SPLIT);
splits_below = 1;
for (i = 0; i < 4; i++) {
int jj = i >> 1, ii = i & 0x01;
if (m[jj * bss * mis + ii * bss].mbmi.sb_type >= sub_subsize) {
splits_below = 0;
}
}
}
// If partition is not none try none unless each of the 4 splits are split
// even further..
if (partition != PARTITION_NONE && !splits_below &&
mi_row + (ms >> 1) < cm->mi_rows &&
mi_col + (ms >> 1) < cm->mi_cols) {
*(get_sb_partitioning(x, bsize)) = bsize;
pick_sb_modes(cpi, mi_row, mi_col, &none_rate, &none_dist, bsize,
get_block_context(x, bsize), INT64_MAX);
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, bsize);
none_rate += x->partition_cost[pl][PARTITION_NONE];
restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
m->mbmi.sb_type = bs_type;
*(get_sb_partitioning(x, bsize)) = subsize;
}
}
pick_sb_modes(cpi, mi_row, mi_col, &last_part_rate, &last_part_dist,
bsize, get_block_context(x, bsize), INT64_MAX);
break;
case PARTITION_HORZ:
*(get_sb_index(xd, subsize)) = 0;
pick_sb_modes(cpi, mi_row, mi_col, &last_part_rate, &last_part_dist,
subsize, get_block_context(x, subsize), INT64_MAX);
if (last_part_rate != INT_MAX &&
bsize >= BLOCK_SIZE_SB8X8 && mi_row + (mh >> 1) < cm->mi_rows) {
update_state(cpi, get_block_context(x, subsize), subsize, 0);
encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize);
*(get_sb_index(xd, subsize)) = 1;
pick_sb_modes(cpi, mi_row + (ms >> 1), mi_col, &rt, &dt, subsize,
get_block_context(x, subsize), INT64_MAX);
if (rt == INT_MAX || dt == INT_MAX) {
last_part_rate = INT_MAX;
last_part_dist = INT_MAX;
break;
}
last_part_rate += rt;
last_part_dist += dt;
}
break;
case PARTITION_VERT:
*(get_sb_index(xd, subsize)) = 0;
pick_sb_modes(cpi, mi_row, mi_col, &last_part_rate, &last_part_dist,
subsize, get_block_context(x, subsize), INT64_MAX);
if (last_part_rate != INT_MAX &&
bsize >= BLOCK_SIZE_SB8X8 && mi_col + (ms >> 1) < cm->mi_cols) {
update_state(cpi, get_block_context(x, subsize), subsize, 0);
encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize);
*(get_sb_index(xd, subsize)) = 1;
pick_sb_modes(cpi, mi_row, mi_col + (ms >> 1), &rt, &dt, subsize,
get_block_context(x, subsize), INT64_MAX);
if (rt == INT_MAX || dt == INT_MAX) {
last_part_rate = INT_MAX;
last_part_dist = INT_MAX;
break;
}
last_part_rate += rt;
last_part_dist += dt;
// Split partition.
last_part_rate = 0;
last_part_dist = 0;
int x_idx = (i & 1) * (ms >> 1);
int y_idx = (i >> 1) * (ms >> 1);
if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
continue;
*(get_sb_index(xd, subsize)) = i;
rd_use_partition(cpi, m + jj * bss * mis + ii * bss, tp, mi_row + y_idx,
mi_col + x_idx, subsize, &rt, &dt, i != 3);
if (rt == INT_MAX || dt == INT_MAX) {
last_part_rate = INT_MAX;
last_part_dist = INT_MAX;
break;
}
last_part_rate += rt;
last_part_dist += dt;
}
break;
default:
assert(0);
}
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, bsize);
if (last_part_rate < INT_MAX)
last_part_rate += x->partition_cost[pl][partition];
if (cpi->sf.adjust_partitioning_from_last_frame
&& partition != PARTITION_SPLIT && bsize > BLOCK_SIZE_SB8X8
&& (mi_row + ms < cm->mi_rows || mi_row + (ms >> 1) == cm->mi_rows)
&& (mi_col + ms < cm->mi_cols || mi_col + (ms >> 1) == cm->mi_cols)) {
BLOCK_SIZE_TYPE split_subsize = get_subsize(bsize, PARTITION_SPLIT);
split_rate = 0;
split_dist = 0;
restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
// Split partition.
for (i = 0; i < 4; i++) {
int x_idx = (i & 1) * (num_4x4_blocks_wide >> 2);
int y_idx = (i >> 1) * (num_4x4_blocks_wide >> 2);
int rt = 0;
int64_t dt = 0;
ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
PARTITION_CONTEXT sl[8], sa[8];
if ((mi_row + y_idx >= cm->mi_rows)
|| (mi_col + x_idx >= cm->mi_cols))
continue;
*(get_sb_index(xd, split_subsize)) = i;
*(get_sb_partitioning(x, bsize)) = split_subsize;
*(get_sb_partitioning(x, split_subsize)) = split_subsize;
save_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
pick_sb_modes(cpi, mi_row + y_idx, mi_col + x_idx, &rt, &dt,
split_subsize, get_block_context(x, split_subsize),
INT64_MAX);
restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
if (rt == INT_MAX || dt == INT_MAX) {
split_rate = INT_MAX;
split_dist = INT_MAX;
break;
}
if (i != 3)
encode_sb(cpi, tp, mi_row + y_idx, mi_col + x_idx, 0,
split_subsize);
split_rate += rt;
split_dist += dt;
set_partition_seg_context(cm, xd, mi_row + y_idx, mi_col + x_idx);
pl = partition_plane_context(xd, bsize);
split_rate += x->partition_cost[pl][PARTITION_NONE];
}
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, bsize);
if (split_rate < INT_MAX) {
split_rate += x->partition_cost[pl][PARTITION_SPLIT];
chosen_rate = split_rate;
chosen_dist = split_dist;
}
}
// If last_part is better set the partitioning to that...
if (RDCOST(x->rdmult, x->rddiv, last_part_rate, last_part_dist)
< RDCOST(x->rdmult, x->rddiv, chosen_rate, chosen_dist)) {
m->mbmi.sb_type = bsize;
if (bsize >= BLOCK_SIZE_SB8X8)
*(get_sb_partitioning(x, bsize)) = subsize;
chosen_rate = last_part_rate;
chosen_dist = last_part_dist;
}
// If none was better set the partitioning to that...
if (RDCOST(x->rdmult, x->rddiv, chosen_rate, chosen_dist)
> RDCOST(x->rdmult, x->rddiv, none_rate, none_dist)) {
if (bsize >= BLOCK_SIZE_SB8X8)
*(get_sb_partitioning(x, bsize)) = bsize;
chosen_rate = none_rate;
chosen_dist = none_dist;
}
restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
// We must have chosen a partitioning and encoding or we'll fail later on.
// No other opportunities for success.
if ( bsize == BLOCK_SIZE_SB64X64)
assert(chosen_rate < INT_MAX && chosen_dist < INT_MAX);
if (do_recon)
encode_sb(cpi, tp, mi_row, mi_col, bsize == BLOCK_SIZE_SB64X64, bsize);
*rate = chosen_rate;
*dist = chosen_dist;
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
static BLOCK_SIZE_TYPE min_partition_size[BLOCK_SIZE_TYPES] =
{ BLOCK_4X4, BLOCK_4X4, BLOCK_4X4, BLOCK_4X4,
BLOCK_4X4, BLOCK_4X4, BLOCK_8X8, BLOCK_8X8,
BLOCK_8X8, BLOCK_16X16, BLOCK_16X16, BLOCK_16X16, BLOCK_16X16 };
static BLOCK_SIZE_TYPE max_partition_size[BLOCK_SIZE_TYPES] =
{ BLOCK_8X8, BLOCK_16X16, BLOCK_16X16, BLOCK_16X16,
BLOCK_32X32, BLOCK_32X32, BLOCK_32X32, BLOCK_64X64,
BLOCK_64X64, BLOCK_64X64, BLOCK_64X64, BLOCK_64X64, BLOCK_64X64 };
// Look at neighbouring blocks and set a min and max partition size based on
// what they chose.
static void rd_auto_partition_range(VP9_COMP *cpi,
BLOCK_SIZE_TYPE * min_block_size,
BLOCK_SIZE_TYPE * max_block_size) {
MACROBLOCKD *const xd = &cpi->mb.e_mbd;
const MODE_INFO *const mi = xd->mode_info_context;
const MB_MODE_INFO *const above_mbmi = &mi[-xd->mode_info_stride].mbmi;
const MB_MODE_INFO *const left_mbmi = &mi[-1].mbmi;
const int left_in_image = xd->left_available && left_mbmi->mb_in_image;
const int above_in_image = xd->up_available && above_mbmi->mb_in_image;
// Frequency check
if (cpi->sf.auto_min_max_partition_count <= 0) {
cpi->sf.auto_min_max_partition_count =
cpi->sf.auto_min_max_partition_interval;
*min_block_size = BLOCK_4X4;
*max_block_size = BLOCK_64X64;
return;
} else {
--cpi->sf.auto_min_max_partition_count;
}
// Check for edge cases
if (!left_in_image && !above_in_image) {
*min_block_size = BLOCK_4X4;
*max_block_size = BLOCK_64X64;
} else if (!left_in_image) {
*min_block_size = min_partition_size[above_mbmi->sb_type];
*max_block_size = max_partition_size[above_mbmi->sb_type];
} else if (!above_in_image) {
*min_block_size = min_partition_size[left_mbmi->sb_type];
*max_block_size = max_partition_size[left_mbmi->sb_type];
} else {
*min_block_size =
min_partition_size[MIN(left_mbmi->sb_type, above_mbmi->sb_type)];
*max_block_size =
max_partition_size[MAX(left_mbmi->sb_type, above_mbmi->sb_type)];
}
}
// TODO(jingning,jimbankoski,rbultje): properly skip partition types that are
// unlikely to be selected depending on previously rate-distortion optimization
// results, for encoding speed-up.
static void rd_pick_partition(VP9_COMP *cpi, TOKENEXTRA **tp, int mi_row,
int mi_col, BLOCK_SIZE_TYPE bsize, int *rate,
int64_t *dist, int do_recon, int64_t best_rd) {
VP9_COMMON * const cm = &cpi->common;
MACROBLOCK * const x = &cpi->mb;
MACROBLOCKD * const xd = &x->e_mbd;
int bsl = b_width_log2(bsize), bs = 1 << bsl;
ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
PARTITION_CONTEXT sl[8], sa[8];
TOKENEXTRA *tp_orig = *tp;
int srate = INT_MAX;
int64_t sdist = INT_MAX;
if (bsize < BLOCK_SIZE_SB8X8)
if (xd->ab_index != 0) {
*rate = 0;
*dist = 0;
return;
}
assert(mi_height_log2(bsize) == mi_width_log2(bsize));
save_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
if (!cpi->sf.auto_min_max_partition_size ||
bsize >= cpi->sf.min_partition_size) {
if (bsize > BLOCK_SIZE_SB8X8) {
int64_t d4 = 0, sum_rd = 0;
subsize = get_subsize(bsize, PARTITION_SPLIT);
for (i = 0; i < 4 && sum_rd < best_rd; ++i) {
int x_idx = (i & 1) * (ms >> 1);
int y_idx = (i >> 1) * (ms >> 1);
if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
continue;
*(get_sb_index(xd, subsize)) = i;
rd_pick_partition(cpi, tp, mi_row + y_idx, mi_col + x_idx, subsize, &r,
&d, i != 3, best_rd - sum_rd);
if (r == INT_MAX) {
r4 = INT_MAX;
sum_rd = INT64_MAX;
} else {
r4 += r;
d4 += d;
sum_rd = RDCOST(x->rdmult, x->rddiv, r4, d4);
}
}
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, bsize);
if (r4 != INT_MAX && i == 4) {
r4 += x->partition_cost[pl][PARTITION_SPLIT];
*(get_sb_partitioning(x, bsize)) = subsize;
assert(r4 >= 0);
assert(d4 >= 0);
srate = r4;
sdist = d4;
best_rd = MIN(best_rd, RDCOST(x->rdmult, x->rddiv, r4, d4));
}
restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
x->fast_ms = 0;
x->pred_mv.as_int = 0;
x->subblock_ref = 0;
// Use 4 subblocks' motion estimation results to speed up current
// partition's checking.
if (cpi->sf.using_small_partition_info) {
// Only use 8x8 result for non HD videos.
// int use_8x8 = (MIN(cpi->common.width, cpi->common.height) < 720) ? 1 : 0;
int use_8x8 = 1;
if (cm->frame_type && !cpi->is_src_frame_alt_ref &&
((use_8x8 && bsize == BLOCK_SIZE_MB16X16) ||
bsize == BLOCK_SIZE_SB32X32 || bsize == BLOCK_SIZE_SB64X64)) {
int ref0 = 0, ref1 = 0, ref2 = 0, ref3 = 0;
PICK_MODE_CONTEXT *block_context = NULL;
if (bsize == BLOCK_SIZE_MB16X16) {
block_context = x->sb8x8_context[xd->sb_index][xd->mb_index];
} else if (bsize == BLOCK_SIZE_SB32X32) {
block_context = x->mb_context[xd->sb_index];
} else if (bsize == BLOCK_SIZE_SB64X64) {
block_context = x->sb32_context;
}
if (block_context) {
ref0 = block_context[0].mic.mbmi.ref_frame[0];
ref1 = block_context[1].mic.mbmi.ref_frame[0];
ref2 = block_context[2].mic.mbmi.ref_frame[0];
ref3 = block_context[3].mic.mbmi.ref_frame[0];
}
// Currently, only consider 4 inter ref frames.
if (ref0 && ref1 && ref2 && ref3) {
int16_t mvr0 = 0, mvc0 = 0, mvr1 = 0, mvc1 = 0, mvr2 = 0, mvc2 = 0,
mvr3 = 0, mvc3 = 0;
int d01, d23, d02, d13; // motion vector distance between 2 blocks
// Get each subblock's motion vectors.
mvr0 = block_context[0].mic.mbmi.mv[0].as_mv.row;
mvc0 = block_context[0].mic.mbmi.mv[0].as_mv.col;
mvr1 = block_context[1].mic.mbmi.mv[0].as_mv.row;
mvc1 = block_context[1].mic.mbmi.mv[0].as_mv.col;
mvr2 = block_context[2].mic.mbmi.mv[0].as_mv.row;
mvc2 = block_context[2].mic.mbmi.mv[0].as_mv.col;
mvr3 = block_context[3].mic.mbmi.mv[0].as_mv.row;
mvc3 = block_context[3].mic.mbmi.mv[0].as_mv.col;
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
// Adjust sign if ref is alt_ref
if (cm->ref_frame_sign_bias[ref0]) {
mvr0 *= -1;
mvc0 *= -1;
}
if (cm->ref_frame_sign_bias[ref1]) {
mvr1 *= -1;
mvc1 *= -1;
}
if (cm->ref_frame_sign_bias[ref2]) {
mvr2 *= -1;
mvc2 *= -1;
}
if (cm->ref_frame_sign_bias[ref3]) {
mvr3 *= -1;
mvc3 *= -1;
}
// Calculate mv distances.
d01 = MAX(abs(mvr0 - mvr1), abs(mvc0 - mvc1));
d23 = MAX(abs(mvr2 - mvr3), abs(mvc2 - mvc3));
d02 = MAX(abs(mvr0 - mvr2), abs(mvc0 - mvc2));
d13 = MAX(abs(mvr1 - mvr3), abs(mvc1 - mvc3));
if (d01 < 24 && d23 < 24 && d02 < 24 && d13 < 24) {
// Set fast motion search level.
x->fast_ms = 1;
// Calculate prediction MV
x->pred_mv.as_mv.row = (mvr0 + mvr1 + mvr2 + mvr3) >> 2;
x->pred_mv.as_mv.col = (mvc0 + mvc1 + mvc2 + mvc3) >> 2;
if (ref0 == ref1 && ref1 == ref2 && ref2 == ref3 &&
d01 < 2 && d23 < 2 && d02 < 2 && d13 < 2) {
// Set fast motion search level.
x->fast_ms = 2;
if (!d01 && !d23 && !d02 && !d13) {
x->fast_ms = 3;
x->subblock_ref = ref0;
}
}
}
}
}
}
if (!cpi->sf.use_max_partition_size ||
bsize <= cpi->sf.max_partition_size) {
// PARTITION_NONE
if ((mi_row + (ms >> 1) < cm->mi_rows) &&
(mi_col + (ms >> 1) < cm->mi_cols)) {
pick_sb_modes(cpi, mi_row, mi_col, &r, &d, bsize,
get_block_context(x, bsize), best_rd);
if (r != INT_MAX && bsize >= BLOCK_SIZE_SB8X8) {
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, bsize);
r += x->partition_cost[pl][PARTITION_NONE];
}
if (r != INT_MAX &&
(bsize == BLOCK_SIZE_SB8X8 ||
RDCOST(x->rdmult, x->rddiv, r, d) <
RDCOST(x->rdmult, x->rddiv, srate, sdist))) {
best_rd = MIN(best_rd, RDCOST(x->rdmult, x->rddiv, r, d));
srate = r;
sdist = d;
if (bsize >= BLOCK_SIZE_SB8X8)
*(get_sb_partitioning(x, bsize)) = bsize;
}
}
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
if (bsize == BLOCK_SIZE_SB8X8) {
int r4 = 0;
int64_t d4 = 0, sum_rd = 0;
subsize = get_subsize(bsize, PARTITION_SPLIT);
for (i = 0; i < 4 && sum_rd < best_rd; ++i) {
int x_idx = (i & 1) * (ms >> 1);
int y_idx = (i >> 1) * (ms >> 1);
int r = 0;
int64_t d = 0;
if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
continue;
*(get_sb_index(xd, subsize)) = i;
rd_pick_partition(cpi, tp, mi_row + y_idx, mi_col + x_idx, subsize, &r,
&d, i != 3, best_rd - sum_rd);
if (r == INT_MAX) {
r4 = INT_MAX;
sum_rd = INT64_MAX;
} else {
r4 += r;
d4 += d;
sum_rd = RDCOST(x->rdmult, x->rddiv, r4, d4);
}
}
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, bsize);
if (r4 != INT_MAX && i == 4) {
r4 += x->partition_cost[pl][PARTITION_SPLIT];
if (RDCOST(x->rdmult, x->rddiv, r4, d4) <
RDCOST(x->rdmult, x->rddiv, srate, sdist)) {
srate = r4;
sdist = d4;
*(get_sb_partitioning(x, bsize)) = subsize;
best_rd = MIN(best_rd, RDCOST(x->rdmult, x->rddiv, r4, d4));
}
}
restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
}
if (!cpi->sf.use_square_partition_only &&
(!cpi->sf.less_rectangular_check ||!larger_is_better)) {
// PARTITION_HORZ
if (bsize >= BLOCK_SIZE_SB8X8 && mi_col + (ms >> 1) < cm->mi_cols) {
int r2, r = 0;
subsize = get_subsize(bsize, PARTITION_HORZ);
*(get_sb_index(xd, subsize)) = 0;
pick_sb_modes(cpi, mi_row, mi_col, &r2, &d2, subsize,
get_block_context(x, subsize), best_rd);
h_rd = RDCOST(x->rdmult, x->rddiv, r2, d2);
if (r2 != INT_MAX && h_rd < best_rd &&
mi_row + (ms >> 1) < cm->mi_rows) {
update_state(cpi, get_block_context(x, subsize), subsize, 0);
encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize);
*(get_sb_index(xd, subsize)) = 1;
pick_sb_modes(cpi, mi_row + (ms >> 1), mi_col, &r, &d, subsize,
get_block_context(x, subsize), best_rd - h_rd);
if (r == INT_MAX) {
r2 = INT_MAX;
} else {
r2 += r;
d2 += d;
}
}
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, bsize);
if (r2 < INT_MAX)
r2 += x->partition_cost[pl][PARTITION_HORZ];
if (r2 != INT_MAX && RDCOST(x->rdmult, x->rddiv, r2, d2)
< RDCOST(x->rdmult, x->rddiv, srate, sdist)) {
best_rd = MIN(best_rd, RDCOST(x->rdmult, x->rddiv, r2, d2));
srate = r2;
sdist = d2;
*(get_sb_partitioning(x, bsize)) = subsize;
}
restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
}
// PARTITION_VERT
if (bsize >= BLOCK_SIZE_SB8X8 && mi_row + (ms >> 1) < cm->mi_rows) {
int r2;
subsize = get_subsize(bsize, PARTITION_VERT);
*(get_sb_index(xd, subsize)) = 0;
pick_sb_modes(cpi, mi_row, mi_col, &r2, &d2, subsize,
get_block_context(x, subsize), best_rd);
v_rd = RDCOST(x->rdmult, x->rddiv, r2, d2);
if (r2 != INT_MAX && v_rd < best_rd &&
mi_col + (ms >> 1) < cm->mi_cols) {
int r = 0;
int64_t d = 0;
update_state(cpi, get_block_context(x, subsize), subsize, 0);
encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize);
*(get_sb_index(xd, subsize)) = 1;
pick_sb_modes(cpi, mi_row, mi_col + (ms >> 1), &r, &d, subsize,
get_block_context(x, subsize), best_rd - v_rd);
if (r == INT_MAX) {
r2 = INT_MAX;
} else {
r2 += r;
d2 += d;
}
}
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, bsize);
if (r2 < INT_MAX)
r2 += x->partition_cost[pl][PARTITION_VERT];
if (r2 != INT_MAX &&
RDCOST(x->rdmult, x->rddiv, r2, d2)
< RDCOST(x->rdmult, x->rddiv, srate, sdist)) {
srate = r2;
sdist = d2;
*(get_sb_partitioning(x, bsize)) = subsize;
}
restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
}
}
}
*rate = srate;
*dist = sdist;
restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
if (srate < INT_MAX && sdist < INT_MAX && do_recon)
encode_sb(cpi, tp, mi_row, mi_col, bsize == BLOCK_SIZE_SB64X64, bsize);
if (bsize == BLOCK_SIZE_SB64X64) {
assert(srate < INT_MAX);
assert(sdist < INT_MAX);
} else {
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
// Examines 64x64 block and chooses a best reference frame
static void rd_pick_reference_frame(VP9_COMP *cpi, TOKENEXTRA **tp, int mi_row,
int mi_col, int *rate, int64_t *dist) {
VP9_COMMON * const cm = &cpi->common;
MACROBLOCK * const x = &cpi->mb;
MACROBLOCKD * const xd = &x->e_mbd;
int bsl = b_width_log2(BLOCK_SIZE_SB64X64), bs = 1 << bsl;
int ms = bs / 2;
ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
PARTITION_CONTEXT sl[8], sa[8];
int pl;
int r;
int64_t d;
save_context(cpi, mi_row, mi_col, a, l, sa, sl, BLOCK_SIZE_SB64X64);
// Default is non mask (all reference frames allowed.
cpi->ref_frame_mask = 0;
// Do RD search for 64x64.
if ((mi_row + (ms >> 1) < cm->mi_rows) &&
(mi_col + (ms >> 1) < cm->mi_cols)) {
cpi->set_ref_frame_mask = 1;
pick_sb_modes(cpi, mi_row, mi_col, &r, &d, BLOCK_SIZE_SB64X64,
get_block_context(x, BLOCK_SIZE_SB64X64), INT64_MAX);
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
set_partition_seg_context(cm, xd, mi_row, mi_col);
pl = partition_plane_context(xd, BLOCK_SIZE_SB64X64);
r += x->partition_cost[pl][PARTITION_NONE];
*(get_sb_partitioning(x, BLOCK_SIZE_SB64X64)) = BLOCK_SIZE_SB64X64;
cpi->set_ref_frame_mask = 0;
}
*rate = r;
*dist = d;
// RDCOST(x->rdmult, x->rddiv, r, d)
restore_context(cpi, mi_row, mi_col, a, l, sa, sl, BLOCK_SIZE_SB64X64);
/*if (srate < INT_MAX && sdist < INT_MAX)
encode_sb(cpi, tp, mi_row, mi_col, 1, BLOCK_SIZE_SB64X64);
if (bsize == BLOCK_SIZE_SB64X64) {
assert(tp_orig < *tp);
assert(srate < INT_MAX);
assert(sdist < INT_MAX);
} else {
assert(tp_orig == *tp);
}
*/
}
static void encode_sb_row(VP9_COMP *cpi, int mi_row, TOKENEXTRA **tp,
int *totalrate) {
VP9_COMMON * const cm = &cpi->common;
// Initialize the left context for the new SB row
vpx_memset(&cm->left_context, 0, sizeof(cm->left_context));
vpx_memset(cm->left_seg_context, 0, sizeof(cm->left_seg_context));
// Code each SB in the row
for (mi_col = cm->cur_tile_mi_col_start; mi_col < cm->cur_tile_mi_col_end;
int dummy_rate;
int64_t dummy_dist;
// Initialize a mask of modes that we will not consider;
// cpi->unused_mode_skip_mask = 0x0000000AAE17F800 (test no golden)
if (cpi->common.frame_type == KEY_FRAME)
cpi->unused_mode_skip_mask = 0;
else
cpi->unused_mode_skip_mask = 0xFFFFFFFFFFFFFE00;
if (cpi->sf.reference_masking) {
rd_pick_reference_frame(cpi, tp, mi_row, mi_col,
&dummy_rate, &dummy_dist);
}
if (cpi->sf.partition_by_variance || cpi->sf.use_lastframe_partitioning ||
cpi->sf.use_one_partition_size_always ) {
const int idx_str = cm->mode_info_stride * mi_row + mi_col;
MODE_INFO *m = cm->mi + idx_str;
MODE_INFO *p = cm->prev_mi + idx_str;
if (cpi->sf.use_one_partition_size_always) {
set_offsets(cpi, mi_row, mi_col, BLOCK_SIZE_SB64X64);
set_partitioning(cpi, m, cpi->sf.always_this_block_size);
rd_use_partition(cpi, m, tp, mi_row, mi_col, BLOCK_SIZE_SB64X64,
&dummy_rate, &dummy_dist, 1);
} else if (cpi->sf.partition_by_variance) {
choose_partitioning(cpi, cm->mi, mi_row, mi_col);
rd_use_partition(cpi, m, tp, mi_row, mi_col, BLOCK_SIZE_SB64X64,
&dummy_rate, &dummy_dist, 1);
if ((cpi->common.current_video_frame
% cpi->sf.last_partitioning_redo_frequency) == 0
|| cm->prev_mi == 0
|| cpi->common.show_frame == 0
|| cpi->common.frame_type == KEY_FRAME
|| cpi->is_src_frame_alt_ref) {
// If required set upper and lower partition size limits
if (cpi->sf.auto_min_max_partition_size) {
rd_auto_partition_range(cpi,
&cpi->sf.min_partition_size,
&cpi->sf.max_partition_size);
}
rd_pick_partition(cpi, tp, mi_row, mi_col, BLOCK_SIZE_SB64X64,
&dummy_rate, &dummy_dist, 1, INT64_MAX);
} else {
copy_partitioning(cpi, m, p);
rd_use_partition(cpi, m, tp, mi_row, mi_col, BLOCK_SIZE_SB64X64,
&dummy_rate, &dummy_dist, 1);
// If required set upper and lower partition size limits
if (cpi->sf.auto_min_max_partition_size) {
rd_auto_partition_range(cpi,
&cpi->sf.min_partition_size,
&cpi->sf.max_partition_size);
}
rd_pick_partition(cpi, tp, mi_row, mi_col, BLOCK_SIZE_SB64X64,
&dummy_rate, &dummy_dist, 1, INT64_MAX);
static void init_encode_frame_mb_context(VP9_COMP *cpi) {
MACROBLOCK *const x = &cpi->mb;
VP9_COMMON *const cm = &cpi->common;
MACROBLOCKD *const xd = &x->e_mbd;
const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
// reset intra mode contexts
if (cm->frame_type == KEY_FRAME)
vp9_init_mbmode_probs(cm);
// TODO(jkoleszar): are these initializations required?
setup_pre_planes(xd, 0, &cm->yv12_fb[cm->ref_frame_map[cpi->lst_fb_idx]],
setup_dst_planes(xd, &cm->yv12_fb[cm->new_fb_idx], 0, 0);
setup_block_dptrs(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
xd->mode_info_context->mbmi.mode = DC_PRED;
xd->mode_info_context->mbmi.uv_mode = DC_PRED;
vp9_zero(cpi->y_mode_count)
vp9_zero(cm->counts.inter_mode)
vp9_zero(cpi->intra_inter_count);
vp9_zero(cpi->comp_inter_count);
vp9_zero(cpi->single_ref_count);
vp9_zero(cpi->comp_ref_count);
vp9_zero(cm->counts.tx);
vp9_zero(cm->counts.mbskip);
// Note: this memset assumes above_context[0], [1] and [2]
// are allocated as part of the same buffer.
vpx_memset(cm->above_context[0], 0,
sizeof(ENTROPY_CONTEXT) * 2 * MAX_MB_PLANE * aligned_mi_cols);
sizeof(PARTITION_CONTEXT) * aligned_mi_cols);
static void switch_lossless_mode(VP9_COMP *cpi, int lossless) {
if (lossless) {
cpi->mb.fwd_txm8x4 = vp9_short_walsh8x4;
cpi->mb.fwd_txm4x4 = vp9_short_walsh4x4;
cpi->mb.e_mbd.inv_txm4x4_1_add = vp9_short_iwalsh4x4_1_add;
cpi->mb.e_mbd.inv_txm4x4_add = vp9_short_iwalsh4x4_add;
cpi->mb.optimize = 0;
cpi->mb.e_mbd.lf.filter_level = 0;
cpi->common.tx_mode = ONLY_4X4;
cpi->mb.fwd_txm8x4 = vp9_short_fdct8x4;
cpi->mb.fwd_txm4x4 = vp9_short_fdct4x4;
cpi->mb.e_mbd.inv_txm4x4_1_add = vp9_short_idct4x4_1_add;
cpi->mb.e_mbd.inv_txm4x4_add = vp9_short_idct4x4_add;
static void switch_tx_mode(VP9_COMP *cpi) {
if (cpi->sf.tx_size_search_method == USE_LARGESTALL &&
cpi->common.tx_mode >= ALLOW_32X32)
cpi->common.tx_mode = ALLOW_32X32;
static void encode_frame_internal(VP9_COMP *cpi) {
MACROBLOCK * const x = &cpi->mb;
VP9_COMMON * const cm = &cpi->common;
MACROBLOCKD * const xd = &x->e_mbd;
// fprintf(stderr, "encode_frame_internal frame %d (%d) type %d\n",
// cpi->common.current_video_frame, cpi->common.show_frame,
// cm->frame_type);
{
FILE *statsfile;
statsfile = fopen("segmap2.stt", "a");
fprintf(statsfile, "\n");
fclose(statsfile);
}
totalrate = 0;
// Reset frame count of inter 0,0 motion vector usage.
cpi->inter_zz_count = 0;
vp9_zero(cm->counts.switchable_interp);
xd->mode_info_context = cm->mi;
xd->prev_mode_info_context = cm->prev_mi;
vp9_zero(cpi->coef_counts);
vp9_zero(cm->counts.eob_branch);
cpi->mb.e_mbd.lossless = cm->base_qindex == 0 && cm->y_dc_delta_q == 0
&& cm->uv_dc_delta_q == 0 && cm->uv_ac_delta_q == 0;
vp9_frame_init_quantizer(cpi);
vp9_initialize_rd_consts(cpi, cm->base_qindex + cm->y_dc_delta_q);
vp9_initialize_me_consts(cpi, cm->base_qindex);
switch_tx_mode(cpi);
if (cpi->oxcf.tuning == VP8_TUNE_SSIM) {
// Initialize encode frame context.
// Build a frame level activity map
build_activity_map(cpi);
}
// re-initencode frame context.
init_encode_frame_mb_context(cpi);
vp9_zero(cpi->rd_comp_pred_diff);
vp9_zero(cpi->rd_filter_diff);
vp9_zero(cpi->rd_tx_select_diff);
vp9_zero(cpi->rd_tx_select_threshes);
set_prev_mi(cm);
// Take tiles into account and give start/end MB
int tile_col, tile_row;
const int tile_cols = 1 << cm->log2_tile_cols;
const int tile_rows = 1 << cm->log2_tile_rows;
for (tile_row = 0; tile_row < tile_rows; tile_row++) {
vp9_get_tile_row_offsets(cm, tile_row);
for (tile_col = 0; tile_col < tile_cols; tile_col++) {
TOKENEXTRA *tp_old = tp;
// For each row of SBs in the frame
vp9_get_tile_col_offsets(cm, tile_col);
for (mi_row = cm->cur_tile_mi_row_start;
encode_sb_row(cpi, mi_row, &tp, &totalrate);
cpi->tok_count[tile_row][tile_col] = (unsigned int)(tp - tp_old);
assert(tp - cpi->tok <= get_token_alloc(cm->mb_rows, cm->mb_cols));
vpx_usec_timer_mark(&emr_timer);
cpi->time_encode_mb_row += vpx_usec_timer_elapsed(&emr_timer);
}
if (cpi->sf.skip_encode_sb) {
int j;
unsigned int intra_count = 0, inter_count = 0;
for (j = 0; j < INTRA_INTER_CONTEXTS; ++j) {
intra_count += cpi->intra_inter_count[j][0];
inter_count += cpi->intra_inter_count[j][1];
}
cpi->sf.skip_encode_frame = ((intra_count << 2) < inter_count);
cpi->sf.skip_encode_frame &= (cm->frame_type != KEY_FRAME);
cpi->sf.skip_encode_frame &= cm->show_frame;
} else {
cpi->sf.skip_encode_frame = 0;
}
// 256 rate units to the bit,
// projected_frame_size in units of BYTES
cpi->projected_frame_size = totalrate >> 8;
// Keep record of the total distortion this time around for future use
cpi->last_frame_distortion = cpi->frame_distortion;
static int check_dual_ref_flags(VP9_COMP *cpi) {
MACROBLOCKD *xd = &cpi->mb.e_mbd;
int ref_flags = cpi->ref_frame_flags;
if (vp9_segfeature_active(&xd->seg, 1, SEG_LVL_REF_FRAME)) {
return (!!(ref_flags & VP9_GOLD_FLAG) + !!(ref_flags & VP9_LAST_FLAG)
+ !!(ref_flags & VP9_ALT_FLAG)) >= 2;
Ronald S. Bultje
committed
}
static int get_skip_flag(MODE_INFO *mi, int mis, int ymbs, int xmbs) {
int x, y;
for (y = 0; y < ymbs; y++) {
for (x = 0; x < xmbs; x++) {
if (!mi[y * mis + x].mbmi.mb_skip_coeff)
return 0;
}
}
return 1;
}
static void set_txfm_flag(MODE_INFO *mi, int mis, int ymbs, int xmbs,
TX_SIZE txfm_size) {
int x, y;
for (y = 0; y < ymbs; y++) {
mi[y * mis + x].mbmi.txfm_size = txfm_size;
}
}
static void reset_skip_txfm_size_b(VP9_COMP *cpi, MODE_INFO *mi, int mis,
TX_SIZE txfm_max, int bw, int bh, int mi_row,
int mi_col, BLOCK_SIZE_TYPE bsize) {
VP9_COMMON * const cm = &cpi->common;
MB_MODE_INFO * const mbmi = &mi->mbmi;
if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
return;
MACROBLOCK * const x = &cpi->mb;
MACROBLOCKD * const xd = &x->e_mbd;
const int ymbs = MIN(bh, cm->mi_rows - mi_row);
const int xmbs = MIN(bw, cm->mi_cols - mi_col);
Loading full blame...